research papers on multilingual

English in the multilingual classroom: implications for research, policy and practice

PSU Research Review

ISSN : 2399-1747

Article publication date: 28 November 2017

The shift in the function of English as a medium of instruction together with its use in knowledge construction and dissemination among scholars continue to fuel the global demand for high-level proficiency in the language. These components of the global knowledge economy mean that the ability of nations to produce multilinguals with advanced English proficiency alongside their mastery of other languages has become a key to global competitiveness. That need is helping to drive one of the greatest language learning experiments the world has ever known. It carries significant implications for new research agendas and teacher preparation in applied linguistics.

Design/methodology/approach

Evidence-based decision-making, whether it pertains to language policy decisions, instructional practices, teacher professional development or curricula/program building, needs to be based on a rigorous and systematically pursued program of research and assessment.

This paper seeks to advance these objectives by identifying new research foci that underscore a student-centered approach.

Originality/value

It introduces a new theoretical construct – multilingual proficiency – to underscore the knowledge that the learner develops in the process of language learning that makes for the surest route to the desired high levels of language proficiency. The paper highlights the advantages of a student-centered approach that focuses on multilingual proficiency for teachers and explores the concomitant conclusions for teacher development.

• Internationalization
• Policy and education

Brutt-Griffler, J. (2017), "English in the multilingual classroom: implications for research, policy and practice", PSU Research Review , Vol. 1 No. 3, pp. 216-228. https://doi.org/10.1108/PRR-10-2017-0042

Emerald Publishing Limited

Copyright © 2017, Janina Brutt-Griffler

Published in PSU Research Review: An International Journal . Published by Emerald Publishing Limited. This article is published under the Creative Commons Attribution (CC BY 4.0) licence. Anyone may reproduce, distribute, translate and create derivative works of this article (for both commercial and non-commercial purposes), subject to full attribution to the original publication and authors. The full terms of this licence may be seen at http://creativecommons.org/licences/by/4.0/legalcode

Introduction

Today’s institutions of higher education are increasingly tied to the global knowledge economy and tasked with preparing students and researchers to succeed transnationally. That includes the universities that are becoming regional centers of learning. Scientists and other professionals require access to the latest research as well as the ability to disseminate their ideas to the widest audience possible to promote economic and social advancement. To advance the project of global interconnectedness and knowledge production, it necessitates that we assess some of the current trends in higher education and policies with respect to systems of communication, English in particular. The ability of nations to produce multilinguals with advanced proficiency has become a key to global competitiveness. To participate fully in the world today means that students will more often need multilingual proficiency – a theoretical construct that I put forward in this account that measures the ability to communicate in a multilingual world. Knowledge of English is, for many, a key component of such proficiency. There is, consequently, a need to enhance preparation of English teachers and advance new research agenda in English language education.

The present day reach of English education is, in some respects, one of the greatest language learning experiments the world has ever known. As I first pointed out in my book World English: A Study of Its Development , for speakers across the globe, English is, by its nature, a language of multilingualism and multilinguals, and English has established itself alongside other languages in many speech communities around the world. This process takes on different forms and intensity and generates a good deal of intellectual debate in the field of applied linguistics ( Brutt-Griffler and Kim, 2016 ; Kramsch, 2016 ; Seidlhofer, 2011 ; Widdowson, 2003 ). My goal in this paper is to look at some of the current processes and consider what drives English learning today, what impact it has on preparing future professionals and students and what kind of new research is needed to understand the needs of the learner.

English in education: the construction and dissemination of knowledge

I locate the shift in the function of English as a medium of instruction as one of the significant processes that impacts English education and the field of applied linguistics today. While English continues to be one of the main languages taught as a subject in many national school systems, English now increasingly serves as a medium of instruction in a growing number of schools and particularly in universities worldwide ( Dearden, 2015 ). In other words, students in many universities outside of what is thought of as English-speaking contexts may pursue their university education in English in content areas such as business, medicine or engineering. We can, for example, see this process unfolding in the European Union, as detailed in a recent study devoted to the topic of English-medium instruction in the 28 EU member states ( Wächter and Maiworm, 2014 ) supported by the European Commission’s Directorate General for Education and Culture and published by the Academic Cooperation Association (ACA). The authors count 8,089 ETPs, a steep growth when compared to the 725 such ETPs in 2001. The study notes that “there is now little doubt that a critical mass of ETPs is on offer across non-English-speaking Europe” ( Wächter and Maiworm, 2014 , p. 16), with The Netherlands, Germany, Sweden, France taking the lead in terms of numbers. It continues, “one of the policy priorities in Europe – and increasingly elsewhere in the world, too – has been to remove or to reduce barriers possibly preventing students from becoming internationally mobile” ( Wächter and Maiworm, 2014 , p. 25). To overcome the “linguistic disadvantage”, the systems set the trend in offering instruction in the most widely taught language in secondary education world-wide, English.

This shift in the function of English in academia is a significant modification of the earlier role of English as a so-called “foreign” language. It carries important implications for getting students ready, ensuring quality instruction in earlier grades, especially at secondary school levels, to equip them with the advanced language proficiencies to study and be assessed in English in a range of subjects at the university level. The use of English as a medium of instruction requires a high level of language proficiencies on the part of the students, faculty and administration to deliver quality curricula. Its use also creates a unique opportunity for many international students to study the language(s) and culture of the host country as well.

A second driver for English learning and use takes the form of its growing role in scientific dissemination. Scholarly publishing in top tier international venues has become almost synonymous with publishing in English. Recent data point out that over 90 per cent of articles in the natural sciences are written in English and more than 70 per cent in the social sciences and humanities ( Hyland, 2015 ; Ferguson et al. , 2011 ; Hamel, 2007 ). We see a slight difference across disciplinary boundaries, with the highest average in English publishing in mathematics and physics, as illustrated in Table I .

Taking the same timeframe, we observe that scientific production and dissemination globally shows a steady and upward progression in English and a corresponding decline in other languages ( Figure 1 ).

Databases are another indicator of scientific production and communication. Humanities databases (e.g. MLA) point to a greater language distribution while social sciences again tend to index English medium sources (see Table II for the language share).

Many national systems have created a reward system that privileges English language publications and, thus, reinforces the need for advanced academic literacy in English. Such academic literacy is not innate but developed over a lengthy process of formal education. There is an evidentiary basis that writing in English can, and does, “impose an additional burden on some non-Anglophone researchers” ( Ferguson et al. , 2011 , p. 43). Specific areas of linguistic difficulty include “a ‘less rich vocabulary’ and ‘less facility in expression’” ( Ferguson et al. , 2011 , p. 43), “word choice and sentence syntax” ( Ferguson et al. , 2011 , p. 43), specificities of scientific discourse and authorial voice and “time needed to learn English to a high level” ( Ferguson et al. , 2011 , p. 44). Beyond the linguistic domain, researchers need to devote additional time to “substantive matters of research design and methodology, focus, narrative, and coherence of argument” ( Ferguson et al. , 2011 , p. 42). Surveying the field, we find a multitude of attitudes on the dominance of English (DoE) in scientific communication. Ferguson et al. ’s (2011) recent study with 300 scientists finds that 83 per cent of the subjects believe that there is a need for one international language of science. Interestingly, the study finds that:

[…] the higher the subjects’ perceived language proficiency, the less likely they are to agree that the DoE is an unjust advantage to English native-speaking academics and the more likely they are to agree that the advantages of English in their work outweigh the disadvantages ( Ferguson et al. , 2011 , p. 54).

Those who take a pragmatic approach believe that English as an international language “facilitates international co-operation, enables scholars to more easily keep abreast of developments in their discipline and generates a wider potential readership for their published outputs” ( Ferguson et al. , 2011 , p. 52). Researchers in physical sciences see as it more advantageous to have one international language of science as opposed to those in social sciences or humanities that are connected more to a national context of research and/or dissemination to communities that may not have equal access to English ( Flowerdew, 2013 ; Hamel, 2007 ). Ferguson et al. ’s (2011 , p. 56) study rightly concludes that preparation in academic writing and “teachers of academic writing in English are important agents in mitigating any disadvantage that flows from it”.

In sum, the role of English in higher education and knowledge dissemination are significant. It determines the need for high levels of proficiency, inclusive of using the language to both understand and produce academic/professional writing aimed at and produced by an international English medium community of experts. There is much more to it than mere disciplinary terminology. To meet the demands of the global market place, including knowledge dissemination, some areas of education, particularly within Science, Technology, Engineering and Mathematics (STEM) as well as in the domain of business and management have turned to English as a medium of instruction in higher education. That reality in turn affects the primary and secondary school levels. In order that STEM and other professionals not face a double burden of both achieving in their field of expertise while struggling with the language demands the need for English places on them, the educational system must establish a base for an exact understanding of national needs, assessment databases and professional English teacher preparation, with a focus on advanced academic literacy skills.

A new theoretical focus for English applied linguistics

To address the processes in higher education and global knowledge economies, the field of English applied linguistics has generated an enormous literature that can provide guidance. There is no more global enterprise than English teaching and learning. It takes place in literally every nation of the world and involves millions of people. Yet, when we look at the picture globally, we often find a disconnect between the typical English learning context and the conditions and assumptions that continue to a large extent to dominate the field of English applied linguistics. The vast and ever-growing majority of English learning takes place outside of the principal English speaking nations. Yet much of the field of English applied linguistics continues to respond to the educational concerns that arise within them.

First among them is the need to educate children who speak another language at home in school systems in which English is the exclusive medium of instruction. With ever-greater frequency, primary school classrooms in English-speaking nations are filled with learners from myriad language backgrounds, part of a broader phenomenon that has come to be known as the “multilingual classroom”. It has been a prominent trend in the USA, UK, Canada and Australia for decades. More recently, it has spread far beyond their borders, now encompassing, for instance, many of the nations that make up the EU. In the careful language of the world of gray literature, such as the European Commission’s (2015a) report Language teaching and learning in multilingual classrooms , the “multilingual classroom” is called a “challenge”, by which is really meant a problem . The approach advocated in such reports, backed up by a body of academic literature and government policy, is to try to mitigate this problem. That consists of the taken-for-granted circumstance that the teacher does not speak the home languages of the multilingual students while the students have only limited proficiency in the language of instruction used in the classroom.

The goal is a “least-worst” outcome. Because, any countervailing research conclusions aside, practical constraints dictate a policy of “mainstreaming” students as quickly as possible into classes taught in the target language, whether they are actually ready for such a transition, the results for students can be anything but optimal. The EC (2015a , p. 13) report notes, “Assessment tools and assessors with negative perceptions of migrant children’s abilities which allocate more of them to lower ability tracks and special education classes”. The students also “lack […] opportunities to develop their mother tongue competences to higher levels” (p. 13). One unfortunate result is that the European Union has effectively dropped all discussions of its policy of “mother tongue +2” – two additional languages in addition to the child’s first language. That laudable goal now seems to promise too much when applied to what the EC (2015a) calls Europe’s “migrant” students, the broad definition of which includes even the grandchildren of persons not born in the country. The policy of mainstreaming prioritizes protecting the role of the language of instruction and insuring that students learn it over the former goal of European-wide multilingualism. It may be a concession to what are seen to be practical impossibilities. It is one, nevertheless, for which children rather than the society as a whole are asked to pay the price.

Even when the outcome measured in terms of learning the language of instruction of the educational system is good, the child’s experience can nevertheless be trying in the extreme. A graduate student of mine from southern Europe whom I shall call Eva found herself mainstreamed in the US school system before she had mastered even the rudiments of the language and with no one able to help her in her first language. Isolated and struggling, to this day she vividly recalls what she considers to be the “trauma” that she experienced as a primary school student. She emphasizes that she, as an aspiring language teacher, would not willingly put any child in that circumstance no matter how favorable the outcome.

To advance the agenda of applied linguistics, the field needs to be liberated from a too-close attachment to language learning in the USA and EU, where the most urgent need is often transitioning students to the language of instruction and leaving their first language behind. Leaving aside for the moment whether such “mainstreaming” is best for the students in these circumstances, we ought to be able to admit that these circumstances peculiar to English-speaking nations are not those upon which we should generalize to the far more prevalent conditions to be found for the most of the world’s hundreds of millions of English learners. The same is true for the other major strand of English applied linguistics research that stems from the ESL classroom at universities in mainly English-speaking nations – a second overemphasized site of research given in its limited share of the English language market. It too features multilingual backgrounds of students, which the instructor does not share with them.

Overgeneralizing from contexts where the teacher does not know the learners’ language to those where the teacher and students share both languages has the tendency to make a virtue of necessity. The result of neglecting the student’s first language in the process of learning and becoming proficient in a second may not produce the kind of trauma that it meant for Eva, but it can result in confusion and frustration, neither of which is necessary or productive. Why would we not make use of all of the knowledge a student brings into the classroom in the effort to help them learn more? From a student-centered perspective the “problem” lies not with the knowledge of the student but often with that of the teacher, who in some cases may lack the language proficiency to unleash the learning potential of the student.

When students study subjects such as math, engineering or technology, we recognize that they develop a body of knowledge. On the other hand, we do not always readily acknowledge the body of knowledge that they develop as they learn multiple languages. Because of the focus on proficiency in the target language (L2), at an earlier juncture in the development of applied linguistics, the tendency grew to view knowledge of the first language (L1) from the standpoint of its negative impact on the learning of the second. There arose an extensive literature on “transfer errors” (for a review, see Odlin, 1989 ). They were held to be of such importance that the learner from the standpoint of the mainstream second language acquisition theoretical frameworks has been viewed as speaking an interlanguage, an incomplete system.

The consequences shaped notions of pedagogy. As a source of “interference”, at the height of the influence of this paradigm, the first language came to be almost regarded with suspicion in the classroom ( De Angelis and Selinker, 2001 ; Selinker, 1983 ). The best way to learn was thought to be to “immerse” the student in the target language with the notion being that assuring error free input would somehow best lead to error free output. It was almost looked at as language learning de novo, the acquisition of a new linguistic “system” to which previous learning had little, if anything, to contribute. It became almost irrelevant whether the teacher spoke the learner’s L1.

Such an approach takes for granted the far more extensive “positive transfer” that goes unnoticed. An L2 can, of course, only be learned at all because the learner previously speaks an L1. We have all experienced this taken-for-granted aspect of language learning. What challenges us most in learning a second language are those components that are missing in our L1 or are so radically different that we struggle to grasp them. If the L1 and L2 share an alphabet or writing system, we breeze through that portion. If not, we laboriously learn that of the target language. It is far more difficult to acquire an L2 with extensive declensions, or a case system of nouns, coming from a first language largely without them. Languages with elaborate morphosyntax of tense, aspect and mood conjugations require tremendous time and effort to master where they differ extensively from our first languages.

We also miss a process every bit as significant: the use of an L2 in learning a third language. Once we have first learned some grammatical forms we have never before encountered, we no longer need to do so if they exist in another language we attempt to learn. The Latin alphabet, with minor modifications, is common to English, French, Spanish and German. A learner of French whose first language uses a different writing system but who has learned English will draw on his knowledge of that language and not the L1 ( Bardel and Falk, 2007 ; Cenoz, 2001 ; Grosjean, 2001 ; Cook, 1995 ).

What happens, if instead of taking for granted in our theories learning always consists of the expansion of the learner’s body of knowledge we make that the theoretical focus? A learner-centric approach, by focusing on the student, leads us to an understanding based on what I will call multilingual proficiency . I define this concept as a person’s total linguistic proficiency across two or more languages. I do not have in mind here mere “awareness” of other languages, as we hear so much about today with respect to the multilingual classroom. I mean knowledge of these languages.

The notion of multilingual proficiency is meant to underscore knowledge that the learner develops in the process of language learning. It recognizes that language learning capacities among students are virtually without limit and conceptually it does not limit itself as a model to one or two languages. Multilingual proficiency recognizes that there is an aspect of language learning, in the form of knowledge of language , that is acquired in the study of multiple languages. That knowledge of language acquired in studying languages in turn aids learning additional languages. Thus, for example, knowing how cases or conjugation are used in one language can aid learning the system in another. A student’s multilingualism is a resource rather than a problem. Unlocking and helping the student to apply their knowledge should be an essential goal of teaching – one that is best activated by direct appeal to their existing multilingual proficiency in helping them acquire still more.

Multilingual proficiency development constitutes an intellectual endeavor in which in the process of language learning a learner uses the knowledge from various languages ( Baker, 2011 ; Lantolf and Thorne, 2006 ; Brutt-Griffler and Varghese, 2004 ; Swain and Lapkin, 2000 ). Thus taking knowledge as its point of departure, multilingual proficiency becomes an objective measure of language learning from its incipient stages all the way through the attainment of advanced level of proficiency in multiple languages.

Neither am I here referring to what has come to be called “translanguaging”, ( García and Wei, 2014 ) the mixing of two languages together. Multilingual proficiency includes the ability to distinguish one language from another. The notion of multilingual proficiency recognizes that proficient speakers are perfectly able to keep their languages distinct. It is incontestably among the most important of skills in a multilingual world, and, of course, one of the driving forces of English learning in the world today.

It goes without saying that learning English is not, as is sometimes falsely assumed from a monolingualist standpoint, a rejection of the advantages of learning other languages. On the contrary, English learners around the world recognize the equal importance of knowing other languages. If we listen, therefore, to the students on whom a student-centered model must be constructed we hear them emphasizing through their actions their own understanding of the need for multilingual proficiency. Their goal is to learn English alongside other languages they grew up speaking, learned from the context around them, or studied in school.

Language is perhaps the only realm in education in which a student’s knowledge is often not credited. It would be unthinkable in mathematics or science education to take no account of a student’s previous knowledge in teaching the subjects. Yet confining an English language classroom, however multilingual, to one language of instruction can have just that effect. Worse, we may look at quite linguistically accomplished multilinguals through something approaching a deficit model, a significant part of the trauma Eva faced, and made worse by the tendency noted by the EU report on the multilingual classroom to place such students into special education. In such circumstances, their very accomplishments as multilinguals are held against them. But this can happen as well in more subtle ways in every English language classroom that replaces multilingual proficiency with English proficiency viewed in isolation. An alternative consists in recognizing the implications that a student-centered approach that focuses on multilingual proficiency holds for teachers.

New roles for English teachers

A Saudi PhD student enrolled in an applied linguistics program in the USA recalls the first time she entered an English classroom in KSA. She began to address the undergraduate students in English, only to have them stop her and say in Arabic, “no, we don’t know what you’re saying […] we don’t know English. Tell us in Arabic so we can understand.” Her surprised reaction was to think “this is my first time teaching […] I’m not going to ruin it for myself […] I’m gonna follow the rules.” Her co-workers told her, “don’t listen to them […] that’s the school policy […] you have to speak in English all the time.” She decided that she had no choice but to use Arabic despite of the policy. The result, she recalls, was immediate: “they were responsive […] they were actively engaged.”

Not all multilingual classrooms are thereby the same. A multilingual classroom may be one where there are multiple languages, but in a state of dormancy. Or it may be one in which students’ multilingual proficiency is activated. The EC (2015b, p. 4) writes in its report on the multilingual classroom, “teaching culture urgently needs to adapt to the presence of several languages in the classrooms”. It is evidence that the pendulum has swung decisively back the other way in acknowledging the place of students’ first language in the second language learning classroom, confirming that the above mentioned English teacher instinctively made a good choice in the context where she taught ( Storch and Aldosari, 2010 ).

It goes without saying that the Saudi PhD student in her stint as a teacher of English in KSA could only make the decision she did because she had the requisite proficiency in the students’ first language. And yet one place we continue to see the influence of the theoretical models to which applied linguistics remains stubbornly attached, and which produce policies like the English-only classroom, is in the lack of attention to the multilingual proficiency of teachers. One adaptation the EC never mentions in the quest to alter “teaching culture” is the training of teachers in multilingual proficiency and the strategic and planned use of students’ language to allow them to access the curriculum or content in the class ( Ferguson, 2003 ). And yet it might easily be supposed that multilingual students require multilingual teachers. In that case, teacher and student have something essential in common: the skills and knowledge of a multilingual, or multilingual proficiency. In that understanding a multilingual classroom would not be one that is simply characterized by students who among them bring two or more home languages different than the medium of instruction. In the more meaningful form of the term, a multilingual classroom is one in which both students and teachers are multilingual and in which they bring their multilingual proficiency to bear on the dual tasks of teaching and learning.

To be fully accurate, the EC (2015a) report does all but admit that it would be better to fully serve students if teachers were multilingual. But the idea is then dismissed as impractical – or, rather, it is not discussed and readers are left to draw that conclusion themselves. What else can we conclude when the EC (2015a , p. 54) admits that students do better with the “adaptation of teaching to provide academic vocabulary in [the] mother tongue”, and that “staff having the same mother tongue and cultural background as the children who can win their trust” (p. 51)? The authors of the report even go as far as to claim that “opportunities for schools to use bilingual […] approaches [to] teaching are available where many children have the same mother tongues” (EC, 2015a, p. 71). Finally, they note, “Having qualified mother tongue teachers in schools and mother tongues included in language curricula and examinations encourage mother tongue learning” (EC, 2015a), p. 71). That is, such teachers promote what I am calling multilingual proficiency.

These conclusions are drawn without being emphasized. They constitute an important and almost surprising admission. When transferred from the EU context to that of English teaching globally, the real advantage is to the laudable goal of building a learner-centered educational system. Multilingual proficiency is best modeled by multilingual teachers, or, put another way, teachers with multilingual proficiency are needed to develop that set of skills within students.

Implications for research agendas

Evidence-based decision-making, whether it pertains to language policy decisions, instructional practices, teacher professional development or curricula/program building, needs to be based on a rigorous and systematically pursued program of research and assessment. First, a research agenda should emerge from the kinds of contexts in which English learning and teaching takes place and should be aligned with the needs of the students and teachers. It should, therefore, be learner centered. It must proceed not from the conceptualization of the multilingual classroom as a “problem” but as a body of knowledge to be leveraged in the interests of the expansion of language learning and developing proficient users of the language(s). It must also be driven by new theoretical models of language learning. In this respect, the notion of multilingual proficiency can help overcome the limitations of many of the monolingualist assumptions held in the field. Pedagogies that are backed by rigorous classroom research that prepare teachers for how using more than one language in the classroom can mutually reinforce each can help address the new trends in higher education.

a learner-centered approach and instructional practices;

teacher professional development;

developing national assessment data on learning outcomes.

With respect to a learner centered approach and for reasons discussed in the first part of this paper, in higher education today learners often face the dual challenge of learning content (e.g. science, STEM, arts and social science) through the second language that they are acquiring. Researchers, therefore, should consider important questions with respect to precise learning goals and teaching practices in school curricula. These should include whether schools should build curricula that incorporate model(s) of a content and language integrated learning (CLIL) approach, where the curricular content is taught through the medium of a second language; and if so, how much explicit language scaffolding should be provided to achieve the desired language learning outcomes in the classroom? Current research from the English CLIL classroom points to many benefits ( Dalton-Puffer, 2011 ). Research also needs to provide evidence of whether existing language programs have the capacities to develop independent writers and readers for tertiary programming demands.

At the core of a learner-centered approach, researchers need to pay attention to student engagement (affective, cognitive and behavioral) in learning ( Brutt-Griffler and Kim, 2017 ). Based on well-grounded and newly emerging evidence, I consider student engagement to be one of the important factors that will mediate the relationship between teachers’ instructional practices and students’ academic outcomes, as illustrated in Figure 2 .

Equally important, teachers are often not sufficiently prepared in instructional practices that capitalize on new technologies and aid learner centered learning (inquiry based, cooperative learning or e-portfolio assessment). A research agenda, therefore, needs to help to identify best instructional practices so that these are modeled and rewarded by school leaders. Research is also needed with respect to the professional development of teachers, including identifying and analyzing the qualities of effective teachers, curricula and course design and integrating these concepts into their teaching and instructional strategies. Enhancing teachers’ instructional practices via engaging them in teacher inclusive educational research and/or study abroad dual degree programming can aid the expectations of excellence in language teaching.

Developing a new agenda in English education also requires steps to building a national database on learning outcomes/assessment and teacher preparation. A data system that efficiently and accurately collects, manages, analyzes and uses education data can be a powerful source of assessment, an essential mechanism for understanding and improving language education in the public and private sectors. It can provide reliable data for longitudinal and large-scale empirical research on academic performance and literacy of the nation’s students. Such a database could be housed in a National Center for English Development and Research (CEDR) and be available to its stakeholders - program administrators, policymakers and researchers.

In World English: A Study of its Development and a number of related publications, I stressed the condition that English around the world has become a language of multilinguals, with important implications of the language. But that important quality of the global English language has equally crucial ramifications for pedagogy. It is my contention that much of this insight will emerge from the kinds of contexts around the world in which most English learning takes place and will do so where pedagogy is adapted most to the needs of the students. A student-centered approach to teaching English makes demands on teachers and educational policy, both of which must look to new frontiers in research for guidance. The question of how to develop teachers for the demands of educating students in English that serves such vital functions for its speakers depends on new understandings of the process of second language acquisition rooted in the experiences of multilinguals and multilingualism. In charting a vision for a research program to establish the direction forward for the field of English applied linguistics, I have introduced the understanding of multilingual proficiency . This new paradigm stems from our need to reverse the usual lens on language learning that makes use of a deficit model of the learner’s knowledge and ask instead what knowledge teachers and students have in common as multilinguals. The research foci I outline above constitute an important starting point.

Share of languages in natural science publications worldwide 1980-1996

Student-centric model: student engagement as a mediator between teacher instructional practices and student academic achievement

Share of languages in several natural sciences in 1996

LLBA: Linguistics & Language Behavior Abstracts; MLA: Modern Language Abstracts

Source: Hamel (2007)

Baker , C. ( 2011 ), Foundations of Bilingual Education and Bilingualism , 5th ed. , Multilingual Matters , Clevedon .

Bardel , C. and Falk , Y. ( 2007 ), “ The role of the second language in third language acquisition: the case of Germanic syntax ”, Second Language Research , Vol. 23 No. 4 , pp. 459 - 484 .

Brutt-Griffler , J. and Varghese , M. (Eds) ( 2004 ), Bilingualism and Language Pedagogy , Multilingual Matters , Clevedon .

Brutt-Griffler , J. and Kim , S. ( 2016 ), “ Closing the gender gap: the role of English ”, in Pitzl , M-L. and Osimk-Teasdale , R. (Eds), English as a Lingua Franca: Perspectives and Prospects , Mouton De Gruyter , Berlin , pp. 245 - 257 .

Brutt-Griffler , J. and Kim , S. ( 2017 ), “ In their own voices: development of English as a gender-neutral language ”, English Today , available at: https://doi.org/10.1017/S0266078417000372 (accessed 15 September 2017 ).

Cenoz , J. ( 2001 ), “ The effect of linguistic distance, L2 status and age on cross-linguistic influence in third language acquisition ”, in Cenoz , J. , Hufeisen , B. and Jessner , U. (Eds), Cross-Linguistic Influence in Third Language Acquisition: Psycholinguistic Perspectives , Multilingual Matters , Clevedon , pp. 8 - 20 .

Cook , V. ( 1995 ), “ Multi-competence and the learning of many languages ”, Language, Culture and Curriculum , Vol. 8 No. 2 , pp. 93 - 98 .

Dalton-Puffer , C. ( 2011 ), “ Content-and-language integrated learning: from practice to principles? ”, Annual Review of Applied Linguistics , Vol. 31 , pp. 182 - 204 .

De Angelis , G. and Selinker , L. ( 2001 ), “ Interlanguage transfer and competing linguistic systems in the multilingual mind ”, in Cenoz , J. , Hufeisen , B. and Jessner , U. (Eds), Cross-Linguistic Influence in Third Language Acquisition: Psycholinguistic Perspectives , Multilingual Matters , Clevedon , pp. 42 - 58 .

Dearden , J. ( 2015 ), English as a Medium of Instruction – A Growing Global Phenomenon , British Council , London , available at: www.britishcouncil.es/sites/default/files/british_council_english_as_a_medium_of_instruction.pdf (accessed 15 August 2017 ).

European Commission ( 2015a ), Language Teaching and Learning in Multilingual Classrooms , Publications Office of the European Union , Luxembourg , available at: http://ec.europa.eu/dgs/education_culture/repository/languages/library/studies/multilingual-classroom_en.pdf (accessed 14 August 2017 ).

European Commission ( 2015b ), Language Teaching and Learning in Multilingual Classrooms: Policy Brief , Publications Office of the European Union , Luxembourg , available at: http://ec.europa.eu/dgs/education_culture/repository/languages/library/policy/policy-brief_en.pdf (accessed 24 June 2017 ).

Ferguson , G. ( 2003 ), “ Classroom code-switching in post-colonial contexts: functions, attitudes and policies ”, AILA Review , Vol. 16 No. 1 , pp. 38 - 51 .

Ferguson , G. , Pérez-Llantada , C. and Plo , R. ( 2011 ), “ English as an international language of scientific publication: a study of attitudes ”, World Englishes , Vol. 30 No. 1 , pp. 41 - 59 .

Flowerdew , J. ( 2013 ), “ English for research publication purposes ”, in Paltridge , B. and Starfield , S. (Eds), The Handbook of English for Specific Purposes , Wiley-Blackwell , Malden, MA , pp. 301 - 321 .

García , O. and Wei , L. ( 2014 ), Translanguaging: Language, Bilingualism and Education , Palgrave Macmillan , New York, NY .

Grosjean , F. ( 2001 ), “ The bilingual’s language modes ”, in Nicol , J.L. (Ed.), One Mind, Two Languages: Bilingual Language Processing , Blackwell , Oxford , pp. 1 - 22 .

Hamel , R.E. ( 2007 ), “ The dominance of English in the international scientific periodical literature and the future of language use in science ”, AILA Review , Vol. 20 No. 1 , pp. 53 - 71 .

Hyland , K. ( 2015 ), Academic Publishing: Issues and Challenges in the Construction of Knowledge , Oxford University Press , Oxford .

Kramsch , C. ( 2016 ), “ Multilingual identity and ELF ”, in Pitzl , M-L. and Osimk-Teasdale , R. (Eds), English as a Lingua Franca: Perspectives and Prospects , Mouton De Gruyter , Berlin , pp. 179 - 186 .

Lantolf , J.P. and Thorne , S.L. ( 2006 ), Sociocultural Theory and the Genesis of Second Language Development , Oxford University Press , Oxford .

Odlin , T. ( 1989 ), Language Transfer: Cross-Linguistic Influence in Language Learning , Cambridge University Press , Cambridge, MA .

Seidlhofer , B. ( 2011 ), Understanding English as a Lingua Franca , Oxford University Press , Oxford .

Selinker , L. ( 1983 ), “ Language transfer ”, in Gass , S. and Selinker , L. (Eds), Language Transfer in Language Learning , Newbury House , Rowley, MA , pp. 33 - 53 .

Storch , N. and Aldosari , A. ( 2010 ), “ Learners’ use of first language (Arabic) in pair work in an EFL class ”, Language Teaching Research , Vol. 14 No. 4 , pp. 355 - 375 .

Swain , M. and Lapkin , S. ( 2000 ), “ Task-based second language learning: the uses of the first language ”, Language Teaching Research , Vol. 4 No. 3 , pp. 251 - 274 .

Wächter , B. and Maiworm , F. (Eds) ( 2014 ), “ English-taught programmes in European higher education. The state of Play in 2014 ”, ACA papers on International Cooperation in Education , Lemmens Medien GmbH , Bonn , available at: www.aca-secretariat.be/fileadmin/aca_docs/images/members/ACA-2015_English_Taught_01.pdf (accessed 15 August 2017 ).

Widdowson , H.G. ( 2003 ), Defining Issues in English Language Teaching , Oxford University Press , Oxford .

Further reading

Brutt-Griffler , J. ( 2002 ), World English: A Study of Its Development , Multilingual Matters , Clevedon .

Corresponding author

Related articles, we’re listening — tell us what you think, something didn’t work….

Report bugs here

All feedback is valuable

Please share your general feedback

Join us on our journey

Platform update page.

Visit emeraldpublishing.com/platformupdate to discover the latest news and updates

Questions & More Information

Answers to the most commonly asked questions here

Multilingual Sentiment Analysis: State of the Art and Independent Comparison of Techniques

• Open access
• Published: 01 June 2016
• Volume 8 , pages 757–771, ( 2016 )

Cite this article

You have full access to this open access article

• Kia Dashtipour 1 ,
• Soujanya Poria 2 ,
• Amir Hussain 1 ,
• Erik Cambria 3 ,
• Ahmad Y. A. Hawalah 4 ,
• Alexander Gelbukh 5 &
• Qiang Zhou 6  

21k Accesses

164 Citations

6 Altmetric

Explore all metrics

An Erratum to this article was published on 30 July 2016

With the advent of Internet, people actively express their opinions about products, services, events, political parties, etc., in social media, blogs, and website comments. The amount of research work on sentiment analysis is growing explosively. However, the majority of research efforts are devoted to English-language data, while a great share of information is available in other languages. We present a state-of-the-art review on multilingual sentiment analysis. More importantly, we compare our own implementation of existing approaches on common data. Precision observed in our experiments is typically lower than the one reported by the original authors, which we attribute to the lack of detail in the original presentation of those approaches. Thus, we compare the existing works by what they really offer to the reader, including whether they allow for accurate implementation and for reliable reproduction of the reported results.

Similar content being viewed by others

A survey on sentiment analysis methods, applications, and challenges

A review on sentiment analysis and emotion detection from text

Sentiment Analysis in the Age of Generative AI

Avoid common mistakes on your manuscript.

Introduction

With the growth of the World Wide Web, the amount of texts available online has been increasing exponentially. In particular, people express their opinions about different subjects and influence each other’s decisions by communicating their sentiments [ 56 , 67 ]. The sentiment towards a brand on the Internet is important for any company concerned about the quality of its product, which makes it crucial for companies to understand people’s sentiments towards products and services [ 60 ]. The past few years have witnessed an explosion of commercial and research interest in the sentiment analysis field [ 4 ]. While information extraction techniques have been developed to deal with the ever-growing amount of texts in Internet, sentiment analysis has its own specific problems and difficulties [ 2 ]. Many approaches have been proposed to classify sentiments expressed in different channels such as Twitter, blogs and user comments.

The majority of current sentiment analysis systems address a single language, usually English; see Figs.  1 and 2 . However, with the growth of the Internet around the world, users write comments in different languages. Sentiment analysis in only single language increases the risks of missing essential information in texts written in other languages. In order to analyse data in different languages, multilingual sentiment analysis techniques have been developed [ 10 ]. With this, sentiment analysis frameworks and tools for different languages are being built.

Number of publications on English sentiment analysis, per year [ 42 ]

Number of publications on multilingual sentiment analysis, per year [ 28 ]

One of the main problems in multilingual sentiment analysis is a significant lack of resources [ 4 ]. Thus, sentiment analysis in multiple languages is often addressed by transferring knowledge from resource-rich to resource-poor languages, because there are no resources available in other languages [ 18 ]. The majority of multilingual sentiment analysis systems employ English lexical resources such as SentiWordNet.

Another approach is to use a machine translation system to translate texts in other languages into English [ 18 ]: the text is translated from the original language into English, and then English-language resources such as SentiWordNet are employed [ 18 ]. Translation systems, however, have various problems, such as sparseness and noise in the data [ 4 ]. Sometimes the translation system does not translate essential parts of a text, which can cause serious problems, possibly reducing well-formed sentences to fragments [ 6 ].

Thus, researchers look for alternative approaches. The field of multilingual sentiment analysis is progressing very fast. In particular, multilingual lexical resources specific to sentiment analysis are being developed. For example, the NTCIR corpus of news articles in English, Chinese, and Japanese contains information on sentiment polarity and opinion holder for news related to the topics such as sport and politics [ 46 ]. However, sentiment analysis corpora and resources, even if created for multiple languages, cannot be used for other languages [ 33 ]. More research is required to improve results in the multilingual sentiment analysis discipline [ 20 ].

In this paper, we discuss existing approaches. More importantly, we report the results of our own experiments with these approaches on the same datasets, which allows direct comparison. For this, we have implemented eleven techniques following as closely as possible their descriptions in the original papers. Our results proved to be lower than the results reported by the original authors, which we attribute in the majority of cases to the lack of detail in their descriptions. Thus, in a way, we measured the real value of the information available on those approaches to the research community: a good approach poorly described is not useful for the community, even if it showed good results in its author’s own experiments, which are not available to the community. Thus, we evaluate what the original papers that we reviewed really offer to the reader, apart from only reporting the results their authors observed.

This paper is organized as follows. Section  2 briefly discusses multilingual sentiment analysis techniques and describes pre-processing, multilingual sentiment analysis resources, tools used in multilingual sentiment analysis, and the features used for machine learning. Sections  3 , 4 , and 5 present an overview of the state-of-the-art corpus-based, lexicon-based, and hybrid sentiment analysis techniques, correspondingly, both for English and for other languages. Section  6 gives a comparison of recently some of those methods in our own experiments on common datasets. Finally, Section  7 concludes the paper.

Sentiment Analysis Framework

In this section, we will discuss the main general techniques used for sentiment analysis, as well as pre-processing procedures, lexical resources, tools, and features typically used in sentiment analysis systems.

Main Techniques

Sentiment analysis systems can be classified into corpus-based approaches using machine learning, lexicon-based approaches, and hybrid approaches. Corpus-based methods use labelled data [ 70 ]; lexicon-based methods rely on lexicons and optionally on unlabelled data [ 57 ]; and hybrid methods are used based on both labelled data and lexicons, optionally with unlabelled data [ 51 ]. A sentiment lexicon is a collection of known sentiment terms [ 32 ].

Pre-processing

The pre-processing task is an important step in multilingual sentiment analysis. It is used to remove irrelevant parts from the data, as well as to transform the text to facilitate its analysis.

Noise Removal

Usually the texts found in Internet have much noise such as HTML tags, scripts, and advertisements. Data pre-processing can reduce noise in the text and improve performance and accuracy of classification. The pre-processing step is crucial for multilingual sentiment analysis. The majority of the proposed approaches to multilingual sentiment analysis employ pre-processing of data to improve performance and accuracy.

Normalization

Often sentiment analysis and opinion mining is performed on texts from social networks and other user-generated contents. Such texts are characterized by very informal language, with grammar and lexicon that greatly differ from the usual language use, especially in Twitter. Such texts need to be transformed into a more grammatical form, more suitable for processing by natural language analysis tools. Such normalization is often performed using specialized lexicons, such as the multilingual Lexicon for pre-processing of social media, social networks, and Twitter texts developed by Posadas-Durán et al. [ 39 ] for English, Spanish, Dutch, and Italian.

Natural Language Analysis

The most important pre-processing tasks performed with natural language analysis techniques are tokenization, sentence splitting, stop-word removal, stemming, and part-of-speech tagging, among others. Tokenization is used to break the text down into words and symbols [ 14 ]. Sentence splitting is used to determine sentence boundaries. Stop words are common words in the given language that do not carry important meaning; their removal usually improves performance of sentiment analysis [ 41 ]. Stemming is a task used to transform words into their root form: for example, the word “working” is changed to its root form “work” [ 42 ].

Sentiment Lexicons

Sentiment lexicons have been used in a number of approaches to multilingual sentiment analysis in order to improve the performance of classification. Sentiment lexicons are used mainly in lexicon-based sentiment analysis.

SenticNet is a lexical resource based on a new multi-disciplinary approach proposed by Cambria et al. [ 11 ] to identify, interpret, and process sentiment in the Internet. SenticNet is used for concept-level sentiment analysis. It is also used to evaluate texts basing on common-sense reasoning tools that require large inputs. However, it is not capable of analysing text with sufficient level of granularity. Sentic computing methodology is used, in particular, to evaluate texts at the page or sentence level. The purpose of SenticNet is to build a collection of concepts, including common-sense concepts, supplied with polarity labels, positive or negative. Unlike SentiWordNet, SenticNet does not assume that a concept can have neutral polarity. SenticNet includes a simple and clear API for its integration in software projects. It can be used with the Open Mind software. It guarantees high accuracy in polarity detection. Multilingual tools are available for SenticNet [ 64 ].

SentiWordNet is a lexical resource that assigns WordNet synsets to three categories: positive, negative, and neutral, using numerical scores ranging from 0.0 to 1.0 to indicate a degree to which the terms included in the synset belong to the corresponding category. SentiWordNet was built using quantitative analysis of glosses for synsets [ 52 ]. While SentiWordNet is an important resource for sentiment analysis, it contains much noise. In addition, it assigns polarity at the syntactic level, but it does not contain polarity information for phrases such as “getting angry” or “celebrate a party”, which correspond to concepts found in the text to express positive or negative opinions [ 11 ].

General Inquirer is a German lexicon supplied with positive and negative labels. For its construction, Google translate was used to translate words and terms into the German language; then, the words without any sentiment were removed from the lexicon. General Inquirer has been employed by Remus et al. [ 44 ]. The main advantage of General Inquirer is its widely used lexicon. Since it includes financial terms, it is used for financial sentiment analysis in the German language. However, its use is limited in other areas such as sport, politics, and product reviews [ 53 ].

SEL is a Spanish emotion lexicon that presents 2036 words supplied with the Probability Factor of Affective use (PFA) as the measure of their expression of basic emotions: joy, anger, fear, sadness, surprise, and disgust, on the scale of null, low, medium, or high. The lexicon was developed manually by 19 annotators, which had to agree above certain threshold for a label on the word to be included in the lexicon. The measure called Probability Factor of Affective use (PFA) was developed by the authors of this lexicon to incorporate agreement between annotators in decision-making on labelling the words: the greater the agreement, the stronger the expression of the emotion by the given word. The lexicon, freely available for download, has been used in opinion mining tasks on Spanish tweets [ 49 ].

Sentiment Corpora

Lexical resources for sentiment analysis include, apart from sentiment lexicons, various corpora developed for sentiment analysis tasks. Sentiment corpora are used mainly for machine learning in corpus-based sentiment analysis.

YouTube dataset is a multimodal sentiment analysis dataset created by Morency et al. [ 35 ] from online social videos. In each clip included in the dataset, a person speaks in the camera expressing an opinion. The dataset has various characteristics challenging for sentiment analysis tasks, such as diversity, multimodal, and ambient noise. The topics discussed in online videos are very diverse. Diversity is important to analyse opinions: people express their opinions in different ways; some people express their opinions in subtle ways. The dataset provided age and gender information on the speakers, as well as topics of the opinions. In order to select best words to identify the sentiment of a sentence, multimodal techniques have been used. Since audio and video data have much noise, these data were recorded by using different cameras and microphones.

Explicit and implicit aspect corpora are used for aspect-based opinion mining. Hu and Liu [ 26 ] developed a corpus widely used in aspect-based sentiment analysis research. The original corpus contained data only for explicit aspect extraction, that is, for work with aspect words explicitly present in the sentence. Cruz-Garcia et al. [ 16 ] developed an implicit aspect corpus based on a subset of the corpus by Hu and Liu. In this new corpus, sentences are labelled with implicit aspects, i.e. aspects not named by any specific word in the sentence, and the corresponding implicit aspect indicators. This corpus, freely available for download, has been used in a number of research works.

MPQA is a subjective lexicon consisting of around eight thousand terms, which have been collected from different sources. The MPQA presents words supplied with part-of-speech tags and polarity (positive, negative or neutral), as well as intensity of polarity [ 59 ].

Machine Learning Tools

WEKA, standing for Waikato Environment for Knowledge Analysis, is a freely available software package built in Java, which provides a large number of machine learning and data mining algorithms. The programme provides pre-processing and performance analysis data [ 25 ].

LIBSVM is a library implementing the support vector machine (LIBSVM) algorithm. It was built in 2000. The main purpose of LIBSVM is to help users to easily include SVM into their applications [ 13 ].

Features Used

Machine learning features typically employed in sentiment analysis approaches include the following classes.

N-grams represent continuous sequences of n items in the text. The n-grams of size one are called unigrams, those of size two are called bigrams, and those of size three are called trigrams. For example, in the sentence “I went to the cinema”, the bigrams (after removing the stop-word “the”) are “I went”, “went to”, “to cinema”, and the trigrams are “I went to” and “went to cinema” [ 40 ].

Document frequency is the total number of documents in the dataset that contain a given word. A threshold is calculated for document frequency of words in the training corpus, and the words with document frequency lower than some threshold or higher than another threshold are removed at the pre-processing stage. This process is important for term selection. Tt is used to scale large datasets to reduce the computation cost of their processing.

Term frequency (TF) is the number of occurrences of an item (such as a word or n-gram) in a given document. It is often used in combination with inverse document frequency (logarithm of the inverse of the share of the documents in the collection that contain the given term) in the form of the TF-IDF feature.

Mutual information (MI) is used to measure the dependence between two different variables [ 36 ]. Mutual information is used in statistical language modelling [ 68 ].

Information gain (IG) measures goodness of features in machine learning. It is used to measure the amount of information contributed the classification process by the absence or presence of a term in the document [ 68 ].

Chi-square test is used to calculate the category of terms [ 68 ]. Chi test measures the divergence from expected distribution based on the features that are independent from the class value [ 58 ].

Corpus-Based Techniques

In this and the next sections, we will discuss the state-of-the-art approaches to sentiment analysis classified into corpus-based, lexicon-based, and hybrid ones, for both English language and other languages. In particular, in this section we present corpus-based techniques, development of which focuses on feature engineering and model selection. The majority of the techniques presented here use annotated corpus and machine learning models to train a suitable sentiment analysis classifier.

Shi and Li [ 47 ] developed a supervised machine learning technique for sentiment analysis of online hotel reviews in English by using unigrams features. They used features such as term frequency and TF-IDF to identify the document polarity as positive or negative. The data were separated into training and testing sets with different data instances. The instances in the training set covered the target values. The support vector machine (SVM) has been used to develop a model able to predict target values of data instances [ 47 ]. The SVM classifier has been chosen because it has been reported to perform better than other classifiers [ 38 ], though Tong and Koller [ 55 ] consider Naive Bayes and SVM the most effective classifiers among machine learning techniques [ 61 ]. The hotel-review corpus contained 4000 (positive and negative) reviews; the reviews have been pre-processed and tagged as positive and negative. Then, the obtained sentiment classification model has been used to classify live information flow into positive and negative documents. The TF-IDF feature performed better than simple term frequency [ 47 ].

Another study [ 10 ] used supervised classification for identification of the sentiment in documents. They applied their method to sentences found in Internet, in particular, in blogs, forums, and reviews. The features of the sentences were extracted using a state-of-the-art algorithm. Sentence parsing has been used for a deeper level of analysis. Finally, the method of active learning has been used to reduce workload in annotation [ 15 ]. After the pre-processing stage, there were different features selected, such as unigrams, stems, negation, and discourse features. The SVM, Maximum Entropy, and multimodal Naïve Bayes classifiers have been employed as machine learning algorithms. For linearly separable data, SVM gives classification results with minimal error. The multimodal Naïve Bayes classifier is very simple to use for efficient classification and with incremental learning [ 31 ]. The Maximum Entropy classifier is efficient in extracting information that leads to good results [ 7 ]. English-language corpora were collected from blogs, reviews, and forum sites such as www.livejournal.com or www.skyrock.com .

The Maximum Entropy classifier showed 83 % accuracy, which is better compared to other classifiers used in this study, namely SVM and multinomial Naïve Bayes; however, other approaches [ 47 ] used SVM to evaluate datasets, and other machine learning techniques have been reported to have accuracy lower than that of SVM.

The main advantage of this approach is that it involves less building effort and is simple to develop. A disadvantage of this approach is the lack of high-quality training data, because data collected from blogs contain many grammatical errors, which negatively affect classification performance [ 10 ].

Other Languages

Habernal et al. [ 23 ] proposed an approach for supervised sentiment analysis in social media for the Czech language. Three different datasets have been employed; first dataset was collected from Facebook, basing on top comments in popular Czech Facebook pages. The Facebook dataset contained positive, negative, neutral, and bipolar information. The second dataset was a movie review dataset downloaded from a Czech movie database. The third dataset contained product review information collected from large online Czech shops. After the data pre-processing step, the n-gram feature has been extracted. The unigrams and bigrams were used as binary features. In addition, the minimum number of occurrences of character n-grams has been established. Part-of-speech (POS) tagging provided characteristics of specific posts. Various POS features have been used, such as adjectives, verbs, and nouns. Two different emoticon lists have been used: one for positive and one for negative sentiment. Another feature used was Delta TF-IDF, a binary word feature, which showed good performance. Delta TF-IDF uses TF-IDF for words, but it treats words as positive or negative.

To evaluate the dataset, two different classifiers were trained: SVM and a Maximum Entropy classifier. The F-measure for combination of features such as bigrams, unigrams, and emoticons was 0.69. The emphasis of this approach was on feature selection. The features that were selected were bigrams, unigrams, POS, and character n-grams. This approach is useful for sentiment analysis in Czech social media. However, it cannot be directly used for other languages, and its results are not very helpful even for Czech social media. Still it can help researchers extend sentiment analysis methods to the Czech language [ 23 ].

Tan and Zhang [ 54 ] introduced an approach for sentiment classification for the Chinese language. First, POS tagging was used; the aim of using POS tagging was to parse and tag the Chinese text. After POS tagging, feature selection was used to determine discriminative terms for classification. Finally, a machine learning approach was used for sentiment classification. Feature selection included four types of information: document frequency, Chi-square feature, mutual information, and information gain. The threshold was defined for the document frequency of words and phrases in the training corpus, and the words with the document frequency lower than a predefined threshold or higher than another predefined threshold were removed. In order to calculate the association between terms, CHI was used. Mutual information was used for statistical language modelling. Information gain measures the amount of information useful for prediction of the category that is contributed by the presence or absence of a given term in the document.

There are various datasets available online for use in Chinese sentiment classification. The Chinese sentiment corpus ChnSentiCorp, collected from online documents, is an online benchmark sentiment analysis database. It includes 1021 documents in three domains: education, movies, and house. For each of these domains, there are positive and negative documents. The centroid classifier, SVM, Naïve Bayes, k -nearest neighbour classifier, and winnow classifier were compared. The overall accuracy of the SVM classifier was better than that of other classifiers.

This approach is unique in comparison with other approaches in that the feature selection scheme is different. The features that are selected are document frequency, mutual information, Chi-square statistic measure, and information gain. Other approaches usually employ such features as bigrams and unigrams. The results of this approach show that of such features as information gain, document frequency, Chi-square statistics, and mutual information, information gain is the best feature and can be recommended for future applications. The main disadvantage of this approach is use of traditional features such as Chi-square statistics, document frequency, and mutual information [ 54 ].

Ghorbel and Jacot [ 21 ] proposed an approach for sentiment analysis of French movie reviews. Their method relies on three types of features, namely lexical, morpho-syntactic, and semantic features. The unigrams were selected as a feature. The goal of this system was to find polarity of the words. The part-of-speech tags were employed to augment unigrams with morpho-syntactic information, in order to reduce word sense ambiguity and to control negation before polarity extraction. SentiWordNet was used to determine polarity of words. This information was used to measure the overall polarity score of the review [ 52 ]. SentiWordNet is an English-language resource; in order to use SentiWordNet, French reviews were translated into English before extraction of polarity. The words were lemmatized before looking them up in a bilingual dictionary; then part-of-speech tags were used for sense selection, to remove uncertain senses, and to predict the correct synset. The dataset of French movie reviews contained 2000 documents: 1000 positive and 1000 negative reviews of ten movies.

The SVM classifier was used for classification. The overall performance on French movie reviews using unigrams, lemmatization, and negation was 92.50 % for positive reviews and 94 % for negative reviews. This approach combined lexical, morpho-syntactic, and semantic orientation of words to improve the results. The accuracy was improved by 0.25 %. The semantic orientation of the words was extracted from SentiWordNet, which further improved the result by 1.75 %.

A disadvantage of this approach is that words need to be translated into English prior to use SentiWordNet, which is an English-language resource. The quality of translation had a negative effect on the performance of the classifier, since translation of words does not preserve the semantic orientation due to differences between languages [ 21 ].

Balahur and Turchi [ 5 ] introduced a hybrid technique for sentiment analysis of Twitter texts. The sentiment analysis tools for various languages were developed to minimize the effort to produce linguistic resources for each of these languages; research on the use of machine translation systems to produce multilingual data was conducted in the context of Twitter texts.

The pre-processing was employed to normalize the texts: at this phase, the linguistic peculiarities of tweets were taken into consideration. Spelling variants, slang, special punctuation, and sentiment-bearing words from the training data were substituted by unique labels. For example, the sentence “I love car” was changed to “I like car”; according to the General Inquirer dictionary, love and like both have positive sentiment.

This approach can be used for various languages with minimal linguistic processing. Only tokenization was used; the method does not require any further processing. The final system should work similarly for all languages.

A standard news translation system was used to obtain data in various languages such as Italian, German, Spanish, and French. The original dictionary was created based on translation of English and Spanish texts into a third language. The dictionary was created for fifteen different languages. This approach includes two main stages: the pre-processing step and the application of a supervised machine learning technique. Support vector machine sequential minimal optimization (SVM SMO) was employed to identify features such as n-grams and bigrams in the training data [ 5 ].

The accuracy on English language was higher than on other languages. The main novelty of this approach was the pre-processing step. The pre-processing of Twitter texts is very important for sentiment analysis, and it significantly affects the accuracy of the classifier. The normalization of tweets at the pre-processing step can improve the accuracy. The main disadvantage of this approach is that on English language better accuracy was obtained in comparison with other languages, while on other languages such as Spanish and Italian the approach did not perform well [ 5 ].

Duwairi and Qarqaz [ 19 ] introduced a supervised technique for sentiment analysis of Arabic tweets. The authors generated a dataset using 10,000 tweets and 500 Facebook reviews in various domains such as news and sport. A number of pre-processing techniques were used in this study including removing duplicated tweets, empty tweets, and emoticon-only reviews. In order to determine the sentiment of collected tweets and Facebook reviews, a number of volunteers were asked to label each tweet or comment as positive, negative, neutral, or other.

A number of pre-processing steps such as tokenization, stemming, forming bi-grams, and detection of negation were then applied to the tweets and Facebook comments. Finally, three supervised machine learning techniques were applied on the prepared dataset, namely k -nearest-neighbour, Naïve Bayes, and SVM classifiers. The tenfold cross-validation method was used for evaluation. It showed that SVM outperformed both k -nearest-neighbour and Naïve Bayes classifiers. A limitation of this study was that the number of trained data was rather small.

Lexicon-Based Techniques

The development of lexicon-based techniques mainly focuses on the different semantic orientation methods. Such techniques use different lexicon resources for sentiment inference.

The unsupervised semantic orientation (SO-PMI-IR) method has been proposed for the sentiment classification of movie reviews. In the semantic orientation, text is classified basing on the score of the chosen sentences. The pointwise mutual (PMI) information for extracted features is calculated as

Here, c denotes the category and t indicates the term [ 69 ]. Pointwise mutual information is used to measure the degree of compatibility of a term and category [ 66 ].

Singh et al. [ 52 ] used the unsupervised semantic orientation with part-of-speech tagging on the Cornell movie review dataset; this approach showed the best results in our own evaluation; see Sect.  6.1 . Feature extraction was done for all reviews. The semantic orientation was calculated for reviews; then adjectives were extracted and the semantic orientation value was assigned to them. Aggregation was done for semantic orientation: each positive term +1 was added to the total document score and for each negative term, –1. Thus, the semantic orientation of each review was the total semantic orientation values for the extracted terms. Then, a threshold of 5 on the absolute value of the score was used to classify a document as positive or negative basing on the aggregation score. This approach was based on SentiWordNet. The features were extracted, and then SentiWordNet was employed to check the scores for the selected features. SentiWordNet provides scores from 0.0 to 1.0 [ 11 ]. Two different datasets were used; one dataset contained one thousand positive and one thousand negative reviews, and another dataset contained seven hundred positive and seven hundred negative reviews. Figure  3 presents the main steps of this approach.

Flowchart of the approach of [ 52 ]

This approach can be easily extended to other languages. In particular, it detects multiword expressions and can handle sarcasm; some languages, such as Persian language, make heavy use of multiword expressions and sarcasm [ 45 ]. In the future, this approach can be improved if different dialects can be detected; for example, Persian language has many different dialects [ 45 ], as do many other languages, such as Arabic, German, and Chinese.

The main disadvantage of this approach was that it required computationally expensive calculation of PMI, which was very time consuming [ 52 ]. The use of PMI in this approach did not improve the performance, which was still below that of other machine learning methods [ 43 ].

In another research, a method for unsupervised sentence classification of product reviews by using tools such as SentiWordNet was introduced. This method consisted of six steps. The first step was to collect different online reviews. The second step was the pre-processing of the reviews. The third step was building lists containing noun features and extracting the noun phrases. The fourth step was to classify sentences into objective and subjective sentences. The fifth step was the opinion sentence detection that calculated the semantic orientation of words related to the weight of the word in the SentiWordNet dictionary. Finally, the last step was to calculate the weight for each sentence and review and determine its polarity. This method obtained regular accuracy. The dataset that was used for evaluation contained online reviews of cameras such as Canon and Nikon. After data collection and pre-processing, the sentences were classified into objective and subjective types. To find semantic orientation of subjective sentences, SentiWordNet was used. The final semantic score was calculated to identify positive and negative statements. However, in other approaches, such as that by Singh et al. [ 52 ], data pre-processing consisted of part-of-speech tagging, the sentences were not classified into objective or subjective types, and an aggregation procedure was used to calculate the semantic orientation score [ 22 ].

The main disadvantage of this approach was the use of SentiWordNet. Its results show that SentiWordNet was ineffective in discovering sentiment words and performing the classification task [ 8 ].

Wan [ 57 ] proposed an approach to leverage English resources to increase performance of Chinese sentiment analysis. The approach included various stages. First, a translation system has been used to translate the Chinese reviews into English. There were various translation systems used, such as Yahoo and Google, to translate Chinese reviews into English. After translation, the semantic orientated approach has been used to calculate the value of reviews. This approach used negation lexicon to reverse the semantic polarity of the words or phrases changing the value of the term to positive or negative. The unsupervised method was very simple. It used positive and negative lexicons; negation lexicon contained different terms used to reverse the semantic polarity of specific terms; intensifier lexicon consisted of words and phrases able to change the degree for the term to positive or negative.

In order to evaluate the performance of the introduced method, one thousand product reviews were collected for Chinese IT products such as mp3 players, mobile phones, laptops, and cameras. Chinese reviews were translated into English and analysed in both languages to obtain better accuracy. The results showed an overall performance improvement. This approach employed the ensemble to improve performance of the classification by 0.25 % [ 57 ].

The advantage of this work was in comparing different translation systems and determining the best system that can be used for future research. A disadvantage of this approach was in that translation of the reviews had a negative effect on performance [ 57 ].

Carroll [ 12 ] developed an innovative unsupervised model for the Chinese product reviews. The approach used comprehensive semantic analysis of words in the Chinese language. Lexical items were sequences of Chinese characters, ignoring punctuation marks. Each zone was classified as positive or negative. The iterative process was able to increase the seed vocabulary into broad vocabulary that consisted of a list of sentiment-bearing lexical item. A classifier was run on Chinese product reviews, giving as outcome positive and negative documents. The sentiment density has been calculated as a proportion of opinion zones in the documents. The sentiment density was not an absolute value, but it was used to compare documents with each other. The sentiment density of 0.5 does not mean half-opinionated document; it can be interpreted as indicating that the review is less opinionated than a review with density of 0.9. The classifier was able to reach 87 % F-measure for sentiment polarity [ 12 ]. A disadvantage of this approach was in using a corpus that did not help to detect the polarity of the words [ 12 ].

Zagibalov and Carroll [ 71 ] used automating seed words for selection in the Chinese language. In unsupervised learning, the training data need not be annotated. The approach did not require word segmentation. The lexical items lexicon was used to treat Chinese characters. In order to improve the classifier to find the seeds automatically, two assumptions have been used: the first assumption was that the attitude was stated by using negation of word items with their opposite meaning; this assumption was used to find negative lexical items from positive seeds. The second assumption concerned polarity of seeds that needed to be identified. To identify the polarity of a seed word, the lexicon was used to reach gold standard for positive lexical item. The sentiment classification and iterative technique were used in the unsupervised method. The method was used to find seeds automatically from raw text. To find positive seeds from the corpus, a special algorithm was developed. It operated over the sequence of characters that should be checked for containing negation or adverbials. This method does not use pre-segmentation or grammar analysis; the unit of processing is a lexical item. Input sequences of Chinese characters did not include punctuation marks and zones. A single zone was classified either as positive or negative, and the corresponding scores were calculated. Then, iterative retaining was used to increase the seed vocabulary in the list of sentiment-bearing lexical items. The latest version of the classifier was used on the corpus to classify documents as positive and negative. The iterative retaining was stopped when there was no modification to the classification of the document. To test the method on the dataset obtained from Chinese product reviews website, the reviews were tagged by polarity and the duplicate reviews were removed.

The main difference of this approach is the seed corpus. To develop the seed corpus, the following algorithm was used:

The sequence of characters should be delimited by non-character symbols;

The number of occurrences of a sequence that follow negated adverbial was counted;

The number of occurrences of a sequence without such construction was counted;

All such sequences were found.

A disadvantage of this approach is that it is very difficult to build and requires extensive parameter tuning [ 24 ].

Zhang et al. [ 72 ] presented a lexicon-based approach for classification of Chinese reviews of different products. This Internet-based method (PMI-IR) consisted in four phases. The first phase was parsing and POS tagging of the reviews; the second phase was extraction of two phrases conforming to a specific pattern in part-of-speech tags; the third phase was to identify phrases and calculate the semantic orientation of SO for all extracted phrases in the reviews. The approach contained different phases that were after the data pre-processing step: the sentiment expression was extracted from the Chinese review, snippet was formed, sentiment orientation of the expression was determined, and finally, sentiment classification for Chinese review was performed. This approach used snippets to identify the sentiment polarity of the phrases. A snippet is a small text from the documents, and it is located below the links returned by search engines. A snippet contains part of query words and allows previewing the query words in the documents. The PMI-IR algorithm was used to calculate the semantic orientation; the words have been estimated by using returned snippets. For example, to calculate the polarity for the word “poor”, the query has been sent to Google and returned snippets were crawled.

In order to evaluate the approach, a mobile phone review dataset, of forty positive and forty negative reviews, was used. The main difference of this approach is the use of snippets. Other approaches usually used online reviews, blogs, and Twitter texts.

Al-Ayyoub et al. [ 3 ] proposed an unsupervised approach to sentiment analysis of Arabic tweets. This approach included two stages: The first stage was collecting and pre-processing the tweets. The pre-processing step included stop-word removal and stemming. The second stage was the development of a sentiment lexicon, with the sentiment scores in the range between zero and one hundred. Scores from zero to forty corresponded to negative sentiment, forty to sixty to neutral, and sixty to one hundred to positive. These values were combined with each other to calculate the sentiment value of the sentence. The overall accuracy of this approach was 86.89 %. A disadvantage of this approach is that it is not able to handle different Arabic dialects [ 3 ].

Hybrid Techniques

In this section, we present resource-hybrid techniques, which combine corpus-based and lexicon-based approaches, focusing on the domain adaption of sentiment analysis for the resource-poor languages or special domains. These techniques mostly use both annotated corpora and lexicon resources for learning more useful sentiment analysis resources.

Mizumoto et al. [ 34 ] introduced unsupervised approach to identify sentiment polarity of the stock market. The polarity of the sentiment for stock news market was identified by using a polarity dictionary that contained words and their polarities. In this method, for a small amount of words, polarity was determined manually. The polarity of new words was then identified automatically. The new dictionary method has been built for unlabelled news. The dictionary contained a small number of words with their polarities such as positive and negative words. If a word was situated in one sentence with both positive and negative words, the co-occurrence of frequency for negative and positive polarity was calculated. The bias of co-occurrence was measured; most of the words were occurring with positive and negative polarities; the rate of co-occurrence of positive and negative polarity of dictionary has been used; then the polarity of those words that were not added was estimated. Finally, the polarity of words was determined. Two different thresholds were introduced, namely thresholdP and thresholdN. The thersholdP value was used to add words to the positive polarity dictionary, and thresholdN was used to add words to the negative polarity dictionary. The threshold values varied from 0.5 to 1. Words with occurrence frequency lower than ten were excluded as not reliable.

An online stock market news dataset has been used for evaluation. It contained 62,478 news items. A polarity dictionary was built automatically with a semi-supervised technique. The method assigned 45 % of correct polarity values for all news items.

The main difference of this approach compared to the supervised and unsupervised learning was in using the bootstrapping approach. The bootstrapping approach is a statistical technique consisting in a very simple procedure based on computer calculations. This approach was used for semi-supervised learning, because it used small amount of labelled data and large amount of unlabelled data [ 34 ].

Zhu et al. [ 73 ] developed a semi-supervised method based on bootstrapping to analyse microblog data. An SVM classifier was trained to classify items as subjective or objective and for polarity classification. The bootstrapping method was automatic classification. This method used a small labelled dataset. Using a corpus with training data, unlabelled data were labelled by the classifier. If a part of samples was integrated into training corpus, bootstrapping can obtain classifier with some labelled data and a large amount of unlabelled data. The features that were selected contained effective characteristics such as word, part-of-speech tags, and emoticon symbols. In order to improve performance, the emoticons have been divided into positive and negative via emoticon lists. The probability to be positive or negative for emoticons was calculated. SVM with default parameters was used for classification of the polarity. The Chinese microblog content was used as a dataset. It was difficult for sentiment analysis because the expression was random. The main problem of this approach was that its accuracy was low. This approach selected different features such as specific symbols and microblogs emoticons set.

Remus et al. [ 44 ] proposed a new approach for semi-supervised German-language sentiment polarity classification. The proposed system was called SentiWS; the dictionary that was used in the SentiWS is freely available online. The weight of entry expression of polarity between –1.0 and +1.0 was calculated. The final stage was to evaluate the performance and accuracy. The part-of-speech tagging was used to build the dictionary, which included positive adverbs, negative adverbs, positive adjectives, negative adjectives, positive nouns, negative nouns, positive verbs, and negative verbs. The SentiWS used several resources to supply words with their semantic orientation. The first resource was the General Inquirer lexicon using Google translator to categorize positive and negative expressions semi-automatically in the German language. The reason for using General Inquirer was that it was widely accepted. The second resource was co-occurrence analysis of rated reviews. The rated reviews can be tagged from strong positive to strong negative. The co-occurrence is important for domain-dependent terminology. The third resource was the German Collocation Dictionary. This dictionary was able to group words that were collated, which were nouns classified by semantic similarity [ 17 , 27 , 63 ]. The German collation dictionary contains 25,288 semantic groups. The pointwise mutual information has been used to calculate the weight of the polarity. The purpose of using pointwise mutual information was to find semantic information from semantic association.

In order to evaluate the method, 2000 sentences were selected from a corpus and manually divided into positive, negative, and neutral. This approach used the General Inquirer lexicon that was not used in other approaches. General Inquirer includes words categorized into positive and negative. Since it has been translated, the translation process may have affected the quality of the process.

This approach contains suffered from missing and ambiguous words, which had a negative effect on the performance [ 44 ].

Guan and Yang [ 29 ] proposed a technique for sentiment analysis in Chinese microblogs in order to develop an approach in analysis of characters for Chinese microblogs compared to traditional online media such as blogs. The purpose of this study was to classify opinion in microblogs into positive or negative. The method required a pre-processing step such as word segmentation and noise symbol filtering. The classification features needed to be extracted for every individual message, and finally, self-training was used to classify the unlabelled data. One of the methods for the semi-supervised learning is self-training, where labelled and unlabelled data together are used as a training corpus. Self-training is a wrapper algorithm that is used in the supervised methods. First, it begins with training labelled data; when the iterations start, it is able to determine unlabelled data that exist with labelled data. The overall performance of the self-training sentiment classification for Chinese is not good compared with supervised learning methods. Reverse self-training is a method that has been used for selecting strategy in labelled and unlabelled learning. The performance can be improved if some of the samples, where the classifier detects low certainty for associated polarity, are labelled. The technique used in the reverse self-training is simple: the classifier determines the unlabelled data, reverses data, and finally adds the most confident unlabelled data and less confident reverse data to the training set. Once this process is completed, the classifier is able to cover the decision space without many majority class samples.

For the evaluation of the Chinese microblogs, the NLP and CC2012 datasets have been employed. They contain twenty topics, 2207 subjective, 407 positive, and 1766 negative items. The sentiment lexicon has been used, provided by HowNet that contains 836 positive sentiment words and 1254 negative sentiment words. The precision for self-training was 0.895, recall was 0.667, and F-measure was 0.765. The precision for reversed self-training was 0.919, recall was 0.683, and F-measure was 0.784.

The main difference of this approach from previous approaches was in using specific domain, such as digital product reviews. The sentiment classification of microblogs contains multi-domain information. The performance of trained model of domain can be very poor when it shifts to another domain.

Mahyoub et al. [ 30 ] proposed an approach for determining sentiment for Arabic text. This study presented a semi-supervised approach to identify Arabic text sentiment by creating an Arabic sentiment lexicon that was able to assign sentiment scores for Arabic words. The Arabic sentiment lexicon was created using the Arabic WordNet. The authors used a small positive and negative Arabic wordlist as a training set, and the main goal was to use it to determine the polarity of all other words in Arabic WordNet. They proposed a semi-supervised algorithm that used the relations between the Arabic WordNet words to spread the sentiment score. The scores in this study were similar to the SentiWordNet ones: a word could be positive, negative, or neutral. The main difference was in that the score was not normalized to be between 0.0 and 1.0. In total, 7500 words were processed, and about 6000 of these words were found to be neutral, while 800 words were found to be positive, and 600 to be negative. The constructed Arabic sentiment lexicon was evaluated using a number of Arabic sentiment corpora, namely the OCA corpus, which contains movie reviews and a book review corpus. A machine learning classifier was applied using both vector space model [ 62 ] and Naïve Bayes model. The technique achieved 96 % classification accuracy. However, its limitation was that most of the Arabic reviews and tweets contained informal words, as well as words in different dialects and special regional words that have not been considered in this study.

Comparison of Multilingual Sentiment Analysis Techniques

In the previous sections, we have described a variety of sentiment analysis techniques. For practical applications and for research work, one would need to choose the best performing approaches. However, direct comparison between those systems is difficult due to a number of factors. First, the original authors report the results on very different datasets, which makes comparison between the reported figures not fair. More importantly, the original authors describe their systems with varying degree of detail and accuracy, which makes the reported results not always reproducible. With this, even if a method showed excellent results in the authors’ own evaluation, lack of detail in their publication may render it unusable in practice for the readers.

To address these two difficulties, we implemented the methods reported in the papers discussed above and applied them to two datasets. In our implementation, we did our best to follow as exactly as possible the descriptions in the respective papers; however, in some cases due to lack of explanations, we had to guess what the authors meant, or had to omit parts of the method when the original paper gave too little clue as to what was meant to be done. For example, Tan and Zhang [ 54 ] mentioned that they implemented four traditional feature selection methods, but did not provide any details on how they were implemented; we had to implement some feature selection approach, which might not coincide with the one used by Tan and Zhang [ 54 ]. Similarly, the original authors often did not specify the tools they used to implement their approaches; in our experiments, we used Java and Python.

With this, our quantitative comparison reflects not the value of the methods as known only to their authors and implemented on their own computers not accessible by anybody else, but the real value of the information on those methods available to the research community through the respective publications—which, unfortunately, is far too often not the same.

In such uniform implementation, we also observed advantages and disadvantages of the methods, such as simplicity of implementation and extensibility, which allowed for qualitative comparison of the methods.

We realize, however, that comparison of approaches on a common dataset may not be fair to the approaches designed for a specific application domain. For example, the system by Shi and Li [ 47 ] was designed for a hotel reviews dataset, which can explain why in our experiments its performance was much lower than the one reported by its authors.

Quantitative Comparison on Common Data

We evaluated the performance of a number of existing multilingual sentiment analysis approaches on two popular datasets that reflect two important application domains of sentiment analysis: a movie review dataset and a product review dataset. As the movie reviews dataset, we used the Cornell movie review data [ 37 ], which contains 1000 reviews labelled as positive and 1000 labelled as negative. As the product reviews dataset, we used the Blitzer dataset [ 9 ], which contains Amazon product reviews. Specifically, we used the reviews on books and DVDs. These datasets, publicly available online, are most commonly used by researches [ 37 ]. On the other hand, these datasets are different enough to test the methods on robustness.

We implemented existing approaches using various tools and programming languages, such as LibSVM, WEKA, Java, and Python. The results of our evaluation of the selected multilingual sentiment analysis approaches are shown in Table  1 . The table shows the accuracy achieved on both datasets, with the better of the two results emphasized. The approaches are presented in the order of the best accuracy they showed in our experiments. The table also shows the accuracy that the authors reported in their corresponding papers.

Performance comparison of state-of-the-art approaches shows a difference between the accuracy reported by the respective authors and the accuracy obtained in our experiments. We attribute this mainly to the lack of detail in the original publications, which did not allow for exact reproduction of the techniques in our implementation.

In some cases, the reported results are not comparable with our results because we used different experiment settings, tools, and datasets. For example, Boiy and Moens [ 10 ] reported 86.35 % accuracy, but we obtained 67.40 %; Habernal et al. [ 23 ] reported 64 % accuracy, but we obtained 59.75 %. Researchers used different datasets, such as the stock market, movie reviews, product reviews, hotel reviews, and tweets. Tan and Zhang [ 54 ] used an online reviews dataset to evaluate the performance of their approach, while we used product reviews, i.e. the Blitzer dataset; Shi and Li [ 47 ] used a hotel reviews dataset, while we used movie reviews, i.e. the Cornell movie review dataset.

In addition, we employed different linguistic resources. For example, Singh et al. [ 52 ] used SentiWordNet, and Mahyoub et al. [ 30 ] and Al-Ayyoub et al. [ 3 ] used Arabic linguistic resources, while we used SentiWordNet. Some of these approaches listed here were developed for languages other than English. For example, Tan and Zhang [ 54 ] developed their approach for sentiment analysis of Chinese texts, and Habernal et al. [ 23 ] for sentiment analysis in Czech. We used an English dataset to evaluate the performance of these approaches. Further, the state-of-the-art approaches employed different tools to build machine learning classifiers, such as SVM Light , WEKA, and LibSVM, while we employed LibSVM and Weka for our experiments.

In our experiments, the approach by Singh et al. [ 52 ] showed the best accuracy. Our experiments also suggest that the SVM classifier usually outperforms by a large margin all other classifiers.

Qualitative Comparison

Different researchers used different experimental settings. Tan and Zhang [ 54 ] selected traditional features such as document frequency, information gain, mutual information, and Chi-square test, while Habernal et al. [ 23 ] used n-grams, emoticons, and part-of-speech features. Some of these features include multiword expressions, which suffer from the data sparsity problem. Due to this, such features are not effective and contain a large amount of noise [ 65 ]. Syntactic n-grams have performed better than traditional linear n-grams because they are more informative and less arbitrary. These features are also more accurate in comparison with information gain, Chi-square test, and n-grams [ 1 , 48 , 50 ].

The approach proposed by Singh et al. [ 52 ] obtained good accuracy, though it requires extensive calculation of many PMI values, which is computationally expensive. The approach proposed by Mizumoto et al. [ 34 ] is only applicable to stock market news; it showed very low accuracy with other types of datasets such as movie reviews or product reviews.

The sentiment analysis approaches have different advantages and disadvantages. Table  2 summarizes the advantages and disadvantages of different approaches.

Conclusions

We gave an overview of state-of-the-art multilingual sentiment analysis methods. We described data pre-processing, typical features, and the main resources used for multilingual sentiment analysis. Then, we discussed different approaches applied by their authors to English and other languages. We have classified these approaches into corpus-based, lexicon-based, and hybrid ones.

The real value of technique for the research community corresponds to the results that can be reproduced with it, not in the results its original authors reportedly obtained with it. To evaluate this real value, we have implemented eleven approaches as closely as we could basing on their descriptions in the original papers, and tested them on the same two corpora. In the majority of the cases, we obtained lower results than those reported by their corresponding authors. We attribute this mainly to the incompleteness of their descriptions in the original papers. In some cases, though, the methods were developed for a specific domain, so in such cases comparison on our test corpora may not be fair. A lesson learnt was that for a method to be useful for the research community, authors should provide sufficient detail to allow its correct implementation by the reader.

According to our results, the approach proposed by Singh et al. [ 52 ] outperforms other approaches. However, this approach is computationally expensive and has been tested only on English-language data. The least accurate approaches of those that we considered were the ones proposed by Zhu et al. [ 73 ], Habernal et al. [ 23 ], and Mizumoto et al. [ 34 ].

The main problem of multilingual sentiment analysis is the lack of lexical resources [ 18 ]. In our future work, we are planning to develop a multilingual corpus, which will include Persian, Arabic, Turkish, and English data, and compare different methods by applying them to this corpus.

Agarwal B, Poria S, Mittal N, Gelbukh A, Hussain A. Concept-level sentiment analysis with dependency-based semantic parsing: a novel approach. Cogn Comput. 2015;7(4):487–99.

Article   Google Scholar  

Ahmad K, Cheng D, Almas Y. Multi-lingual sentiment analysis of financial news streams. In: Proceedings of the 1st international conference on grid in finance; 2006.

Al-Ayyoub M, Essa SB, Alsmadi I. Lexicon-based sentiment analysis of arabic tweets. Int J Soc Netw Min. 2015;2:101–14.

Balahur A, Turchi M. Multilingual sentiment analysis using machine translation? In: Proceedings of the 3rd workshop in computational approaches to subjectivity and sentiment analysis. Association for Computational Linguistics; 2012, p. 52–60.

Balahur A, Turchi M. Improving sentiment analysis in twitter using multilingual machine translated data. In: RANLP; 2013, p. 49–55.

Bautin M, Vijayarenu L, Skiena S. International sentiment analysis for news and blogs. In: ICWSM; 2008.

Berger AL, Pietra VJD, Pietra SAD. A maximum entropy approach to natural language processing. Comput Linguist. 1996;22:39–71.

Google Scholar  

Bhaskar J, Sruthi K, Nedungadi P. Enhanced sentiment analysis of informal textual communication in social media by considering objective words and intensifiers. In: Recent advances and innovations in engineering (ICRAIE), 2014. IEEE; 2014, p. 1–6.

Blitzer J, Dredze M, Pereira F, et al. Biographies, bollywood, boom-boxes and blenders: domain adaptation for sentiment classification. In: ACL; 2007, p. 440–47.

Boiy E, Moens M-F. A machine learning approach to sentiment analysis in multilingual Web texts. Inf. Retr. 2009;12:526–58.

Cambria E, Speer R, Havasi C, Hussain A. SenticNet: a publicly available semantic resource for opinion mining. In: AAAI fall symposium: commonsense knowledge. 2010, p. 02.

Carroll TZJ. Unsupervised classification of sentiment and objectivity in Chinese text. In: Third international joint conference on natural language processing. 2008, p. 304.

Chang C-C, Lin C-J. LIBSVM: a library for support vector machines. ACM Trans Intell Syst Technol TIST. 2011;2:27.

Chikersal P, Poria S, Cambria E. SeNTU: sentiment analysis of tweets by combining a rule-based classifier with supervised learning. In: Proceedings of the international workshop on semantic evaluation (SemEval 2015). 2015.

Croft WB, Lafferty J. Language modeling for information retrieval. Berlin: Springer; 2003.

Book   Google Scholar  

Cruz-Garcia IO, Gelbukh A, Sidorov G. Implicit aspect indicator extraction for aspect based opinion mining. Int J Comput Linguist Appl. 2014;5(2):135–52.

Das N, Ghosh S, Gonçalves T, Quaresma P. Comparison of different graph distance metrics for semantic text based classification. Polibits. 2014;49:51–7.

Denecke K. Using sentiwordnet for multilingual sentiment analysis. In: IEEE 24th international data engineering workshop, 2008. ICDEW 2008. IEEE; 2008, p. 507–12.

Duwairi RM, Qarqaz I (2014) Arabic sentiment analysis using supervised classification. In: 2014 international conference on future internet of things and cloud (FiCloud). IEEE; 2014.

Evans DK, Ku L-W, Seki Y, Chen H-H, Kando N. Opinion analysis across languages: an overview of and observations from the NTCIR6 opinion analysis pilot task. In: Applications of fuzzy sets theory. Berlin, Heidelberg: Springer; 2007, p. 456–63.

Ghorbel H, Jacot D. Further experiments in sentiment analysis of french movie reviews. In: Advances in Intelligent Web Mastering–3. Berlin, Heidelberg: Springer; 2011, p. 19–28.

Ghosh M, Kar A. Unsupervised linguistic approach for sentiment classification from online reviews using SentiWordNet 3.0. Int J Eng Res Technol. 2013.

Habernal I, Ptácek T, Steinberger J. Sentiment analysis in Czech social media using supervised machine learning. In: Proceedings of the 4th workshop on computational approaches to subjectivity, sentiment and social media analysis. 2013, p. 65–74.

He Y, Zhou D. Self-training from labeled features for sentiment analysis. Inf Process Manag. 2011;47:606–16.

Holmes G, Donkin A, Witten IH. Weka: a machine learning workbench. In: Proceedings of the 1994 Second Australian and New Zealand conference on intelligent information systems. IEEE; 1994, p. 357–61.

Hu M, Liu B. Mining and summarizing customer reviews. In: Proceedings of the tenth ACM SIGKDD international conference on knowledge discovery and data mining. ACM; 2004, p. 168–77.

Jimenez S, Gonzalez FA, Gelbukh A. Soft cardinality in semantic text processing: experience of the SemEval international competitions. Polibits. 2015;51:63–72.

Liu B. Sentiment analysis: mining opinions, sentiments, and emotions. Cambridge: Cambridge University Press; 2015.

Liu Z, Dong X, Guan Y, Yang J. Reserved self-training: a semi-supervised sentiment classification method for Chinese microblogs. In: Proceedings of IJCNLP; 2013.

Mahyoub FHH, Siddiqui MA, Dahab MY. Building an Arabic sentiment lexicon using semi-supervised learning. J King Saud Univ Comput Inf Sci. 2014;26(4):417–24.

Manning CD, Raghavan P, Schütze H. Introduction to information retrieval. Cambridge: Cambridge University Press; 2008.

Medhat W, Hassan A, Korashy H. Sentiment analysis algorithms and applications: a survey. Ain Shams Eng. J. 2014;5:1093–113.

Mirchev U, Last M. Multi-document summarization by extended graph text representation and importance refinement. Innov Doc Summ Tech Revolut Knowl Underst Revolut Knowl Underst. 2014; 28.

Mizumoto K, Yanagimoto H, Yoshioka M. Sentiment analysis of stock market news with semi-supervised learning. In: 2012 IEEE/ACIS 11th international conference on computer and information science (ICIS). IEEE, 2012; p. 325–28.

Morency L-P, Mihalcea R, Doshi P. Towards multimodal sentiment analysis: harvesting opinions from the web. In: Proceedings of the 13th international conference on multimodal interfaces. ACM; 2011, p. 169–76.

Narayanan V, Arora I, Bhatia A. Fast and accurate sentiment classification using an enhanced Naive Bayes model. In: Intelligent data engineering and automated learning–IDEAL 2013. Berlin: Springer; 2013, p. 194–201.

Pang B, Lee L. A sentimental education: sentiment analysis using subjectivity summarization based on minimum cuts. In: Proceedings of the 42nd annual meeting on association for computational linguistics. Association for Computational Linguistics; 2004, p. 271.

Pang B, Lee L, Vaithyanathan S. Thumbs up?: sentiment classification using machine learning techniques. In: Proceedings of the ACL-02 conference on empirical methods in natural language processing, vol 10. Association for Computational Linguistics, 2002; p. 79–86.

Posadas-Durán J-P, Markov I, Gómez-Adorno H, Sidorov G, Batyrshin I, Gelbukh A, Pichardo-Lagunas O. Syntactic N-grams as features for the author profiling task. Notebook for PAN at CLEF 2015. CEUR Workshop Proceedings 1391; 2015.

Raina P. Sentiment analysis in news articles using sentic computing. In: 2013 IEEE 13th International Conference on Data Mining Workshops (ICDMW). IEEE; 2013, p. 959–62.

Rajagopal D, Cambria E, Olsher D, Kwok K. A graph-based approach to commonsense concept extraction and semantic similarity detection. In: Proceedings of the 22nd international conference on world wide web companion. International World Wide Web Conferences Steering Committee; 2013, p. 565–70.

Ravi K, Ravi V. A survey on opinion mining and sentiment analysis: tasks, approaches and applications. Knowl Based Syst. 2015.

Read J. Recognising affect in text using pointwise-mutual information. Unpubl. M Sc Diss. Univ. Sussex UK; 2004.

Remus R, Quasthoff U, Heyer G. SentiWS-a publicly available German-language resource for sentiment Analysis. In: LREC. 2010.

Saraee M, Bagheri A. Feature selection methods in Persian sentiment analysis. In: Natural Language Processing and Information Systems. Springer; 2013, p. 303–308.

Seki Y, Evans DK, Ku L-W, Sun L, Chen H-H, Kando N, Lin C-Y. Overview of multilingual opinion analysis task at NTCIR-7. In: Proceedings of the 7th NTCIR workshop meeting on evaluation of information access technologies: information retrieval, question answering, and cross-lingual information access. 2008, p. 185–203.

Shi H-X, Li X-J. A sentiment analysis model for hotel reviews based on supervised learning. In: 2011 international conference on machine learning and cybernetics (ICMLC). IEEE; 2011, p. 950–54.

Sidorov G. Should syntactic n-grams contain names of syntactic relations? Int J Comput Linguist Appl. 2014;5(2):25–47.

Sidorov G, Miranda-Jiménez S, Viveros-Jiménez F, Gelbukh A, Castro-Sánchez N, Velásquez F, Díaz-Rangel I, Suárez-Guerra S, Treviño A, Gordon J. Empirical study of opinion mining in Spanish tweets. MICAI 2012. Lect Notes Comput Sci. 2012;7629:1–14.

Sidorov G, Velasquez F, Stamatatos E, Gelbukh A, Chanona-Hernández L. Syntactic n-grams as machine learning features for natural language processing. Expert Syst Appl. 2014;41(3):853–60.

Sindhwani V, Melville P. Document-word co-regularization for semi-supervised sentiment analysis. In: Eighth IEEE international conference on data mining, 2008. ICDM’08. IEEE; 2008, p. 1025–30.

Singh VK, Piryani R, Uddin A, Waila P, et al. Sentiment analysis of textual reviews; Evaluating machine learning, unsupervised and SentiWordNet approaches. In: 2013 5th international conference on knowledge and smart technology (KST). IEEE; 2013, p. 122–27.

Stone PJ, Dunphy DC, Smith MS. The general inquirer: a computer approach to content analysis; 1966.

Tan S, Zhang J. An empirical study of sentiment analysis for Chinese documents. Expert Syst Appl. 2008;34:2622–9.

Tong S, Koller D. Support vector machine active learning with applications to text classification. J Mach Learn Res. 2002;2:45–66.

Tromp E. Multilingual sentiment analysis on social media. Master’s Thesis, Dep. Math. Comput. Sci. Eindh. Univ. Technol.; 2011.

Wan X. Using bilingual knowledge and ensemble techniques for unsupervised Chinese sentiment analysis. In: Proceedings of the conference on empirical methods in natural language processing. Association for Computational Linguistics; 2008, p. 553–61.

Wang S, Li D, Song X, Wei Y, Li H. A feature selection method based on improved fisher’s discriminant ratio for text sentiment classification. Expert Syst Appl. 2011;38:8696–702.

Wiebe J, Mihalcea R. Word sense and subjectivity. In: Proceedings of the 21st international conference on computational linguistics and the 44th annual meeting of the Association for Computational Linguistics. Association for Computational Linguistics; 2006, p. 1065–72.

Wong K-F, Xia Y, Xu R, Wu M, Li W. Pattern-based opinion mining for stock market trend prediction. Int J Comput Process Orient Lang. 2008;21(4):347–61.

Xia Y, Cambria E, Hussain A, Zhao H. Word polarity disambiguation using Bayesian model and opinion-level features. Cogn Comput. 2015;7(3):369–80.

Xia Y, Wang L, Wong K-F. Sentiment vector space model for lyric-based song sentiment classification. Int J Comput Process Orient Lang. 2008;21(4):331–45.

Xia Y, Zhao T, Yao J, Jin P. Measuring Chinese-English cross-lingual word similarity with HowNet and parallel corpus. In: Computational linguistics and intelligent text processing, 12th international conference, CICLing 2011, vol. 2. 2011, p. 221–33.

Xia Y, Li X, Cambria E, Hussain A. A localization toolkit for SenticNet. In: 2014 IEEE international conference on data mining workshop (ICDMW). 2014, p. 403–8.

Xia R, Zong C. Exploring the use of word relation features for sentiment classification. In: Proceedings of the 23rd international conference on computational linguistics: posters. Association for Computational Linguistics; 2010, p. 1336–44.

Xu Y, Jones GJ, Li J, Wang B, Sun C. A study on mutual information-based feature selection for text categorization. J Comput Inf Syst. 2007;3:1007–12.

Xu R, Wong K-F, Lu Q, Xia Y, Li W. Learning knowledge from relevant webpage for opinion analysis. In: IEEE/WIC/ACM international conference on web intelligence and intelligent agent technology, WI-IAT ‘08. 2008, p. 307–13.

Yang Y, Pedersen JO. A comparative study on feature selection in text categorization. In: ICML; 1997, p. 412–20.

Ye Q, Shi W, Li Y. Sentiment classification for movie reviews in Chinese by improved semantic oriented approach. In: Proceedings of the 39th annual Hawaii international conference on system sciences, HICSS’06. IEEE; 2006, p. 53b–53b.

Ye Q, Zhang Z, Law R. Sentiment classification of online reviews to travel destinations by supervised machine learning approaches. Expert Syst Appl. 2009;36:6527–35.

Zagibalov T, Carroll J. Automatic seed word selection for unsupervised sentiment classification of Chinese text. In: Proceedings of the 22nd international conference on computational linguistics, vol. 1. Association for Computational Linguistics; 2008, p. 1073–80.

Zhang Z-Q, Li Y-J, Ye Q, Law R. Sentiment classification for Chinese product reviews using an unsupervised Internet-based method. In: International conference on management science and engineering, 2008. ICMSE 2008. 15th annual conference proceedings. IEEE; 2008, p. 3–9.

Zhu S, Xu B, Zheng D, Zhao T. Chinese microblog sentiment analysis based on semi-supervised learning. In: Semantic web and web science. New York: Springer; 2013, p. 325–31.

Download references

Author information

Authors and affiliations.

Department of Computing Science and Mathematics, University of Stirling, Stirling, FK9 4LA, Scotland, UK

Kia Dashtipour & Amir Hussain

Temasek Laboratory, Nanyang Technological University, Singapore, Singapore

Soujanya Poria

School of Computer Engineering, Nanyang Technological University, Singapore, Singapore

Erik Cambria

Taibah University, Madina, Saudi Arabia

Ahmad Y. A. Hawalah

CIC, Instituto Politécnico Nacional, 07738, Mexico City, Mexico

Alexander Gelbukh

Tsinghua University, Beijing, China

You can also search for this author in PubMed   Google Scholar

Corresponding author

Correspondence to Kia Dashtipour .

Ethics declarations

Conflict of interest.

Kia Dashtipour, Soujanya Poria, Amir Hussain, Erik Cambria, Ahmad Y. A. Hawaleh, Alexander Gelbukh, and Qiang Zhou declare that they have no conflict of interest.

Informed Consent

All procedures followed were in accordance with the ethical standards of the responsible committee on human experimentation (institutional and national) and with the Helsinki Declaration of 1975, as revised in 2008 (5). Additional informed consent was obtained from all patients for which identifying information is included in this article.

Human and Animal Rights

This article does not contain any studies with human or animal subjects performed by the any of the authors.

Rights and permissions

Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License ( http://creativecommons.org/licenses/by/4.0/ ), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.

Reprints and permissions

About this article

Dashtipour, K., Poria, S., Hussain, A. et al. Multilingual Sentiment Analysis: State of the Art and Independent Comparison of Techniques. Cogn Comput 8 , 757–771 (2016). https://doi.org/10.1007/s12559-016-9415-7

Download citation

Received : 18 December 2015

Accepted : 10 May 2016

Published : 01 June 2016

Issue Date : August 2016

DOI : https://doi.org/10.1007/s12559-016-9415-7

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Artificial intelligence
• Natural language processing
• Opinion mining
• Sentic computing
• Sentiment Analysis
• Find a journal
• Publish with us
• Track your research

Help | Advanced Search

Computer Science > Computation and Language

Title: a survey of multilingual models for automatic speech recognition.

Abstract: Although Automatic Speech Recognition (ASR) systems have achieved human-like performance for a few languages, the majority of the world's languages do not have usable systems due to the lack of large speech datasets to train these models. Cross-lingual transfer is an attractive solution to this problem, because low-resource languages can potentially benefit from higher-resource languages either through transfer learning, or being jointly trained in the same multilingual model. The problem of cross-lingual transfer has been well studied in ASR, however, recent advances in Self Supervised Learning are opening up avenues for unlabeled speech data to be used in multilingual ASR models, which can pave the way for improved performance on low-resource languages. In this paper, we survey the state of the art in multilingual ASR models that are built with cross-lingual transfer in mind. We present best practices for building multilingual models from research across diverse languages and techniques, discuss open questions and provide recommendations for future work.

Submission history

Access paper:.

• Other Formats

References & Citations

• Google Scholar
• Semantic Scholar

BibTeX formatted citation

Bibliographic and Citation Tools

Code, data and media associated with this article, recommenders and search tools.

• Institution

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs .

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

• View all journals
• Explore content
• About the journal
• Publish with us
• Sign up for alerts
• CORRESPONDENCE
• 01 August 2018

Make research-paper databases multilingual

• Daniel Prieto 0

Instituto de Investigaciones Biologicas Clemente Estable, Montevideo, Uruguay.

You can also search for this author in PubMed   Google Scholar

It is absurd to put effort and public resources into research that has already been published. This will continue to be a risk as long as papers in non-English journals are not routinely indexed in the international databases (see also J. Lebel and R. McLean Nature 559 , 23–26; 2018 ).

Access options

Access Nature and 54 other Nature Portfolio journals

Get Nature+, our best-value online-access subscription

24,99 € / 30 days

cancel any time

Subscribe to this journal

Receive 51 print issues and online access

185,98 € per year

only 3,65 € per issue

Rent or buy this article

Prices vary by article type

Prices may be subject to local taxes which are calculated during checkout

Nature 560 , 29 (2018)

doi: https://doi.org/10.1038/d41586-018-05844-0

Mount Etna’s spectacular smoke rings and more — April’s best science images

News 03 MAY 24

Plagiarism in peer-review reports could be the ‘tip of the iceberg’

Nature Index 01 MAY 24

How reliable is this research? Tool flags papers discussed on PubPeer

News 29 APR 24

A guide to the Nature Index

Nature Index 13 MAR 24

Decoding chromatin states by proteomic profiling of nucleosome readers

Article 06 MAR 24

‘All of Us’ genetics chart stirs unease over controversial depiction of race

News 23 FEB 24

Seeking Global Talents, the International School of Medicine, Zhejiang University

Welcome to apply for all levels of professors based at the International School of Medicine, Zhejiang University.

Yiwu, Zhejiang, China

International School of Medicine, Zhejiang University

Assistant, Associate, or Full Professor

Athens, Georgia

University of Georgia

Associate Professor - Synthetic Biology

Position Summary  We seek an Associate Professor in the department of Synthetic Biology (jcvi.org/research/synthetic-biology). We invite applicatio...

Rockville, Maryland

J. Craig Venter Institute

Associate or Senior Editor (microbial genetics, evolution, and epidemiology) Nature Communications

Job Title: Associate or Senior Editor (microbial genetics, evolution, and epidemiology), Nature Communications Locations: London, New York, Philade...

New York (US)

Springer Nature Ltd

Postdoctoral Research Fellow

Two postdoctoral positions offered at HMS to study the role hypothalamic leptin resistance in control of neuron activity

Boston, Massachusetts (US)

Boston Children’s Hospital-Ozcan Lab

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

Quick links

• Explore articles by subject
• Guide to authors
• Editorial policies

Multilingual machine translation : An analytical study

Ieee account.

• Change Username/Password
• Update Address

Purchase Details

• Payment Options
• Order History
• View Purchased Documents

Profile Information

• Communications Preferences
• Profession and Education
• Technical Interests
• US & Canada: +1 800 678 4333
• Worldwide: +1 732 981 0060
• Contact & Support
• About IEEE Xplore
• Accessibility
• Terms of Use
• Nondiscrimination Policy
• Privacy & Opting Out of Cookies

A not-for-profit organization, IEEE is the world's largest technical professional organization dedicated to advancing technology for the benefit of humanity. © Copyright 2024 IEEE - All rights reserved. Use of this web site signifies your agreement to the terms and conditions.

Navigation Menu

Search code, repositories, users, issues, pull requests..., provide feedback.

We read every piece of feedback, and take your input very seriously.

Saved searches

Use saved searches to filter your results more quickly.

To see all available qualifiers, see our documentation .

• Notifications