Split (1)

train · 2k rows

text stringlengths 41 31.4k
<s>and key phrases depending on semantic and syntactic relationship among them [11].Moreover, the complexity of Bengali grammar and ambiguities in sentence structures raise thechallenges for working in Bengali. However, recent interest of the researchers in the field ofBengali opens a wide variety of scopes to work.Domain dependency is inherited to any language [12]. It is more significant in determiningthe sentiment at concept or sentence level. For instant, a single concept may containmultiple sentiments depending on the context of the text. As a consequence, domainspecific methodologies offer better performance than generalized method in sentiment analysis.Knowledge bases used for NLP tasks still provide very limited coverage for domain specificwords. On the other hand, training data mainly contributes to domain specification of anymethod. Prior to any NLP task parsing the text or sentences is very crucial for the subsequentstages of the work. This solicits an independent parser in Bengali for deconstructing text intophrases hence, extract the concepts.The performance of NLP approaches can vary due to the nature and quality of the data set.Preparation of the training data is the most important process in a classification method as ithighly influences the performance of the classifier. On the other hand, the generation of a goodstandard corpus is highly important to evaluate the method. Though a notable number of goodstandard corpus is available in rich languages like English, Chinese, French; there is hardly anyBengali corpus to be mentioned [13]. Therefore, construction of well annotated corpus is aprerequisite for an NLP task, especially, in Bengali.1.2. PROBLEM DEFINITION 31.2 Problem DefinitionConcept-based sentiment analysis is intended to infer the semantic and affective informationassociated with natural language opinions. There is hardly any existing method for sentimentanalysis at the concept level in Bengali. A concept can take different polarities based on thecontext domain. To address the fact, domain specific methods are preferable in sentimentanalysis. Moreover, the concepts are not just the terms within a sentence rather, refers to agroup of terms, key phrases or even clauses that contain an idea. Therefore, concept extractionfrom the sentence is subtle.Prior to performing the polarity detection, generation of an independent semantic parser forBengali sentences is a requirement. The influence of modifiers and the dependency among theconcepts need to be addressed with proper attention to increase the accuracy of the polaritydetection. Integration of rule based and machine learning method facilitates resolving thelimitations of existing approaches, hence improve the performance.Construction of the data set demands special care as it affects the performance of the NLP taskin subsequent stages of the work. The challenges for Bengali are more for the scarcity of NLPtools needed for annotation of unstructured text data. Annotation of the text with the appropriatePOS tagging to each token in a sentence presents an idle corpus for parse tree generation. In thiscontext, the accuracy of the POS tagger is very crucial as it affects the subsequent stages of thework.1.3 Research Aim and ObjectiveBased on the problem definition and the applications, objectives for the thesis are formulatedwith specific aims. The objectives of this thesis with specific aim</s>
<s>are listed as followings:1. To generate a rule based parse tree generator for the Bengali language to break the text intoclauses and deconstruct the clauses into key phrases that will be used as input to conceptextractor.2. To develop a method to extract the concept from the parse tree that is likely to expresssentiment for polarity detection.3. To derive a classification model to determine the concept valence and hence, determinethe sentiment based on the cognitive and affective information associated with naturallanguage.4. To develop a simple method for polarity detection at sentence level without using so manyNLP resources to detect the sentiment of a sentence.1.4. OVERVIEW OF THE WORK 4Figure 1.1: An overview of the proposed work.1.4 Overview of the WorkIn this work, a classification model is proposed to detect the polarity of the concept; hence,the sentence polarity is determined. Before proceeding to the classification of concepts, amethodology is developed for extracting the concepts from the annotated data set. Experimentaldata set will be generated from popular daily newspapers and online blogs on the field of atargeted domain. The data set is annotated with POS tags. Then, a rule based semantic parserdecomposes the sentence into a parse tree representation. Concepts are extracted from the parsetree using some tree traversal rules developed for this purpose. The polarity of the concept isdetected using the classification model. A pictorial overview of the proposed work is presentedin Figure 1.1. A stepwise enumeration of the proposed methodology is summarized as follows:1. At first, the data acquisition is carried out from available resources on a targeted domain.The annotated data set is generated through tokenization and POS tagging using standardNLP tools like; tokenizer, POS tagger, etc. Depending on the syntactic and semanticrelations among the tokens, a rule based parse tree is generated to decompose a sentenceinto clauses, key phrases, modifiers and other constituents. The parse tree generatedfor each sentence also exhibits the dependency of the modifiers and the tokens withina sentence.1.4. OVERVIEW OF THE WORK 52. Next, the concepts containing sentiment are extracted from the parse tree. The methodfor extracting the concepts is developed relying on the syntactic and semantic associationamong the key phrases and terms within a sentence. Semantic relations are mostlydependent on linguistic morphology and language grammar. Whereas, the syntacticrelations explore POS information and the stemming varieties. Identification of thisrelationship is very effective for detecting the dependency among the tokens andmodifiers.The dependencies of the modifiers being a vital factor in determining the sentence polarityare also be extracted in this phase.3. Concepts are the root constituents for the polarity detection model proposed in this work.AffectSpace is used as a knowledge base in this model to explore the affective reasoningamong the concepts. Linear discriminant analysis is used as a classification model withthree classes named as positive, negative and neutral. The model uses the training dataon the targeted domain to initiate the classifier which in turn, accepts the concepts anddetermines the polarity class.4. Finally, the polarity of the sentence is determined to analyze the relationship of theconcepts and modifiers. A tree</s>
<s>traversal in the reverse order is executed starting from theconcepts to the root of the parse tree of the sentence. Here, the rules are highly sensitiveto the dependencies that were detected in the earlier stage. The modifiers are very crucialto handle as these can alter the polarity of the sentence at any stage of the tree traversal inthe sentence.5. On completion of the classification, the evaluation of the experimental result is carried outusing test data. Performance of each individual method along with the overall outcomeof the model is explained through result analysis. We use the precision and recall metricsfor evaluating the output of the model. In case of parse tree and concept extractor, wecompare the output of our methods with the expected result. Evaluating metrics such as;precision, recall and F1 score values are calculated for each of the three classes. Moreover,the accuracy of every class and the overall accuracy is also determined for the test data toestimate the overall performance of the model. An endeavor is made to explain the causesof exceptions in the performance of the methods presented in this work. In addition, acomparison is made between binary classification and classification with tree classes forbetter analysis of the outcome.The proposed work is based on Bengali language. However, some resources and knowledgebases from English are also been used. As the linguistic morphology and language grammarfor each language are different from each other, therefore, the parse tree generation and conceptextraction are language dependent. On the other hand, the classification method highly relies onthe dictionary meaning of the concepts which can be considered language independent. Since1.5. THESIS CONTRIBUTION AND FINAL OUTCOME 6the meaning of the concept may take different form based on the literature or sentence context,they are domain dependent.1.5 Thesis Contribution and Final OutcomeThe contribution of this thesis is two-fold. Firstly, we explore the language dependencies inthe field of NLP and focus on a less privileged language like Bengali. In this context, ourconcentration focus on sentence parsing, which is language dependent and is a prerequisite toanyNLP task. The parser is capable of generating parse tree through decomposition of a sentencethat is very rare of this kind in Bengali. In addition, the dependencies among the constituentsof the sentence and the modifiers are also detected in the process of parsing. Though the rulebased parser is generated focusing the concept extraction for the proposed classification model,it can be used for other NLP task as well. Thereby, an NLP tool of this kind can facilitate widevarieties of opportunities in the field of Bengali language processing works.Secondly, we introduce the concept as a constituent to be used in polarity detection. Conceptlevel sentiment analysis is a novel integration to the field of Bengali of this kind. In performingthe classification task, we utilize the domain dependencies of the concepts to improve theaccuracy of the polarity detection model. The novelty of this model is the blending of thesemantic and affective information associated with natural language. In addition, the concept-based approaches step away from the blind use of keywords and</s>
<s>word co-occurrence counts butrely on the emotion categorization to detect polarity.The contributions can be more visualized over the possible outcomes of the thesis. A set of toolsand models with their specific capabilities are developed based on this work. The outcome ofthis thesis are listed as follows:1. A semantic parser to generate the parse tree for Bengali sentence with the capabilities toidentify dependencies the constituents of the sentence.2. A concept extractor which can generate the concept containing sentiment from the parsetree.3. A novel model for concept level polarity detection using affect space for the specificdomain of context. The domain can be altered through the generation of training dataonly.4. Finally, A polarity detection method for determining the sentence polarity which utilizesthe interrelationship and dependencies among the concepts and modifiers.Though the classification method is independent of the language, the domain dependenciesimprove the performance of the model to detect the polarity at the concept level. Use of1.6. THESIS OUTLINE 7AffectiveSpace as knowledge base improve the accuracy of the model by exploring the affectiveassociations among the concepts.1.6 Thesis OutlineThis thesis is further organized in four more chapters. Chapter 2 discusses the existing worksrelated to our problem definition. Through a detail literature review, it highlights the scopeof the works based on the limitation found in the state of the art. At the end of this chapter,the research questions for the thesis are enumerated. With a view to better understanding themethodology of the proposed work, we explain some of the NLP fundamentals along with theNLP techniques used for this work in Chapter 3. We also provide a precursory idea on the NLPresources and knowledge bases related to our model. The mathematical models used for theclassification scheme are also explained in this chapter. In Chapter 4, we briefly represent themethods developed in the proposed work. The working procedure for each method is explainedthrough examples and applications. Finally, the experimental analysis is performed in Chapter 5to evaluate the performance of the model using the test data-set. Finally, we will conclude thethesis by summarizing the proposed work and highlighting some future scope of the work.Chapter 2Related WorksEarly works on Natural Language Processing (NLP) [14] are focused on the construction ofNLP tools like parts of speech (POS) tagger, named entity recognizer (NER) etc. With theadvent of World Wide Web, the research on NLP has been diversified to multiple branches liketext summarization, subjectivity detection, opinion mining and sentiment analysis. Opinionmining or sentiment analysis is an active area of study in the field of NLP, to find an automatedway to determine the expression or emotion from text. Research on sentiment analysis has twomain directions: 1. feature-based opinion mining which identifies the features within a reviewsentence to determine the orientation of the opinion, 2. sentiment classification to determinethe polarity of opinionated sentence or document. This chapter will discuss the present stateof the related works on sentiment analysis specially focused on polarity detection. Section 2.1will discuss the different methods of sentiment analysis and Section 2.2 will discuss the relatedworks on semantic parsing and concept extraction.</s>
<s>Works on sentiment analysis in Bengaliwill be discussed in Section 2.3 and scope of the works found through related works will behighlighted in Section 2.4. Finally, an endeavor will be made to formulate the research questionsin Section 2.5 to be address in this work.2.1 Sentiment AnalysisExisting works in sentiment analysis [8] can be broadly grouped as rule based approaches andmachine learning approaches. The former one offers high syntactic sensitivity, but performspoor in context or domain dependent classification whereas; the later is time-consuming andexpensive to manage the large amounts of training data necessary for good performance.Concept-based approaches [15] are the recent evolution in sentiment analysis which are intendedto infer the contextual and affective information associated with natural language opinions.2.1. SENTIMENT ANALYSIS 92.1.1 Rule Based ApproachesMost of the rule based approaches [8] are lexicon based, though some corpus based methodsare also introduced. Lexicon based approaches explore the syntactic and semantic relationshipamong the lexicons within a corpus for sentiment classification at sentence or documentlevel. Lexicon-based approaches mainly rely on sentiment lexicon, i.e. a collection ofpre-generated sentiment terms, phrases and even idioms, developed for traditional genres ofcommunication. These methods identify the sentiment lexicons applying some rules based onlanguage syntax. POS information is most commonly used to exploit syntactic relation amongthe lexicons within a sentence. Polarity of these sentiment words are determined by exploringthe semantic orientation of the word. Adjectives have been employed as features by a numberof researchers [16, 17].One of the earliest proposals for the data-driven prediction of the semantic orientation of wordswas developed for adjectives [18]. Here [19], a method is proposed for determining the semanticorientation of terms through gloss classification [19]. Sentence or document level polarity isdetermined using the semantic association among the lexicons. Therefore, they rely primarilyon the underlying sentiment or opinion words. These sentiment words are generally labeledaccording to their semantic orientation as either positive or negative [20]. NLP resourceslike SentiWordNet [9], Affectnet [21] are very popular to acquire the semantic orientationat word or phrase level. Semantic orientation of sentiment lexicons is annotated as positive,negative or neutral using SentiWordNet for sentiment/ opinion classification in [22]. Sometimes,classification at sentence or document level is simply subjected to keywords and word co-occurrence counts within the corpus. In [23], a lexicon-based method is proposed to determinethe orientation of opinion bearing words in a review sentence. The work basically counts thepositive and negative words in the corpus and classifies the sentence as positive if the numberof positive words is more than the number of negative words, else classifies as negative. Amore sophisticated method based on relaxation labeling has been introduced in [24]. A lexicon-enhanced method for the sentiment analysis of user generated reviews has been introducedin [1] which incorporates many NLP tools and resources along with the rule-based classificationscheme (Figure 2.1).Rule based approaches have some major shortcomings. Firstly, it cannot handle the contextand domain dependent words. The accuracy of sentiment classification can be influenced bythe context of the items to which it is applied [25]. This is because, the same word can takedifferent sentiment</s>
<s>polarity in different domains or contexts. For instance, an observation waspresented in [26] that “unpredictable” is a positive description for a movie plot, but a negativedescription for a car’s steering abilities. Without prior knowledge of context, there is probablyno way to determine the semantic orientation of a context dependent opinion word by looking atonly the word itself. Obtaining such knowledge on the huge number of opinion words froma domain expert or user is also not scalable. A holistic lexicon-based approach is propose2.1. SENTIMENT ANALYSIS 10Figure 2.1: Rule based classification scheme integrated with different NLP resources(collected from [1]).to deal with the limitations of lexicon based approaches in [27]. Contrary to locked in thecurrent sentence, it exploits the external information and evidence in other sentences withinthe document. It also infers the linguistic attributes in natural language expressions to inferorientations of opinion words. Thereby, they also incorporate the semantic relations amongthe opinion words within the corpus. Secondly, occurrences of multiple conflicting words in asentence raise dubiety in determining sentence polarity. For instance, the sentence, “he is honestbut crabby”, contains two conflicting words and makes the sentence polarity inexplicable. Oneof the pertinent solutions to address these problems is machine learning approaches.2.1.2 Machine Learning ApproachesSeveral works on sentiment analysis and opinion mining have been done towards using machinelearning approaches. These approaches are popular for wide range of coverage to the domain orcontext specific words. They also reduce the syntactic and semantic sensitivity of the lexiconswithin the corpus and improve the performance in sentence or document level. On the contrary,large amount of training data is required for good performance. Therefore, learning the model2.1. SENTIMENT ANALYSIS 11is time consuming and expensive for any NLP task. Some machine learning methods likeNaive Bayes Classifier (NBC), Conditional Random Field (CRF), and Support Vector Machine(SVM) have been used for sentiment classification. NBC is a probabilistic classification methodbased on Bayes’ theorem and used for both classifications as well as training purposes. SVManalyzes data and defines decision boundaries by having hyper-planes. In binary classificationproblem, the hyper-plane separates the document vector in one class from other class. CRF is adiscriminating model, which can easily integrate various features.The aspect of sentiment classification is categorization with positive and negative sentimentsin [28]. Three different machine learning algorithms have been applied, such as: NBC, SVM,and Maximum Entropy (ME). In this framework, bag-of-words has been used as features toimplement the machine learning algorithms. According to the analysis, SVM outperforms everyother classifier in predicting the sentiment of a review. A comparison of results obtained byapplying NBC and SVM classification algorithm is presented in [29] and almost the same resultis observed.The proposed work in [30], presents the classification of Chinese comments based on word-to-vector and SVM. It improves the performance of SVM by using word-to-vector tool to clustersimilar features in order to acquire the semantic features in selected domain. In the later part, thelexicon based and POS based feature selection approaches are adopted to generate the trainingdata in a specific domain. Though the method is very useful for domain specific sentimentanalysis,</s>
<s>it fails to handle the influence of modifiers which can change the valence of the featureswithin the domain.Existing works using CRF, demonstrate that it performs well for aspect based sentiment analysis.For example, with CRF, sentiment classification has been performed in sentence and documentlevel in [31]; the work [32] identifies opinion expressions from news-wire documents. CRFis well known for sequence labeling tasks. To conform, the context information of the aspectterms are sequentially leveled in aspect based sentiment analysis [33]. Text representation andadditional positional features of CRF have been improved by MFE-CRF [34] that introducesMulti-Feature Embedding (MFE) clustering based on the CRF to improve the effect of aspectterm extraction.The performance of these approaches can vary due to the nature of the data sets used, thequality of the training data set and the various technical approaches used internally. Sincethe performance of a machine learning methodology heavily depends on the choice of datarepresentation, much effort is given for data analysis to build powerful feature extractor. Variantin mathematical model can also improve the performance of these methodologies. Recently,fuzzy logic and deep learning [35] approaches are emerging as powerful computational modelsthat discover intricate semantic representation of text automatically from data without featureengineering.2.1. SENTIMENT ANALYSIS 122.1.3 Concept Based ApproachesOpinions and sentiments do not occur only at document-level, nor are they limited to a singlevalence or target. In recent works, text analysis granularity has been taken down to phraselevel. Rule based methods are efficient for lexicon based sentiment analysis, whereas, atdocument-level, machine learning techniques have reached to a satisfactory level of efficiency.Unfortunately, these approaches are still far from being able to infer the cognitive and affectiveinformation associated with natural language. Approaches of these kinds are mostly dependenton knowledge bases that are still too limited to efficiently process text at document and sentence-level. These methods, however, are not semantically very strong because of their reliance onparts of text in which opinions are explicitly expressed, i.e. verbs, adjectives and adverbs.Moreover, such text analysis granularity might still not be enough as a single sentence or wordsmay contain different opinions about different context or domain. In order to properly infer themeaning of the text within a specified context or domain, a NLP system must be developed toextract concepts rather than specific opinion words or phrases within the sentence.A novel paradigm to sentiment analysis has been introduced in [2] that merges linguistics,common-sense computing, and machine learning for properly deconstructing natural languagetext into concepts and, hence, for improving the accuracy of polarity detection. ConceptNet [36],a freely available large-scale commonsense knowledge base is introduced as an integratednatural-language-processing tool-kit that supports many practical textual-reasoning tasks overreal-world documents. In ConceptNet, the nodes of WordNet are extended from purely lexicalitems (words and simple phrases with atomic meaning) to higher-order compound concepts,which can devise an action verb with its relevant arguments (e.g. ‘buy food’, ‘drive to store’).ConceptNet is a directed graph in which the concepts are represented as nodes and the edges arelabeled with attributes that are sometime composed through common sense and interconnectedthe concepts (Figure 2.2). A linguistic</s>
<s>resource named WordNet-Affect [7] has been presentedfor the lexical representation of affective knowledge from WordNet.Efficiency of concept-level sentiment analysis depends on anticipating the affective valenceof unknown multi-word expressions. In the significant work [3], a new data-set is built upby applying blending technique on ConceptNet and WordNet-Affect to represent as a singlematrix. Truncated singular value decomposition (TSVD) is performed on the matrix fordimension reduction that yield a 50-dimensional vector space, which they namedAffectiveSpace(Figure 2.3). AffectiveSpace is a major step out in concept based sentiment analysis in whichcommon sense and affective knowledge coexist. In AffectiveSpace, different vectors representaffective valence, ways of making binary distinctions among concepts or emotions. Affectivestates of the concepts go from strongly positive to null to strongly negative within the space.Therefore, concepts with the same affective valence are likely to have similar features. Likewise,concepts expressing the same type of emotion tend to position closer to each other in affectivespace. However, similarity or affinity among the concepts in the affective space does not depend2.1. SENTIMENT ANALYSIS 13Figure 2.2: Fragment of a ConceptNet (collected from [2]).Figure 2.3: A sketch of AffectiveSpace [3]. Affectively positive concepts (in the bottom-left corner) and affectively negative concepts (in the up-right corner) are floating in themulti-dimensional vector space.on absolute position of the concepts rather, on their relative position.A new paradigm known as Sentic Computing is presented in [4]. In the Sentic Computing, fourdimensions are taken as basis to classify the affective states: Sensitivity, Attention, Pleasantnessand Aptitude. The transition between different emotional states within the same affective2.1. SENTIMENT ANALYSIS 14Figure 2.4: The 3D model and the net of the Hourglass of Emotions [4]. Since affectivestates go from strongly positive to null to strongly negative, the model assumes anhourglass shape.dimension, when depicted, generates a symmetric inverted bell curve shape. Mapping of thiscurves of possible emotions leads to an hourglass shape (Figure 2.4), hence, the name the modelis hourglass of emotion [37]. In this framework, the vertical dimension represents the intensity ofthe different affective dimensions, while the radial dimension represents emotion categorization.Though the Hourglass model is designed for emotion detection, it also used for polarity detectiontasks. Since polarity is strongly connected to attitudes and feelings, it is defined in terms of thefour affective dimensions in [4], according to the formula:p =i=1Pleasentness(ci) + \|Attention(ci)\| − \|Sensitivity(ci)\|+ Aptitude(ci)(2.1)Where, ci is an input concept, N the total number of concepts. The major observation indetermining the polarity using hourglass emotions through the equation-1 are: firstly, measuringthe emotion is very complex and costly, secondly, emotions like attention and sensitivity alwayscontribute positively and negatively respectively, which may not be true in all cases.The concept based approaches step away from blind use of keywords and word co-occurrencecounts and dependency on dictionary based NLP resources. They allow the aggregation ofconceptual and affective information, but rely on the emotion categorization, for example, hourglass of emotion to detect polarity. Some emotions do not provide any valence, and sometimes,2.2. SENTENCE PARSING AND CONCEPT EXTRACTION 15can be universally positive or negative. This kind of emotion increases the complexity of thepolarity detection</s>
<s>method using hourglass emotions. On the contrary, affective space provide auniform association among the concepts within the space. Therefore, determining the conceptvalence from affective information rather than emotion categorization will provide a simplerand efficient polarity detection model. Moreover, an attempt to find a domain specific featuresubspace of affect space will further enhance class separability.2.2 Sentence Parsing and Concept ExtractionIn the scientific paper [38], a high performance syntactic and semantic dependency parser ispresented. This system takes a sentence as input and performs syntactic and semantic annotationusing the CoNLL 2009 format [39]. The dependency parser is based on Carrerass algorithm [40]and second order spanning trees. Overall performance of the semantic analyzer consists ofseveral sub-tasks, such as: tokenization, lemmatizing, tagging, sentence parsing and dependencydetection. Application of context free grammar rule is very crucial to explore the dependencyamong the tokens.A rule based technique to parse Bengali sentence using context free grammar rules is presentedin [11]. This work analyzes Bengali sentences in three phases. Firstly, lexical analysisphase performs sequential scanning of the characters’ stream and groups them into tokens orlexicons. Secondly, Syntax analysis phase groups the lexicons having a collective meaning.Finally, semantic analyzer ensures that the discrete input components fit together meaningfully.This phase is highly application-dependent and is regulated by the norms and rules of theconcerned natural language. One of the achievements of their work is the capability of handlingthe complex and compound Bengali sentences. However, they still failed to determine thedependency among the sentences as well as modifiers.One of the notable works in Sentic Computing [2] presents a semantic parser with a view tofacilitating concept extraction from the text. The role of the semantic parser is to break textinto clauses and hence deconstruct such clauses into concepts, to be fed later to affect space.Knowledge on the semantics associated with text and some affective information associatedwith such semantics are often sufficient to perform tasks such as emotion recognition and polaritydetection. In this connection, the parser uses the POS information to break the text into clauses,then, in the next step, verb and noun chunks are separated following the Stanford Chunker(Figure 2.5) [41]. Finally, a POS based bi-gram algorithm is applied to extract the conceptimplementing some association rules.The major disadvantage of this parser is the language dependency. Against this backdrop,efficiency of NLP tools like POS tagger plays an important role in the overall performanceof the parser, hence, extraction of concepts. Therefore, it is necessary to generate a separateindependent semantic parser, particularly for a less privileged language like Bengali.2.3. SENTIMENT ANALYSIS IN BENGALI 16Figure 2.5: Parse Tree generated by Stanford Parser [5].2.3 Sentiment Analysis in BengaliBengali is the national language of Bangladesh and the second most spoken language in Indiaas well as the primary language of West Bengal. It is also the sixth most spoken language inthe world [10]. But, there are very few achievements in the field of NLP for Bengali. Withthe advent of World Wide Web, online contents are increasing a lot over the Bengali blogs,newspapers and social media. Researchers have also shown their interest in</s>
<s>Bengali NLP taskover the years. Most of the works focus on generation of NLP tools like Bengali POS tagger,NER, sentence parser and to some extent of text categorization. A few number of NLP resourcesare also available, for example: Bengali SentiWordNet, WordNetaffect, stop word list, etc.A computational technique for developing Bengali SentiWordNet is proposed using English-Bengali bilingual dictionary and English sentiment lexicons in [42]. In that endeavour,transformation of the WordNetAffect in Bengali was carried out in [43]. The lists are updatedwith the synsets retrieved from the English SentiWordNet to make adequate number of emotionword entries. A few number of works on development of Bengali parser is also available. Asmentioned earlier, a Bengali sentence parser was presented applying the context free grammarrules in [11]. Some works on subjectivity detection and sentiment classification are alsoavailable, which are mostly learning base computational methods. An effort is made forsubjectivity detection and opinion polarity identification from Bengali news text using CRF [43]and SVM [44] respectively.Though the interest of NLP researchers in Bengali is increasing, there is an acute scarcity ofNLP tools and resources. Complexity of Bengali grammatical structures is a substantial hurdlefor the parsing of Bengali sentences, such as, the verb can be placed in any position in thesentence. Sometimes, more than one or two verbs can form a single verb. For example,2.4. SCOPE OF THE WORK 17the sentence “অিভেযাগ করা যেত পাের ” ([obHidZog kra dZete pare], Can be complained),contains three auxiliary verbs. In this breath, the adjective and adverb can also be positionedas ‘pre’ or ‘post’ modifier to the noun or verb respectively. Heterogeneity among the literature,communicative writing and spoken form is another ambiguity for Bangle language. For instance,the literature form of the sentence “ মাহা দ একজন ভাল ছেল ” ([moHammOd EkdZn bHalOtShele], Mohammad is a good boy) can take the form “ মাহা াদ ছেলিট ভাল ” ([moHammOdtSheleti bHalO], Mohammad is a good boy) in communicative writing. Whereas, the spokenform can be much crisp as “ মাহা াদ ভাল ” ([moHammOd bHalO], Mohammad is a good boy),where the verb is concealed within the sentence. These types of heterogeneity is very complex tohandle in NLP task. Unlike other languages domain dependency is also a challenge in sentimentclassification or polarity detection in Bengali. For example, consider the following sentence:“ কউ ইভ িটিজংেয়র িশকার হেল কােছর থানায় িগেয় িলিখত অিভেযাগ করা উিচত ” ([keu ibHtidZiNerSikar Hole katSher tHanae gie obHidZog kra utSit ], If someone is victim of eve teasing, sheshould complain to nearby police station).In this context, the phrase “অিভেযাগ করা ” ([obHidZog kra], complain against) carries a positivesentiment though the phrase is generally used as negative sense. However, the integration ofsemantic association along with the syntactic parsing can handle those more efficiently.2.4 Scope of the WorkRule based approaches are popular for their accessibility and simplicity in lexicon basedmethods. On the other hand, feature based classifications are efficient in sentence or documentlevel and provide limited coverage to domain dependency. There are some major weaknessesof these methods as discussed in early sections of this chapter.</s>
<s>Firstly, the poor recognitionof opinion words when modifiers are involved in influencing the sentiment valence of thesentence. A modifier can turn the polarity of an opinion word or phrase from neutral tonegative or positive. Secondly, reliance on surface features ignoring the affective associationamong the words within the sentence. Concept based approach can be an effective solution tosentiment analysis resolving the above-mentioned limitations. Concept level sentiment analysisis intended to infer the semantic and affective information associated with natural languageopinions. Concept can be categorized within the affective space generated fromWordNetAffect,and allows the aggregation of conceptual and affective information. Moreover, an attempt to finda domain specific feature subspace of AffectiveSpace will further enhance class separability. Italso offer the scope of building a classification model to determine the concept valence andhence, determine the sentiment associated with natural language.Prior to any NLP task, parsing the sentence is very crucial for the performance in subsequentsteps. Considering the complexity of the Bengali language structure and diversity in semantic2.5. RESEARCH QUESTIONS 18composition, an independent semantic parser is necessary for deconstructing Bengali text intokey phrases. It will also enable to determine the dependency among the key phrases as well asmodifiers for Bengali language, and hence, generate a parse tree. Extraction of concepts fromthe parse tree before performing the classification will make the model more economic as well.To keep it simple and efficient, a rule based method can be developed to extract the conceptsthat are likely to express sentiment for polarity detection from the parse tree.With the rapid expansion of the Internet, human interaction with the web is rising every moment.Users from different corners of the world are generating a huge volume of data over the WorldWide Web. Along with English, varying amounts of data are growing in many other languages,particularly, over the social media sites like Facebook, Twitter and various web portals. Onlineblogs and newspapers in native languages are rapidly gaining popularity as they allow people toshare and express their views about topics in public platform. In this regard, Bengali as the sixthmost spoken language in the world, is also moving forward. Therefore, Bengali newspapers orblogs can be used to generate Bengali corpus for various NLP tasks.2.5 Research QuestionsBringing an end to the discussion on related works, we can formulate the following issues asour research questions to be addressed in this thesis.[RQ1] What is the significance of developing an independent semantic parser for differentlanguages in NLP?Efficiency of any NLP task is highly influenced by the parser to extract the desiredconstituent of the text. The syntactic parser, no doubt, is language dependent as the syntaxfor every language is unique to each other. On the contrary, the uniqueness of semanticsof the languages is inconclusive to the NLP researchers. Related works reveal that: (a) aunified semantic parser can be developed based on the idea of universal language [45] toreduce the complexity of language dependency. (b) an independent semantic parser canbe developed to increase the task efficiency . Some of the researchers come up with theidea to generate independent task specific semantic parser</s>
<s>within the language.[RQ2] What genetic factors should be considered in determining the constituents of the languagefor sentiment analysis and, how can those be adopted to extract from text?Researchers have introduced different levels of text constituents like lexicons, keywords,concepts, key phrases, even clauses and sentences. With their merits and demerits,selection of appropriate constituent of the text has immense importance. Particularly, insentiment analysis where the sentiment or opinion can vary a lot with the level of languagegranularity.2.5. RESEARCH QUESTIONS 19[RQ3] How can the rule based and machine learning methods be integrated to develop anefficient classification model, eliminating the shortcomings of both?From the study of related works it was found that, most of the works are focused oneither rule based or machine learning methods in carrying out sentiment analysis tasks. Itwill be very interesting as well as motivating to improve the efficiency of the sentimentclassifier by integrating bothmethods with their positive contributions. In the same breath,determining a befitting mathematical model is very essential for its performance.We try to overcome the limitations in the state of the art and explore the scope of the work indeveloping the thesis. The subsequent stages of this thesis will try to focus on these researchquestions in developing the methodology, generating the data set, and evaluating the model.Thereby, it enables to answer these questions at the end of this thesis.Chapter 3PreliminariesThis chapter will discuss the preliminary knowledge on NLP which will be helpful to developthe methodology of our proposed work. Section 3.1 will focus on the fundamentals of NLPresearch with the emphasis on defining the basic terminologies and primitives to sentimentanalysis. Section 3.2 and Section 3.3 will provide a brief idea on NLP techniques and resourcesrespectively, used throughout this thesis. Finally, some mathematical model used in thesubsequent chapters to this thesis will be explained in Section 3.4.3.1 NLP FundamentalsThis section will define various terminologies that are widely used in the field of NLP.Knowledge on these topics is primitive for understanding and developing our methodology. Inthe process of discussion, our aim is to disclose the abstractions prevailed in various terms likelexicon, concept, corpus, opinion, sentiment and semantic dependency.3.1.1 LexiconA lexicon is the vocabulary or a coordinate term of the vocabulary of a language or subject. Itcan be referred to as a dictionary that includes or focuses on lexemes. Items in the lexicon arecalled lexemes, or lexical items. In the context of NLP, it can be defined as Definition 1.Definition 1. (Lexicon) A lexicon is the unit semantic constituent in the linguisticanalysis which must contain meaning or expression in a language.It differs from the token in the sense that the lexicon can be one or a set of tokens. Whereas,the tokens are the unit granularity in the textual representation which may not express anymeaning. An ordinary dictionary is an example of a lexicon. As the outline or format of adictionary is presented for human use, it is inappropriate for computational use like NLP. A3.1. NLP FUNDAMENTALS 21notable difficulty is the explications of the meaning of each word are themselves enumeratedin natural</s>
<s>language. Lexicon resources are illustrated in a machine readable format that can beinterpreted by the computational application, such as,WordNet. Instead of a definition,WordNetuses the synonymy to represent the relationships among the words or lexicons, as between shutand close or car and automobile. Synonyms are grouped into unordered sets called synsets.3.1.2 ConceptAccording to the Oxford dictionary [46], “Concept is an idea or a principle that is connected withsomething abstract”. It can be referred to mental representations that are used to discriminatebetween objects, events, relationships, or other states of affairs. In NLP, a concept can be viewedas a single or group of words that incorporate the common sense knowledge with the dictionarymeaning, applying commonsense reasoning to natural language processing and understanding,thereby, disclose the abstraction expressed within the text. In a work of cognitive science [47],Concept has been stated as the constituent of thought, and in principle, the thought is unbounded.It can be formally defined as Definition 2.Definition 2. (Concept) Concept is the linguistic constituent extracted from the sentenceor documents that blend the common sense knowledge with dictionary meaning applyingcommonsense reasoning to infer the semantic and affective information associated withnatural language.Concepts are learned inductively from the sparse and noisy data applying some common senseknowledge bases which are usually developed by exploring the semantic synset and theirapplication in the different context of the language. For instant [36], from the sentence “I gotfired today”, the computer reader would not know what to think. Someone can not infer anysubstantial meaning from the lexicons “got fired”. However, the common sense knowledgebase can reason out something things about “getting fired”; someone gets fired for lack of skill.As a consequence of getting fired, someone will not be able to meet his daily expenses. Likewise,in the sentence “I had a long day”, the concept, “long day” does refer to the length of the day,rather; will infer the knowledge as a hectic day, which causes tiredness.3.1.3 CorpusIn linguistics andNLP, corpus (literally Latin for body) refers to a large and structured set of textsthat are used to do statistical analysis and hypothesis testing, checking occurrences or validatinglinguistic rules within a specific language territory. A formal definition can be as Definition 3.Definition 3. (Corpus) A usually large collection of documents that can be used to inferand validate linguistic rules, as well as to do statistical analysis and hypothesis testing.3.1. NLP FUNDAMENTALS 22Construction of a gold standard annotated corpus is a colossal task which may be referred to dataacquisition phase of any NLP task. The text in a corpus may be annotated with various syntacticand semantic information for specific language which assists in performing the various NLPtask. The performance of the NLP task like sentiment analysis greatly depends on the precisionof the annotation. Several corpora are available for the most widely used languages. To bementioned, Brown Corpus 1 contains a wide collection of texts of different genres includingnewspapers, fiction, scientific text, legal text, and others. Lancaster Oslo Bergen (LOB) corpusis the British interpretation of Brown corpus. As an example, a fragment of an English</s>
<s>BrownCorpus annotated with POS tag is highlighted below:Daniel/np personally/rb led/vbd the/at fight/nn for/in the/at measure/nn,/, which/wdt he/pps had/hvd watered/vbn down/rp considerably/rb since/inits/pp$ rejection/nn by/in two/cd previous/jj Legislatures/nns-tl ,/,in/in a/at public/jj hearing/nn before/in the/at House/nn-tl Committee/nnon/in-tl Revenue/nn-tl and/cc-tl Taxation/nn-tl ./.3.1.4 Semantic DependencyIn the book [48], linguistics dependency is categorized into three major types: syntactic,semantic and morphological dependency. Syntactic dependency understandably depends onsentence structure and language grammar. Few languages like English, French, Chinese,etc. have already achieved some satisfactory level of efficiency in developing syntacticdependency grammar. While working in sentiment analysis, we are mostly interested insemantic dependency. Semantic dependency can be effectively inferred through the conceptionof ‘predicate’ and ‘argument’ of the sentence. Reasonably, the argument is semanticallydependent on the predicate where predicate can be recognized through the language syntax.In some instances, the semantic dependency can overlap the syntactic dependency. However,it can be sometimes opposite or even entirely independent of each other. One of the simplestdefinition of semantic dependency, found in the literature [48] is stated as Definition 4:Definition 4. (Semantic Dependency) The word form w2 is said to be ‘SemanticallyDependent’ on the word form w1 in the given utterance if and only if the meaning of w1is or (includes) a predicate and the meaning of w2 is an argument to this predicate in thisutterance (defined in the book [48]).For example, in the sentence “I saw[w1] allen’s[w2] wife[w3] going[w4] to market[w5]”,w1 doesnot carry semantic dependency to w2, though, it has syntactical dependency shown as w1 → w3.However, the dependency w3 → w2 is overlapping between syntactic and semantic. In the1A tagged corpus of about a million words put together at Brown University during the 1960s and1970s.3.1. NLP FUNDAMENTALS 23context of sentiment analysis, concept can be extracted exploring these dependencies, such as,the dependency w3 → w2 can form a concept ‘allen-wife’. Likewise, w4 → w5 can forma concept ‘go-market’[c2]. In the same context, dependencies among the concept can also begenerated as ‘go-market’[c2] semantically depends on ‘allen-wife’[c1] not onw1 and representedas c2 → c1.One of the popular semantic dependency Extraction tool for English is Stanford dependencydeveloped by “Stanford Natural Language Processing Group” 2. Stanford dependency finds thedependency from the sentence discussed in the previous example as followings:[nsubj](saw-2, I-1)[nmod:poss](wife-5, allen-3)[dobj](saw-2, wife-5)[acl](wife-5, going-6)[nmod](going-6, market-8)Here, dependencies are extracted through textual relation depending on the grammatical rules.3.1.5 OpinionOpinion is one’s views or judgment on a specific aspect of a thing or object. It is mostly usedagainst review of product or things. InNLP, opinion has structured representationwith its severalcomponents extracted from unstructured text. Definition 5 presents the formal definition ofopinion as stated in [8].Definition 5. (Opinion) An opinion (or regular opinion) is a quintuple, (ei; aij; ooijkl;hk; tl), where ei is the name of an entity, aij is an aspect of ei, ooijkl is the orientation ofthe opinion about aspect aij of entity ei, hk is the opinion holder, and tl is the time whenthe opinion is expressed by hk. The opinion orientation ooijkl can be positive, negativeor neutral, or be expressed with different strength/intensity levels. When an opinion ison the</s>
<s>entity itself as a whole, we use the special aspect GENERAL to denote it[Stated at[8]].In NLP, an opinion generally consists of five components mentioned in Definition 5. However,it can vary depending on the application. For example, sometimes, it is not necessary to knowthe opinion holder if the task is to summarize opinion from different people. On the contrary,sometimes, we want to know more information like the sex of the opinion holder. One of themost effective contributions of the definition is its capability to present a structured text fromthe unstructured one. Thereby, the NLP task like sentiment analysis, opinion summarizationbecome more simple and efficient.2The Natural Language Processing Group at Stanford University is a team of faculty, postdocs,programmers and students who work together on algorithms that allow computers to process andunderstand human languages.3.2. NLP TECHNIQUES 243.1.6 SentimentNot surprisingly, there is some confusion among researchers and professionals in differentiatingbetween opinion and sentiment. In the sense of dictionary meaning, these two are almostcomplementary to each other and sometimes used as a synonym. However, distinct differencescan be found in its applications in NLP, where, Sentiment referred to attitude, thoughts or moodof a text and opinion is defined as view or judgment on a specific aspect of a matter.In the context of NLP, opinion mining extracts and analyze people’s opinion about an entitythrough converting the unstructured text into structured one while sentiment analysis search forthe sentiment words/expression in a text and then analyze it. Sentiment analysis most preciselydepends on the identification of sentiment or opinion bearing words and their affective relationto finding the mood of the text. Then, various classification methods are used to reach a decisionabout the valence of the mood of the text. Sentiment analysis can be carried out in document,sentence or even phrase level.3.2 NLP TechniquesNLP task requires a series of subtasks to perform, most of which are common to all. Varioustools have been developed using different techniques for performing NLP task in differentstages of language evaluation. This section will discuss different integrals tools for NLP andtheir techniques with a view to finding the suitability to use in our proposed methodology forsentiment analysis.3.2.1 TokenizationThe first step to any NLP task is the tokenization of the sentences. Tokenization is a fundamentaltechnique to split a sentence or document into tokens. Tokens are the smallest constituent of thetext of a language, which, can be used later on to form the semantic constituents of the languagesuch as, keywords, phrases, and concepts. It can be formally defined as definition 6.Definition 6. (Tokenization) Tokenization is the process of splitting up the text into unitscalled tokens. The tokens may be words or number or punctuation mark. Tokenizationdoes this task by locating word boundaries. The terminating point of a word and thebeginning of the next word is called word boundaries. Tokenization is also known as wordsegmentation (collected from [49]).The process of tokenization relies on language structure and conventions. For example, tokens oflanguage such as English, Bengali are separated by space delimiter. Tokens are not necessarilyalways to be meaningful words, as the</s>
<s>punctuation marks, articles, and prepositions, such as3.2. NLP TECHNIQUES 25‘the’, ‘a’, ‘and’ etc. These tokens are sometimes, called stop words and removed from thecorpus, as they do not generally contribute greatly to the semantic evaluation of the content.As a fundamental technique, many tokenization tools are available, such as Stanford Tokenizer,OpenNLP, Natural Language Toolkit (NLTK) tokenizer etc.3.2.2 POS TaggingPOS tagging is the technique that analyzes the syntactic and semantic information depending onlanguage grammar. According to Wikipedia, “In corpus linguistics, POS tagging is the processof marking up a word in a text (corpus) as corresponding to a particular part of speech, basedon both its definition and its context i.e., its relationship with adjacent and related words in aphrase, sentence, or paragraph”. POS tagging is also a sequential labeling problem. It plays animportant role to distinguish opinion or sentiment words within a sentence as opinion words areusually adjectives, adverbs, and sometimes nouns or combination of nouns.Though the POS tagging technique mostly relies on the semantic association among the wordswithin the sentence. It can be important for semantic categorization of the tokens as well. Foran instant, the word ‘can’ may take several meanings based on its relative use in the sentence.It may be used as an auxiliary verb to form a question, another form, may be a container forholding food or liquid. Another variation is being a verb denoting the ability to do something.Distinguishing such a specific meaning facilitates to explore the sentiment of the word in thecontext of the text more efficiently. Most of the POS taggers follow the rule based methods andmostly depends on a dictionary to determine the tag to be used. Besides, statistical methodslike hidden Markov model, CRF model are also contributing a lot to improve the performanceof the taggers. Several POS taggers are available, especially, for language like English, French,Chinese. NLTK provides a very handful POS tagger. For example, an online demonstration ofNLTK POS tagger output the following version of a plain text after performing POS tagging:Important/JJ gains/NNS have/VBP been/VBN made/VBN in/IN the/DT past/JJtwo/CD decades/NNS in/IN the/DT participation/NN of/IN women/NNS in/INscience/NN ,/, engineering/NN ,/, and/CC biomedical/JJ disciplines/NNSat/IN the/DT undergraduate/NN and/CC graduate/VB levels/NNS in/INthe/DT United/NNP States/NNPSVery few works found in POS tagging for Bengali languages. Though some POS taggers areavailable, their performance is also not very encouraging [50].3.2.3 ParsingWhile POS tagging provides lexical information, parsing obtains syntactic information. Parsingrepresents the grammatical structure of a given sentence with the corresponding relationship of3.3. NLP RESOURCES 26Figure 3.1: Generation of a Parse Tree using POS information.different constituents. Comparing to POS tagging, parsing provides richer structural information[51]. The outcome of the parsing would be a parse tree. As a simple example, the sentence “Tomate an apple” is the root, intermediate nodes such as noun phrase and verb phrase are the childnodes of the root, hence, called non-terminals. Finally, the leaves of the tree ‘Tom’, ‘ate’, ‘an’,‘apple’ are called terminals. The tree is shown in (Figure 3.1):Apart from its capabilities of syntactic and grammatical annotation, the parser also providesdependency among the terms of the sentence.</s>
<s>In the context of sentiment analysis, theparser plays a very influential role to form the concepts from the corpus, hence, extract thedependencies among them as discussed in Section 3.1.4. One of the enhanced version of it issemantic dependency parser. To be mentioned, the Stanford dependency parser is an examplefor one of this kind which produces a dependency tree from the annotated corpus. For instance,the Stanford dependency [6] of the sentence, “Bills on ports and immigration were submitted bySenator Brownback, Republican of Kansas” are as following (Figure 3.2):3.3 NLP Resources3.3.1 Concept NetConceptNet was first introduced in [36], motivated with the significance of large-scalecommonsense knowledge bases to textual information management. It is a freely availablecommonsense knowledge base and natural-language-processing tool-kit which supports manypractical textual-reasoning tasks over real-world documents including topic-gisting, analogy-making, and other context oriented inferences. The knowledge base is optimized for makingpractical context-based inferences over real-world text. It is a graphical representation ofa semantic network, where the nodes represent the concepts with their attributes as edgesassociated with each other. Presently, it consists of over 1.6 million edges connecting more than3.3. NLP RESOURCES 27Figure 3.2: Stanford Dependency Tree [6].300 000 nodes. A partial snapshot of actual knowledge in ConceptNet is given in Figure 2.2 ofChapter 2.In the field of NLP, the conceptNet has opened up new diversity in sentiment analysis. Insteadof depending on the key phrases and the dictionary biased knowledge bases, ConceptNet utilizesthe commonsense knowledge base to extract the concept which carries related information onthe context of the text to be analyzed. Thereby, it excels the contextual commonsense reasoningover real-world texts. The concept level sentiment analysis also offers more efficiency andaccuracy in the classification of sentiment.3.3.2 WordNet-AffectWordNet-Affect, first proposed in [7], is a linguistic resource for the lexical representation ofaffective knowledge, developed starting from WordNet [52]. The knowledge base is built byassigning a number of WordNet synsets to one or more affective labels (a-labels). In particular,the affective concepts representing emotional states, are identified by synsets marked with thea-label ‘emotion’, but there are also other a-labels for concepts representing moods, situationseliciting emotions or emotional responses. Figure 3.3 shows an example of a list of a-levelemotion and their corresponding synset.The purpose of the WordNet-Affect was to incorporate the lexical information with the affectiveinformation. Lexical information includes the correlation between English and Italian terms,POS, definitions, synonyms, and antonyms. Affective information is a reference to one or moreof the three main kinds of theories on emotion representation: discrete theories (based on the3.3. NLP RESOURCES 28Figure 3.3: A-Labels and corresponding example synsets (collected from [7])concept of cognitive evaluation), basic emotion theories and dimensional theories.To build the WordNet-Affect, an initial set of affective words directly or indirectly referringto mental (e.g. emotional) states was build. By mapping the senses of affective words to theirrespective synsets, the “affective core” was identified. Then, a subset ofWORDNET containingall synsets in which there are at least one word of the affective word list, and rejected thosesynsets that are not recognized as affective concepts were prepared. Finally, an automatic checkfor the coherence of</s>
<s>the affective information inside the synsets was performed. The resultshave shown that the synsets are a good model for the representation of affective concepts.3.3.3 AffectiveSpaceThe latest form of the knowledge base in the field of emotion reasoning is the AffectiveSpace.In [3], a blending technique was applied on ConceptNet and WordNet-Affect to build asuitable knowledge base for emotive reasoning. Blending performs inference over both sourcessimultaneously to combine two sparse matrices linearly into a single matrix, so that, it can all berepresented in the same matrix. Singular value decomposition (SVD) is applied on the blend ofConceptNet andWordNet-Affect to form the affective space (illustrated in Figure 3.4), in whichcommon sense and affective knowledge are in fact combined, not just concomitant.Though affective space is mainly generated for sentic computing [4], it is a very effectiveresource for opinion mining and sentiment analysis. Information sharing properties of truncatedSVD, suggest that concepts with the same affective valence are likely to have similar features,i.e. concepts concerning the same emotion tend to fall near each other in AffectiveSpace.For instances, ‘love’ and ‘affection’ tend to be nearer to each other, in contrast, ‘honesty’and ‘corruption’ will fall far from each other. Therefore, affective space will enhance theperformance of sentiment classification. An enhanced version is proposed as AffectiveSpace-2[53]. It represents a novel vector space model for concept-level sentiment analysis that allowsreasoning by analogy on natural language concepts, even when these are represented by highlydimensional semantic features.3.4. MATHEMATICAL MODELS 29Figure 3.4: A Sketch of the AffectiveSpace (collected from [3]).3.4 Mathematical ModelsThere are several mathematical models used in different machine learning or even rule basedmethods for NLP tasks like sentiment classification, subjectivity prediction, sentic computingetc. This section will highlight the basics of the model singular value decomposition (SVD) andlinear discriminant analysis as our proposed model will use these methods for the classificationof concepts.3.4.1 Singular Value Decomposition (SVD)Singular value decomposition takes a rectangular matrix of gene expression data (defined as A,where A is a n×pmatrix) in which the n rows represent the genes, and the p columns representthe experimental conditions. The SVD theorem states as Equation (3.1):An×p = Un×nSn×pVp×p (3.1)Where:• U and V are orthogonal• the columns of U are the left singular vectors and UTU = In×n ,• V T has rows that are the right singular vectors and V TV = Ip×p.• S (the same dimensions as A) has singular values and is diagonal and3.4. MATHEMATICAL MODELS 30The SVD represents an expansion of the original data in a coordinate system where thecovariance matrix is diagonal and vice versa. The eigenvalues and eigenvectors of AAT andATA make up the columns of V and U respectively. These values need to be Calculated toperform the SVD. Also, the singular values in S are square roots of eigenvalues from AAT orATA and always real numbers. Equation (3.1) can be expressed using summation notation asEquation (3.2).aij =k=1uikskvjk (3.2)One of the popular application of SVD is dimension reduction. The singular values are thediagonal entries of the S matrix and are arranged in descending order, where the larger valuescorrespond to</s>
<s>the vectors in U and V that are more significant components of the initial Amatrix. A tranculated SVD can be created with a reduced number of singular value p followingEquation (3.3):aij ≈k=1uikskvjk (3.3)where p < n is the number of singular values that will remain.Equation (3.3) can be used for data compression by storing the truncated forms of U , S, andV in place of A and for variable reduction by replacing A with U . The principal componentsof A form a low-rank approximation of the original data by often discarding all but the firstk components. This factorization allows the row space of A and the column space of A to beprojected into a common space by the transformations U and V . We can think of these spaces ascontaining two types of objects, which we can represent as row and column vectors ofA, whichare related to each other by the values where they meet.3.4.2 Linear Discriminant Analysis (LDA)LDA is a probabilistic classification method used in statistics, pattern recognition and machinelearning to find a linear combination of features that characterizes or separates two or moreclasses of objects or events. The particularity of LDA is its capability to model the distributionof predictors separately in each of the response classes, and then, use Bayes’ theorem to estimatethe probability.For instant3, an observation needs to be classified into one ofK classes, where K≥ 2 andΠk bethe overall probability that an observation is associated with the kth class. Then, let fk(x) denotethe density function ofX for an observation that comes from the kth class. That is, fk(x) is large3Mathematical equations and their explanation used in this section are mostly followed to the platform“Towards Data Science” sharing concepts, ideas and code (https://towardsdatascience.com/).3.4. MATHEMATICAL MODELS 31if the probability that an observation from the kth class hasX = x. Then, Bayes’ theorem statesthe Equation (3.4) to predict the class (Y = k) for the observation (X = x).Pr(Y = k \| X = x) =Πkfkx∑kl=1 Πlfl(x)(3.4)The challenge here is to estimate the density function fk(x). It consists of statistical propertiesof data, calculated for each class. For a single predictor, the density function can be assumed asa normal distribution (Equation (3.5)):fk(x) =exp(− 12σ2(x−µ)2)(3.5)Now, to maximize Pk(x) for the observationX = x, it will plugin the density function in Pk(x)and take the log to generate Equation (3.6):δk(x) = x− µ22σ2+ log(Πk) (3.6)The Equation (3.6) above is called the discriminant function for class k given input x. As it islinear, hence, the name linear discriminant analysis. The boundary equations of the two classeswith equal distributions are represented in Figure 3.5. In reality, LDA cannot exactly calculatethe boundary line rather, makes use of the following approximation:• For the average of all training observations, called as mean, µkµk =i:y=kxi (3.7)• For the weighted average of sample variances for each class, called as σ2σ2 =n− kk=1i:y=k(x− µk)2 (3.8)For multiple predictors, the same properties calculated over the multivariate Gaussiandistribution with a class-specific mean vector, and a common covariance matrix. Figure 3.6shows an example of a</s>
<s>correlated and uncorrelated Gaussian distribution.To accommodate multiple predictors, the discriminant equation remains the same but it isexpressed using vector notation as Equation (3.9):δk(x) = xT−1∑µk −−1∑µk + log(Πk) (3.9)3.4. MATHEMATICAL MODELS 32Figure 3.5: Boundary line to separate 2 classes using LDA.Figure 3.6: Uncorrelated (left) and correlated (right) normal distribution.In sentiment classification, LDA can be utilized as an effective classifier, especially, in featurebased text classification. The features of the concept can be taken as independent variables tocategorize them in different classes like positive, negative or neutral, which can be considered asdependent variables. LDA is also a very effective tool for dimension reduction. In the process ofgenerating classification function coefficients, it automatically discards the features which haveno effective contribution to determine the classes of the dependent variables. Thereby, LDAoffers a composed classification function with fewer parameters.Chapter 4Proposed MethodologyIn this chapter, we present our proposed methodology. At the very outset, we will present anoverview of the methodology in Section 4.1. Then, the construction of our methodology willbe discussed in four steps in four subsequent sections. To step forward to our methodology,firstly, we will discuss the process of generating annotated data-set in Section 4.2. Thereafter,a rule based semantic parser will be presented in Section 4.3. In Section 4.4, we will discussthe technique of concept extraction from the parse tree and, determine the dependencies of themodifiers. Finally, we will present the concept level polarity detection model and explain theprocess of determining the polarity of the sentence in Section 4.5.4.1 Overview of the MethodologyProposed methodology presents a polarity detection model to detect polarity at the concept leveland eventually, determines the polarity of the sentence in a specific domain or context in Bengali.This model will use the AffectiveSpace (discussed in Chapter 3) as a knowledge base. On theother hand, concepts on the desired domain are selected from AffectiveSpace and grouped aspositive, negative and neutral to be used as training data for the model. Corpus is prepared withannotated text which is used as input to the model. The output of the model are the conceptswith their polarity. Workflow diagram of the proposed methodology is presented in Figure 4.1which highlights the overview of methodology.To start with, a rule based semantic parser takes the annotated data as input which produces aparse tree for each sentence. The parse tree also contains the dependencies among the child nodesof a single root node. The outcome of the semantic parser is fed to the concept extractor whichexploits the semantic dependencies among the tokens based on POS and extracts the conceptsfrom the parse tree. Thereby, we get all possible concepts from out corpus which are constituentscontaining some opinion words for the sentiment of the sentence. On the other end, the trainingdata along with the AffectiveSpace are fed to the model, in which the concepts in training4.2. DATA ACQUISITION 34Figure 4.1: Workflow diagram of the proposed methodology.data are mapped in AffectiveSpace to know a set of features known as ‘eigenmood’. LDA,the mathematical classifier used by the model generates the classification function coefficientfor each class corresponding to</s>
<s>positive, negative and neutral. This step is termed as functiongenerator for concept level polarity detection which generates the polarity detection functionusing the classification function coefficient values produced by LDA for each class. Finally,concepts extracted from the corpus are mapped to affective space to get the correspondingegenmood values for the polarity detection function. These values in turn, fed to the functionfor each class to produce the classification value. Comparing the classification values of theclasses of a concept, the polarity for the concept is determined. This step is termed as conceptlevel polarity detector which produces the concepts with their polarity. In the end, the sentencepolarity is determined through a back propagation technique on the parse tree. To gain betterinsight, the proposed methodology is described broadly in four steps in four subsequent sections.4.2 Data AcquisitionUnlike any other research work, our foremost initiative was to generate a data set. As it wasmentioned in Chapter 2 about the scarcity of annotated data set, especially, for the less privilegedlanguage like Bengali, we develop an annotated data set as a part of our research work. Firstly,we build a collection of sentences in natural Bengali language crawled from news editorials andblog. Then, we apply NLTK POS tagger for Bengali on each sentence. The POS tagging processinvolves assigning the right POS marker to all the words in a sentence (or corpus). Accordingto Bengali grammar, it has five kinds of POS with several sub-kinds of each. Though there aresome deviations in the POS category of Bengali from the English, most of the tagger availablefor Bengali uses the English categorization as a universal POS tags. To achieve better syntacticrepresentation, particularly for sentiment analysis, we have also used the universal POS tag setgenerally used for English. For instant, the sentence “ইভ িটিজং বা উ তার িশকার হেল কারওসংেকাচ করা উিচত নয় ” ([ibHtidZiNer ba uttktOtar Sikar HOle karO SNkotS kra utSit nOi], If4.2. DATA ACQUISITION 35someone is victim of eve teasing or harassing, she should not be ashamed of that) annotatedwith POS using the Bengali NLTK POS tagger is as follows:[ ইভ/JJ িটিজং/NN বা/CC উ তার/NN িশকার/NN হেল/VM কারও/CC সংেকাচ/NN করা/VMউিচত/JJ নয়/VM ]To achieve gold standard annotated corpus, improvement of accuracy of the tagger is essential,particularly when the performance of the tagger is not very encouraging. To be mentioned here,“কারও /CC” ([karo], someone) is tagged as conjunction based on detecting the ‘ও ’ ([o], and)as conjunction. Whereas, it is clearly a pronoun. Likewise, The NLTK POS tagger has no‘VAUX’ (auxiliary verb) tag for Bengali, rather it considers the words ‘হেল [HOle]/VM’ and‘করা [kra]/VM’ as main verb (VM), which are actually contemporary to auxiliary verb usedin English. Although Bengali has several forms of the verb, we categorize them in two types,such as main verb (VM) and auxiliary verb (VAUX), to facilitate the parsing. Therefore, were-tag the above mentioned tokens as ‘হেল [HOle]/VAUX’ and করা [kra]/VAUX. In the samecontext, the POS tag used for ‘িশকার /NN’ ([Sikar], hunt) and ‘সংেকাচ /NN’ ([SNkotS], ashamed)are reformed as ‘িশকার /VM’ [Sikar] and ‘সংেকাচ /VM’ [SNkotS]. On the</s>
<s>other hand, the word tag‘নয় /VM’ ([nOi], not) is wrongly tagged as ‘VM’ which should be a negation(NEG) annotation.Thereby, we get a more comprehensive annotation of the aforementioned sentence as under:[ ইভ/JJ িটিজং/NN বা/CC উ তার/NN িশকার/VM হেল/VAUX কারও/PRP সংেকাচ/VMকরা/VAUX উিচত/RB নয়/NEG ]To limit the scope of our work, we improve the accuracy of the POS tagging with manualannotation for the false tagging as some of themmentioned above. Doing so, we have introducedthe tag set for our work as listed in Table 4.1. For the simplicity; quantifiers, intensifiers,and interjections are categorized as modifiers and further subcategorized as pre-modifier andpost-modifier depending on their position related to modified words as they play an importantrole in the performance of dependency parser. As negation has a very effective contribution indetermining the sentiment polarity, therefore, we use the tag ‘NEG’ for all negation words inthe corpus. In addition, we recognize the named entity (NE) early on and tag them as NE inthe corpus, with the intention to make our corpus application specific to our methodology. Asthe NE has no significant contribution in determining the sentiment polarity, therefore, taggingthem early on will reduce the computational complexity.4.3. PARSE TREE GENERATION 36Table 4.1: Tag list used in annotated data with examples.Category Tag ExampleNoun NNিতি য়া ([pr9tikriE], reaction), উ য়ন([unnOn], development)Pronoun PRPআমােদর ([amader], us), তােদর ([tader],their)Main Verb VMবলা ([bOla], say), শানা ([Sona], listen)AuxiliaryVerbAUXহয় ([HÔi], is), হেব ([HObe], will)Adjective JJভােলা ([bHalO], good), র ([SundOr],beautiful)Adverb RBত ([drutO], quickly), চা ভােব([SutSarubHabe], skillfully)Conjunction CCএবং ([ebON], and), অথবা ([OtHOba], or)Preposition PSPজ ([dZOnnO], for), সােথ ([SatHe], with)Pre Modifier PREMODঅিত ([Oti], too), খুব ([kHub], very)Post Modifier POSMODবেটই ([btêi], indeed), তর ([tOrO], more)Named Entity NEবগম রােকয়া [begOm rokea], বাংলােদশ([baNladeS]Negation NEGনা ([na], no), নাই ([nâi], not)Besides generating the corpus to be used in the proposed model, we use the ‘AffectiveSpace’ asa resource to incorporate the semantic affectivity among the concepts. To generate the trainingdata, we extract a set of high frequency concepts on the targeted domain from the news portals,blogs, and literature available and label as positive, negative and neutral. While preparingtraining data, our focus was on those words of the targeted domain which may change thesentiment polarity in other domain.4.3 Parse Tree GenerationIn the proposed work, a semantic parse tree is developed depending on the syntactic affinityamong the POS of the token. The parser takes the annotated corpus as input and generates aparse tree for each sentence as its output. The aim of the parse tree generator is to deconstruct4.3. PARSE TREE GENERATION 37the text into tokens exploring the syntactic dependencies. The parse tree will be later on fedto the concept extractor to construct the concepts. Doing so, we have formulated the followingrules depending on the syntactic association through the POS tagged with each token:[RULE- 1] Parse the complex or compound sentence into simple sentences: If more than one verbclause parse is available then each verb clause with its preceding words forms a simplesentence.[RULE- 2] Parse the simple sentence into subject and predicate:(a) Add token to subject until first NN/PRP occurs.(b) Check the next token: If NN/NPN/</s>
<s>(MOD/NEG)/CC add to subject and repeat theprocess, else Parse as subject.(c) Parse the rest of the sentence except subject as object.[RULE- 3] Parse the Subject into tokens:(a) If the subject is a single word parse into a single token.(b) If any MOD/NEG is present parse associated token with MOD/NEG, else parse intoa single token.(c) For each token with MOD/NEG parse into toke and MOD/NEG marked as tokenMOD/NEG :[RULE- 4] Parse the Object into tokens:(a) Parse object into the verb clause and others along with associated MOD/NEG(b) For parsing the verb clause and others follow the parsing rule for subjectThe first step to parse the sentence is to decompose the complex and compound sentence intosimple sentences. Bengali simple sentence usually contains a single verb clause and ends with afinite closing auxiliary verb. There is hardly any verb represented with a single token. Almost,in all the cases, the verb clause ends with a finite closing auxiliary verb, such as, থাকা [tHaka],করা [kra], হওয়া [HOa], পারা [para] which can be used within or at the end of a sentence.For example, in the sentence “ মেয়িট ইভিটিজং এর িশকার হওয়াই থানায় অিভেযাগ কের ” ([meetiibHtidZiN er Sikar HOai tHane ObHidZog kre], The girl complain to police for being victim ofeve-teasing), the auxiliary verb ‘হওয়াই ’ [HOai] is used inside the sentence, whereas, ‘কের ’ [kre]is used at end of the complex sentence. Table 4.2 shows some examples of verb phrase with afinite closing auxiliary verb with their different forms based on the position in the sentence. Theparser detects the verb clause in the complex and compound sentence and applies the rule-1 todeconstruct simple sentences as a child of the complex or compound sentence.4.4. CONCEPT EXTRACTION AND DEPENDENCY DETECTION 38Table 4.2: Various forms of auxiliary verb depending on the position within the sentence.Verb Phrase Auxiliary Verb Inside theSentenceAuxiliary Verb at the endof the Sentenceদাঁিড়েয় থাকা ([dar̃ie tHaka],stand)থাকােল, থাকায় [tHakle, tHakâi] থািক, থাক, থােক [tHaki,tHak, tHake]অিভেযাগ করা ([obHidZogkra], complain)করেল, করায় [kOrle, krâi] কির, কর, কের [kOri, KorO,kOre]িশকার হওয়া ([Sikar Hoa],victim)হেল, হওয়াই [Hole, HOai] হই, হও, হয় [HÔi, HOô, Hôi]বলেত পারা (bOlte para],say)পারেল, পারাই [parle, parâi] পাির, পার, পাের [pari,parO, pare]The next step is to deconstruct the simple sentence into subject and predicate. As subject mustcontain one or more noun or pronoun, therefore, it detects the first noun or pronoun in thesentence to form the subject and concatenate the succeeding noun, pronoun, conjunction and ifany modifier or negation token is present. As these succeeding elements form the noun clause,we consider those to be part of the subject. In the same context, the preseeding tokens of thesubject are actually dependent on the subject and included in the subject as per the rule-2. Afterdetecting the subject rest of the sentence are considered to be the object. Rule-3 is generated todeconstruct the subject into dependent tokens. Here, all the tokens will be independent exceptany negation or modifier is present. In case of negations and modifiers, the token will be parsedalong with its modifiers. It is important to explore the dependency</s>
<s>of the modifiers or negationas it will reverse the polarity of the sentiment of the associated concept. On the other hand, ifthe negation is associated with the verb clause then, it will negate the overall sentiment of thesentence. Otherwise, if it is associated with an adjective or any other opinion word, it will onlyreverse the polarity of the concern concept. Therefore, we generate the rule-4 which parse theobject into verb clause and other first, then parse the subsequent phases.Figure 4.2 shows a parse tree generated following the rule we discussed in this section for acomplex sentence. However, the negations in the parse tree are associated with the concerntokens only. In the next section, while generating the concept we will see how negation ormodifiers are associated with the concepts or sentences.4.4 Concept Extraction and Dependency DetectionIn this section, we proposed a technique for concept extraction and in the process, we willdetect the dependency as well. A concept can be a single word or a chank of words dependingon the semantic association of the terms in the sentence. As we discussed in Chapter 3,4.4. CONCEPT EXTRACTION AND DEPENDENCY DETECTION 39Figure 4.2: Rule based parse tree of a complex sentence generated from annotated data.the concept can be formed combining two or more words which add some abstraction overthe dictionary meaning of the constituents. Combining the words sometimes brings massivesemantic metamorphosis, especially, for a language like Bengali with semantic diversity. Forinstant, “িশকার করা ” ([SIkar kra], hunt) may be considered as a neutral sentiment. On the otherhand, if we use ‘হওয়া ’ [HOa] instead of ‘করা ’ [kra], then “িশকার হওয়া ” ([SIkar HOa], victim)is used to express sentiment like “someone is victim of something”, which exhibits a negativesentiment. Therefore, the right choice of constituents of the text is very important in conceptextraction particularly, in sentiment analysis.POS of the tokens have a semantic association among them depending on their syntactic relationwithin the sentence. For an instant, when a noun is preceded by an adjective, the combinationis generally a single entity. For Example, ‘কমজীবী ’ ([krmOdZibi], working) is an adjective and‘নারী ’ ( [nari], woman) is a noun. When the noun is preceded by the adjective, it became asingle concept like “কমজীবী নারী ” ([krmOdZibi nari], working woman). In this context, theconcept does not simply indicates the meaning as “the woman who works” rather it representsthe knowledge as “the woman who earns wages through regular employment outside the home”and generally exhibits a positive sentiment. On the contrary, “noun + adjective” combinationwill not be considered as a concept as there will be hardly any dependency. In the same context,consecutive or dependent nouns can form a single concept. Such as, ‘নারী ’ ( [nari], woman)and ‘িনরাপ া ’ ( [nirapOtta], safety) form a concept “নারী িনরাপ া ” [nari nirapOtta] whichmeans “woman safety” that indicates some specific measures for the safety of woman. If thereis any stem in the preceding noun, like, “নারীর িনরাপ া ” [narir nirapOtta] means the safety of awoman. Since the combination does</s>
<s>neither enhance the original meaning nor indicate any otherknowledge, therefore, it will not be considered as a concept. However, some of the tokens withPOS like a noun, adjective, and VM remaining stand lone can also be considered as a concept.4.4. CONCEPT EXTRACTION AND DEPENDENCY DETECTION 40The rules based on the POS combination that are taken into consideration to form a concept arelisted below:1. NOUN + NOUN: If two nouns are consecutive and the preceding one has no steamingthen it indicates a single entity and added to concept list.2. ADJECTIVE + NOUN: If the adjective is preceding the noun then the combination isadded to concept list as the adjective is obviously dependent on the noun.3. MAIN VERB + AUXILIARY VERB: In Bengali auxiliary verb can influence thesemantic orientation of the main verb. Hence, they are considered a concept as a verbphrase.4. ADVERB + VERB PHRASE: As the adverb will modify the action of verb therefore, thecombination is added to the concept.5. NOUN, ADJECTIVE: Besides the combinations above a stand-alone noun and adjectivecan have an impact on the sentiment of the sentence, hence, taken as concept.After generating the parse tree, we traverse the parse tree from leaf nodes towards the rootand extract concepts applying the rules listed above. While identifying the concept we try toextend the combination of tokens to the highest level. However, there is no use of going beyondtwo levels up from the child node in traversing the parse tree as we will not find any POScombination at the sentence level. Concept extracted from the parse tree in the Figure 4.2 areshown in Figure 4.3. The concepts are highlighted in bold and shown their association withinthe sentence.One of the most important tasks in this stage is to identify the dependencies. For the simplicity,we are mostly concentrated on the negation. If any negation is available with any token thenit will be considered to be associated with the concept that contains the token. However, ifthe negation is associated with a verb phrase then it will be associated with the concern simplesentence. Because, the verb contains the momentum of the sentence and the negation of theverb indicates the negation of the sentence polarity. The dependency identification rule can berepresented as follows:[RULE- 1] If the root of the negation is a verb parse, level the negation as “Sentence NEG”[RULE- 2] If the root of the negation is a concept without a verb phrase , level the negation as“Concept NEG”At this juncture, we discuss the process of the concept extraction and dependency detection fromthe parse tree in Figure 4.2. The concept extracted are as follows:4.5. POLARITY DETECTION 41Figure 4.3: Concepts extracted from parse tree and their association within the sentence.(“ ইভ _িটিজং ” [ibHtidZiN], eve-teasing) (“উ তা ” [uttktOta], harassment) (“িশকার _হওয়া” [Sikar hOa], to be victim of) (“সংেকাচ _করা ” [SNkotS kra], to be ashamed of)The first concept was taken following the rule “ADJECTIVE+NOUN” where two differenttokens with independent meaning form a concept that represents a specific meaning as“harassing the girls or woman”. Likewise</s>
<s>“িশকার _হওয়া ” [Sikar hOa] and “সংেকাচ _করা ” [SNkotSkora] have been formed following the rule “MAIN VERB + AUXILIARY VERB”. Both theconcept represent some action like “to be victim of” and “to be ashamed of” respectively andcarry negative sentiment, whereas, ‘িশকার ’ ([Sikar], hunt) and ‘সংেকাচ ’ ([SNkotS], shame)stand-alone may contain neutral sentiment. On the other hand, ‘উ তা ’ [uttktOta] is a standalone concept which carries the default meaning with its negative sentiment. To identify thedependency of the negation, we can see that the root of the negation in the Figure 4.2 verbphrase. Therefore, we consider it to be negation for the simple sentence “কারও সংেকাচ করা উিচতনয় ” ([karO SNkotS kra utSit nOi], someone should not be ashamed of) and alter the valence ofthe sentence while determining the sentence polarity.4.5 Polarity DetectionIn this section, we will present our proposed “concept level polarity detection model”. In theprocess, we will discuss the construction method and integration of affective space followed bythe working procedure of the model in detecting the polarity of the concept. At later part, wewill discuss the procedure to detect the sentence polarity using the parse tree in the examplefrom the previous section.4.5. POLARITY DETECTION 42Figure 4.4: Overview of the concept based polarity detection model.4.5.1 Construction of the Polarity Detection ModelThe proposed model for detection of the polarity of a concept is inspired through the affectivecategorization model widely known as Hourglass of Emotion [37]. AffectiveSpace is usedas a knowledge base for the proposed model. The affective relation among the conceptsin AffectiveSpace is the inspiration for the proposed methodology. We use the LDA as amathematical model to train the model with training data. On the other hand, the conceptsextracted from the corpus are considered to be input to the model and output the concept withits sentiment polarity. Figure 4.4 represent the overview of the proposed model which can bedescribed in two phases.To start with, it takes the training data that includes the list of positive, negative and neutralconcepts on the targeted domain. The concepts are then mapped on AffectiveSpace to find theappropriate matching within the affective space. In affective space, each concept contains 100‘eigenmood’ values leveled as (e0, e2, e3, ...., e99). After the mapping, we get a training data setas a matrix in which, each row represents the concept with the eigenvalues as a column. Thevalues in the last column are class level. Table 4.3 shows the data representation used by themodel. The last column is the value for the class in the training data where positive, negativeand neutral class are denoted by 1, 2 and 3 respectively.Table 4.3: Concepts with their eigenvalues and class labels to train the model.Concept e0 e1 ..... e99 ClassSocial_development 0.9351 -0.0514 ..... 0.0143 1Sexual_victim 0.7651 -0.0904 ..... 0.0214 2Working_girl 0.8331 -0.0674 ..... 0.0035 3..... ..... ..... ..... ..... .....4.5. POLARITY DETECTION 43As we discussed in Chapter 3, the AffectiveSpace is constructed through the blending ofConceptNet and WordNetAffect to combine the commonsense and affective knowledge. Sincethe blending process explores the information sharing</s>
<s>properties of truncated SVD to build theaffective space, the ‘eigenmood’ values of the concept nearer to each other within the space tendto be similar. To categorize the sentiment, our aim is to discriminate among the positive, negativeand neutral concepts based on their eigenvalues. In this context, LDA is an appropriate choicefor its property of maximization of class separability. As we mentioned in preliminaries, LDAhas the capability of modeling the distribution of predictors separately in each of the responseclasses, thereby, formulate a class specific mean vector and covariance matrix which is used todetermine the significance of the features of the independent variable. In the proposed method,three sentiment classes are considered as dependent variables and the ‘eigenmood’ are taken asthe independent variable for LDA. The model explores the dimension reduction properties todiscard the less significant ‘eigenmood’ that contribute very less in discriminant function. Atthe same time, as the eigenvalues are sorted from strong to weak, it increases the efficiencyof the dimension reduction. LDA generates two discriminant function which divides the spaceinto three groups according to the three classes. Finally, the classification function coefficientfor each class is generated by LDA as shown in Table 4.4.Table 4.4: Classification function coefficient matrix.Eigenmoods(E)Classification Function CoefficientClass-1 Class-2 Class-3ei a01 a02 a03−− −− −− −−ei ai1 ai2 ai3Constant Constant1 Constant2 Constant3From the classification function coefficient generated by the classifier, we formulate theclassification function for each of positive, negative and neutral as Equation (4.1), Equation (4.2),Equation (4.3). These functions are fed to the polarity detector to determine the concept polarity(Figure 4.4).Scorepositive(C) =i∈Eeiai1 + Constant1 (4.1)ScoreNegative(C) =i∈Eeiai2 + Constant2 (4.2)ScoreNeutral(C) =i∈Eeiai3 + Constant3 (4.3)Where:4.5. POLARITY DETECTION 44• C is the concept with its eigenvalue• ei is the ith eigenvalue of the concept C• ai1, ai2, ai3 are the ith coefficient values for class-1 (positive), class-2 (negative) and class-3 (neutral) respectively• Constant1, Constant1, Constant1 are the constant value for class-1 (positive), class-2(negative) and class-3 (neutral) respectively• E is the set of significant eigenmood in for classification function generatedThe next step is to determine the polarity of the concept that is extracted from the corpus. Thepolarity detector receives a concept as input. The concept is mapped in the AffectiveSpaceto find the best match. The rows of the AffectiveSpace contain the eigenvalues for eachconcept. Eigenvalues of the concept are used as parameters to Equation (4.1), Equation (4.2),Equation (4.3). The output of the equations are positive, negative and neutral score for theconcept. Finally, the polarity of the concept is determined as the polarity class with themaximumscore.4.5.2 Determination of the Polarity of the SentenceThe polarity of the sentence is strongly depends on the association of the concepts within thesentence. However, the dependency of the negation and modifiers can highly influence thesentence polarity. At the same time, the syntactic relation and POS information are also playsan important role in determining the sentiment polarity of the sentence. Therefore, we use theparse tree with concepts in Figure 4.3.The technique for determining the polarity of the sentence can be described in the Figure 4.5.First of all, we label all</s>
<s>the concepts with their polarity determined by the concept level polaritydetector. Then, we traverse the tree from the concept node to the root following the rulesdescribed below.1. If any node contains one concept the node will be labeled the polarity of the concept.2. If any node contains two concepts then the polarity of the node will be labeled as follows:(a) If both are positive then the node is Positive.(b) If they are of opposite polarity then the node is negative.(c) If both are negative and the concepts are associated with conjunction then the nodewill be labeled as negative, else it will be positive.4.5. POLARITY DETECTION 453. If any node contains more the two concepts then the polarity label of the node will bedetermined pairwise following the rule 2.4. If any sentence negation is available it will reverse the polarity of the simple sentenceat the child node of the root. The token negation will reverse the polarity label of theimmediate concept that contains the token.5. From child node to parent node the label is determined as follows:(a) POSITIVE + POSITIVE = POSITIVE(b) POSITIVE + NEGATIVE = NEGATIVE(c) NEGATIVE + NEGATIVE = NEGATIVEFigure 4.5: Sentence level polarity detection through tree traversal.In Figure 4.5, both the concepts “ ইভ _িটিজং ” ([ibHtidZiN], eve-teasing) and “উ তা” ([uttktOta], harassment) are labeled as negative (N). AS they are associated throughconjunction their parent node will be labeled as (N) as well. On the other, stand-alone conceptwill label their parent node same as their label. As the negation is a sentence negation, thereforeit has reversed the polarity of the simple sentence “কারও সংেকাচ করা উিচত নয় ” ([karO SNkotSkra utSit nOi], someone should not be ashamed) from negative to positive. Finally, the polarityof the sentence is labeled as (N) as the child node are opposite to each other.In summarizing the methodology, the data set was built by semi-automated annotation by POStagging followed by using the NE recognizer. The annotated data is used as input to the parse4.5. POLARITY DETECTION 46tree generator to build the parse tree. The concepts are extracted from the parse tree by theconcept extractor. Then, the concepts are mapped to the AffectiveSpace to be used as inputto the classification method named as concept level polarity detection model. Mapping of theconcept is carried out through translation and matching of words that form the concept. LDAis used as classification and dimension reduction method to the model. The model generates apolarity detection function which takes the ‘eigenmood’ values of each concept to output thescore of each polarity class. Finally, the sentence polarity is determined following a rule basedtree traversal.Chapter 5Experimental AnalysisIn this chapter, we will discuss the experimental analysis process and evaluate the performanceof our proposed model through result analysis using our data-set. Initially, we will discuss theexperimental data set in Section 5.2 followed by presenting an outline of the experimental setupin Section 5.1. Section 5.3 will give an overview on evaluation method and performance metricsused for evaluating the result and for analyzing the data-set. In Section</s>
<s>5.4, we will evaluate theperformance of the model and represent some statistical data on experimental analysis. Finally,Section 5.5 will recapitulate the experimental analysis through an inquisitive discussion onperformance evaluation of the model.5.1 Experimental SetupIn this work, we implement our experiment using Python as a programming platform and carryout the statistical computation in SPSS. Doing so, we use the Python-2.3.15 version in windowsset up. To perform the NLP tasks, we use the NLTK1 plugin for Python, a leading platform forbuilding Python programs to work with human language data. We utilize the natural languagetext processing API such as tagging and chunk extraction, phrase extraction and NE recognition.To our knowledge, NLTK is the only platform which offers the POS tagger and some otherlimited NLP tools for Bengali. However, the performance of this requires improvements to beused in an organized framework.For implementing the classification model, we use the IBM SPSS software2 platform to takeadvantage of the readily executable mathematical model. This platform offers all of thewidely used classifiers with their traditional variation as well. We also incorporate the PythonIntegration Package for IBM SPSS Statistics with the SPSS. It allows us to build pythonprograms that control the flow of command syntax jobs, read and write data, and procreate1https://www.nltk.org/2https://www.ibm.com/analytics/spss-statistics-software5.2. EXPERIMENTAL DATA SET 48Table 5.1: Statistics of the training data set.TypeSourceNumberdocumentNumber ofPositiveConceptNumber ofNegativeConceptNumber ofNeutralConceptTotal NumberConceptBlog 5 35 44 39 118News Portal 3 24 25 20 69Total 8 59 69 59 187custom procedures. This integration facilitates the handling of large data set and carrying outexperimental analysis more efficiently.5.2 Experimental Data SetWe build a training data-set containing three sets of the concept named as positive, negative andneutral. The domain dependency of the proposed model relies on the domain of the trainingdata. In this work, our targeted domain is “woman harassment”. To generate the data set, wecrawl through the various Bengali blogs, online news portals especially the editorials on womanharassment, as these sources mostly provides some opinions.Our aim is to determine the most frequently used and the most influencing concepts fromthe available resources in the targeted domain. We use an improved term frequency-inversedocument frequency (tf-idf) scheme proposed in [54] as a weighting factor to the terms in thecorpus which can form a concept. In the process, the most weighting terms are selected andmapped to the affective space to determine the concept within the space. Then, the concepts aregrouped according to the valence manually. Table 5.1 shows the overall statistics of the trainingdata used in this work.In the statistics of the training data, we observe that the number of the negative concepts is morethan the positive. In reality, we also find that the number of negative terms is more than that ofpositive in various lexicon resources like SentiWordNet. Moreover, the NLP researchers alsoreached in consensus that the number of negative sentiment words is more than the positive inany language. As mentioned in methodology (Chapter 4), we develop a new domain specificcorpus to evaluate the proposed model. Three paragraphs are considered, two forms the mostpopular online Bengali newspaper “Prothom Alo” and another</s>
<s>one from a Bengali news portalon law “LawyersClub Bangladesh.com”. The statistics of the corpus is highlighted in Table 5.2.As the method finally finds the polarity at sentence level using the polarity of the conceptsdetermined through the concept level polarity detection model, therefore, variation in theparagraph does not affect the performance of the model. However, they influence the contextualsense in the sentence and concept level.5.3. EVALUATION METHOD 49Table 5.2: Statistics of the corpus to evaluate the model.Paragraph Number ofSentenceNumber ofSimple SentenceNumber ofconceptParagraph 1 10 17 49Paragraph 2 9 23 44Paragraph 3 13 21 73Total 32 61 166Table 5.3: Confusion Matrix for precision and recall.XXXXXXXXXXXXPredictedActual Positive NegativePositive True Positive False PositiveNegative False Negative True Negative5.3 Evaluation MethodSince the proposed model classifies the concepts as positive, negative and neutral, therefore wewant to know the answer to the following question for evaluating the method:1. What proportion of concepts of specific polarity in the corpus are classified correctly.2. What portion of the classified concepts of specific polarity are correctly classified.3. Finally, what percentage of the total number of concepts in the corpus are classified totheir actual classes.Precision and recall [55] are the perfect metrics to get the answers to the above question forevaluating the proposed model. For binary classification, precision means the percentage ofthe result that is relevant. On the other hand, recall means the percentage of total relevantresult that are predicted correctly in comparison to their actual class. These metrics are realizedmore specifically with the mathematical notation using the confusion matrix given at Table 5.3.The true positive or true negative denotes the instances where the positive or negative case ispredicted correctly to their actual classes. Whereas, the false positive means that an instanceother than the positive is predicted as positive and the false negative means that an instanceother then the negative is predicted as negative comparing to their actual class.From the Table 5.3, we can find the precision and recall value with the Equation (5.1) andEquation (5.2) respectively.Precision =True PositiveActual ResultTrue PositiveTrue Positive+ False Positive(5.1)Recall =True PositivePredicted ResultTrue PositiveTrue Positive+ False Negative(5.2)5.4. RESULT ANALYSIS 50Accuracy is another metric for evaluating a classification model. It provides an idea on theoverall performance of themodel by detecting the percentage of the correctly identified instanceswith respect to the total number of instances in the corpus. It can be expressed in terms ofthe parameters of the confusion matrix as Equation (5.3) for a binary classification scheme.However, we also find the overall accuracy of the model by finding the total number of correctlyclassified concept comparing to the total number of concept in the test corpus. The overallaccuracy of the model can be calculated by applying Equation (5.4).Accuracy =True Positive+ True NegativeTrue Positive+ False Positive+ True Negative+ False Negative(5.3)Accuracy of the model =Number of correctly classified conceptTotal Number of concept in the corpus(5.4)There is a trade-off between recall and precision based on a threshold value in maximizationbetween the two. If we recall everything, we need to keep generating results which are notaccurate, hence lowering the precision. A simplermetric known as F1</s>
<s>score, takes both precisionand recall into account, and represents the balance between recall and precision. F1 Score is alsoviewed as the weighted average of precision and recall which is measured by the Equation (5.5).When one of the precision and recall is given emphasize over the other, the F1 score decreases.On the other hand, a higher value of F1 score expresses the desired harmonic balance betweenprecision and recall.F1 Score = 2× Precision× recallPrecision+ recall(5.5)Although the above metrics are suitable for binary classification, it is also used for multiclassclassification. In that case, metric values for each class are determined considering a binaryclassification as one being the class itself and another is its complement. For instances, whilebuilding the confusion matrix for the neutral class, we consider neutral to be one of theclassification states. Whereas, classification other the neutral is another state which can becompared to positive and negative for a binary classification scheme.5.4 Result AnalysisIn this section, we will evaluate the proposed method through result analysis based on the valuesof the evaluating metrics. To start with, we want to evaluate the performance of the parse treeand the concept extractor. As there is no traditional metric to evaluate the parse tree, we focusedon evaluating the concept extractor. Then, we evaluate the performance of the classificationmodel used for classifying the training data. Finally, we will analyze the accuracy of the polaritydetection at concept level as well as sentence level.5.4. RESULT ANALYSIS 51Table 5.4: Performance of the concept extractor.Number ofSentenceNumber ofConceptPerfectlyExtractedPartiallyExtractedNotExtracted31 166 87 68 115.4.1 Performance Analysis on Concept ExtractionParse tree generation is the precursory step for concept extraction. The performance of conceptextractor depends on the proficiency of the parse tree, thereby, concept extractor exhibits theperformance of the parse tree as well. the concept being a newly introduced entity in the fieldof NLP, very few works are available in the state of art. We hardly find any works at conceptlevel in the Bengali language. Therefore, we evaluate the performance of the concept extractorby comparing the extracted concepts with the expected concepts within a sentence. In this work,about 53 percent of the concepts are extracted which match to our expected concepts and 41percent of the concept was extracted partially.Partial extraction is elucidated with an example such as, in the sentence “ মেয়িট ইভিটিজং বাউ তার িশকার হওয়াই থানায় অিভেযাগ কের ” ([meeti ibHtidZiN ba uttktOtar Sikar HOai tHaneObHidZog kre], The girl complain to police for being victim of eve-teasing), both “ইভিটিজং এরিশকার ” ([ibHtidZiN er Sikar], eve-teasing victim) and “উ তারর িশকার ” ([uttktOtar Sikar],harassment) are expected to be extracted as concept. Whereas, the first one was partiallyextracted as “ইভিটিজং ” ([ibHtidZiN], eve-teasing) even though the second one was accuratelyextracted as “উ তারর িশকার ” ([uttktOtar Sikar], harassment). However, 6 percent of theterms are not extracted, though they may form a potential concept. These terms need to form aconcept along with some other terms which are already used in another concept. Therefore, theterms are ignored as the proposed concept extractor does not allow multiple uses of a</s>
<s>term in asentence. Table 5.4 represents the performance of the concept extractor.Proposed method failed to extract very few concepts of which most of them are adverb notpositioned adjacent to the verb. Some of the auxiliary verbs sometime express the sentiment,hence, expected to be extracted as a concept. However, this performance can not be comparedwith other existing concept extractors as the proposed extractor is domain and applicationspecific. Therefore, it may not perform well for other applications and languages to extractconcepts from the sentence.5.4.2 Evaluation on Classification of Training DataThe accuracy of the proposed method depends highly on the accuracy of the classifier whileperforming supervised classification of the training data. Wewill try to evaluate the performanceof the classification of training data set with the integral metrics that are provided by the LDA inSPSS. SPSS provides some of the statistical as well as graphical analysis on the performance of5.4. RESULT ANALYSIS 52Figure 5.1: Positional overview of the data points based on the canonical discriminantfunction.the classifier. Figure 5.1 shows the positional overview of the data point based on the canonicaldiscriminant functions.Based on the accuracy of the classification, SPSS provides a statistical summary of classificationresult in Figure 5.2. The summary result represents the performance of the classifier to classifythe concepts correctly as per the classes assigned to the concepts. The overall accuracy of theclassifier is 88.8 percent which means that only 11.2 percent of concepts have not been classifiedto their original classes. We can observe from the summary result that only one concept withnegative class is classified as positive and five concepts of the positive class is classified asnegative which are very negligible in the context of NLP.A significant number of concepts from positive and negative class are classified as neutral andvice versa. In practical terms, some of the neutral concepts may shift their valence to positiveor negative sentiment and contrariwise. As a matter of fact, there is a very low semanticgap between neutral and positive or neutral and negative. As a consequence, the positionaldistance between them in the affective space are also marginal and sometimes overlap each other.Therefore, most of the miss classification occurs between neutral and positive class or neutraland negative class in the affective space. Nevertheless, there is hardly any miss classificationfound between positive and negative. This is because the affective strength of the positiveand negative concepts are usually very strong and position them opposite to each other in the5.4. RESULT ANALYSIS 53Figure 5.2: Summary result of classification of training data with three polarity classes.Figure 5.3: Summary result of classification of training data with two polarity classes.affective space.This can be further analyzed through a classification scheme with a training data set with twopolarity classes. Figure 5.2 shows the statistical summary of the classification result with twopolarity classes. 95.3 percent of the concepts are correctly classified to their assigned class. Thereason for the improvement in accuracy is already mentioned above. The neutral concepts areclassified into positive and negative depending on their inclination towards positive and negativeuse in the proposed domain. Despite the better</s>
<s>efficiency of binary classification, the exclusionof neutral class may affect the efficiency in determining the sentence valence as many neutralconcepts will be considered as either positive or negative. Sometimes, it is cumbersome to dealwith the concepts which are neither positive nor negative.However, the binary classification is very effective to handle the concepts which shift theirvalence depending on the context of the sentence. Depending on the domain context, theseconcepts are classified into desired classes which also contribute to function coefficient matrixused to build the polarity detection function. Therefore, the binary classification scheme issometimes preferable where the detection of sentence polarity is the primary task, even though,it is not effective for polarity detection at the concept level.5.4. RESULT ANALYSIS 54Figure 5.4: Fraction of classification function coefficient matrix.5.4.3 Analysis of the Polarity Detection ModelAt this stage, we focus on evaluating our proposed model for polarity detection at the conceptlevel. As mentioned in methodology (Chapter 4), the performance of the classification functionto determine the polarity of a concept depends on the classification function coefficient matrixgenerated by SPSS. Figure 5.4 shows a fraction of the classification function coefficient matrixgenerated based on our training data. It provides coefficient values for classification functionfor each of the three classes. It contains forty one eigenmoods along with values of the constant.As we discussed early that the strength of the eigenmood in the affective space decreases as theorder increases. Therefore, the higher ordered eigenmoods have very minimal significance indetermining the concept class. Therefore, LDA carries out tolerance test for each variable andignores the eigenmoods failing the test to qualify for the coefficient matrix.Performance of the proposed polarity detection model mainly relies on the efficiency of thepolarity detection function. As the proposed model classifies the polarity in three classesas positive, negative and neutral, we will use the multiple class evaluation technique for theprecision and recall method. Therefore, we generate the confusionmatrix for each of the polarityclasses of concepts to evaluate the performance of the polarity detector. Table 5.5, Table 5.6 andTable 5.7 show the confusion matrix for the positive, negative and neutral polarity classes of5.4. RESULT ANALYSIS 55Table 5.5: Confusion Matrix for the concept with positive polarity.XXXXXXXXXXXXPredictedActual Positive Not PositivePositive 22 10Not Positive 15 06Table 5.6: Confusion Matrix for the concept with negative polarity.XXXXXXXXXXXXPredictedActual Negative Not NegativeNegative 35 15Not Negative 10 19Table 5.7: Confusion Matrix for the concept with neutral polarity.XXXXXXXXXXXXPredictedActual Neutral Not NeutralNeutral 33 17Not Neutral 15 60concept respectively. Here, the polarity class of concepts, determine by the polarity detectoris considered as predicted class, whereas, the actual class is determined by manual annotation.In the confusion matrix of each class, the class itself considered as positive, whereas, the classother than itself is considered as negative to conform with binary classification metric as shownin Table 5.3.To evaluate the performance of the proposed model; precision, recall, F1 score and the accuracyfor each of the polarity classes are calculated from the confusion matrix. We use the equationsmentioned in Section 5.3 to calculate the values of the metrics for each polarity class as shownin Table</s>
<s>5.8. A system with high recall but low precision returns a high number of instancesfor desired class but many of the predictions are incorrect compared to their actual class. Asystem with high precision but low recall is just the opposite, returning very few results for thedesired class, but most of which are predicted correctly as their actual class. An ideal systemwith high precision and high recall will maximize the instances of the desired class with a correctprediction to their actual class.In observing the result of our polarity detection model, we find that the accuracy in classifyingthe positive class is lower than the other two classes. On the other hand, the negative class showsa very high recall value. This can be explained through semantic gaps among the concepts andthe affective strength of the concepts in the affective space. As the affective strength of negativeconcepts usually very strong, therefore, they have very low semantic overlapping with otherclass and shows high recall value. On the contrary, the semantic gap between positive andnegative class is very low and reduces the recall value for positive and neutral class. The spatialdistribution of positive and neutral concepts in the affective space is also overlapping each other.5.4. RESULT ANALYSIS 56Table 5.8: Values of the metrics for the polarity classes of concept.Polarity Class Precision Recall F1 Score Accuracy Model AccuracyPositive 68.75 59.46 61.97 52.8370.24Negative 70.00 77.77 73.68 68.35Neutral 66.00 68.75 67.35 74.40Table 5.9: Performance evaluation for polarity detection at sentence level.SimpleSentenceComplex/ CompoundSentenceCount Accuracy Count Accuracy61 73.77 32 65.63This spatial overlapping influences to miss classify between positive and negative class whichreduces the precision as well.However, the precision values for all the classes are almost equal and are also well accepted inthe context of polarity classification as an NLP task for a less privileged language like Bengali.The balance between recall and precision can be observed through the F1 score. The F1 scorefor the negative class is higher than the other classes. This is because of the high precision andrecall and the maintenance of balance between the twomeasures. The F1 scores for other classesbeing greater than 0.50 are also in an acceptable range. However, the F1 score is not convenientfor evaluating the proposed method, as it completely ignores the true negatives. The overallaccuracy of the model shows that 70.40 percent of the total number of concepts are classifiedcorrectly. Acquirement this performance is considered promising, especially, for an ambiguouslanguage like Bengali.In order to evaluate the performance of polarity detection at the sentence level, we compare thepolarity of a sentence determined by the proposed model with the actual polarity of the sentence.However, discovering of the actual polarity of a sentence is sometimes cumbersome, especially,for the complex and compound sentence where different parts or clauses of the sentence expressdifferent polarity. In case of the simple sentences, these ambiguities are very limited as theynormally inherit a single context. Therefore, we evaluate the accuracy both for simple sentencesas well as the original sentences as presented in the documents. Table 5.9 shows the performanceof polarity detection at sentence</s>
<s>level both for simple sentences and the complex or compoundsentences.We observe that the accuracy for polarity detection of the simple sentences is higher than theaccuracy of complex or compound sentence. This is due to the complexity raised in the instanceswhere different parts of the complex or compound sentence contain opposite polarity. In mostcases, the model transmutes the polarity of a complex or compound sentence to negative if anypart of the sentence contains a negative simple sentence. In the same context, we find that thenumber of false negative is much higher than the number of false positive out of the total miss5.5. DISCUSSION 57classification.5.5 DiscussionTo our knowledge, there is hardly any work that finds the polarity at concept level in the Bengalilanguage. Fewworks determine the polarity of the keywords or key phrases using various lexicalresources which are mainly dictionary based approaches. In addition, the proposed model isdomain dependent and language specific. Comparison of the model to the other methods mayraise some miss leading result. Therefore, we evaluate our model by comparing the predictedresult with the actual or desired result. The overall accuracy for polarity detection at conceptlevel being above 70 percent is considered to be an acceptable range for a less privilegedlanguage like Bengali. However, some of the concept level polarity detection model in richlanguages like English, French, Chinese shows accuracy up to 80 percent. This is possible dueto the availability of the highly efficient NLP tools for those languages.Though the accuracy for polarity detection of simple sentences is very high, the efficiency isreduced in case of the complex and compound sentences. Variations in polarity of the simplesentences as parts of the complex and compound sentence raise the ambiguities. An efficientdependency parser can solve the problem and improve the efficiency of the polarity detection atthe sentence level. This requires a detail linguistic analysis incorporating the morphological andgrammatical influences which are beyond the scope of this thesis. However, the performance ofthe parse tree in accordance with the concept extractor shows a satisfactory result especially inour application to determine the concept level polarity in the domain of woman harassment.Chapter 6ConclusionThis chapter will conclude the thesis with a brief summery of the works in Section 6.1 includingthe performance evaluation. Section 6.2 will highlight some challenges encountered in thisthesis. Finally, scope of the future works will be enumerated in Section 6.3.6.1 Contribution of the WorkWith the increasing importance of autonomous information retrieval from online content,sentiment analysis becomes a popular area of research in the field of NLP. Bengali beingone of the most significant languages in the world, has a scarcity of NLP tools and resource.Particularly, there is hardly any appropriate parser that explores the Bengali language structurecomprehensively. This provides themotivation for developing an independent parser for Bengalisentences. The methodology follows a rule based approach which explores the languagemorphology and syntactical structure. This parser is capable of decomposing the Bengalisentence into clauses, key phrases and terms. The final outcome of the parser is a parse treewhich also exhibits the dependencies among the constituents of the sentence.In</s>
<s>the latter part of this thesis, a concept level polarity detection model is introduced using alearning method which uses the AffectiveSpace as a knowledge base. The motivation behindusing the AffectiveSpace is to infer the semantic and affective information associated withnatural language opinions. Moreover, concept-based approaches are the recent evolution insentiment analysis that step away from blind use of keywords and word co-occurrence counts todetect polarity. Since the concept usually shifts the valence depending on the contextual domain,therefore, themodel is developed as domain specific to increase the classification accuracywhiledetecting the polarity.The proposed model finds the polarity of the concepts that are extracted from the parse treegenerated by the rule based semantic parser. Doing so, it determines a polarity detection functionfor each of the positive, negative and neutral classes through classification model. LDA is6.2. CHALLENGES 59used as a mathematical model in which the training data are applied to generate the polaritydetection function. Prior to the classification, the concepts are mapped to the AffectiveSpaceto get the eigenmood values which are considered to be the classification features for eachconcept. The polarity detection function finds the positive, negative and neutral score for eachconcept applying the eigenvalues as the parameters to the functions. Comparing the scores of thepolarity detection functions, the final polarity of a concept is determined. Finally, the polarityof the sentence is determined through the tree traversal in reverse order. In this process, theresultant polarity for a sentence highly depends on the mutual relationship among the conceptsand dependencies of them on the modifiers.Both the training data and the test data were developed using traditional NLP tools. However, theannotation of the data set requires some major improvement as it highly affects the subsequentstages of this work. The improvement was done through manual annotation. A comprehensiveevaluation through result analysis based on precision and recall is carried out using the testdata set. The evaluation shows a satisfactory performance of the work for polarity detection atconcept level as well as sentence level.6.2 ChallengesIn the implementation phase of the thesis, one of the major challenges was to perform NLP taskin Bengali. There is an acute scarcity of appropriate NLP tools with good performance. Thecomplexity of Bengali language structure and grammatical directives is also required specialattention. Since we use the AffectiveSpace that is built in English, therefore, translating theconcepts from Bengali to English and mapping them in the space was sometime ambiguous.Some of the concepts may be translated to different English concepts which yield the confusionin selecting the appropriate one.6.3 Scope of Future WorkFrom the challenges and the explanation of result analysis, we can highlight some of the scopeof the future works to step forward from this thesis and improve the performance as well. Thefuture works are listed as follows:1. An independent dependency parser can be generated for Bengali, which is capable ofidentifying the dependency among the concepts as well as the modifiers. It requires anexhaustive study of Bengali grammar and linguistic structure to explore the relationshipamong the constituents of the sentence. The dependencies among the POS and theirvariation within</s>
<s>the sentence based on stem categories need to address to increase the6.3. SCOPE OF FUTURE WORK 60efficiency of the parser. Integration of morphological knowledge can have a great impacton the performance of the dependency parser.2. The concept extractor presented in this thesis extracts the concepts containing only theterms found within the sentence. To achieve a higher level abstraction while extractingthe concepts from a sentence, the new word, terms or phrases may be of great utilization.Thereby, the performance of the concept extractor might be improved, hence, improvesthe accuracy of the polarity detection model.3. Besides the difficulties in translating the concepts that are mentioned earlier, it sometimesdrops out some linguistic information as well. This may have some significant impact inthe context of sentiment analysis at the sentence level. Therefore, an affective space inBengali can be generated for an exact representation of concepts extracted from sentences.This requires a huge effort in collecting the data set to build the knowledge base. However,a transformation method can be applied to mirror the English AffectiveSpace to Bengaliwhere the approximation must be made with great care.4. There is a lot of room for improvement in polarity detection at the sentence level. Oncethe polarity of the sentence elements is known, the sentence polarity can be determinedmore efficiently without using any NLP resources. Here, the approximation of resultantpolarity of the clauses or phrases from the polarities of the child node is very crucial.Semantic affinity along with the syntactic dependencies can improve the performance ofsentence level polarity detection.The concept level sentiment analysis presented in this thesis is a new integration of this typeto the field of NLP tasks in Bengali. Having a number of limitations, it requires a lot ofimprovement to achieve the desired abstraction level of a concept. Besides the sentimentanalysis, the concept based analysis can be implemented in NLP tasks like opinion mining,document summarizing, subjectivity detection, etc.References[1] M. Z. Asghar, A. Khan, S. Ahmad, M. Qasim, and I. A. Khan, “Lexicon-enhancedsentiment analysis framework using rule-based classification scheme,” PloS one, vol. 12,no. 2, p. e0171649, 2017.[2] S. Poria, E. Cambria, G. Winterstein, and G.-B. Huang, “Sentic patterns: Dependency-based rules for concept-level sentiment analysis,” Knowledge-Based Systems, vol. 69,pp. 45–63, 2014.[3] E. Cambria, A. Hussain, C. Havasi, and C. Eckl, “AffectiveSpace: blending common senseand affective knowledge to perform emotive reasoning,” WOMSA at CAEPIA, Seville,pp. 32–41, 2009.[4] E. Cambria and A. Hussain, Sentic computing: Techniques, tools, and applications, vol. 2.Springer Science & Business Media, 2012.[5] C.Manning, M. Surdeanu, J. Bauer, J. Finkel, S. Bethard, and D.McClosky, “The StanfordCoreNLP natural language processing toolkit,” in Proceedings of 52nd annual meeting ofthe association for computational linguistics: system demonstrations, pp. 55–60, 2014.[6] M.-C. DeMarneffe and C. D. Manning, “The stanford typed dependencies representation,”inColing 2008: proceedings of the workshop on cross-framework and cross-domain parserevaluation, pp. 1–8, Association for Computational Linguistics, 2008.[7] C. Strapparava et al., “Wordnet affect: an affective extension of wordnet.,” in Lrec, vol. 4,pp. 1083–1086, Citeseer, 2004.[8] K. Ravi and V. Ravi, “A survey on opinion mining and sentiment analysis: tasks,approaches and applications,” Knowledge-Based Systems, vol.</s>
<s>89, pp. 14–46, 2015.[9] S. Baccianella, A. Esuli, and F. Sebastiani, “Sentiwordnet 3.0: an enhanced lexicalresource for sentiment analysis and opinion mining,” in Lrec, pp. 2200–2204, 2010.[10] G. F. Simons and C. D. Fennig, “Ethnologue: Language of the World.” https://www.ethnologue.com/ethnoblog/gary-simons/welcome-21st-edition, Feb.2018. (last accessed <Jan 19, 2019>).https://www.ethnologue.com/ethnoblog/gary-simons/welcome-21st-editionhttps://www.ethnologue.com/ethnoblog/gary-simons/welcome-21st-editionREFERENCES 62[11] L. Mehedy, S. N. Arifin, and M. Kaykobad, “Bangla syntax analysis: A comprehensiveapproach,” in Proceedings of International Conference on Computer and InformationTechnology (ICCIT), Dhaka, Bangladesh, pp. 287–293, 2003.[12] G. Qiu, B. Liu, J. Bu, and C. Chen, “Expanding domain sentiment lexicon through doublepropagation,” in Twenty-First International Joint Conference on Artificial Intelligence,2009.[13] N. S. Dash and L. Ramamoorthy, “Corpus and technical termbank,” in Utility andApplication of Language Corpora, pp. 173–191, Springer, 2019.[14] S. Sun, C. Luo, and J. Chen, “A review of natural language processing techniques foropinion mining systems,” Information Fusion, vol. 36, pp. 10–25, 2017.[15] B. Agarwal, S. Poria, N. Mittal, A. Gelbukh, and A. Hussain, “Concept-level sentimentanalysis with dependency-based semantic parsing: a novel approach,” CognitiveComputation, vol. 7, no. 4, pp. 487–499, 2015.[16] T. Mullen and N. Collier, “Sentiment analysis using support vector machines with diverseinformation sources,” in Proceedings of the 2004 conference on empirical methods innatural language processing, 2004.[17] C. Whitelaw, N. Garg, and S. Argamon, “Using appraisal groups for sentiment analysis,”in Proceedings of the 14th ACM international conference on Information and knowledgemanagement, pp. 625–631, ACM, 2005.[18] V. Hatzivassiloglou and K. R. McKeown, “Predicting the semantic orientation ofadjectives,” inProceedings of the 35th annual meeting of the association for computationallinguistics and eighth conference of the european chapter of the association forcomputational linguistics, pp. 174–181, Association for Computational Linguistics, 1997.[19] A. Esuli and F. Sebastiani, “Determining the semantic orientation of terms through glossclassification,” in Proceedings of the 14th ACM international conference on Informationand knowledge management, pp. 617–624, ACM, 2005.[20] C. H. E. Gilbert, “Vader: A parsimonious rule-based model for sentiment analysis of socialmedia text,” in Eighth International Conference on Weblogs and Social Media (ICWSM-14). Available at (20/04/16) http://comp. social. gatech. edu/papers/icwsm14. vader. hutto.pdf, 2014.[21] A. Mollahosseini, B. Hasani, and M. H. Mahoor, “Affectnet: A database for facialexpression, valence, and arousal computing in the wild,” arXiv preprint arXiv:1708.03985,2017.REFERENCES 63[22] H. Saggionα and A. Funk, “Interpreting SentiWordNet for opinion classification,” inProceedings of the seventh conference on international language resources and evaluationLREC10, 2010.[23] M. Hu and B. Liu, “Mining and summarizing customer reviews,” in Proceedings of thetenth ACM SIGKDD international conference on Knowledge discovery and data mining,pp. 168–177, ACM, 2004.[24] A.-M. Popescu, B. Nguyen, and O. Etzioni, “Opine: Extracting product features andopinions from reviews,” in Proceedings of HLT/EMNLP on interactive demonstrations,pp. 32–33, Association for Computational Linguistics, 2005.[25] B. Pang and L. Lee, “Opinion mining and sentiment analysis,” Foundations and Trends®in Information Retrieval, vol. 2, no. 1–2, pp. 1–135, 2008.[26] P. D. Turney, “Thumbs up or thumbs down?: semantic orientation applied to unsupervisedclassification of reviews,” in Proceedings of the 40th annual meeting on association forcomputational linguistics, pp. 417–424, Association for Computational Linguistics, 2002.[27] X. Ding, B. Liu, and P. S. Yu, “A holistic lexicon-based approach to opinion mining,”</s>
<s>inProceedings of the 2008 international conference on web search and data mining, pp. 231–240, ACM, 2008.[28] B. Pang, L. Lee, and S. Vaithyanathan, “Thumbs up?: sentiment classificationusing machine learning techniques,” in Proceedings of the ACL-02 conference onEmpirical methods in natural language processing-Volume 10, pp. 79–86, Association forComputational Linguistics, 2002.[29] A. Tripathy, A. Agrawal, and S. K. Rath, “Classification of Sentimental Reviews UsingMachine Learning Techniques,” Procedia Computer Science, vol. 57, pp. 821–829, 2015.[30] D. Zhang, H. Xu, Z. Su, and Y. Xu, “Chinese comments sentiment classification based onword2vec and SVMperf,” Expert Systems with Applications, vol. 42, no. 4, pp. 1857–1863,2015.[31] J. Zhao, K. Liu, and G. Wang, “Adding redundant features for CRFs-based sentencesentiment classification,” inProceedings of the conference on empirical methods in naturallanguage processing, pp. 117–126, Association for Computational Linguistics, 2008.[32] E. Breck, Y. Choi, and C. Cardie, “Identifying Expressions of Opinion in Context,” inIJCAI, vol. 7, pp. 2683–2688, 2007.REFERENCES 64[33] B. G. Patra, S. Mandal, D. Das, and S. Bandyopadhyay, “Ju_cse: A conditional randomfield (crf) based approach to aspect based sentiment analysis,” in Proceedings of the 8thInternational Workshop on Semantic Evaluation (SemEval 2014), pp. 370–374, 2014.[34] Y. Xiang, H. He, and J. Zheng, “Aspect Term Extraction Based on MFE-CRF,”Information, vol. 9, no. 8, p. 198, 2018.[35] B. Kratzwald, S. Ilić, M. Kraus, S. Feuerriegel, and H. Prendinger, “Deep learningfor affective computing: Text-based emotion recognition in decision support,” DecisionSupport Systems, vol. 115, pp. 24–35, 2018.[36] H. Liu and P. Singh, “ConceptNet—a practical commonsense reasoning tool-kit,” BTtechnology journal, vol. 22, no. 4, pp. 211–226, 2004.[37] E. Cambria, A. Livingstone, and A. Hussain, “The hourglass of emotions,” in Cognitivebehavioural systems, pp. 144–157, Springer, 2012.[38] A. Björkelund, B. Bohnet, L. Hafdell, and P. Nugues, “A high-performance syntactic andsemantic dependency parser,” in Proceedings of the 23rd International Conference onComputational Linguistics: Demonstrations, pp. 33–36, Association for ComputationalLinguistics, 2010.[39] J. Hajič, M. Ciaramita, R. Johansson, D. Kawahara, M. A. Martí, L. Màrquez, A. Meyers,J. Nivre, S. Padó, and J. Štěpánek, “The CoNLL-2009 shared task: Syntactic and semanticdependencies in multiple languages,” in Proceedings of the Thirteenth Conference onComputational Natural Language Learning: Shared Task, pp. 1–18, Association forComputational Linguistics, 2009.[40] T. Koo, X. Carreras, and M. Collins, “Simple semi-supervised dependency parsing,”Proceedings of ACL-08: HLT, pp. 595–603, 2008.[41] C. D. Manning, “Part-of-speech tagging from 97% to 100%: is it time for somelinguistics?,” in International conference on intelligent text processing and computationallinguistics, pp. 171–189, Springer, 2011.[42] A. Das and S. Bandyopadhyay, “Sentiwordnet for Bangla,” Knowledge Sharing Event-4:Task, vol. 2, pp. 1–8, 2010.[43] D. Das and S. Bandyopadhyay, “Developing Bengali WordNet affect for analyzingemotion,” in International Conference on the Computer Processing of Oriental Languages,pp. 35–40, 2010.[44] A. Das and S. Bandyopadhyay, “Opinion-polarity identification in Bengali,” in Interna-tional Conference on Computer Processing of Oriental Languages, pp. 169–182, 2010.REFERENCES 65[45] J. Wieting, M. Bansal, K. Gimpel, and K. Livescu, “Towards universal paraphrasticsentence embeddings,” arXiv preprint arXiv:1511.08198, 2015.[46] “Oxford Learner’s Dictionaries.” https://www.oxfordlearnersdictionaries.com.[47] N. D. Goodman, J. B. Tenenbaum, J. Feldman, and T. L. Griffiths, “A rational analysis ofrule-based concept learning,” Cognitive science, vol. 32, no.</s>
<s>1, pp. 108–154, 2008.[48] A. Polguère et al., Dependency in linguistic description, vol. 111. John BenjaminsPublishing, 2009.[49] Robin, “Article on Natural Language Processing.” http://language.worldofcomputing.net/category/tokenization, Nov. 2009. (last accessed<Jan 19, 2019>).[50] A. Ekbal, R. Haque, and S. Bandyopadhyay, “Bengali part of speech tagging usingconditional random field,” in Proceedings of the seventh International Symposium onNatural Language Processing, SNLP-2007, 2007.[51] T. Koo, X. Carreras, and M. Collins, “Simple Semi-supervised Dependency Parsing,” inProceedings of ACL-08: HLT, Association for Computational Linguistics, 2008.[52] C. Fellbaum, “Wordnet,” in Theory and applications of ontology: computer applications,pp. 231–243, Springer, 2010.[53] E. Cambria, J. Fu, F. Bisio, and S. Poria, “Affectivespace 2: Enabling affective intuitionfor concept-level sentiment analysis.,” in AAAI, pp. 508–514, 2015.[54] N. Wang, P. Wang, and B. Zhang, “An improved TF-IDF weights function based oninformation theory,” in 2010 International Conference on Computer and CommunicationTechnologies in Agriculture Engineering, vol. 3, pp. 439–441, IEEE, 2010.[55] K. M. Ting, “Precision and recall,” Encyclopedia of machine learning, 2010.http://language.worldofcomputing.net/category/tokenizationhttp://language.worldofcomputing.net/category/tokenizationGenerated using Postgraduate Thesis LATEX Template, Version 0.97. Department ofComputer Science and Engineering, Bangladesh University of Engineering and Technology,Dhaka, Bangladesh.This thesis was generated on March 12, 2019 at 10:22am. List of Figures List of Tables Abstract Introduction Background Problem Definition Research Aim and Objective Overview of the Work Thesis Contribution and Final Outcome Thesis Outline Related Works Sentiment Analysis Rule Based Approaches Machine Learning Approaches Concept Based Approaches Sentence Parsing and Concept Extraction Sentiment Analysis in Bengali Scope of the Work Research Questions Preliminaries NLP Fundamentals Lexicon Concept Corpus Semantic Dependency Opinion Sentiment NLP Techniques Tokenization POS Tagging Parsing NLP Resources Concept Net WordNet-Affect AffectiveSpace Mathematical Models Singular Value Decomposition (SVD) Linear Discriminant Analysis (LDA) Proposed Methodology Overview of the Methodology Data Acquisition Parse Tree Generation Concept Extraction and Dependency Detection Polarity Detection Construction of the Polarity Detection Model Determination of the Polarity of the Sentence Experimental Analysis Experimental Setup Experimental Data Set Evaluation Method Result Analysis Performance Analysis on Concept Extraction Evaluation on Classification of Training Data Analysis of the Polarity Detection Model Discussion Conclusion Contribution of the Work Challenges Scope of Future Work References</s>
<s>Semantic Textual Similarity in Bengali TextSee discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/329394770Semantic Textual Similarity in Bengali TextConference Paper · September 2018DOI: 10.1109/ICBSLP.2018.8554940CITATIONSREADS3082 authors:Some of the authors of this publication are also working on these related projects:Image captioning View projectunsupervised learning View projectMd ShajalalHajee Mohammad Danesh Science and Technology University12 PUBLICATIONS 24 CITATIONS SEE PROFILEMasaki AonoToyohashi University of Technology133 PUBLICATIONS 2,436 CITATIONS SEE PROFILEAll content following this page was uploaded by Md Shajalal on 09 December 2018.The user has requested enhancement of the downloaded file.https://www.researchgate.net/publication/329394770_Semantic_Textual_Similarity_in_Bengali_Text?enrichId=rgreq-1cea3bcadcc9b1a4c7cb9b25e45b65fb-XXX&enrichSource=Y292ZXJQYWdlOzMyOTM5NDc3MDtBUzo3MDE4MDk5Mjg3MDgwOTdAMTU0NDMzNTkzNDU3OA%3D%3D&el=1_x_2&_esc=publicationCoverPdfhttps://www.researchgate.net/publication/329394770_Semantic_Textual_Similarity_in_Bengali_Text?enrichId=rgreq-1cea3bcadcc9b1a4c7cb9b25e45b65fb-XXX&enrichSource=Y292ZXJQYWdlOzMyOTM5NDc3MDtBUzo3MDE4MDk5Mjg3MDgwOTdAMTU0NDMzNTkzNDU3OA%3D%3D&el=1_x_3&_esc=publicationCoverPdfhttps://www.researchgate.net/project/Image-captioning?enrichId=rgreq-1cea3bcadcc9b1a4c7cb9b25e45b65fb-XXX&enrichSource=Y292ZXJQYWdlOzMyOTM5NDc3MDtBUzo3MDE4MDk5Mjg3MDgwOTdAMTU0NDMzNTkzNDU3OA%3D%3D&el=1_x_9&_esc=publicationCoverPdfhttps://www.researchgate.net/project/unsupervised-learning?enrichId=rgreq-1cea3bcadcc9b1a4c7cb9b25e45b65fb-XXX&enrichSource=Y292ZXJQYWdlOzMyOTM5NDc3MDtBUzo3MDE4MDk5Mjg3MDgwOTdAMTU0NDMzNTkzNDU3OA%3D%3D&el=1_x_9&_esc=publicationCoverPdfhttps://www.researchgate.net/?enrichId=rgreq-1cea3bcadcc9b1a4c7cb9b25e45b65fb-XXX&enrichSource=Y292ZXJQYWdlOzMyOTM5NDc3MDtBUzo3MDE4MDk5Mjg3MDgwOTdAMTU0NDMzNTkzNDU3OA%3D%3D&el=1_x_1&_esc=publicationCoverPdfhttps://www.researchgate.net/profile/Md_Shajalal?enrichId=rgreq-1cea3bcadcc9b1a4c7cb9b25e45b65fb-XXX&enrichSource=Y292ZXJQYWdlOzMyOTM5NDc3MDtBUzo3MDE4MDk5Mjg3MDgwOTdAMTU0NDMzNTkzNDU3OA%3D%3D&el=1_x_4&_esc=publicationCoverPdfhttps://www.researchgate.net/profile/Md_Shajalal?enrichId=rgreq-1cea3bcadcc9b1a4c7cb9b25e45b65fb-XXX&enrichSource=Y292ZXJQYWdlOzMyOTM5NDc3MDtBUzo3MDE4MDk5Mjg3MDgwOTdAMTU0NDMzNTkzNDU3OA%3D%3D&el=1_x_5&_esc=publicationCoverPdfhttps://www.researchgate.net/institution/Hajee_Mohammad_Danesh_Science_and_Technology_University?enrichId=rgreq-1cea3bcadcc9b1a4c7cb9b25e45b65fb-XXX&enrichSource=Y292ZXJQYWdlOzMyOTM5NDc3MDtBUzo3MDE4MDk5Mjg3MDgwOTdAMTU0NDMzNTkzNDU3OA%3D%3D&el=1_x_6&_esc=publicationCoverPdfhttps://www.researchgate.net/profile/Md_Shajalal?enrichId=rgreq-1cea3bcadcc9b1a4c7cb9b25e45b65fb-XXX&enrichSource=Y292ZXJQYWdlOzMyOTM5NDc3MDtBUzo3MDE4MDk5Mjg3MDgwOTdAMTU0NDMzNTkzNDU3OA%3D%3D&el=1_x_7&_esc=publicationCoverPdfhttps://www.researchgate.net/profile/Masaki_Aono?enrichId=rgreq-1cea3bcadcc9b1a4c7cb9b25e45b65fb-XXX&enrichSource=Y292ZXJQYWdlOzMyOTM5NDc3MDtBUzo3MDE4MDk5Mjg3MDgwOTdAMTU0NDMzNTkzNDU3OA%3D%3D&el=1_x_4&_esc=publicationCoverPdfhttps://www.researchgate.net/profile/Masaki_Aono?enrichId=rgreq-1cea3bcadcc9b1a4c7cb9b25e45b65fb-XXX&enrichSource=Y292ZXJQYWdlOzMyOTM5NDc3MDtBUzo3MDE4MDk5Mjg3MDgwOTdAMTU0NDMzNTkzNDU3OA%3D%3D&el=1_x_5&_esc=publicationCoverPdfhttps://www.researchgate.net/institution/Toyohashi_University_of_Technology?enrichId=rgreq-1cea3bcadcc9b1a4c7cb9b25e45b65fb-XXX&enrichSource=Y292ZXJQYWdlOzMyOTM5NDc3MDtBUzo3MDE4MDk5Mjg3MDgwOTdAMTU0NDMzNTkzNDU3OA%3D%3D&el=1_x_6&_esc=publicationCoverPdfhttps://www.researchgate.net/profile/Masaki_Aono?enrichId=rgreq-1cea3bcadcc9b1a4c7cb9b25e45b65fb-XXX&enrichSource=Y292ZXJQYWdlOzMyOTM5NDc3MDtBUzo3MDE4MDk5Mjg3MDgwOTdAMTU0NDMzNTkzNDU3OA%3D%3D&el=1_x_7&_esc=publicationCoverPdfhttps://www.researchgate.net/profile/Md_Shajalal?enrichId=rgreq-1cea3bcadcc9b1a4c7cb9b25e45b65fb-XXX&enrichSource=Y292ZXJQYWdlOzMyOTM5NDc3MDtBUzo3MDE4MDk5Mjg3MDgwOTdAMTU0NDMzNTkzNDU3OA%3D%3D&el=1_x_10&_esc=publicationCoverPdfInternational Conference on Bangla Speech and Language Processing (ICBSLP), 21-22 September, 2018Semantic Textual Similarity in Bengali TextMd ShajalalDepartment of Computer Science & EngineeringBangladesh Army University of Science and TechnologySaidpur, Nilphamari, Bangladeshshajalal@baust.edu.bdMasaki AonoDepartment of Computer Science & EngineeringToyohashi University of TechnologyToyohashi, Aichi, Japanaono@tut.jpAbstract—Measuring the textual similarity is indis-pensable in many information retrieval applications.Researchers proposed numerous similarity measuresto compute the semantic similarity between texts formonolingual and multilingual texts. But methods formeasuring similarity for Bengali text segments are notso commonly available. In this paper, we propose anapproach to estimate the semantic similarity betweenBengali text segments. The similarity score is computedwith the help of word-level semantics from a pre-trainedword-embedding model trained on Bengali Wikipediatexts. In this regard, we employ an algorithm to mea-sure the semantic similarity of Bengali texts. To test theperformance of our method, we conducted experimentson a dataset for semantic textural similarity for Bengalitexts. We prepare the dataset using the same approachas SemEval applied in the STS 2017. The experimen-tal results in terms of Pearson correlation coefficientconclude that our method achieves a state-of-the-artperformance for semantic textual similarity in Bengalitexts.Key–Words: Bengali Textual Similarity;Semantic Sim-ilarity; Word-level Semantics; Word-embedding.I. IntroductionThe textual similarity between texts is an importantand mandatory task in many applications in informationretrieval. The performance of many natural language pro-cessing (NLP) applications such as text summarization,machine translation, plagiarism detection, sentiment anal-ysis etc. is also dependent on the textual and semanticsimilarity. There are also some other applications thatused the similarity such as relevance feedback, text classi-fication, word sense disambiguation, subtopic mining, websearch and so on [1]–[3]. The similarity measures for manylanguages like English, Arabic, Spanish are available andthere are some research tasks to compute the similaritybetween multilingual and monolingual texts organized bySemEval STS organizer [4]–[6].One of the typical approaches to compute the similarityis lexical matching between texts. The similarity scoreis computed based on the number of terms belong toboth text segments. But these measures are not able tocompute the similarity beyond a trivial level. Moreover,this matching can only estimate the textual similarity butnot semantic. Let us consider two texts, “িতিন েতামােদরবাংলা িশক্ষক”(meaning, He is your Bengali teacher) and “িতিনেতামােদর বাংলা িশক্ষকেক িপিটেয়েছন” (meaning, He has beatenyour Bengali teacher). According to the lexical matching,there are four lemmatized terms (“িতিন”, “েতামােদর”, “বাংলা”,and “িশক্ষক”) exist in both sentences. That means thesimilarity is nearly 0.80(on a scale of 1.0). But thereis no semantic connection between these two texts. Ifwe consider another example sentence-pair, “আমার একিটেপাষা পৰ্ানী আেছ” (meaning, I have a pet) and “আিম িবড়ালপুিষ”(meaning, I own cat). There is no single</s>
<s>terms existin these two sentences but there is an obvious semanticsimilarity. The table I depicts the scenario. However, thesetwo examples conclude that the lexical measures are notenough to capture the similarity.TABLE I: The weekness of lexical matching in capturingsemantic similaritySentence 1 Sentence 2 Similarityিতিন েতামােদর বাংলা িশক্ষক িতিন েতামােদর বাংলািশক্ষকেক িপিটেয়েছনLexically similar butnot semanticallyআমার একিট েপাষা পৰ্ানীআেছআিম িবড়াল পুিষ Semantically similarbut not lexicallyEstimating similarity between Bengali texts is morechallenging as compared to the English texts. One of themain reasons is that the resources for the Bengali languageare not even comparable with the resources of Englishlanguage. To preprocess English texts there are well knowntokenizer, stemmer, lemmatizer, that are applied in almostevery NLP task and their performance is even arguablybetter. But on contrary, these kind of tools are not socommonly available to canonicalize Bengali texts. Further-more, there are also well-organized resources such as Word-Net, NLP POS tagger etc. that amplify the performance ofany similarity estimation method. Therefore the methodsfor Bengali texts lack such kind of tools and resources.In this paper, we try to overcome the challenges incapturing the semantic similarity between Bengali text-pair. We introduced a method for measuring the semanticsimilarity for Bengali texts. In this regard, an efficientBengali semantic textual similarity measuring algorithm isemployed based on a pre-trained word-embedding modelthat is trained on Wikipedia texts. We also prepare athe dataset that can be used to test the performance ofsemantic similarity measure in Bengali texts. We followthe same procedure as the SemEval STS 2017 task [6]. Theexperimental results clearly demonstrate that our methodis effective to measure the semantic similarity in Bengalitexts. The contributions of this research are:1) A semantic textual similarity estimation algo-rithm for Bengali texts, and978-1-5386-8207-4/18/$31.00 ©2018 IEEE2) A dataset that can be used for measuring theperformance of any Bengali semantic similaritymeasure.The rest of the paper is structured as follows: InSection II, we present the working principle of word-embedding. Section III summarizes the related work onsemantic textual similarity. In Section IV, we presentour proposed method to incorporate the challenges onsemantic similarity for Bengali texts. The experiments andevaluation results are presented to show the effectivenessof our proposed method in Section V. Some concludedremarks and future directions are described in SectionVI.II. Word-EmbeddingThe word-embedding (word2vec) can predict a word ina certain context for a set of given words. The frameworkfor learning word vectors is shown in the Fig. 1. Here thecontext of three words (“the”, “cat”, and “sat”) is used topredict the word (“on”). The input words are mapped tocolumns of the matrix W to predict the output word [7].W W Wthe cat satClassifierAverage/ConcatenationWord MatrixFig. 1: A framework for learning word vectors [7].Every word has a unique vector representation wherethe vector is represented by a column in a matrix W . Thecolumn is indexed by the position of the word in the vocab-ulary. To predict the next word in a sentence, the concate-nation or sum of the vectors is employed as features [7].Given a sequence of training words w1, w2, w3, ...WT , themain objective of word2vec model is to maximize theaverage log probabilityT−k∑t=klog</s>
<s>p(wt\|wt−k, ..., wt+k)A multiclass classifier is used for the prediction task,such as softmax [7].p(wt\|wt−k, ..., wt+k) =eywt∑i ewhere each of yi is un-normalized log-probability for eachoutput i, computed asy = b+ Uh(wt−k, ..., wt+k;W )where U , b are the softmax parameters. h is constructedby a concatenation or average of word vectors extractedfrom W .The word vector based on neural networks are usu-ally trained using stochastic gradient descent wherethe gradient obtained by backpropagation. An algo-rithm for training word vectors is publicly available atcode.google.com/p/word2vec [8]. In a trained word vectormodel, words with similar meaning are mapped to a similarposition in a vector space.A pre-trained word-embedding model1 trained in Ben-gali Wikipedia texts [22] is used in our method to estimatethe similarity. The dimension of the feature vector perword is 300. The other parameters to train the model areexplain in [22].III. Related WorkThe similarity measures to compute the semantic tex-tual similarity between multilingual and monolingual textshave been proposed in recent past [4]–[6], [9]–[11]. Butthe similarity measures for Bengali texts are not so com-monly available. Researchers proposed different methodsand techniques to capture the semantic similarity betweentext segments using different resources [4]–[6], [9]–[11].But these methods are mostly for English textual simi-larity. SemEval Semantic Textual Similarity (STS) taskswere organized for monolingual and multilingual texts [4]–[6]. The participating methods employed a large numberof features using a wide variety of resources [11], [12].Additionally, they applied some handcrafted rules thatdeal with currency values, negation, compounds, numberoverlap and literal matching [4]–[6]. Different multiple re-sources such as WordNet, Wikipedia, a dependency parser,NER tools, lemmatizer, POS tagger, stop word list etcwere leveraged to extract features. Based on the contentinformation of the text segments and external resourcesmultiple syntactic, semantic, and structural features werealso used to capture the similarity [13].However, the similarity measures for Bengali texts arehardly available. Sinha et al. [14] proposed a new lexiconand similarity measure for measuring the similarity be-tween Bengali texts. A distinct lexical organization usingthe semantic association between Bengali words than canbe accessed efficiently by different applications. Rudrapalet al. [15] proposed a method for measuring the semanticsimilarity of Bengali tweets using WordNet. Mihalcea etal. [9] suggested a method for measuring the semanticsimilarity of texts by exploiting the information that canbe drawn from the similarity of the component words.Specifically, the researchers used numerous corpus-basedand knowledge-based measures of word semantic similarityand used them to derive a text-to-text similarity metric.1https://github.com/facebookresearch/fastText/blob/master/pretrained-vectors.mdSome structured semantic knowledge like Wikipediaand WordNet are also employed to estimate the similarity.In some prior works [16]–[18], the methods are very similarto one another, using pairings of words and WordNet-based measures for semantic similarity. Researchers alsoused corpus-based methods combining with WordNet-based measures [9], [19]. In [9], they introduced an IDF-weighted alignment approach, based on WordNet-basedand corpus-based similarities. They applied the similaritiesbetween words which are identical in terms of their part-of-speech (POS) tag. Then a single score is calculated usingthe average over the maximum similarities. In [19], thesimilarity score has been measured by combining the wordorder score and a WordNet similarity measure. Recently,researchers tried word-embedding based techniques forsemantic similarity [20], [21].IV. Semantic Textual</s>
<s>Similarity MeasuringAlgorithm for Bengali TextsThis section presents our introduced algorithm formeasuring the semantic similarity between two sentencesS1 and S2. The Algorithm 1 presents the pseudo-code.The table II summarizes the basic notation used in ourproposed algorithm.TABLE II: Basic notation used in Algorithm 1Symbol DescriptionS_terms[ ] List of words after processing sentenceAVS 300 dimensional average feature vec-tor for each sentence Stc_S Total number of words contain in vo-cabulary of w2v_model for Sw2v_model Trained word-embedding modelvocab(w2v_model) Vocabulary of w2v_modeladd(t, AVS, w2v_model) Adding the 300 dimensional vector forterm t with AVSdivide(AVS, tc_S) Dividing each value ofAVS with tc_Ssim The semantic textual similarity scorebetween S1 and S2For a given pair of sentences S1 and S2, we first removedifferent types of punctuation marks, Bengali digits, etc.The preprocessed sentences are split into the list of words.The list is denoted by S_terms[ ]. Two list of wordsS1_terms[ ], and S2_terms[ ](in step 1 & 2 in Algo-rithm 1) are used to compute the similarity between twocorresponding sentences S1 and S2, respectively. Then wecompute the average feature vector AVS for each sentence.Word-embedding model returns a 300-dimensional vectorfor each term. Therefore we retrieve the feature vectorfor each term t belongs to S_terms[ ]. The vectors arecomputed only for words those belong to the vocabularyof the Word2Vec model, vocab(w2v_model). The featurevectors for each word belongs to a particular sentence arethen added.This addition is done in scope (from step 7 to step12) and scope (from step 14 to step 19) for S1 and S2,respectively. The vectors after the addition are stored inAVS1 and AVS2 for the two sentences. Each values inAVS1 and AVS2 are then divided by total number ofwords tc_S1 and tc_S2, respectively for correspondingsentences S1 and S2. The division is done in step 13 andAlgorithm 1: Semantic textual similarity estima-tion algorithm for Bengali texts: BSTS(S1, S2,w2v_model)Input: Sentence1 S1, Sentence2 S2, and Word2Vecmodel, w2v_modelOutput: Similarity score, sim(S1, S2) between S1and S21 S1_terms[ ]← Preprocess(S1)2 S2_terms[ ]← Preprocess(S2)3 AVS1← [0, · · · , 0]4 AVS2← [0, · · · , 0]5 tc_S1← 06 tc_S2← 07 for for each term, t ∈ S1_terms do8 if t in vocab(w2v_model) then9 AVS1← add(t,AVS1, w2v_model)10 tc_S1++11 end12 end13 AVS1← divide(AVS1, tc_S1)14 for for each term, t ∈ S2_terms do15 if t in vocab(w2v_model) then16 AVS2← add(t,AVS2, w2v_model)17 tc_S2++18 end19 end20 AVS2← divide(AVS2, tc_S2)21 sim(S1, S2)← AVS1·AVS2\|\|AVS1\|\|·\|\|AVS2\|\|20, respectively. The average feature vectors AVS1 andAVS2 are then used to calculate the similarity. We ap-plied the cosine similarity to compute the similarity scorebetween S1 and S2. The cosine similarity is computed instep 21. The equation of cosine similarity can be elaboratedas follows:sim(S1, S2) =AVS1 ·AVS2\|\|AVS1\|\| · \|\|AVS2\|\|∑300i=0 AV S1i ·AV S2i√∑300i=0 AV S12i√∑300i=0 AV S22iwhere AV S1i and AV S2i denote the i-th feature value ofvector AVS1 and AVS2, respectively.V. Experiments and EvaluationA. Dataset PreparationWe prepared the dataset following the same approachis done by SemEval Semantic Textual Similarity task,STS2017 [6]. In the STS2017 task, the organizers provided250 pairs of sentences. We translated those sentencesmanually by the human assessor. The STS2017 organizersprovided the similarity score per sentence-pair that arecalculated</s>
<s>by human assessors’ judgments. We employedtheir provided gold-standard judgment as a ground truthin this research. Their human assessors have given thesimilarity score using the following similarity label rangesfrom [0, 5].• Label 0: On different topics• Label 1: Not similar but share few common details• Label 2: Not similar but share some commondetails• Label 3: Roughly similar• Label 4: Similar• Label 5: Completely similarThe distribution of the similarity labels after the annota-tion is mentioned elsewhere in [6]. The human assessorsare instructed to assign the labels as followings [6]:1) Assign labels as precisely as possible accordingto the underlying meaning of the two sentencesrather than their superficial similarities or differ-ences.2) Be careful of wording differences that have animpact on what is being said or described.3) Ignore the grammatical errors and awkward word-ing as long as they do not obscure what is beingconveyed.4) Avoid over labeling pairs with middle rangescores.5) Be careful of over-reliance on an extreme scorelike 0 or 5.We applied a pre-trained model which trained on Ben-gali Wikipedia texts [22]. The dimension of the featurevector per words in the model is 300. Continuous-bag-of-words, cbow algorithm is used to train the model. Theother parameters are described elsewhere in [22]. Themodel was trained on whole Bengali Wikipedia texts [22].B. Evaluation MetricThe performance of our method has been tested basedon Pearson Correlation Coefficient2. This evaluation met-ric has also been used as an official metric to test theperformance of a method in SemEval STS2017 [6].Let X = {x1, x2, x3...xn} and Y = {y1, y2, y3...yn} bethe two sets of scores for n pairs of sentences generated bythe system and human assessors’ judgment, respectively.Each element xi or yi in set X and Y , respectivelyrepresents the semantic textual similarity between i-thsentence-pair. The Pearson Coefficient Correlation r isdefined as follows:r =i=1(xi − x̄)(yi − ȳ)√∑ni=1(xi − x̄)2√∑ni=1(yi − ȳ)2where n is the number of sentence pairs and xi, yi are thesimilarity scores given by participant and human assessors,respectively indexed with i. The arithmetic mean of theelements of X is defined by x̄ = 1i=1 xi and analogouslyfor ȳ.2https://en.wikipedia.org/wiki/Pearson_correlation_coefficientC. Experimental ResultsWe conducted experiments with different experimentalsettings. We first employed the edit distance (using theterms as the lexical unit) based lexical similarity betweensentences as the baseline. We then applied our proposedalgorithm (Algorithm 1), BSTS to compute the semanticsimilarity for Bengali texts. The experimental results aresummarized in Fig. 2.0.10.20.30.40.50.6Baseline BSTSAlgorithmFig. 2: The performance of our porpoised algorithm andthe baseline method in terms of Pearson’s CorrelationCoefficient r.The Fig. 2 illustrated that our method can capture afar better semantic and textual similarity as compared tothe traditional lexical similarity. The figure also indicatesthat the vector representation of each word in the word-embedding model is useful and effective to measure thesemantic similarity between texts.TABLE III: Walk-trough exampleSentence 1 Sentence 2 Lexical BSTSিতিন েতামােদর বাংলা িশক্ষক িতিন েতামােদর বাংলািশক্ষকেক িপিটেয়েছন0.8000 0.3377আমার একিট েপাষা পৰ্ানীআেছআিম িবড়াল পুিষ 0.0000 0.6138Table III illustrates the performance of our proposedalgorithm and the lexical similarity between the two ex-ample sentences-pairs (given in table I). As we noted thatthe traditional lexical matching</s>
<s>has given 80% similarityfor first sentence-pair (“িতিন েতামােদর বাংলা িশক্ষক” and “িতিনেতামােদর বাংলা িশক্ষকেক িপিটেয়েছন”) but, the performance ofour method concludes that they are 33% similar. Based onthe semantic meaning of these two sentences, they are notsimilar but share few common information. Our proposedalgorithm has given less similarity as compared to the lexi-cal similarity. On contrary, for the second example (“আমারএকিট েপাষা পৰ্ানী আেছ” and “আিম িবড়াল পুিষ”), the lexicalView publication statsView publication statshttps://www.researchgate.net/publication/329394770measure conclude that they are not even similar. But thesemantic meaning reflects that they are obviously similarwith the same meaning. Our proposed algorithm achieved61% similarity between them. Therefore, we can concludethat the proposed algorithm is effective to compute thesemantic similarity using the word-level semantics.VI. Conclusion and Future DirectionsThis paper introduced a method using word-embeddingto compute the semantic similarity between Bengali texts.The method utilized a pre-trained word-embedding modelto compute the word-level semantics. We overcome thechallenges to capture the semantic similarity with high di-mension vector representation of words. We also prepareda dataset that can be used to measure the performanceof any semantic similarity measure for Bengali texts. Theexperimental results based on Pearson’s correlation co-efficient r demonstrated the effectiveness of our methodin measuring semantic textual similarity. In future, wehave a plan to apply Long short-term memory (LSTM) tointroduce a new similarity measure for semantic similarity.References[1] R. M. Aliguliyev, “A new sentence similarity measure andsentence based extractive technique for automatic text sum-marization,” Expert Systems with Applications, vol. 36, no. 4,pp. 7764–7772, 2009.[2] D. Metzler, S. Dumais, and C. Meek, “Similarity measures forshort segments of text,” in European conference on informationretrieval. Springer, 2007, pp. 16–27.[3] H. Li, J. Xu et al., “Semantic matching in search,” Foundationsand Trends® in Information Retrieval, vol. 7, no. 5, pp. 343–469, 2014.[4] E. Agirre, C. Banea, C. Cardie, D. Cer, M. Diab, A. Gonzalez-Agirre, W. Guo, I. Lopez-Gazpio, M. Maritxalar, R. Mihal-cea et al., “Semeval-2015 task 2: Semantic textual similarity,english, spanish and pilot on interpretability,” in Proceedingsof the 9th international workshop on semantic evaluation (Se-mEval 2015), 2015, pp. 252–263.[5] E. Agirre, C. Banea, D. Cer, M. Diab, A. Gonzalez-Agirre,R. Mihalcea, G. Rigau, and J. Wiebe, “Semeval-2016 task1: Semantic textual similarity, monolingual and cross-lingualevaluation,” in Proceedings of the 10th International Workshopon Semantic Evaluation (SemEval-2016), 2016, pp. 497–511.[6] D. Cer, M. Diab, E. Agirre, I. Lopez-Gazpio, and L. Spe-cia, “Semeval-2017 task 1: Semantic textual similarity-multilingual and cross-lingual focused evaluation,” arXivpreprint arXiv:1708.00055, 2017.[7] Q. V. Le and T. Mikolov, “Distributed representations of sen-tences and documents.” in ICML, vol. 14, 2014, pp. 1188–1196.[8] T. Mikolov, K. Chen, G. Corrado, and J. Dean, “Efficientestimation of word representations in vector space,” arXivpreprint arXiv:1301.3781, 2013.[9] R. Mihalcea, C. Corley, C. Strapparava et al., “Corpus-basedand knowledge-based measures of text semantic similarity,” inAAAI, vol. 6, 2006, pp. 775–780.[10] Z. Zhang and V. Saligrama, “Zero-shot learning via semanticsimilarity embedding,” in Proceedings of the IEEE internationalconference on computer vision, 2015, pp. 4166–4174.[11] F. Šarić, G. Glavaš, M. Karan, J. Šnajder, and B. D. Bašić,“Takelab: Systems for measuring semantic text similarity,”in Proceedings of the First Joint Conference on Lexical</s>
<s>andComputational Semantics-Volume 1: Proceedings of the mainconference and the shared task, and Volume 2: Proceedingsof the Sixth International Workshop on Semantic Evaluation.Association for Computational Linguistics, 2012, pp. 441–448.[12] D. Bär, C. Biemann, I. Gurevych, and T. Zesch, “Ukp: Comput-ing semantic textual similarity by combining multiple contentsimilarity measures,” in Proceedings of the First Joint Confer-ence on Lexical and Computational Semantics-Volume 1: Pro-ceedings of the main conference and the shared task, and Volume2: Proceedings of the Sixth International Workshop on SemanticEvaluation. Association for Computational Linguistics, 2012,pp. 435–440.[13] H. Hassanzadeh, T. Groza, A. Nguyen, and J. Hunter, “Uqere-search: semantic textual similarity quantification,” in Proceed-ings of the 9th International Workshop on Semantic Evaluation(SemEval 2015), 2015, pp. 123–127.[14] M. Sinha, A. Jana, T. Dasgupta, and A. Basu, “A new semanticlexicon and similarity measure in bangla,” in Proceedings of the3rd Workshop on Cognitive Aspects of the Lexicon (CogALex-III), 2012, pp. 171–182.[15] D. Rudrapal, A. Das, and B. Bhattacharya, “Measuring seman-tic similarity for bengali tweets using wordnet,” in Proceedingsof the International Conference Recent Advances in NaturalLanguage Processing, 2015, pp. 537–544.[16] M. C. Lintean and V. Rus, “Measuring semantic similarity inshort texts through greedy pairing and word semantics.” inFLAIRS Conference, 2012.[17] R. Ferreira, R. D. Lins, F. Freitas, S. J. Simske, and M. Riss, “Anew sentence similarity assessment measure based on a three-layer sentence representation,” in Proceedings of the 2014 ACMsymposium on Document engineering. ACM, 2014, pp. 25–34.[18] S. Fernando and M. Stevenson, “A semantic similarity approachto paraphrase detection,” in Proceedings of the 11th AnnualResearch Colloquium of the UK Special Interest Group forComputational Linguistics, 2008, pp. 45–52.[19] Y. Li, D. McLean, Z. A. Bandar, K. Crockett et al., “Sentencesimilarity based on semantic nets and corpus statistics,” IEEETransactions on Knowledge & Data Engineering, no. 8, pp.1138–1150, 2006.[20] L. Han, A. L. Kashyap, T. Finin, J. Mayfield, and J. Weese,“Umbc_ebiquity-core: semantic textual similarity systems,” inSecond Joint Conference on Lexical and Computational Seman-tics (* SEM), Volume 1: Proceedings of the Main Conferenceand the Shared Task: Semantic Textual Similarity, vol. 1, 2013,pp. 44–52.[21] T. Kenter and M. De Rijke, “Short text similarity with wordembeddings,” in Proceedings of the 24th ACM international onconference on information and knowledge management. ACM,2015, pp. 1411–1420.[22] P. Bojanowski, E. Grave, A. Joulin, and T. Mikolov, “Enrich-ing word vectors with subword information,” arXiv preprintarXiv:1607.04606, 2016.</s>
<s>untitledSee discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/269297101New approach of solving semantic ambiguity problem of Bangla Root wordsusing universal networking language (UNL)Conference Paper · May 2014DOI: 10.1109/ICIEV.2014.6850778CITATIONSREADS8803 authors:Some of the authors of this publication are also working on these related projects:Identification of Expectancy, Proximity, and Compatibility of the Bengali Language View projectData Extraction from Natural Text View projectM. Firoz Mridha Ph. D.Bangladesh University of Business and Technology (BUBT)60 PUBLICATIONS 108 CITATIONS SEE PROFILEDr. Aloke Kumar SahaUniversity of Asia Pacific27 PUBLICATIONS 54 CITATIONS SEE PROFILEJ. K. DasJahangirnagar University26 PUBLICATIONS 61 CITATIONS SEE PROFILEAll content following this page was uploaded by M. Firoz Mridha Ph. D. on 26 September 2017.The user has requested enhancement of the downloaded file.https://www.researchgate.net/publication/269297101_New_approach_of_solving_semantic_ambiguity_problem_of_Bangla_Root_words_using_universal_networking_language_UNL?enrichId=rgreq-7f88d2fdcdb99060469f32ff05f01d75-XXX&enrichSource=Y292ZXJQYWdlOzI2OTI5NzEwMTtBUzo1NDI3MDU0MjYzMzc3OTJAMTUwNjQwMjQ2NDgyMw%3D%3D&el=1_x_2&_esc=publicationCoverPdfhttps://www.researchgate.net/publication/269297101_New_approach_of_solving_semantic_ambiguity_problem_of_Bangla_Root_words_using_universal_networking_language_UNL?enrichId=rgreq-7f88d2fdcdb99060469f32ff05f01d75-XXX&enrichSource=Y292ZXJQYWdlOzI2OTI5NzEwMTtBUzo1NDI3MDU0MjYzMzc3OTJAMTUwNjQwMjQ2NDgyMw%3D%3D&el=1_x_3&_esc=publicationCoverPdfhttps://www.researchgate.net/project/Identification-of-Expectancy-Proximity-and-Compatibility-of-the-Bengali-Language?enrichId=rgreq-7f88d2fdcdb99060469f32ff05f01d75-XXX&enrichSource=Y292ZXJQYWdlOzI2OTI5NzEwMTtBUzo1NDI3MDU0MjYzMzc3OTJAMTUwNjQwMjQ2NDgyMw%3D%3D&el=1_x_9&_esc=publicationCoverPdfhttps://www.researchgate.net/project/Data-Extraction-from-Natural-Text?enrichId=rgreq-7f88d2fdcdb99060469f32ff05f01d75-XXX&enrichSource=Y292ZXJQYWdlOzI2OTI5NzEwMTtBUzo1NDI3MDU0MjYzMzc3OTJAMTUwNjQwMjQ2NDgyMw%3D%3D&el=1_x_9&_esc=publicationCoverPdfhttps://www.researchgate.net/?enrichId=rgreq-7f88d2fdcdb99060469f32ff05f01d75-XXX&enrichSource=Y292ZXJQYWdlOzI2OTI5NzEwMTtBUzo1NDI3MDU0MjYzMzc3OTJAMTUwNjQwMjQ2NDgyMw%3D%3D&el=1_x_1&_esc=publicationCoverPdfhttps://www.researchgate.net/profile/M_Ph_D?enrichId=rgreq-7f88d2fdcdb99060469f32ff05f01d75-XXX&enrichSource=Y292ZXJQYWdlOzI2OTI5NzEwMTtBUzo1NDI3MDU0MjYzMzc3OTJAMTUwNjQwMjQ2NDgyMw%3D%3D&el=1_x_4&_esc=publicationCoverPdfhttps://www.researchgate.net/profile/M_Ph_D?enrichId=rgreq-7f88d2fdcdb99060469f32ff05f01d75-XXX&enrichSource=Y292ZXJQYWdlOzI2OTI5NzEwMTtBUzo1NDI3MDU0MjYzMzc3OTJAMTUwNjQwMjQ2NDgyMw%3D%3D&el=1_x_5&_esc=publicationCoverPdfhttps://www.researchgate.net/profile/M_Ph_D?enrichId=rgreq-7f88d2fdcdb99060469f32ff05f01d75-XXX&enrichSource=Y292ZXJQYWdlOzI2OTI5NzEwMTtBUzo1NDI3MDU0MjYzMzc3OTJAMTUwNjQwMjQ2NDgyMw%3D%3D&el=1_x_7&_esc=publicationCoverPdfhttps://www.researchgate.net/profile/Dr_Aloke_Saha?enrichId=rgreq-7f88d2fdcdb99060469f32ff05f01d75-XXX&enrichSource=Y292ZXJQYWdlOzI2OTI5NzEwMTtBUzo1NDI3MDU0MjYzMzc3OTJAMTUwNjQwMjQ2NDgyMw%3D%3D&el=1_x_4&_esc=publicationCoverPdfhttps://www.researchgate.net/profile/Dr_Aloke_Saha?enrichId=rgreq-7f88d2fdcdb99060469f32ff05f01d75-XXX&enrichSource=Y292ZXJQYWdlOzI2OTI5NzEwMTtBUzo1NDI3MDU0MjYzMzc3OTJAMTUwNjQwMjQ2NDgyMw%3D%3D&el=1_x_5&_esc=publicationCoverPdfhttps://www.researchgate.net/institution/University_of_Asia_Pacific?enrichId=rgreq-7f88d2fdcdb99060469f32ff05f01d75-XXX&enrichSource=Y292ZXJQYWdlOzI2OTI5NzEwMTtBUzo1NDI3MDU0MjYzMzc3OTJAMTUwNjQwMjQ2NDgyMw%3D%3D&el=1_x_6&_esc=publicationCoverPdfhttps://www.researchgate.net/profile/Dr_Aloke_Saha?enrichId=rgreq-7f88d2fdcdb99060469f32ff05f01d75-XXX&enrichSource=Y292ZXJQYWdlOzI2OTI5NzEwMTtBUzo1NDI3MDU0MjYzMzc3OTJAMTUwNjQwMjQ2NDgyMw%3D%3D&el=1_x_7&_esc=publicationCoverPdfhttps://www.researchgate.net/profile/J_Das3?enrichId=rgreq-7f88d2fdcdb99060469f32ff05f01d75-XXX&enrichSource=Y292ZXJQYWdlOzI2OTI5NzEwMTtBUzo1NDI3MDU0MjYzMzc3OTJAMTUwNjQwMjQ2NDgyMw%3D%3D&el=1_x_4&_esc=publicationCoverPdfhttps://www.researchgate.net/profile/J_Das3?enrichId=rgreq-7f88d2fdcdb99060469f32ff05f01d75-XXX&enrichSource=Y292ZXJQYWdlOzI2OTI5NzEwMTtBUzo1NDI3MDU0MjYzMzc3OTJAMTUwNjQwMjQ2NDgyMw%3D%3D&el=1_x_5&_esc=publicationCoverPdfhttps://www.researchgate.net/institution/Jahangirnagar_University?enrichId=rgreq-7f88d2fdcdb99060469f32ff05f01d75-XXX&enrichSource=Y292ZXJQYWdlOzI2OTI5NzEwMTtBUzo1NDI3MDU0MjYzMzc3OTJAMTUwNjQwMjQ2NDgyMw%3D%3D&el=1_x_6&_esc=publicationCoverPdfhttps://www.researchgate.net/profile/J_Das3?enrichId=rgreq-7f88d2fdcdb99060469f32ff05f01d75-XXX&enrichSource=Y292ZXJQYWdlOzI2OTI5NzEwMTtBUzo1NDI3MDU0MjYzMzc3OTJAMTUwNjQwMjQ2NDgyMw%3D%3D&el=1_x_7&_esc=publicationCoverPdfhttps://www.researchgate.net/profile/M_Ph_D?enrichId=rgreq-7f88d2fdcdb99060469f32ff05f01d75-XXX&enrichSource=Y292ZXJQYWdlOzI2OTI5NzEwMTtBUzo1NDI3MDU0MjYzMzc3OTJAMTUwNjQwMjQ2NDgyMw%3D%3D&el=1_x_10&_esc=publicationCoverPdfNew Approach of Solving Semantic Ambiguity Problem of Bangla Root Words Using Universal Networking Language (UNL) M. F Mridha Department of CSE, University of Asia Pacific, Dhaka, Bangladesh E-mail: mdfirozm@yahoo.com Aloke Kumar Saha Department of CSE, University of Asia Pacific, Dhaka, Bangladesh E-mail: aloke71@yahoo.com Jugal Krishna Das Department of CSE, Jahangirnagar University, Savar, Dhaka, Bangladesh E-mail:drdas64@yahoo.com Abstract—Ambiguity problems of Bangla ROOT words are the difficult problems in the field of natural language processing of Bangla sentence. In this work we looked at it through the knowledge based perspective and language phenomena in terms of rules and dictionary features. This work is more focused on solving the problems of Bangla sentences with semantic problem, thereby improving the accuracy of analysis process. The problem needs, particularly, a knowledge intensive solution. We have used insights from linguistics, towards solving this problem. Also, the usefulness of automatic extraction of features for words in the dictionary becomes evident through the work. Keywords - Syntax and Semantic analysis, Bangla Root Word, Morphology and UNL. I. INTRODUCTION Syntax as part of grammar is a description of how words grouped and connected to each other in a sentence. There is a good definition of syntax for programming languages: “… syntax usually entails the transformation of a linear sequence of tokens (a token is key to an individual word or punctuation mark in a natural language) into a hierarchical syntax tree”. Later we will see that the same definition also can be used for NL. Main problems on this level are: part of speech tagging (POS tagging), chunking or detecting syntactic categories (verb, noun phrases) and sentence assembling (constructing syntax tree). Semantics and its understanding as a study of meaning covers most complex tasks like: finding synonyms, word sense disambiguation, constructing question-answering systems, translating from one NL to another, populating base of knowledge. Basically one needs to complete morphological and syntactical analysis before trying to solve any semantic problem. In this paper we present the Root Word analysis of Bangla Sentences for UNL system. The major components of our research works touches upon i) Different types of Ambiguity that is caused by Bangla Root Word ii) UNL Expression of the Bangla Root word and iii) Bangla sentences analysis. In section 2 we describe the Bangla Sentence structure. In sections 3 and</s>
<s>4, we present our main works that include all the above three components. II. PROBLEM DEFINITION A. Structural Ambiguity A word, phrase, sentence or other communication is called ambiguous if it can be interpreted in more than one way. If the ambiguity is because of a multiple meanings of a word, it is called lexical ambiguity. One type of ambiguity, called structural ambiguity, arises due to more than one possible structure for the sentence. In Bangla: “সময় তীেরর মত চেল” In English: Time flies like an arrow. This sentence has two possible interpretations. In first case, flies named ‘time’ loves an arrow, whereas the other interpretation is ‘time’ passes like an arrow (as fast as an arrow). In this case, both the meanings of a sentence are semantically valid and acceptable. Such sentences are said to be inherently ambiguous. Even human being needs a context to select the appropriate meaning. But there is other kind of ambiguity which concerns NLP. Certain sentences are interpreted only in one way by human being but multiple parses of such sentences are possible for a machine. Moreover, a machine cannot select the meaningful interpretation out of the given possible parses because of the lack of world knowledge. B. Attachment ambiguity This is a specific type of structural ambiguity in which a clause or a phrase has more than one possible association in the tree structure of the sentence of which it is a part. If the ambiguity is about the attachment of a clause then it is called clause attachment and if it is about attachment of prepositional phrase it is called prepositional phrase 3rd INTERNATIONAL CONFERENCE ON INFORMATICS, ELECTRONICS & VISION 2014978-1-4799-5180-2/14/$31.00 ©2014 IEEEattachment. Depending on the site of attachment there are at least two possibilities, noun attachment or verb attachment. For example, In Bangla: েস বiিট েটিবেলর uপের রাখল। In English:�He�kept�the�book�on�the�table.�The sentence can be explained in two ways. First the prepositional phrase ‘on the table’ is attached to the noun ‘the book’ making a noun phrase ‘the book on the table’ which is semantically incorrect. Because, if we replace the NP ‘the book on the table’ in the sentence with just ‘the book’ the sentence becomes incomplete. The other suggests that the prepositional phrase ‘on the table’ is attached to the main verb ‘kept’ of the sentence. That is, ‘the book’ is not directly associated with ‘the table’. However, ‘the book’ is related to ‘the table’ through the verb ‘kept’. Thus ‘on the table’ is attached to the verb ‘kept’ Table 1:Bangla Ambiguous Sentence of Root Word “পড়” Root word Meaning when used sentence UNL Relation Example fall agt েস পেড় েগল। He has fallen down. read Agt,obj বাcািট বi পেড় The kid is reading book. go Agt,to,puবাcািট পড়েত িশkেকর কােছ েগেছ। The kid has gone to the teacher for tuition. raining rain বিৃ পড়েছ। It is raining. finish Agt,pos,gol েছেলিটপড়ােশানােশষকেরেছ। The boy has finished his study. study Aoj,tim েমেয়িটপড়ােশানায়aেনকভাল। The girl is good at study. forgive Aoj,mod,pos ছাtিট ভুেলর জনয্ িশkক</s>
<s>eর পােয় পড়ল। Student teacher's legs were mistakes. rests Agt,man মা ঘিুমেয় পেড়েছ। Mother has fallen asleep. broken Obj,met ঘিড়িট পেড় েভে েগেছ। The clock has broken by falling down teach agt িশkক পড়ােcন। The teacher is teaching. Table 2: Bangla Ambiguous Sentence of Root Word “ধর ” Root word Meaning when used in sentence UNL Relation Example catch Agt,obj েছেলিট মাছ ধরেছ। The boy is catching fish. caught obj েচারিট ধরা পেড়েছ। The thief has been caught. hold Agt,pos,obj বাcািট মােক ধের েরেখেছ। The kid is holding his mother. walk Aoj,obj,oেমেয়িটর চলার ধরণ ভাল না। Walking style of the girl is not good. hate Agt,pos,obj েসতারহাতধরলনা। He did not hold her hand. burn Agt,obj বািড়িটেত আগনু The house is on fire. took Aoj,pos,mod,obj েস তার িনেজর পথ ধরল। He took his own way. support Agt,man,pos,obj আমারহাতধেরiেসeতদরূআসেছ। He reached so far taking my support. turn Obj,gol,oেস মসুিলম ধমর্ gহন করল। He turned into Muslim. fitting Aoj,plc ঝুিড়েত আম ধরেছ না। Mangoes are not fitting in basket. C. Why Semantic analysis of Root word need? Verb is the main part of any sentence for any native language. Any Sentence can complete without subject or object. But without verb no sentence is complete. So verb analysis is need to converting from Bangla to UNL. And verb is the combination of root word and suffixes. And root word is titled as entry node when converting any native language to UNL. And not only verb but also other word is derived from a root word that may have the different transformations. This happens because different morphemes are added with it as suffixes. Therefore, the meaning of the word varies for its different transformations. We developed the following rules to this problem. III. AMBIGUITY OF BANGLA WORD IN SENTENCES It is necessary to make universal words in the context of bangle sentence and their usage. Converting to the English sentence and then make the universal word for Bangla language will not be semantically correct. In that case enconversion and deconversion will not be correct also. We have some dictionary entries for special Bangla words which are called ÒGK K_vq cÖKvkÓ\| This means to express a group of words in a single word shortly. w`‡b †h GKevi Avnvi K‡i- GKvnvix\| ‡Q‡jwU w`‡b GKevi Avnvi K‡i\| ‡Q‡jwU GKvnvix\| Both the sentences are same in meaning. So, the Enconversion as well as Deconversion system should know this. 3rd INTERNATIONAL CONFERENCE ON INFORMATICS, ELECTRONICS & VISION 2014978-1-4799-5180-2/14/$31.00 ©2014 IEEE G Mv‡Qi মলূ Mfx‡i hvq\| Here, root word [hv] ‘go’and the immediate previous word [Mfx‡i]. It needs to make a relation / dependency checking among these two words as meaning of root depends on its previous word. So, there should have a technique of matching which retrieves appropriate word from the dictionary. G ধােন †ekxw`b hv‡e bv\| Here, root word [hv] and the immediate previous word [‡ekxw`b]. েpন AvR hv‡e bv\| Here, root word [hv] and the immediate previous word [AvR]. ‡m বািড় hvq\| Here, root word [hv] and the</s>
<s>immediate previous word [বািড়]. bZzb ােট K‡e hv‡eb? Here, root word [hv] and the immediate previous word [K‡e]. It needs to know semantically the use of root word in each sentence same root carry different meaning. For this it needs to analyze the words in the sentences before and after the main root/verb. The parser need to know which particular dictionary entry has to retrieve to make the universal word. If the meaning of the root word is: [hv]- ‘go’; - then the dictionary entry will be [hv]{}”go(icl>do)”(ROOT, BANJANT) when it is of [hv]- spread than [hv]{}”spread(icl>occur, equ>distribute, obj>thing, scn>thing)” (ROOT, BANJANT) The dictionary entry should be of all possible of dimensions for each root words from which the rules will select required combination according to the meaning of sentence. Consider the following examples. Arif won’t invite Maria. Arif won’t come Maria. Here, the verbs are ‘invite’ and ‘come’ respectively. In the first sentence the verb ‘invite’ results in a grammatical sentence when followed by an noun phrase (NP) whereas in the second sentence, it seems that the verb ‘come’ does not allow an noun phrase to follow it. Hence, it can be said that a verb like ‘invite’ takes an noun phrase complement, whereas a verb like ‘come’ does not. There does not seem to be any general way from which one can predict whether a given verb does or does not take a following noun phrase. This is not dependent on the meaning as well. Rules for solving ambiguity of Bangla Root word[13][14][17]. Rule sets 1: (a) Some Suffixes (িবভিk) are used immediately after the root (ধাতু) for sadhu (সাধ)ু & cholito (চিলত) both languages. (b) Some Suffixes (িবভিk) are changed according to Person (পুরষু) for sadhu (সাধ)ু & cholito (চিলত) both languages. (c) For sadhu (সাধ)ু the suffixes are [�(শূনয্ িবভিk),iya(iয়া), ite(iেত) ] . (d) For cholito ( চিলত ) the suffixes are [ �(শূনয্ িবভিk)and e (e) ] . (e) For Person ( পুরষু ) the suffixes are [ lam(লাম) , le(েল), l(ল), ch(ছ), che(েছ), chi(িছ), bo(ব), be(েব)] . Rule sets 2: sadhu (সাধ ু (and cholito (চিলত (language for different Tense (কাল) (a) If the suffixes (িবভিk) for sadhu (সাধ)ু language is [�(শূনয্ িবভিk) ] then the corresponding suffix (িবভিk) for cholito (চিলত) is [�(শূনয্ িবভিk) ]. (b) If the suffixes (িবভিk) for sadhu (সাধ)ু language is ite (iেত) then the corresponding suffix (িবভিk) for cholito (চিলত) is [�(শূনয্ িবভিk) ]. (c) If the suffix (িবভিk) for sadhu (সাধ)ু language is iya (iয়া) then the corresponding suffix for cholito (চিলত) is e [(e)]. Rule set 3: Person )পুরষু ((1st , 2nd and 3rd ) (singular and plural) (a) If the suffixes (িবভিk) for 1st person is i(i) then the corresponding suffix (িবভিk)for 2nd person and 3rd person is [�(শূনয্ িবভিk) ] or O(o) and e [ ( e ) ] or y(য়) respectively. (b) If the suffixes (িবভিk) for 1st person is chi(িছ) then the corresponding suffix(িবভিk)for 2nd person and 3rd person is ch(ছ) and</s>
<s>che(েছ) respectively. (c) If the suffixes (িবভিk) for 1st person is lam(লাম) or chilam (িছলাম) then the corresponding suffix (িবভিk) for 2nd person and 3rd person is l(ল) or chil(িছল) and le(েল) or chile(িছেল) respectively. (d) If the suffixes (িবভিk) for 1st person is bo(ব) then the corresponding suffix (িবভিk) for 2nd person and 3rd person is be(েব) and be(েব) respectively. (e) ffixes (িবভিk) for 1st person is ai(আi) then the corresponding suffix (িবভিk)for 2nd person and 3rd person is ao(আo) and ay(আয়) respectively. A. Dictionary Entry[12][15][7] Shadhu Suffix [“”]{}“” (PROT,KBIVOKTI,SHADHU,INDIFINIT)<B,0,0> [i]{} “ ” (PROT,KBIVOKTI,SHADHU,INDIFINIT)<B,0,0> [iেত]{}“” (PROT,KBIVOKTI,SHADHU,CONTINUOUS)<B,0,0> 3rd INTERNATIONAL CONFERENCE ON INFORMATICS, ELECTRONICS & VISION 2014978-1-4799-5180-2/14/$31.00 ©2014 IEEE[iয়া]{}“ ” (PROT,KBIVOKTI,SHADHU,PERFECT)<B,0,0> Cholito Suffix [“”]{}“ ” (PROT,KBIVOKTI,CHOLITO,INDIFINIT)<B,0,0> [“ ”]{}“ ” (PROT,KBIVOKTI,CHOLITO,CONTINUOUS)<B,0,0> [e]{}“ ” (PROT,KBIVOKTI,CHOLITO,PERFECT)<B,0,0> Person Suffix [i]{}“ ” (PROT,KBIVOKTI ,1P)<B,0,0> [0]{}“ ” (PROT,KBIVOKTI, 2P)<B,0,0> [o]{}“ ” (PROT,KBIVOKTI, 2P)<B,0,0> [e]{}“ ” (PROT,KBIVOKTI, 3P)<B,0,0> [য়]{}“ ” (PROT,KBIVOKTI, 3P)<B,0,0> [িছ]{}“ ”(PROT,KBIVOKTI,1P)<B,0,0> [ছ]{}“ ”(PROT,KBIVOKTI,2P)<B,0,0> [েছ]{}“ ”(PROT,KBIVOKTI,3P)<B,0,0> [লাম]{}“ ”(PROT,KBIVOKTI,1P)<B,0,0> [লা]{}“ ”(PROT,KBIVOKTI,2P)<B,0,0> [েল]{}“ ”(PROT,KBIVOKTI,3P)<B,0,0> [ব]{}“ ”(PROT,KBIVOKTI,1P)<B,0,0> [েব]{}“ ”(PROT,KBIVOKTI,2P)<B,0,0> [েব]{}“ ”(PROT,KBIVOKTI,3P)<B,0,0> IV. OUR PROPOSED METHOD There have two reasons to get better performance than other methods. The first reason is the effect of the approach in case of Bangla language. For example, the existing way to find the UNL expression uses three dimensions to find the enconverted or deconverted output which are i) Dictionary Entry Look-up, ii) Rules of morphological analysis and iii) Semantic Analysis. Since component nodes are created by using these steps, they may be less accurate sue which may not expresses semantically correct output as there are different language constraints. As a result, the converted expression may be grammatically correct one but not be meaningfully correct. The second reason is the determination of the number of component nodes path for constructing desired output. Although the nature of input is meaningfully different as seen in Table-1 and Table 2, existing approach uses the grammatical attributes to select component nodes for all problems. However, Attribute Analysis Approach uses different options to get actual path in the component networks for different input sentence based on the nature of input. Flow Chart of our proposed program Architecture is shown below: Figure 1 Proposed program Architecture Here, Input sentence: The Bengali sentence which will be converted to UNL expression. This sentence is given as string. Example: “আিম আম খাi” Tokenizer: In here the input sentence “String” is dividing into tokens. Example: “আিম” “আম” “খাi” Token Token Token Validator: It check that is the is the tokens are arranged in right order or check is there any grammatical mistakes in the given sentence or “String” or “Tokens”. Example: “আিম তুিম খাi” - Invalid “আিম আম খাi” – Valid En-Converter: It convert the given sentence or “String” or “Tokens” in UNL expression. Example: “আিম আম খাi” 1. In put Bangla sentence: “আিম আম খাi” [S:00] {org:en} I eat mango {/org} {unl} agt(eat(icl>consume>do,agt>living_thing,obj>concrete_thing,ins>thing).@entry.@present,i(icl>person)) obj(eat(icl>consume>do,agt>living_thing,obj>concrete_thing,ins>thing).@entry.@present,mango(icl>edible_fruit>thing)) {/unl} [/S] 3rd INTERNATIONAL CONFERENCE ON INFORMATICS, ELECTRONICS & VISION 2014978-1-4799-5180-2/14/$31.00 ©2014 IEEE2. In put Bangla sentence: “সময় তীেরর মত চেল” [S:00] {org:en} Time flies like an arrow. {/org} {unl}</s>
<s>obj(fly(icl>occur,equ>pass,obj>thing).@entry.@present,time(icl>abstract_thing,equ>occasion)) man(fly(icl>occur,equ>pass,obj>thing).@entry.@present,like(icl>how,obj>thing)) obj(like(icl>how,obj>thing),arrow(icl>mark>thing).@indef) {/unl} [/S] 3. In put Bangla sentence: েস বiিট েটিবেলর uপের রাখল” [S:00] {org:en} He kept the book on the table. {/org} {unl} aoj(keep(icl>be,equ>continue,obj>action,aoj>thing).@entry.@past,he(icl>person)) obj(keep(icl>be,equ>continue,obj>action,aoj>thing).@entry.@past,book(icl>publication>thing).@def) plc(book(icl>publication>thing).@def,on(icl>how,com>surface,obj>concrete_thing,plc<uw)) obj(on(icl>how,com>surface,obj>concrete_thing,plc<uw),table(icl>place>abstract_thing,com>restaurant).@de{/unl} [/S] Some Rules of insertion of Ambiguous Root words: 1. +{V:+.@future.@sg.@male:null}{[কিরেব]:null:null} 2. +{V:+.@future.@sg.@female:null}{[কিরেব]:null:null} 3. +{V:+.@past.@custom.@sg.@female:null}{[কের]:null:null} 4. +{V:+.@past.@custom.@pl.@female:null}{[কেরিছ]:null} 5. +{V:+.@past.@custom.@pl.@male:null}{[কেরিছ]:null:null} 6. >{N,INANI,ACT,TCL:null:agt}{V,^VOCURR,^COO:+AGTRES,+TCL:null} 7. >{N,INANI,ACT:null:agt}{N,KAR,^VOCURR,^COO:+AGTRES:null} 8. >{N,INANI,MACH:null:agt}{V,^VOCURR,^COO:+AGTRES:null} 9. >{N,INANI,MACH:null:agt}{N,KARFSG,^VOCURR,^COO:+AGTRES:null} 10. >{N,INANI,MACH:null:agt}{N,KARFPL,^VOCURR,^COO:+AGTRES:null} 11. >{N,INANI,ACT,TCL:null:agt}{V,^VOCURR,^COO:+AGTRES,+TCL:null} 12. >{N,INANI,ACT:null:agt}{N,KAR,^VOCURR,^COO:+AGTRES:null} 13. >{N,CASE,^CAG:null:agt}{V,^AGTRES,^VOCURR,^COO:+AGTRES:null} 14. >{PRON,BEN:null:ben}{V:+BENRES:null} 15. +{V:+.@present.@sg.@male:null}{[খাi]:null:null} 16. +{V:+.@present.@sg.@male:null}{[খাi]:null:null} 17. +{V:+.@custom.@present.@sg.@female:null}{খাi ]:null} 18. +{V:+.@future.@sg.@male:null}{[খােব]:null:null} 19. +{V,m,sg:+.@future.@sg.@male:null}{খাiব]:null:null} 20. +{V:+.@future:null}{[খাব]:null:null} V. CONCLUSION Semantic Analysis method improves correct method of enconversion of UNL expression of Bangla language. A new technique has been proposed in this paper. Here the new technique used a constructive approach to determine the universal words of Bangla language. The novelty of this method is that, it used straightforward and simple technique to determine the ambiguity of Bangla word as well as the diversified usage of words in sentences for a given Bangla sentence. Semantic Analysis Approach first tried to solve the given problem by some example sentences, than it finds out required approaches to get semantically valid equivalence to get actual meaning of the sentence. Semantic Analysis Approach explores a new era in universal word construction, i.e., determining number of paths by analyzing the dictionary entries; which leads to creates good options to find appropriate meaning of the input sentence for proper enconversion and deconversion process. REFERENCES [1] H. Uchida, M. Zhu. The Universal Networking Language (UNL) Specification Version 3.0 Edition 3 ,Technical Report, UNU, (2005/6-UNDL Foundation, International Environment House, Tokyo, 2004) [2] H. Uchida, M. Zhu, “The Universal Networking Language (UNL) Specification Version 3.0”, Technical Report, United Nations University, Tokyo, 1998 [3] S. Abdul-Rahim, A.A. Libdeh, F. Sawalha, M. K. Odeh, “Universal Networking Language(UNL) a Means to Bridge the Digital Divide”, Computer Technology Training and Indistrial Studies Center, Royal Scientific Sciety, March 2002. [4] M. M. Asaduzzaman, M. M. Ali, “Morphological Analysis of Bengali Words for Automatic Machine Translation”, International Conference on Computer, and Information Technology (ICCIT), Dhaka, 2003, pp.271-276 [5] Bengali Academy (2004), Bengali-English Dictionary, Dhaka. [6] Enconverter Specifications, version 3.3, UNL Center/ UNDL Foundation, Tokyo, Japan 2002. [7] Enconverter Specification Version 3.3, (UNU Centre, Tokyo 150-8304, Japan 2002) [8] DeConverter Specification, Version 2.7, (UNL Center, UNDL Foundation, Tokyo 150-8304, Japan 2002) [9] D.M. Shahidullah. Bengali Baykaron, (Ahmed Mahmudul Haque of Mowla Brothers prokashani, Dhaka 2003) [10] Zakir Hossain, Shahid Al Noor, Muhammad Firoz Mridha Some Proposed Standard Models for Bengali Dictionary Entries of Bengali Morphemes for Universal Networking Language. IJCSNS International Journal of Computer Science and Network Security, V OL.12 No.11, November 2012 . [11] Bouguslavsky, I., Frid, N. and Iomdin, L. (2000). Creating a Universal Networking Module within an Advanced NLP system. Proceedings of the 18th International Conference on Computational Linguistics, pp. 83-89. 3rd INTERNATIONAL CONFERENCE ON INFORMATICS, ELECTRONICS & VISION 2014978-1-4799-5180-2/14/$31.00 ©2014 IEEE[12] Aloke Kumar Saha, Muhammad F. Mridha, Manoj Banik, and Jugal Krishna Das. Specification of UNL Deconverter for Bengali Language. International Journal of Scientific & Engineering Research, Volume 3, Issue</s>
<s>9, September-2012 ISSN 2229-5518. [13] Muhammad Firoz Mridha, Md. Zakir Hossain, Shahid Al Noor, “Development of Morphological Rules for Bangla Words for Universal Networking Language” IJCSNS International Journal of Computer Science and Network Security, VOL.10 No.10, October 2010. [14] Muhammad Firoz Mridha, Kamruddin Md. Nur, Manoj Banik and Mohammad Nurul Huda, “Structure of Dictionary Entries of Bangla Morphemes for Morphological Rule Generation for Universal Networking Language”. International Journal of Computer Information Systems and Industrial Management Applications (IJCISIM) 2011. [15] Muhammad Firoz Mridha, Kamruddin Md. Nur, Manoj Banik and Mohammad Nurul Huda, “Generation of Attributes for Bangla Words for Universal Networking Language(UNL)”. International Journal of Computer Information Systems and Industrial Management Applications (IJCISIM) 2011. [16]. Md. Sadequr Rahman, Sangita Rani Poddar, Muhammad Firoz Mridha, Mohammad Nurul Huda, “Open Morphological Machine Translation: Bangla to English”. NWESP 2010, page, 460-465 , ISBN: 978-1-4244-7817-0 [17] Muhammad Firoz Mridha, Md. Nawab Yousuf Ali, Manoj Banik3, Mohammad Nurul Huda,Chowdhury Mofizur Rahman, Jugal Krishna Das, "Conversion of Bangla Sentences to Universal Networking Languages, " SKIMA'10, Paro, Bhutan, August 2010. 3rd INTERNATIONAL CONFERENCE ON INFORMATICS, ELECTRONICS & VISION 2014978-1-4799-5180-2/14/$31.00 ©2014 IEEEView publication statsView publication statshttps://www.researchgate.net/publication/269297101 /ASCII85EncodePages false /AllowTransparency false /AutoPositionEPSFiles true /AutoRotatePages /None /Binding /Left /CalGrayProfile (Gray Gamma 2.2) /CalRGBProfile (sRGB IEC61966-2.1) /CalCMYKProfile (U.S. Web Coated \050SWOP\051 v2) /sRGBProfile (sRGB IEC61966-2.1) /CannotEmbedFontPolicy /Error /CompatibilityLevel 1.7 /CompressObjects /Off /CompressPages true /ConvertImagesToIndexed true /PassThroughJPEGImages true /CreateJobTicket false /DefaultRenderingIntent /Default /DetectBlends true /DetectCurves 0.0000 /ColorConversionStrategy /LeaveColorUnchanged /DoThumbnails false /EmbedAllFonts true /EmbedOpenType false /ParseICCProfilesInComments true /EmbedJobOptions true /DSCReportingLevel 0 /EmitDSCWarnings false /EndPage -1 /ImageMemory 1048576 /LockDistillerParams true /MaxSubsetPct 100 /Optimize true /OPM 0 /ParseDSCComments false /ParseDSCCommentsForDocInfo false /PreserveCopyPage true /PreserveDICMYKValues true /PreserveEPSInfo false /PreserveFlatness true /PreserveHalftoneInfo true /PreserveOPIComments false /PreserveOverprintSettings true /StartPage 1 /SubsetFonts true /TransferFunctionInfo /Remove /UCRandBGInfo /Preserve /UsePrologue false /ColorSettingsFile () /AlwaysEmbed [ true /AbadiMT-CondensedLight /ACaslon-Italic /ACaslon-Regular /ACaslon-Semibold /ACaslon-SemiboldItalic /AdobeArabic-Bold /AdobeArabic-BoldItalic /AdobeArabic-Italic /AdobeArabic-Regular /AdobeHebrew-Bold /AdobeHebrew-BoldItalic /AdobeHebrew-Italic /AdobeHebrew-Regular /AdobeHeitiStd-Regular /AdobeMingStd-Light /AdobeMyungjoStd-Medium /AdobePiStd /AdobeSansMM /AdobeSerifMM /AdobeSongStd-Light /AdobeThai-Bold /AdobeThai-BoldItalic /AdobeThai-Italic /AdobeThai-Regular /AGaramond-Bold /AGaramond-BoldItalic /AGaramond-Italic /AGaramond-Regular /AGaramond-Semibold /AGaramond-SemiboldItalic /AgencyFB-Bold /AgencyFB-Reg /AGOldFace-Outline /AharoniBold /Algerian /Americana /Americana-ExtraBold /AndaleMono /AndaleMonoIPA /AngsanaNew /AngsanaNew-Bold /AngsanaNew-BoldItalic /AngsanaNew-Italic /AngsanaUPC /AngsanaUPC-Bold /AngsanaUPC-BoldItalic /AngsanaUPC-Italic /Anna /ArialAlternative /ArialAlternativeSymbol /Arial-Black /Arial-BlackItalic /Arial-BoldItalicMT /Arial-BoldMT /Arial-ItalicMT /ArialMT /ArialMT-Black /ArialNarrow /ArialNarrow-Bold /ArialNarrow-BoldItalic /ArialNarrow-Italic /ArialRoundedMTBold /ArialUnicodeMS /ArrusBT-Bold /ArrusBT-BoldItalic /ArrusBT-Italic /ArrusBT-Roman /AvantGarde-Book /AvantGarde-BookOblique /AvantGarde-Demi /AvantGarde-DemiOblique /AvantGardeITCbyBT-Book /AvantGardeITCbyBT-BookOblique /BakerSignet /BankGothicBT-Medium /Barmeno-Bold /Barmeno-ExtraBold /Barmeno-Medium /Barmeno-Regular /Baskerville /BaskervilleBE-Italic /BaskervilleBE-Medium /BaskervilleBE-MediumItalic /BaskervilleBE-Regular /Baskerville-Bold /Baskerville-BoldItalic /Baskerville-Italic /BaskOldFace /Batang /BatangChe /Bauhaus93 /Bellevue /BellGothicStd-Black /BellGothicStd-Bold /BellGothicStd-Light /BellMT /BellMTBold /BellMTItalic /BerlingAntiqua-Bold /BerlingAntiqua-BoldItalic /BerlingAntiqua-Italic /BerlingAntiqua-Roman /BerlinSansFB-Bold /BerlinSansFBDemi-Bold /BerlinSansFB-Reg /BernardMT-Condensed /BernhardModernBT-Bold /BernhardModernBT-BoldItalic /BernhardModernBT-Italic /BernhardModernBT-Roman /BiffoMT /BinnerD /BinnerGothic /BlackadderITC-Regular /Blackoak /blex /blsy /Bodoni /Bodoni-Bold /Bodoni-BoldItalic /Bodoni-Italic /BodoniMT /BodoniMTBlack /BodoniMTBlack-Italic /BodoniMT-Bold /BodoniMT-BoldItalic /BodoniMTCondensed /BodoniMTCondensed-Bold /BodoniMTCondensed-BoldItalic /BodoniMTCondensed-Italic /BodoniMT-Italic /BodoniMTPosterCompressed /Bodoni-Poster /Bodoni-PosterCompressed /BookAntiqua /BookAntiqua-Bold /BookAntiqua-BoldItalic /BookAntiqua-Italic /Bookman-Demi /Bookman-DemiItalic /Bookman-Light /Bookman-LightItalic /BookmanOldStyle /BookmanOldStyle-Bold /BookmanOldStyle-BoldItalic /BookmanOldStyle-Italic /BookshelfSymbolOne-Regular /BookshelfSymbolSeven /BookshelfSymbolThree-Regular /BookshelfSymbolTwo-Regular /Botanical /Boton-Italic /Boton-Medium /Boton-MediumItalic /Boton-Regular /Boulevard /BradleyHandITC /Braggadocio /BritannicBold /Broadway /BrowalliaNew /BrowalliaNew-Bold /BrowalliaNew-BoldItalic /BrowalliaNew-Italic /BrowalliaUPC /BrowalliaUPC-Bold /BrowalliaUPC-BoldItalic /BrowalliaUPC-Italic /BrushScript /BrushScriptMT /CaflischScript-Bold /CaflischScript-Regular /Calibri /Calibri-Bold /Calibri-BoldItalic /Calibri-Italic /CalifornianFB-Bold /CalifornianFB-Italic /CalifornianFB-Reg /CalisMTBol /CalistoMT /CalistoMT-BoldItalic /CalistoMT-Italic /Cambria /Cambria-Bold /Cambria-BoldItalic /Cambria-Italic /CambriaMath /Candara /Candara-Bold /Candara-BoldItalic /Candara-Italic /Carta /CaslonOpenfaceBT-Regular /Castellar /CastellarMT /Centaur /Centaur-Italic /Century /CenturyGothic /CenturyGothic-Bold /CenturyGothic-BoldItalic /CenturyGothic-Italic /CenturySchL-Bold /CenturySchL-BoldItal /CenturySchL-Ital</s>
<s>/CenturySchL-Roma /CenturySchoolbook /CenturySchoolbook-Bold /CenturySchoolbook-BoldItalic /CenturySchoolbook-Italic /CGTimes-Bold /CGTimes-BoldItalic /CGTimes-Italic /CGTimes-Regular /CharterBT-Bold /CharterBT-BoldItalic /CharterBT-Italic /CharterBT-Roman /CheltenhamITCbyBT-Bold /CheltenhamITCbyBT-BoldItalic /CheltenhamITCbyBT-Book /CheltenhamITCbyBT-BookItalic /Chiller-Regular /Cmb10 /CMB10 /Cmbsy10 /CMBSY10 /CMBSY5 /CMBSY6 /CMBSY7 /CMBSY8 /CMBSY9 /Cmbx10 /CMBX10 /Cmbx12 /CMBX12 /Cmbx5 /CMBX5 /Cmbx6 /CMBX6 /Cmbx7 /CMBX7 /Cmbx8 /CMBX8 /Cmbx9 /CMBX9 /Cmbxsl10 /CMBXSL10 /Cmbxti10 /CMBXTI10 /Cmcsc10 /CMCSC10 /Cmcsc8 /CMCSC8 /Cmcsc9 /CMCSC9 /Cmdunh10 /CMDUNH10 /Cmex10 /CMEX10 /CMEX7 /CMEX8 /CMEX9 /Cmff10 /CMFF10 /Cmfi10 /CMFI10 /Cmfib8 /CMFIB8 /Cminch /CMINCH /Cmitt10 /CMITT10 /Cmmi10 /CMMI10 /Cmmi12 /CMMI12 /Cmmi5 /CMMI5 /Cmmi6 /CMMI6 /Cmmi7 /CMMI7 /Cmmi8 /CMMI8 /Cmmi9 /CMMI9 /Cmmib10 /CMMIB10 /CMMIB5 /CMMIB6 /CMMIB7 /CMMIB8 /CMMIB9 /Cmr10 /CMR10 /Cmr12 /CMR12 /Cmr17 /CMR17 /Cmr5 /CMR5 /Cmr6 /CMR6 /Cmr7 /CMR7 /Cmr8 /CMR8 /Cmr9 /CMR9 /Cmsl10 /CMSL10 /Cmsl12 /CMSL12 /Cmsl8 /CMSL8 /Cmsl9 /CMSL9 /Cmsltt10 /CMSLTT10 /Cmss10 /CMSS10 /Cmss12 /CMSS12 /Cmss17 /CMSS17 /Cmss8 /CMSS8 /Cmss9 /CMSS9 /Cmssbx10 /CMSSBX10 /Cmssdc10 /CMSSDC10 /Cmssi10 /CMSSI10 /Cmssi12 /CMSSI12 /Cmssi17 /CMSSI17 /Cmssi8 /CMSSI8 /Cmssi9 /CMSSI9 /Cmssq8 /CMSSQ8 /Cmssqi8 /CMSSQI8 /Cmsy10 /CMSY10 /Cmsy5 /CMSY5 /Cmsy6 /CMSY6 /Cmsy7 /CMSY7 /Cmsy8 /CMSY8 /Cmsy9 /CMSY9 /Cmtcsc10 /CMTCSC10 /Cmtex10 /CMTEX10 /Cmtex8 /CMTEX8 /Cmtex9 /CMTEX9 /Cmti10 /CMTI10 /Cmti12 /CMTI12 /Cmti7 /CMTI7 /Cmti8 /CMTI8 /Cmti9 /CMTI9 /Cmtt10 /CMTT10 /Cmtt12 /CMTT12 /Cmtt8 /CMTT8 /Cmtt9 /CMTT9 /Cmu10 /CMU10 /Cmvtt10 /CMVTT10 /ColonnaMT /Colossalis-Bold /ComicSansMS /ComicSansMS-Bold /Consolas /Consolas-Bold /Consolas-BoldItalic /Consolas-Italic /Constantia /Constantia-Bold /Constantia-BoldItalic /Constantia-Italic /CooperBlack /CopperplateGothic-Bold /CopperplateGothic-Light /Copperplate-ThirtyThreeBC /Corbel /Corbel-Bold /Corbel-BoldItalic /Corbel-Italic /CordiaNew /CordiaNew-Bold /CordiaNew-BoldItalic /CordiaNew-Italic /CordiaUPC /CordiaUPC-Bold /CordiaUPC-BoldItalic /CordiaUPC-Italic /Courier /Courier-Bold /Courier-BoldOblique /CourierNewPS-BoldItalicMT /CourierNewPS-BoldMT /CourierNewPS-ItalicMT /CourierNewPSMT /Courier-Oblique /CourierStd /CourierStd-Bold /CourierStd-BoldOblique /CourierStd-Oblique /CourierX-Bold /CourierX-BoldOblique /CourierX-Oblique /CourierX-Regular /CreepyRegular /CurlzMT /David-Bold /David-Reg /DavidTransparent /Dcb10 /Dcbx10 /Dcbxsl10 /Dcbxti10 /Dccsc10 /Dcitt10 /Dcr10 /Desdemona /DilleniaUPC /DilleniaUPCBold /DilleniaUPCBoldItalic /DilleniaUPCItalic /Dingbats /DomCasual /Dotum /DotumChe /EdwardianScriptITC /Elephant-Italic /Elephant-Regular /EngraversGothicBT-Regular /EngraversMT /EraserDust /ErasITC-Bold /ErasITC-Demi /ErasITC-Light /ErasITC-Medium /ErieBlackPSMT /ErieLightPSMT /EriePSMT /EstrangeloEdessa /Euclid /Euclid-Bold /Euclid-BoldItalic /EuclidExtra /EuclidExtra-Bold /EuclidFraktur /EuclidFraktur-Bold /Euclid-Italic /EuclidMathOne /EuclidMathOne-Bold /EuclidMathTwo /EuclidMathTwo-Bold /EuclidSymbol /EuclidSymbol-Bold /EuclidSymbol-BoldItalic /EuclidSymbol-Italic /EucrosiaUPC /EucrosiaUPCBold /EucrosiaUPCBoldItalic /EucrosiaUPCItalic /EUEX10 /EUEX7 /EUEX8 /EUEX9 /EUFB10 /EUFB5 /EUFB7 /EUFM10 /EUFM5 /EUFM7 /EURB10 /EURB5 /EURB7 /EURM10 /EURM5 /EURM7 /EuroMono-Bold /EuroMono-BoldItalic /EuroMono-Italic /EuroMono-Regular /EuroSans-Bold /EuroSans-BoldItalic /EuroSans-Italic /EuroSans-Regular /EuroSerif-Bold /EuroSerif-BoldItalic /EuroSerif-Italic /EuroSerif-Regular /EuroSig /EUSB10 /EUSB5 /EUSB7 /EUSM10 /EUSM5 /EUSM7 /FelixTitlingMT /Fences /FencesPlain /FigaroMT /FixedMiriamTransparent /FootlightMTLight /Formata-Italic /Formata-Medium /Formata-MediumItalic /Formata-Regular /ForteMT /FranklinGothic-Book /FranklinGothic-BookItalic /FranklinGothic-Demi /FranklinGothic-DemiCond /FranklinGothic-DemiItalic /FranklinGothic-Heavy /FranklinGothic-HeavyItalic /FranklinGothicITCbyBT-Book /FranklinGothicITCbyBT-BookItal /FranklinGothicITCbyBT-Demi /FranklinGothicITCbyBT-DemiItal /FranklinGothic-Medium /FranklinGothic-MediumCond /FranklinGothic-MediumItalic /FrankRuehl /FreesiaUPC /FreesiaUPCBold /FreesiaUPCBoldItalic /FreesiaUPCItalic /FreestyleScript-Regular /FrenchScriptMT /Frutiger-Black /Frutiger-BlackCn /Frutiger-BlackItalic /Frutiger-Bold /Frutiger-BoldCn /Frutiger-BoldItalic /Frutiger-Cn /Frutiger-ExtraBlackCn /Frutiger-Italic /Frutiger-Light /Frutiger-LightCn /Frutiger-LightItalic /Frutiger-Roman /Frutiger-UltraBlack /Futura-Bold /Futura-BoldOblique /Futura-Book /Futura-BookOblique /FuturaBT-Bold /FuturaBT-BoldItalic /FuturaBT-Book /FuturaBT-BookItalic /FuturaBT-Medium /FuturaBT-MediumItalic /Futura-Light /Futura-LightOblique /GalliardITCbyBT-Bold /GalliardITCbyBT-BoldItalic /GalliardITCbyBT-Italic /GalliardITCbyBT-Roman /Garamond /Garamond-Bold /Garamond-BoldCondensed /Garamond-BoldCondensedItalic /Garamond-BoldItalic /Garamond-BookCondensed /Garamond-BookCondensedItalic /Garamond-Italic /Garamond-LightCondensed /Garamond-LightCondensedItalic /Gautami /GeometricSlab703BT-Light /GeometricSlab703BT-LightItalic /Georgia /Georgia-Bold /Georgia-BoldItalic /Georgia-Italic /GeorgiaRef /Giddyup /Giddyup-Thangs /Gigi-Regular /GillSans /GillSans-Bold /GillSans-BoldItalic /GillSans-Condensed /GillSans-CondensedBold /GillSans-Italic /GillSans-Light /GillSans-LightItalic /GillSansMT /GillSansMT-Bold /GillSansMT-BoldItalic /GillSansMT-Condensed /GillSansMT-ExtraCondensedBold /GillSansMT-Italic /GillSans-UltraBold /GillSans-UltraBoldCondensed /GloucesterMT-ExtraCondensed /Gothic-Thirteen /GoudyOldStyleBT-Bold /GoudyOldStyleBT-BoldItalic /GoudyOldStyleBT-Italic /GoudyOldStyleBT-Roman /GoudyOldStyleT-Bold /GoudyOldStyleT-Italic /GoudyOldStyleT-Regular /GoudyStout /GoudyTextMT-LombardicCapitals /GSIDefaultSymbols /Gulim /GulimChe /Gungsuh /GungsuhChe /Haettenschweiler /HarlowSolid /Harrington /Helvetica /Helvetica-Black /Helvetica-BlackOblique /Helvetica-Bold /Helvetica-BoldOblique /Helvetica-Condensed /Helvetica-Condensed-Black /Helvetica-Condensed-BlackObl /Helvetica-Condensed-Bold /Helvetica-Condensed-BoldObl /Helvetica-Condensed-Light /Helvetica-Condensed-LightObl /Helvetica-Condensed-Oblique /Helvetica-Fraction /Helvetica-Narrow /Helvetica-Narrow-Bold /Helvetica-Narrow-BoldOblique /Helvetica-Narrow-Oblique /Helvetica-Oblique /HighTowerText-Italic /HighTowerText-Reg /Humanist521BT-BoldCondensed /Humanist521BT-Light /Humanist521BT-LightItalic /Humanist521BT-RomanCondensed /Imago-ExtraBold /Impact /ImprintMT-Shadow /InformalRoman-Regular /IrisUPC /IrisUPCBold /IrisUPCBoldItalic /IrisUPCItalic /Ironwood /ItcEras-Medium /ItcKabel-Bold /ItcKabel-Book /ItcKabel-Demi /ItcKabel-Medium /ItcKabel-Ultra /JasmineUPC /JasmineUPC-Bold /JasmineUPC-BoldItalic /JasmineUPC-Italic /JoannaMT /JoannaMT-Italic /Jokerman-Regular /JuiceITC-Regular /Kartika /Kaufmann /KaufmannBT-Bold /KaufmannBT-Regular /KidTYPEPaint /KinoMT /KodchiangUPC /KodchiangUPC-Bold /KodchiangUPC-BoldItalic /KodchiangUPC-Italic /KorinnaITCbyBT-Regular /KozGoProVI-Medium /KozMinProVI-Regular /KristenITC-Regular /KunstlerScript</s>
<s>/Latha /LatinWide /LetterGothic /LetterGothic-Bold /LetterGothic-BoldOblique /LetterGothic-BoldSlanted /LetterGothicMT /LetterGothicMT-Bold /LetterGothicMT-BoldOblique /LetterGothicMT-Oblique /LetterGothic-Slanted /LetterGothicStd /LetterGothicStd-Bold /LetterGothicStd-BoldSlanted /LetterGothicStd-Slanted /LevenimMT /LevenimMTBold /LilyUPC /LilyUPCBold /LilyUPCBoldItalic /LilyUPCItalic /Lithos-Black /Lithos-Regular /LotusWPBox-Roman /LotusWPIcon-Roman /LotusWPIntA-Roman /LotusWPIntB-Roman /LotusWPType-Roman /LucidaBright /LucidaBright-Demi /LucidaBright-DemiItalic /LucidaBright-Italic /LucidaCalligraphy-Italic /LucidaConsole /LucidaFax /LucidaFax-Demi /LucidaFax-DemiItalic /LucidaFax-Italic /LucidaHandwriting-Italic /LucidaSans /LucidaSans-Demi /LucidaSans-DemiItalic /LucidaSans-Italic /LucidaSans-Typewriter /LucidaSans-TypewriterBold /LucidaSans-TypewriterBoldOblique /LucidaSans-TypewriterOblique /LucidaSansUnicode /Lydian /Magneto-Bold /MaiandraGD-Regular /Mangal-Regular /Map-Symbols /MathA /MathB /MathC /Mathematica1 /Mathematica1-Bold /Mathematica1Mono /Mathematica1Mono-Bold /Mathematica2 /Mathematica2-Bold /Mathematica2Mono /Mathematica2Mono-Bold /Mathematica3 /Mathematica3-Bold /Mathematica3Mono /Mathematica3Mono-Bold /Mathematica4 /Mathematica4-Bold /Mathematica4Mono /Mathematica4Mono-Bold /Mathematica5 /Mathematica5-Bold /Mathematica5Mono /Mathematica5Mono-Bold /Mathematica6 /Mathematica6Bold /Mathematica6Mono /Mathematica6MonoBold /Mathematica7 /Mathematica7Bold /Mathematica7Mono /Mathematica7MonoBold /MatisseITC-Regular /MaturaMTScriptCapitals /Mesquite /Mezz-Black /Mezz-Regular /MICR /MicrosoftSansSerif /MingLiU /Minion-BoldCondensed /Minion-BoldCondensedItalic /Minion-Condensed /Minion-CondensedItalic /Minion-Ornaments /MinionPro-Bold /MinionPro-BoldIt /MinionPro-It /MinionPro-Regular /MinionPro-Semibold /MinionPro-SemiboldIt /Miriam /MiriamFixed /MiriamTransparent /Mistral /Modern-Regular /MonotypeCorsiva /MonotypeSorts /MSAM10 /MSAM5 /MSAM6 /MSAM7 /MSAM8 /MSAM9 /MSBM10 /MSBM5 /MSBM6 /MSBM7 /MSBM8 /MSBM9 /MS-Gothic /MSHei /MSLineDrawPSMT /MS-Mincho /MSOutlook /MS-PGothic /MS-PMincho /MSReference1 /MSReference2 /MSReferenceSansSerif /MSReferenceSansSerif-Bold /MSReferenceSansSerif-BoldItalic /MSReferenceSansSerif-Italic /MSReferenceSerif /MSReferenceSerif-Bold /MSReferenceSerif-BoldItalic /MSReferenceSerif-Italic /MSReferenceSpecialty /MSSong /MS-UIGothic /MT-Extra /MT-Symbol /MT-Symbol-Italic /MVBoli /Myriad-Bold /Myriad-BoldItalic /Myriad-Italic /MyriadPro-Black /MyriadPro-BlackIt /MyriadPro-Bold /MyriadPro-BoldIt /MyriadPro-It /MyriadPro-Light /MyriadPro-LightIt /MyriadPro-Regular /MyriadPro-Semibold /MyriadPro-SemiboldIt /Myriad-Roman /Narkisim /NewCenturySchlbk-Bold /NewCenturySchlbk-BoldItalic /NewCenturySchlbk-Italic /NewCenturySchlbk-Roman /NewMilleniumSchlbk-BoldItalicSH /NewsGothic /NewsGothic-Bold /NewsGothicBT-Bold /NewsGothicBT-BoldItalic /NewsGothicBT-Italic /NewsGothicBT-Roman /NewsGothic-Condensed /NewsGothic-Italic /NewsGothicMT /NewsGothicMT-Bold /NewsGothicMT-Italic /NiagaraEngraved-Reg /NiagaraSolid-Reg /NimbusMonL-Bold /NimbusMonL-BoldObli /NimbusMonL-Regu /NimbusMonL-ReguObli /NimbusRomDGR-Bold /NimbusRomDGR-BoldItal /NimbusRomDGR-Regu /NimbusRomDGR-ReguItal /NimbusRomNo9L-Medi /NimbusRomNo9L-MediItal /NimbusRomNo9L-Regu /NimbusRomNo9L-ReguItal /NimbusSanL-Bold /NimbusSanL-BoldCond /NimbusSanL-BoldCondItal /NimbusSanL-BoldItal /NimbusSanL-Regu /NimbusSanL-ReguCond /NimbusSanL-ReguCondItal /NimbusSanL-ReguItal /Nimrod /Nimrod-Bold /Nimrod-BoldItalic /Nimrod-Italic /NSimSun /Nueva-BoldExtended /Nueva-BoldExtendedItalic /Nueva-Italic /Nueva-Roman /NuptialScript /OCRA /OCRA-Alternate /OCRAExtended /OCRB /OCRB-Alternate /OfficinaSans-Bold /OfficinaSans-BoldItalic /OfficinaSans-Book /OfficinaSans-BookItalic /OfficinaSerif-Bold /OfficinaSerif-BoldItalic /OfficinaSerif-Book /OfficinaSerif-BookItalic /OldEnglishTextMT /Onyx /OnyxBT-Regular /OzHandicraftBT-Roman /PalaceScriptMT /Palatino-Bold /Palatino-BoldItalic /Palatino-Italic /PalatinoLinotype-Bold /PalatinoLinotype-BoldItalic /PalatinoLinotype-Italic /PalatinoLinotype-Roman /Palatino-Roman /PapyrusPlain /Papyrus-Regular /Parchment-Regular /Parisian /ParkAvenue /Penumbra-SemiboldFlare /Penumbra-SemiboldSans /Penumbra-SemiboldSerif /PepitaMT /Perpetua /Perpetua-Bold /Perpetua-BoldItalic /Perpetua-Italic /PerpetuaTitlingMT-Bold /PerpetuaTitlingMT-Light /PhotinaCasualBlack /Playbill /PMingLiU /Poetica-SuppOrnaments /PoorRichard-Regular /PopplLaudatio-Italic /PopplLaudatio-Medium /PopplLaudatio-MediumItalic /PopplLaudatio-Regular /PrestigeElite /Pristina-Regular /PTBarnumBT-Regular /Raavi /RageItalic /Ravie /RefSpecialty /Ribbon131BT-Bold /Rockwell /Rockwell-Bold /Rockwell-BoldItalic /Rockwell-Condensed /Rockwell-CondensedBold /Rockwell-ExtraBold /Rockwell-Italic /Rockwell-Light /Rockwell-LightItalic /Rod /RodTransparent /RunicMT-Condensed /Sanvito-Light /Sanvito-Roman /ScriptC /ScriptMTBold /SegoeUI /SegoeUI-Bold /SegoeUI-BoldItalic /SegoeUI-Italic /Serpentine-BoldOblique /ShelleyVolanteBT-Regular /ShowcardGothic-Reg /Shruti /SimHei /SimSun /SimSun-PUA /SnapITC-Regular /StandardSymL /Stencil /StoneSans /StoneSans-Bold /StoneSans-BoldItalic /StoneSans-Italic /StoneSans-Semibold /StoneSans-SemiboldItalic /Stop /Swiss721BT-BlackExtended /Sylfaen /Symbol /SymbolMT /Tahoma /Tahoma-Bold /Tci1 /Tci1Bold /Tci1BoldItalic /Tci1Italic /Tci2 /Tci2Bold /Tci2BoldItalic /Tci2Italic /Tci3 /Tci3Bold /Tci3BoldItalic /Tci3Italic /Tci4 /Tci4Bold /Tci4BoldItalic /Tci4Italic /TechnicalItalic /TechnicalPlain /Tekton /Tekton-Bold /TektonMM /Tempo-HeavyCondensed /Tempo-HeavyCondensedItalic /TempusSansITC /Times-Bold /Times-BoldItalic /Times-BoldItalicOsF /Times-BoldSC /Times-ExtraBold /Times-Italic /Times-ItalicOsF /TimesNewRomanMT-ExtraBold /TimesNewRomanPS-BoldItalicMT /TimesNewRomanPS-BoldMT /TimesNewRomanPS-ItalicMT /TimesNewRomanPSMT /Times-Roman /Times-RomanSC /Trajan-Bold /Trebuchet-BoldItalic /TrebuchetMS /TrebuchetMS-Bold /TrebuchetMS-Italic /Tunga-Regular /TwCenMT-Bold /TwCenMT-BoldItalic /TwCenMT-Condensed /TwCenMT-CondensedBold /TwCenMT-CondensedExtraBold /TwCenMT-CondensedMedium /TwCenMT-Italic /TwCenMT-Regular /Univers-Bold /Univers-BoldItalic /UniversCondensed-Bold /UniversCondensed-BoldItalic /UniversCondensed-Medium /UniversCondensed-MediumItalic /Univers-Medium /Univers-MediumItalic /URWBookmanL-DemiBold /URWBookmanL-DemiBoldItal /URWBookmanL-Ligh /URWBookmanL-LighItal /URWChanceryL-MediItal /URWGothicL-Book /URWGothicL-BookObli /URWGothicL-Demi /URWGothicL-DemiObli /URWPalladioL-Bold /URWPalladioL-BoldItal /URWPalladioL-Ital /URWPalladioL-Roma /USPSBarCode /VAGRounded-Black /VAGRounded-Bold /VAGRounded-Light /VAGRounded-Thin /Verdana /Verdana-Bold /Verdana-BoldItalic /Verdana-Italic /VerdanaRef /VinerHandITC /Viva-BoldExtraExtended /Vivaldii /Viva-LightCondensed /Viva-Regular /VladimirScript /Vrinda /Webdings /Westminster /Willow /Wingdings2 /Wingdings3 /Wingdings-Regular /WNCYB10 /WNCYI10 /WNCYR10 /WNCYSC10 /WNCYSS10 /WoodtypeOrnaments-One /WoodtypeOrnaments-Two /WP-ArabicScriptSihafa /WP-ArabicSihafa /WP-BoxDrawing /WP-CyrillicA /WP-CyrillicB /WP-GreekCentury /WP-GreekCourier /WP-GreekHelve /WP-HebrewDavid /WP-IconicSymbolsA /WP-IconicSymbolsB /WP-Japanese /WP-MathA /WP-MathB /WP-MathExtendedA /WP-MathExtendedB /WP-MultinationalAHelve /WP-MultinationalARoman /WP-MultinationalBCourier /WP-MultinationalBHelve /WP-MultinationalBRoman /WP-MultinationalCourier /WP-Phonetic /WPTypographicSymbols /XYATIP10 /XYBSQL10 /XYBTIP10 /XYCIRC10 /XYCMAT10 /XYCMBT10 /XYDASH10 /XYEUAT10 /XYEUBT10 /ZapfChancery-MediumItalic /ZapfDingbats /ZapfHumanist601BT-Bold /ZapfHumanist601BT-BoldItalic /ZapfHumanist601BT-Demi /ZapfHumanist601BT-DemiItalic /ZapfHumanist601BT-Italic /ZapfHumanist601BT-Roman /ZWAdobeF /NeverEmbed [ true /AntiAliasColorImages false /CropColorImages true /ColorImageMinResolution 200 /ColorImageMinResolutionPolicy /OK /DownsampleColorImages true /ColorImageDownsampleType /Bicubic /ColorImageResolution 300 /ColorImageDepth -1 /ColorImageMinDownsampleDepth 1 /ColorImageDownsampleThreshold 2.00333 /EncodeColorImages true /ColorImageFilter /DCTEncode /AutoFilterColorImages true /ColorImageAutoFilterStrategy /JPEG /ColorACSImageDict << /QFactor 0.76 /HSamples [2 1 1 2] /VSamples [2 1 1 2]</s>
<s>/ColorImageDict << /QFactor 1.30 /HSamples [2 1 1 2] /VSamples [2 1 1 2] /JPEG2000ColorACSImageDict << /TileWidth 256 /TileHeight 256 /Quality 10 /JPEG2000ColorImageDict << /TileWidth 256 /TileHeight 256 /Quality 10 /AntiAliasGrayImages false /CropGrayImages true /GrayImageMinResolution 200 /GrayImageMinResolutionPolicy /OK /DownsampleGrayImages true /GrayImageDownsampleType /Bicubic /GrayImageResolution 300 /GrayImageDepth -1 /GrayImageMinDownsampleDepth 2 /GrayImageDownsampleThreshold 2.00333 /EncodeGrayImages true /GrayImageFilter /DCTEncode /AutoFilterGrayImages true /GrayImageAutoFilterStrategy /JPEG /GrayACSImageDict << /QFactor 0.76 /HSamples [2 1 1 2] /VSamples [2 1 1 2] /GrayImageDict << /QFactor 1.30 /HSamples [2 1 1 2] /VSamples [2 1 1 2] /JPEG2000GrayACSImageDict << /TileWidth 256 /TileHeight 256 /Quality 10 /JPEG2000GrayImageDict << /TileWidth 256 /TileHeight 256 /Quality 10 /AntiAliasMonoImages false /CropMonoImages true /MonoImageMinResolution 400 /MonoImageMinResolutionPolicy /OK /DownsampleMonoImages true /MonoImageDownsampleType /Bicubic /MonoImageResolution 600 /MonoImageDepth -1 /MonoImageDownsampleThreshold 1.00167 /EncodeMonoImages true /MonoImageFilter /CCITTFaxEncode /MonoImageDict << /K -1 /AllowPSXObjects false /CheckCompliance [ /None /PDFX1aCheck false /PDFX3Check false /PDFXCompliantPDFOnly false /PDFXNoTrimBoxError true /PDFXTrimBoxToMediaBoxOffset [ 0.00000 0.00000 0.00000 0.00000 /PDFXSetBleedBoxToMediaBox true /PDFXBleedBoxToTrimBoxOffset [ 0.00000 0.00000 0.00000 0.00000 /PDFXOutputIntentProfile (None) /PDFXOutputConditionIdentifier () /PDFXOutputCondition () /PDFXRegistryName () /PDFXTrapped /False /CreateJDFFile false /Description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> /CHS <FEFF4f7f75288fd94e9b8bbe5b9a521b5efa7684002000410064006f006200650020005000440046002065876863900275284e8e5c4f5e55663e793a3001901a8fc775355b5090ae4ef653d190014ee553ca901a8fc756e072797f5153d15e03300260a853ef4ee54f7f75280020004100630072006f0062006100740020548c002000410064006f00620065002000520065006100640065007200200035002e003000204ee553ca66f49ad87248672c676562535f00521b5efa768400200050004400460020658768633002> /CHT <FEFF4f7f752890194e9b8a2d7f6e5efa7acb7684002000410064006f006200650020005000440046002065874ef69069752865bc87a25e55986f793a3001901a904e96fb5b5090f54ef650b390014ee553ca57287db2969b7db28def4e0a767c5e03300260a853ef4ee54f7f75280020004100630072006f0062006100740020548c002000410064006f00620065002000520065006100640065007200200035002e003000204ee553ca66f49ad87248672c4f86958b555f5df25efa7acb76840020005000440046002065874ef63002> /CZE <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> /DAN <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> /DEU <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> /ESP <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> /ETI <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> /FRA <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> /GRE <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> /HUN <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> /ITA <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> /JPN <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> /KOR <FEFFc7740020c124c815c7440020c0acc6a9d558c5ec0020d654ba740020d45cc2dc002c0020c804c7900020ba54c77c002c0020c778d130b137c5d00020ac00c7a50020c801d569d55c002000410064006f0062006500200050004400460020bb38c11cb97c0020c791c131d569b2c8b2e4002e0020c774b807ac8c0020c791c131b41c00200050004400460020bb38c11cb2940020004100630072006f0062006100740020bc0f002000410064006f00620065002000520065006100640065007200200035002e00300020c774c0c1c5d0c11c0020c5f40020c2180020c788c2b5b2c8b2e4002e> /LTH <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> /LVI <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> /NLD (Gebruik deze instellingen om Adobe PDF-documenten te maken die zijn geoptimaliseerd voor weergave op een beeldscherm, e-mail en internet. De gemaakte PDF-documenten kunnen worden geopend met Acrobat en Adobe Reader 5.0 en hoger.) /NOR <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> /POL <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> /PTB <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> /RUM <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> /RUS <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> /SKY <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> /SLV <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> /SUO <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> /SVE <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> /TUR <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> /UKR <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> /ENU (Use these settings to create Adobe PDF documents best suited for on-screen display, e-mail, and the Internet. Created PDF documents can be opened with Acrobat and Adobe Reader 5.0 and later.) /Namespace [ (Adobe) (Common) (1.0) /OtherNamespaces [ /AsReaderSpreads false /CropImagesToFrames true /ErrorControl /WarnAndContinue /FlattenerIgnoreSpreadOverrides false /IncludeGuidesGrids false /IncludeNonPrinting false /IncludeSlug false /Namespace [ (Adobe) (InDesign) (4.0) /OmitPlacedBitmaps false /OmitPlacedEPS false /OmitPlacedPDF false /SimulateOverprint /Legacy /AddBleedMarks false /AddColorBars false /AddCropMarks false /AddPageInfo false /AddRegMarks false /ConvertColors /ConvertToRGB /DestinationProfileName (sRGB IEC61966-2.1) /DestinationProfileSelector /UseName /Downsample16BitImages true /FlattenerPreset << /PresetSelector /MediumResolution /FormElements false /GenerateStructure false /IncludeBookmarks false /IncludeHyperlinks false /IncludeInteractive false /IncludeLayers false /IncludeProfiles true /MultimediaHandling /UseObjectSettings /Namespace [ (Adobe) (CreativeSuite) (2.0) /PDFXOutputIntentProfileSelector /NA /PreserveEditing false /UntaggedCMYKHandling /UseDocumentProfile /UntaggedRGBHandling /UseDocumentProfile /UseDocumentBleed false>> setdistillerparams /HWResolution [600 600] /PageSize [612.000 792.000]>> setpagedevice</s>
<s>See discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/328333982A Deep Recurrent Neural Network with BiLSTM model for SentimentClassificationPreprint · September 2018CITATIONSREADS2,1291 author:Some of the authors of this publication are also working on these related projects:Fake News Detection View projectAuthorship Attribution in Bengali Literature View projectMd Saiful IslamShahjalal University of Science and Technology70 PUBLICATIONS 237 CITATIONS SEE PROFILEAll content following this page was uploaded by Md Saiful Islam on 17 October 2018.The user has requested enhancement of the downloaded file.https://www.researchgate.net/publication/328333982_A_Deep_Recurrent_Neural_Network_with_BiLSTM_model_for_Sentiment_Classification?enrichId=rgreq-9226e3746c7d1b343dd0a885289f32c4-XXX&enrichSource=Y292ZXJQYWdlOzMyODMzMzk4MjtBUzo2ODI2MTk0NDAyMDU4MjlAMTUzOTc2MDU2NTc4MA%3D%3D&el=1_x_2&_esc=publicationCoverPdfhttps://www.researchgate.net/publication/328333982_A_Deep_Recurrent_Neural_Network_with_BiLSTM_model_for_Sentiment_Classification?enrichId=rgreq-9226e3746c7d1b343dd0a885289f32c4-XXX&enrichSource=Y292ZXJQYWdlOzMyODMzMzk4MjtBUzo2ODI2MTk0NDAyMDU4MjlAMTUzOTc2MDU2NTc4MA%3D%3D&el=1_x_3&_esc=publicationCoverPdfhttps://www.researchgate.net/project/Fake-News-Detection-20?enrichId=rgreq-9226e3746c7d1b343dd0a885289f32c4-XXX&enrichSource=Y292ZXJQYWdlOzMyODMzMzk4MjtBUzo2ODI2MTk0NDAyMDU4MjlAMTUzOTc2MDU2NTc4MA%3D%3D&el=1_x_9&_esc=publicationCoverPdfhttps://www.researchgate.net/project/Authorship-Attribution-in-Bengali-Literature?enrichId=rgreq-9226e3746c7d1b343dd0a885289f32c4-XXX&enrichSource=Y292ZXJQYWdlOzMyODMzMzk4MjtBUzo2ODI2MTk0NDAyMDU4MjlAMTUzOTc2MDU2NTc4MA%3D%3D&el=1_x_9&_esc=publicationCoverPdfhttps://www.researchgate.net/?enrichId=rgreq-9226e3746c7d1b343dd0a885289f32c4-XXX&enrichSource=Y292ZXJQYWdlOzMyODMzMzk4MjtBUzo2ODI2MTk0NDAyMDU4MjlAMTUzOTc2MDU2NTc4MA%3D%3D&el=1_x_1&_esc=publicationCoverPdfhttps://www.researchgate.net/profile/Md_Saiful_Islam14?enrichId=rgreq-9226e3746c7d1b343dd0a885289f32c4-XXX&enrichSource=Y292ZXJQYWdlOzMyODMzMzk4MjtBUzo2ODI2MTk0NDAyMDU4MjlAMTUzOTc2MDU2NTc4MA%3D%3D&el=1_x_4&_esc=publicationCoverPdfhttps://www.researchgate.net/profile/Md_Saiful_Islam14?enrichId=rgreq-9226e3746c7d1b343dd0a885289f32c4-XXX&enrichSource=Y292ZXJQYWdlOzMyODMzMzk4MjtBUzo2ODI2MTk0NDAyMDU4MjlAMTUzOTc2MDU2NTc4MA%3D%3D&el=1_x_5&_esc=publicationCoverPdfhttps://www.researchgate.net/institution/Shahjalal_University_of_Science_and_Technology?enrichId=rgreq-9226e3746c7d1b343dd0a885289f32c4-XXX&enrichSource=Y292ZXJQYWdlOzMyODMzMzk4MjtBUzo2ODI2MTk0NDAyMDU4MjlAMTUzOTc2MDU2NTc4MA%3D%3D&el=1_x_6&_esc=publicationCoverPdfhttps://www.researchgate.net/profile/Md_Saiful_Islam14?enrichId=rgreq-9226e3746c7d1b343dd0a885289f32c4-XXX&enrichSource=Y292ZXJQYWdlOzMyODMzMzk4MjtBUzo2ODI2MTk0NDAyMDU4MjlAMTUzOTc2MDU2NTc4MA%3D%3D&el=1_x_7&_esc=publicationCoverPdfhttps://www.researchgate.net/profile/Md_Saiful_Islam14?enrichId=rgreq-9226e3746c7d1b343dd0a885289f32c4-XXX&enrichSource=Y292ZXJQYWdlOzMyODMzMzk4MjtBUzo2ODI2MTk0NDAyMDU4MjlAMTUzOTc2MDU2NTc4MA%3D%3D&el=1_x_10&_esc=publicationCoverPdfInternational Conference on Bangla Speech and Language Processing(ICBSLP), 21-22 September, 2018A Deep Recurrent Neural Network with BiLSTMmodel for Sentiment ClassificationAbdullah Aziz SharfuddinDepartment of CSEShahjalal University of Science andTechnologySylhet, Bangladeshabdullahshakkhor@gmail.comMd. Nafis TihamiDepartment of CSEShahjalal University of Science andTechnologySylhet, Bangladeshnafistiham@student.sust.eduMd. Saiful IslamDepartment of CSEShahjalal University of Science andTechnologySylhet, Bangladeshsaiful-cse@sust.eduAbstract—In the field of sentiment classification, opinionsor sentiments of the people are analyzed. Sentiment analysissystems are being applied in social platforms and in almost everybusiness because the opinions or sentiments are the reflectionof the beliefs, choices and activities of the people. With thesesystems it is possible to make decisions for businesses to politicalagendas. In recent times a huge number of people share theiropinions across the Internet using Bengali. In this paper a newway of sentiment classification of Bengali text using RecurrentNeural Network(RNN) is presented. Using deep recurrent neuralnetwork with BiLSTM, the accuracy 85.67% is achieved.Index Terms—Bengali text; Deep learning; Sentiment Classifi-cation; RNN; LSTM; BiLSTM; Facebook; NLPI. INTRODUCTIONWith the elevation in the communication technology i.e.world wide web, a huge number of people from all lineagesacross the world take part in social networks and express theiremotions or opinions on a wide range of topics. Now it is adire need to summarize the data created by people over thesocial networks and see the insights from them. Besides, inthe field of NLP, it has become a topic of enormous interest.Because, it is needed to make smart recommending systems,anticipating the results of political elections, understanding thefeedback of people on public events and movements.SA is a method of finding and classifying opinions ex-pressed in a piece of text basing on computation technologies,especially in order to find out whether the writer’s behaviortowards a specific topic, product, etc. is positive, negative,or neutral. SA also refers to the administration of opinions,sentiments and subjective text [1]. It also gives the compre-hensive data associated to public views, as it goes throughall the different kinds of tweets, reviews and comments. Itis a verified mechanism for the prediction of a numerousmomentous circumstances, for instance movie ratings at boxoffice and public or regional elections [2]. Public opinions areused to value a certain matter, i.e. person, product or placeand might be found at different websites like Amazon andYelp. The sentiments can be specified into negative, positive orneutral and even more classes. SA can automatically find outthe expressive direction of user reviews or opinions whetherthe users have a good or positive impression or a negativeimpression. [3]The usage of SA is broad and powerful. Its demand hasgrown due to the escalating need of extracting and inspectinghidden information from data coming through</s>
<s>social medias.Different organizations around the world are using the abilityto extract hidden data now-a-days. The change in sentimentscan be connected to the change in stock market. The Obamaadministration used opinion mining in the 2012 presidentialelection to detect the public opinion before the announcementof policy. [4]Deep learning has shown great performance in SA. In thisarticle, a way of SA using deep recurrent neural network withBiLSTM is presented.II. RELATED WORKSSentiment analysis is not new for English language. Asignificant number of research works have been done withinthis scope. Arvind et al. [5] applied Skip-Gram model where asParamveer et al. [6] applied CCA(Canonical Correlation Anal-ysis) for in-depth vectorization. Duyu Tang et al. [7] did theirwork on sentiment analysis of tweets. These contained not onlythe information but also the syntactic context.Bengali hasn’tyet progressed in this particular research area. Approaches thathave been already made are dependent on machine learningmostly. Dipankar Das [8] used Parts of speech tagger forto tag emotion. He achieved 70% accuracy. K.M AzharulHasan et al. [9] used contextual valency analysis in Banglatext for SA. In their approach, using parts of speech tagger,they calculated the three classes (positivity, negativity andneutrality).The percentages was summed up to get result.They got an accuracy of 75%. K.M Azharul Hasan et. al[10] used unsupervised learning algorithm and created phrasepattern. Shaika Chowdhury [11] used SVM and MaximumEntropy on Bengali micro blog posts. They compared theseclassifiers using the accuracy metric. M Al-Amin et. al [12]used word2vec model for their base of sentiment analysis inBangla. They used the positions of vector representation ofthe Bengali words to create a positive or negative score foreach word and determine the sentiment of a specific text. Theyachieved a 72.5% accuracy. Sentiment analysis on the rohingyaissue was done by Hemayet et al. [13] Similar works were alsodone by Al-Amin et al. [14].978-1-5386-8207-4/18/$31.00 c©2018 IEEEIII. METHODOLOGYThis section describes the training and architecture of themodels that are proposed for the classification of sentimentsfrom Bengali texts.A. Recurrent Neural NetworksRecurrent neural networks (RNN) are a type of networkwhich form memory through recurrent connections. In feedforward networks, inputs are independent of each other. Butin RNN, all inputs are connected to each other. This letsthe network to exhibit vigorous temporal behavior for a timesequence which makes it favorable for sequential classificationlike sentiment analysis. As it can be seen in the figure, at first,Fig. 1. RNN loopit takes the x0 from the sequence of input and then it outputsh0 which together with x1 is the input for the next step. So,the h0 and x1 is the input for the next step. Similarly, h1 fromthe next is the input with x2 for the next step and so on. Thisway, it keeps remembering the context while training.ht = f(Whxxt +Whhht−1 + b) (1)yt = g(Wyhht + c) (2)B. Long Short Term Memory UnitsIn 1997, an alteration of RNN with Long Short-TermMemory units, or LSTM units [15], was proposed by theSepp Hochreiter and Juergen Schmidhuber.Some errors back-propagate through time in general RNN. These LSTM unitshelp to bypass these errors. While keeping a more consistenterror, they let RNNs</s>
<s>keep on learning over several time steps.LSTMs consist of information outside the basic flow of thernn in a valved block [16].Fig. 2. LSTM UnitNeural network’s notes get triggered by the notes they get.Likewise, LSTM’s gates pass on or block the data based onits weight. After that, these signals are grated with their ownsets of weights. Subsequently, RNN’s learning process modifythese weights that control hidden states and input. Ultimately,these cells learn when to let information to get in, get outor be deleted through the consequent steps of taking guesses,back propagating error, and modulating weights via gradientdescent. [17]ft = σg(Wfxt + Ufht−1 + bf ) (3)it = σg(Wixt + Uiht−1 + bi) (4)ot = σg(Woxt + Uoht−1 + bo) (5)ct = ftoct−1 + itoσc(Wcxt + Ucht−1 + bc) (6)ht = otoσh(ct) (7)C. Bidirectional LSTMBiLSTMs are proven especially helpful in the occasionswhere the context of the input is needed. It is very useful forworks like sentiment classification. In unidirectional LSTMinformation flows from the backward to forward. On thecontrary in Bi-directional LSTM information not only flowsbackward to forward but also forward to backward using twohidden states. Hence Bi-LSTMs understand the context better.BiLSTMs were used to escalate the chunk of input informationusable to the network. Structure of RNN with LSTM andRNN with BiLSTM. [18] Basically, BRNN follows such aprocess where the neurons of a normal RNN are broken intobidirectional ways. One is for backward states or negativetime direction, and another for forward states or positivetime direction. The inputs of the reverse direction states arenot connected to these two states’ results. The structure orBiLSTM is shown in the diagram below. By utilizing two timeFig. 3. Bidirectional LSTMdirections, input data from the past and future of the currenttime frame can be used. Whereas standard RNN requires thedelays for including future data. [18]IV. EXPERIMENT SETUP AND RESULTSA. Sentiment Analysis DatasetIt is very hard to find a benchmark dataset for BengaliSentiment Classification. Because researchers do not publishtheir datasets along with their work. But we have managed adataset from Md. Al-Amin et al. [12] Which contained about15000 comments fetched from Facebook. But the dataset wasnot as good as we expected. so needed to reproduce the datasetby ourselves. We have fetched comments from Facebook andlabeled them by hand before training our models. Total numberof fetched comments was around 30,000. Among them about30% were usable. There are in total of 10000 comments(5000positive comments and 5000 negative comments). The datasetis evenly distributed. All the emojis, symbols, numbers, stick-ers were deleted. All English letters were removed. The datasetonly contained plain Bengali text.TABLE IINFORMATION OF WORDSFig. 4. Length vs number of commentsIn the Fig:4, the distribution in the aspects of length of thecomments can be seen. Retrospectively, we have seen that, thecomments on various news article posts are rather short. In thegraph there is a high peak in the area of approximately 6. andthe peak steeply goes down to 20. And then a gradual fall.Nearly all of the words are shorter than length of 50 words.The rest are negligible. It should also be taken into</s>
<s>considera-tion that, some one word comments have been dropped whichdo not express a specific sentiment. Some other informationscan be found in table I. If we have a look at the number ofunique words and the number of most occurrences, we cananticipate the distribution of words against their occurrences.On average every word has occurred 5.73 times.B. Results and AnalysisOur models have been trained on both datasets. Despite ofbeing small, it performed better on the reproduced dataset. Italso outperformed the previous work [12] with the proposeddeep learning model.TABLE IIACCURACY COMPARISON OF MODELS ON DATASET [12]Model Accuracy(%) Train(%) Test(%)Model [12] 75.50 90 10Proposed Approach 82.53 67 33TABLE IIIACCURACY COMPARISON WITH TRADITIONAL CLASSIFYING MODELSModel Accuracy(%)Support Vector Machine 68.77%Decision Tree Classifier 67.50%Logistic Linear Regression 60.94%Proposed Approach 85.67%And for the dataset with 10000 comments using the samemetrics the model achieved 85.67% accuracy. The dataset wasalso trained with the traditional models like support vectormachine, decision tree classifier and logistic linear regressionand accuracy obtained were respectively 68.77%, 67.50% and60.94%.True Positive 44%True Negative 42%False Positive False Negative CONFUSION MATRIXFig. 5. Confusion MatrixTABLE IVRESULT OF TESTING ON 3300 DATAPredicted PredictedPositive NegativeActually Positive 44.15 8.13Actually Negative 41.52 6.20TABLE VTEST AGAINST CUSTOM SAMPLESThen both models were tested (trained with dataset [12] andtrained with small dataset) against some custom context basedsamples that were not in any of the datasets. The predictionsare shown in table 4. In the table the mis-predictions aremarked in block fonts. Out of six samples the model trainedwith 15000 comments got 4 wrong. And on the other handmodel trained with 10000 comments got all correct.Moreover, dataset [12] was not polished. It contained emo-jis, numbers, English letters punctuations and symbols. Inspite of being smaller the reproduced dataset with 10000comments is more fine-grained compared to the one with15000 comments. The model was able to understand thesentiment from the context of the sentences. Hence the modelperformed better.V. CONCLUSIONSA has become very important for the business owners.Because, with sentiment analyzers, it is now possible tounderstand the user activities and choices. Many works havebeen done for English. In contrast of that, work done inBengali is very little. This research is a little step forwardto fill the void.Despite being one of the most used languages in the world,Bengali lacks in both benchmark datasets and a well furnishedmodel for sentiment analysis. Moreover, researchers usuallydo not publish their dataset. The dataset that was made forthis research is clearly step ahead since it will be enrichedand published for research purposes. Moreover, the datasetwas not stemmed for our purpose. In future, we can stem thedataset and our result might improve.It is clear that deep learning models are the future here.Using the recurrent deep learning models, it is now possibleto achieve state of the art performance in Bengali sentimentclassification. In future it will be interesting to see the businessapplications or sentiment analyzers for Bengali text using deeprecurrent models. Also hybrid deep learning models can beapplied to the task of sentiment classification.VI. ACKNOWLEDGEMENTFor our work, we are greatly thankful to the SUST NLPgroup and also thankful to the previous researchers who haveput</s>
<s>their efforts in Bengali sentiment analysis. Also, we aregrateful to the researchers who have progressed the field ofNLP and Neural Networks. We acknowledge M. Al-Amin etal. for providing us with their dataset without which our workwould not be complete.REFERENCES[1] A. V. Yeole, P. Chavan, and M. Nikose, “Opinion mining for emotionsdetermination,” in Innovations in Information, Embedded and Commu-nication Systems (ICIIECS), 2015 International Conference on. IEEE,2015, pp. 1–5.[2] B. Heredia, T. M. Khoshgoftaar, J. Prusa, and M. Crawford, “Cross-domain sentiment analysis: An empirical investigation,” in InformationReuse and Integration (IRI), 2016 IEEE 17th International Conferenceon. IEEE, 2016, pp. 160–165.[3] F. Luo, C. Li, and Z. Cao, “Affective-feature-based sentiment analysisusing svm classifier,” in Computer Supported Cooperative Work inDesign (CSCWD), 2016 IEEE 20th International Conference on. IEEE,2016, pp. 276–281.[4] K. Bannister. (2018) Sentiment analysis uses @ONLINE. [Online].Available: https://www.brandwatch.com/blog/understanding-sentiment-analysis/[5] A. Neelakantan, J. Shankar, A. Passos, and A. McCallum, “Efficientnon-parametric estimation of multiple embeddings per word in vectorspace,” arXiv preprint arXiv:1504.06654, 2015.[6] P. Dhillon, D. P. Foster, and L. H. Ungar, “Multi-view learning of wordembeddings via cca,” in Advances in neural information processingsystems, 2011, pp. 199–207.[7] D. Tang, F. Wei, N. Yang, M. Zhou, T. Liu, and B. Qin, “Learningsentiment-specific word embedding for twitter sentiment classification,”in Proceedings of the 52nd Annual Meeting of the Association forComputational Linguistics (Volume 1: Long Papers), vol. 1, 2014, pp.1555–1565.[8] D. Das, “Analysis and tracking of emotions in english and bengali texts:a computational approach,” in Proceedings of the 20th internationalconference companion on World wide web. ACM, 2011, pp. 343–348.[9] K. A. Hasan, M. Rahman et al., “Sentiment detection from banglatext using contextual valency analysis,” in Computer and InformationTechnology (ICCIT), 2014 17th International Conference on. IEEE,2014, pp. 292–295.[10] K. A. Hasan, S. Islam, G. Mashrur-E-Elahi, and M. N. Izhar, “Sentimentrecognition from bangla text,” in Technical Challenges and DesignIssues in Bangla Language Processing. IGI Global, 2013, pp. 315–327.[11] S. Chowdhury and W. Chowdhury, “Performing sentiment analysis inbangla microblog posts,” in 2014 International Conference on Informat-ics, Electronics & Vision (ICIEV). IEEE, 2014, pp. 1–6.[12] M. Al-Amin, M. S. Islam, and S. D. Uzzal, “Sentiment analysis ofbengali comments with word2vec and sentiment information of words,”in Electrical, Computer and Communication Engineering (ECCE), In-ternational Conference on. IEEE, 2017, pp. 186–190.[13] H. A. Chowdhury, T. A. Nibir, M. Islam et al., “Sentiment analysis ofcomments on rohingya movement with support vector machine,” arXivpreprint arXiv:1803.08790, 2018.[14] M. S. Islam, M. Al-Amin, and S. D. Uzzal, “Word embedding withhellinger pca to detect the sentiment of bengali text,” in Computer andInformation Technology (ICCIT), 2016 19th International Conferenceon. IEEE, 2016, pp. 363–366.[15] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neuralcomputation, vol. 9, no. 8, pp. 1735–1780, 1997.[16] J. Wang, L.-C. Yu, K. R. Lai, and X. Zhang, “Dimensional sentimentanalysis using a regional cnn-lstm model,” in Proceedings of the54th Annual Meeting of the Association for Computational Linguistics(Volume 2: Short Papers), vol. 2, 2016, pp. 225–230.[17] F. A. Gers, J. Schmidhuber, and F. Cummins, “Learning to forget:Continual prediction with lstm,” 1999.[18] M. Schuster and K. K. Paliwal, “Bidirectional recurrent neural net-works,” IEEE Transactions on</s>
<s>Signal Processing, vol. 45, no. 11, pp.2673–2681, 1997.View publication statsView publication statshttps://www.researchgate.net/publication/328333982</s>
<s>Bengali VADER: A Sentiment Analysis Approach Using Modified VADER2019 International Conference on Electrical, Computer andCommunication Engineering (ECCE), 7-9 February, 2019Bengali VADER: A Sentiment Analysis ApproachUsing Modified VADERAl Amin, Imran Hossain, Aysha Akther* and Kazi Masudul AlamDGTED Lab, Computer Science and Engineering DisciplineKhulna UniversityKhulna-9208, Bangladesh{alamin1512, imran1503, aysha, kazi}@cseku.ac.bdAbstract—Sentiment analysis is an essential field of naturallanguage processing (NLP) that classifies the opinion expressed ina text according to its polarity (e.g., positive, negative or neutral).Bengali NLP research is lagging behind English NLP, wherethere are very few works on Bengali sentiment analysis. In thispaper, we approach this issue by modifying a popular Englishtool VADER to support Bengali sentiment polarity identification.We have compiled a Bengali polarity lexicon from the Englishpolarity lexicon of VADER. Furthermore, we have modified thefunctionalities of English VADER, so that it can directly classifyBengali text sentiments without the requirement of Bengalito English translation using tools such as Google Translator,MyMemory Translator, etc. Our experiments demonstrate thatthe modified Bengali VADER significantly improves the senti-ment analysis result of Bengali text over the current model.Index Terms—Sentiment analysis, Bengali VADER, BengaliLexicon, Bengali to English Translation.I. ISentiment is a term related to sensitiveness or emotionalfeelings. Sentiment is a genuine and refined sensibility, whichis affected by feeling as opposed to reason or reality [1].Whereas, analysis is the detailed examination that is requiredto understand the nature or to determine the essential featuresof something complex through analytical study [2]. Accord-ing to [3], sentiment analysis determines the tendency ofindividual’s opinions through information mining by usingnatural language processing (NLP), computational linguisticsand textual analysis on emotional data gathered from theInternet, online social media and similar sources. On the otherhand, Oxford dictionary describes sentiment analysis as theprocess that identifies and categorizes opinions expressed ina piece of text, in order to determine whether the user’s viewis positive, negative, or neutral towards a specific theme, itemor so forth [4].Business organizations, governments, brand managementcompanies, and data scientists are demonstrating increasinglyhigher interest about sentiment analysis and its various appli-cations. Data analytics can provide significant insight aboutthe users perception about any topic that can: improve busi-ness strategies, improve brand impression, improve campaign*Corresponding Authormanagement, modify marketing strategies, generate businessleads, test business KPIs, etc. Related opinion data is availablein the comments of news articles, comment/status sectionof social medias (i.e. Facebook, Twitter, Instagram), blogs,review of products, etc. We can easily conclude that businessorganizations can improve their bottom line using sentimentanalysis in different data segments. Similarly, it is importantfor the governments to understand the sentiments of thecitizens to address important public opinions as well as updatepolitical agendas to avoid unwanted circumstances. In all thesecases, sentiment analysis can be an important game changingtool.Multiple international languages such as English, Frenchare exploring the sentiment analysis field for a while. ThoughBengali is one of the most used languages in the world, itis well behind in the race. As Bengali is used everyday bymore than 250 million people of the world, primary languagein Bangladesh and secondary language in India [5] [6] [7], ithas huge potential for the business as well as governments.In 2015, the British Council published a report</s>
<s>on the mostimportant languages for the future. They considered severaldifferent factors, one of them is languages spoken in thefastest-growing emerging economies by 2050. Out of theseemerging economies, Bengali is expected to be the third mostcommonly spoken language, lower than Chinese and Hindi[8].As a result, researchers should focus more on Bengali textanalysis. Since there are very few works on sentiment analysisof Bengali text, in this paper we have put some light in thatdirection. Mobile adoption as well as Internet growth is veryhigh in Bangladesh, which is contributing rich sentimental datain the online world. From this perspective, we have modifieda popular English sentiment classification tool VADER [9]and upgraded it to support Bengali sentiment analysis withoutusing intermediate Bengali to English translation.Rest of the paper is organized as Section II describes state ofthe art works in the field of Bengali text analysis. Section II-Cbriefly discusses about the methodology of VADER operationin English text. In Section III, we discuss our approach to buildthe Bengali polarity lexicon from VADER English lexicon andthe operation methodology of the proposed Bengali VADER,which is followed by Results and Discussion in Section IV.Finally, we conclude the paper in Section V with possible978-1-5386-9111-3/19/$31.00 ©2019 IEEEfuture works.II. L RA. LexiconA lexicon is the vocabulary of a man, dialect, or a depart-ment of expertise that stocks the lexemes in that linguistics.Polarity lexicons are that which have a listing of words withinitial level of polarities. It is one of the main resources foranalyzing the sentiments and opinion expressed in texts in ancomputerized way. There are primarily three ways to buildpolarity lexicons: interpreting existing lexicons from differentlanguages, extracting polarity lexicons from corpora [10], andannotating sentiments Lexical Knowledge Base [11] [12] [13][14].For major languages there are well known manually con-structed lexicons such as General Inquirer [15], OpinionFinder[16], SO-CAL [17], etc. In [18] and [19] authors analyzedthe methodology of translating English resources to Romanianand Spanish respectively. There are several English polaritylexicons available online such as SentiWordNet1, VADER buthardly any Bengali polarity lexicon.B. English LanguageEnglish language is rich for notable works on sentimentanalysis such as VADER. In this model, researchers com-bined qualitative and quantitative methods to produce a gold-standard sentiment lexicon, which was later empirically val-idated against especially receptive microblog type contexts.VADER combines important lexical features obtained fromfive generalized rules that embody grammatical and syntacticalconventions of human speech. It also retains the advantagesof conventional sentiment lexicons such as LIWC [20] [21].In [14] authors explored three strategies to build polaritylexicons: interpreting existing lexicons from other languages,explaining sentiments from lexical knowledge bases and ex-tracting polarity lexicons from corpora. All of the models re-quire different degrees of human effort. Here, Spanish lexicon[22] has been translated by means of the Elhuyar Spanish-Basque2 dictionary, where every Spanish word incorporatesfive interpretations.C. VADERVADER combines qualitative and quantitative methods toproduce a gold-standard sentiment lexicon. A good number ofsentiment analysis models rely significantly on basic sentiment(or opinion) lexicons. A sentiment lexicon is a listing of lexicalcapabilities (e.g. words) which can be commonly categorizedwith their semantic orientation as either positive or negative[23]. After producing and validating lexicon, they have</s>
<s>usedthe architecture shown in Figure 1 to evaluate the sentimentof texts. The architecture mainly follows the steps describedbelow to calculate the polarity of the sentences:1http://sentiwordnet.isti.cnr.it/2http://hiztegiak.elhuyar.eus3VADER Sentiment, https://github.com/cjhutto/vaderSentimentEnglish TextPreprocessWords and emoticons, is_cap_diffWords_plus_punctuationValence CalculationScore_valenceNormalizationSeparate_sentiment_scoresBoostingIs_cap_diffBi-gram, Tri-gramIdioms_checkBut_checkTokenFigure 1. Implementation architecture of English VADER31) Preprocessing: In preprocessing step, input English textis tokenized. At first, words, emoticons and all capitalizedwords are tokenized. Later, words and punctuation marksare tokenized. Some punctuation marks affect the valence ofwords, which is kept with the word.2) Boosting: Once tokenization is complete all the tokensare checked for valence boosting purpose. VADER uses bi-gram and tri-gram model for boosting. If a boosting word suchas “extremely, very, great” are found then the valence of theword is boosted. Then all capital words are considered and thevalence is further boosted. The text is also checked for idiomsand phrases, if found then the valence is boosted again. Later,it is looks for ‘but’ word, if found then the sentence is dividedinto two parts and valence is calculated on two different parts.Next, overall valence is calculated for such a sentence.3) Valence Calculation: In this step, valence of a sentenceis measured, which is between -4 to +4. This calculated valueis further normalized to range between -1 to +1. In this way,every sentence is assigned with their respective polarity.D. Bengali LanguageThough Bengali language sentiment analysis is far behindthan English language sentiment analysis, Bengali polaritylexicon based works can learn from popular works such asSentiWordNet [24], VADER [9] etc. In [25], authors classifysentiment from sentence or an entire document as positive,negative or neutral. Firstly, they have used the WordNet toget the senses of words according to their respective partsof speech and later applied SentiWordNet to get their priorvalence or polarity.The authors of [26] have translated the SentiWordNet inBengali and used its polarity to calculate the valence ofBengali texts. They have developed training corpus usingsemi-supervised bootstrapping method; used Support VectorMachine (SVM) and Maximum Entropy (MaxEnt) for classi-fication; and later combined various set of features to comparethe performance of these two machine learning models. Theyalso built a Twitter-specific Bengali sentiment lexicon forprocedure-based classification using binary features.Das et. al. [27] constructed an opinion outline system basedon Bengali news corpus. The system identifies sentiment infor-mation from each document, which is further aggregated usingtopic-sentiment model and later summarized. Topic-sentimentmodel uses theme clustering (k-means), which further usesdocument level theme relation graph to achieve discourse leveltheme identification and topic sentiment aggregation. Finallystandard page rank algorithms are utilized for informationretrieval purpose.III. P B VADERWe have developed a model to identify sentiments fromBengali text. Figure 2 represents our developed system thatincludes preprocessing, boosting and Bengali valence creation.This paper focuses on two interrelated things: 1) developmentand validation of Bengali polarity lexicon and 2) extractionof sentiment intensity from Bengali text using our methods.These processes are described accordingly in the followingsections.A. DictionaryWe have created two dictionaries of negation words andbooster words4. They are used for negation and boosting of asentence respectively.1) Negation List: We have created a list of words that arecommon in Bengali to represent negativity of the texts. If anyof these words are</s>
<s>present in the text, then the polarity of thetext is reversed to either positive or to negative. Some examplewords are:(1) `না', `িন', `নয়', `নাই', `েনই'2) Booster Dictionary: Booster dictionary contains Bengaliwords that are used to boost valence of any text. If any oneof them are present in the text then polarity will be increasedin our model. The dictionary contains words such as:(2) `েবিশ', `খুব', `অেনক', `অিত', `অিতশয়', `বহুত', `অিধক',`অিধকতর'4https://github.com/mkazi078/bengalisentimentB. Bengali LexiconWe have constructed a lexicon translating VADER [9] lexi-con using a bilingual dictionary and given their correspondingpolarity. The valences of the words are given in a range of-4 to +4, where -4 represents most negativity, +4 representsmost positivity, and 0 represents neutral. There are more than3000 words in our compiled lexicon, which is continuouslygrowing.C. PreprocessingIn the preprocessing stage, we have removed punctuationas well as unnecessary stop words and conducted stemmingon the words. The steps are:1) Punctuation removal: The text is tokenized i.e. split intowords and punctuation marks are removed from the texts. Forexample, the following text(3) েস খুব েবিশ ভয় পায়িন ।will appear as(4) `েস', `খুব', `েবিশ', `ভয়', `পায়িন' `।'After punctuation removal will be presented as:(5) `েস', `খুব', `েবিশ', `ভয়', `পায়িন'2) Bengali stop-words removal: Stop words are thosewhich are ignored during sentiment analysis. We have createda list of Bengali stop words. The list contains words suchas `েস', `আিম', `এবং', etc. Whenever these words are found inthe token list, they are removed from the token list. Afterremoving the stop-words the token list of `েস খুব েবিশ ভয় পায়িন'appears as:(6) `খুব', `েবিশ', `ভয়', `পায়িন'3) Stemming: Natural language processing uses stemmingprocess to reduce derived words into their original form orstem. In our proposed Bengali VADER model, we stemmed aBengali word to its root form so that it can be easily comparedwith the lexicon. We first verify if the word needs to bestemmed. If the word has `ে◌', `◌া', `ি◌', ` ◌ী', `ে◌া' etc. at theend, then the word is stemmed. Then it is analyzed whetherthe word has `ে◌র', `টা', `িট' etc. at the end, if found they areremoved. In the third step, we check for `না', `িন' at the end ofthe word and if found the word is split there.Example of stemming `ে◌':(7) `ফাঁেদ', `পা', `িদেবন', `না'(8) `ফাঁদ', `পা', `িদেবন', `না'Example of stemming `এর':(9) `িবপেদর', `সময়', `না'(10) `িবপদ', `সময়', `না'Example of stemming `িন':(11) `খুব', `েবিশ', `ভয়', `পায়িন'(12) `খুব', `েবিশ', `ভয়', `পায়', `িন'D. Boosting word valenceIn this step, we search if the text has any booster word. Ifthe text contains any booster word included in the boosterdictionary, then it intensifies its valence according to theposition of the booster word in the sentence. To identify theposition of a booster word following three processes have beenused:1) Bigram: We have selected bigram or digram tokens i.e.two adjacent tokens in a sequence from a sentence. This pairof two adjacent tokens is again used to boost word valence.Example of bigram:(13) `খুব েবিশ', `েবিশ ভয়', `ভয় পায়', `পায় িন'2) Trigram: Trigram is a special case of the n-gram wheren=3. In this case, a pair of three consecutive tokens is againused</s>
<s>to boost a word valence. Example of trigram:(14) `খুব েবিশ ভয়', `েবিশ ভয় পায়', `ভয় পায় িন'For any word, if the boosting word is found in the booster dic-tionary, then for bigram the valence of the token is multipliedby 0.9, and for trigram, it is multiplied by 0.75 [9].3) Negation: Usually, overall sentiment of a sentence isaffected by negation words. But, use of negation words isdifferent in Bengali from English. In case of English, it is usedin the middle of the sentences, whereas in Bengali it is usednormally at the end of the sentences. At this step, negationword is searched according to the negation list constructedbefore. If a negation word is found, then the valence of thesentence is multiplied by -1 i.e. existing valence is reversed.E. Valence CalculationAfter completing all the previous steps, the final step is tocalculate the valence. This model guides in finding the correctvalence for any text. At any time, the system outputs sentimentscores in three classes: positive, negative and neutral. It furtherrequires normalization.1) Normalization: Compound score is computed by sum-ming the valence of each word in the lexicon, adjusted withrules, and then normalized to be generally between -1 (extremenegative) and +1 (extreme positive). This is an useful metricto get a single unidimensional measurement about sentiment.It is actually called “normalized weighted composite score”.To normalize the score, we use the equation:Normalizedscore =score√(score ∗ score) + alpha(1)where alpha = 15 is approximated maximum expected valueand score is the calculated score to be normalized.2) Separation of sentences: If the valence of the text is lessthan 0 and upto -1 then it expresses negativity. If it is 0 then itrepresents neutral and if the valence is greater than 0 upto +1then the text expresses positivity. Every sentence is markedas positive, negative and neutral according to their calculatedpolarity.IV. R DAfter construction of polarity lexicon and applying methodsaccordingly which has been described in the previous section,we get the result of a text if it is positive, negative orneutral depending on the generated polarity of that sentence.A sentence is called positive if its polarity is between 0 to +1,where +1 represents the highest intensity of positivity. On theBengali TextValence CalculationScore_valenceNormalizationSeparate_sentiment_scoresBoostingBi-gram, Tri-gramNegation HandlingTokenPreprocessTokeniztionStop_words RemovalStemmingPunctuation RemovalFigure 2. System Architecture of proposed Bengali VADERcontrary, if the polarity is between 0 to -1 then it representsthe negativity of a sentence, where -1 is the highest intensityof negative valence. And if the polarity is 0, then the sentenceis considered neutral in our Bengali VADER model.We have evaluated the sentiment for a sentence usingVADER, where VADER first translates the given Bengali textto English and gives the polarity using its English polaritylexicon. To translate the Bengali into English VADER usesMyMemory translator. Then we have used Google translatorand Python translator to translate the text to English forVADER and analyzed the polarity of the sentences. Thoseresults, as well as polarity analyzed by our system usingBengali polarity lexicon, are shown in Table I.From Table I we can see that for the Bengali sentence `েসপির মী না' VADER gives positive score using all the three</s>
<s>of itstranslators though in Bengali it is a negative sentence. Whereasour proposed Bengali VADER gives the correct polarity ofthe sentence and determines it as a negative sentence. For theBengali sentence `িবেদিশ সাহােয র গিত বা েল আর রাজ আদােয়রবৃি বা েল ঋেণর পিরমাণ কেম আসেব ।' VADER detects thesentence as negative by using its translator, whereas ourproposed Bengali VADER correctly determines it as positive.Table IC B VADER O P B VADERSentences VADER Our Proposed SystemPython Translator Google Translator MyMemory Translator Bengali VADERআিম আমার মত ভাল আিছ 0.5106 0.5106 0.5106 0.6597ব থতা সাফেল র চািব । 0.1027 0.1027 0.1027 0.1027িনেজর উপর আ িব াস থাকাভাল ।0.4404 0.4404 0.4404 0.7096িতিন একজন আদশ িশ ক 0.5267 0.5267 0.5267 0.5267েস ভূত ভয় পায় । -0.4404 -0.4404 -0.4404 -0.7003আমার েসৗভাগ হয়িন । -0.3412 -0.3412 -0.3412 -0.3818েস পির মী না 0.0762 0.0762 0.0762 -0.4767িক দা ণ খবর । 0.6249 0.6249 0.6249 0.4767ফাঁেদ পা িদেবন না । -0.3182 -0.3182 -0.3182 0.3182েস েলখাপ ায় মনেযাগী নয় -0.2584 0.0 -0.2584 -0.3818বুি ম ার সােথ েলেগ থাকেলআপিনও সফল হেবন ।0.7783 0.7783 0.7783 0.7003আিম আমার মা েক অেনকভালবািস ।0.6369 0.6369 0.6369 0.6697িবেদিশ সাহােয র গিত বা েলআর রাজ আদােয়র বৃিবা েল ঋেণর পিরমাণ কেমআসেব ।-0.0516 -0.0516 -0.3612 0.4215েস খুব েবিশ ভয় পায়িন । 0.0 0.0 0.0 0.5034চকচক করেলই েসানা হয় না 0.0 0.0 0.0 0.0Another Bengali sentence `েস েলখাপ ায় মনেযাগী নয়' is anegative sentence and is correctly detected by VADER usingMyMemory or Python translator but Google translator failsin this case. So, we can conclude that intermediate translatorcan not be an efficient way to translate Bengali sentencesinto English and measure their sentiment polarity as the resultlargely depend on the translator and how they are translated.From Table I we can see that in most of the cases our proposedsystem gives better results than VADER’s method to analyzesentiment from Bengali language. At the same time, since ourprocess avoids intermediate Bengali to English translation, itperforms faster than other processes.V. C F WIn this paper, we present the development of Bengali polar-ity lexicon from existing English lexicon of VADER. We havemodified the existing VADER architecture to accommodateBengali sentiment polarity detection. In our methodology, weuse stemming, enlist Bengali boosting words, apply bigram-trigram to combine with the system so that it can give a betterperformance. The results are encouraging as the system givesbetter analytical results than readily available translation basedmodel. From our experimental analysis, we demonstrate thatour proposed Bengali VADER performs efficiently than theother systems. Since, our system avoids intermediate Bengalito English translation it also performs faster. Bengali polaritydetection system is useful to understand users perception aboutany product, observation about political scenarios, etc. Byimproving the Bengali lexicon, bigram-trigram list we canproduce much better results. In future, we will apply machinelearning approaches so that it can give better performancein some sentences which consists of words that have botha positive and negative meaning.[1] T. Dictionary. Sentiment. Accessed 2018-06-12. [Online]. Available:http://www.dictionary.com/browse/sentiment[2] M. Webster. Analysis. Accessed 2018-06-12. [Online]. Available:https://www.merriam-webster.com/dictionary/analysis[3] Technopedia. Sentiment Analysis. Accessed 2018-06-12.[Online]. Available: https://www.techopedia.com/definition/29695/sentiment-analysis[4] O. Dictionary. Sentiment Analysis. Accessed 2018-06-12. [On-line]. Available: https://en.oxforddictionaries.com/definition/sentiment_analysis[5] K. Hasan, A. Mondal, A. Saha</s>
<s>et al., “Recognizing bangla grammarusing predictive parser,” arXiv preprint arXiv:1201.2010, 2012.[6] M. A. Islam, K. A. Hasan, and M. M. Rahman, “Basic hpsg structurefor bangla grammar,” in Computer and Information Technology (ICCIT),2012 15th International Conference on. IEEE, 2012, pp. 185–189.[7] K. A. Hasan, A. Mondal, and A. Saha, “A context free grammar andits predictive parser for bangla grammar recognition,” in Computer andInformation Technology (ICCIT), 2010 13th International Conferenceon. IEEE, 2010, pp. 87–91.[8] T. W. Post. The future of language. Accessed 2018-07-19. [Online].Available: https://www.washingtonpost.com/news/worldviews/wp/2015/09/24/the-future-of-language/?utm_term=.5158d11a583a[9] C. H. E. Gilbert, “Vader: A parsimonious rule-based model for sentimentanalysis of social media text,” in Eighth International Conferenceon Weblogs and Social Media (ICWSM-14). Available at (20/04/16)http://comp. social. gatech. edu/papers/icwsm14. vader. hutto. pdf, 2014.[10] V. Hatzivassiloglou and K. R. McKeown, “Predicting the semanticorientation of adjectives,” in Proceedings of the 35th annual meetingof the association for computational linguistics and eighth conferenceof the european chapter of the association for computational linguistics.Association for Computational Linguistics, 1997, pp. 174–181.[11] S.-M. Kim and E. Hovy, “Determining the sentiment of opinions,”in Proceedings of the 20th international conference on ComputationalLinguistics. Association for Computational Linguistics, 2004, p. 1367.[12] J. Kamps, M. Marx, R. J. Mokken, M. De Rijke et al., “Using wordnet tomeasure semantic orientations of adjectives.” in LREC, vol. 4. Citeseer,2004, pp. 1115–1118.[13] H. Liu and P. Singh, “Conceptnet—a practical commonsense reasoningtool-kit,” BT technology journal, vol. 22, no. 4, pp. 211–226, 2004.[14] I. San Vicente and X. Saralegi, “Polarity lexicon building: to what extentis the manual effort worth?” in LREC, 2016.[15] P. J. Stone, D. C. Dunphy, and M. S. Smith, “The general inquirer: Acomputer approach to content analysis.” 1966.[16] T. Wilson, P. Hoffmann, S. Somasundaran, J. Kessler, J. Wiebe, Y. Choi,C. Cardie, E. Riloff, and S. Patwardhan, “Opinionfinder: A systemfor subjectivity analysis,” in Proceedings of hlt/emnlp on interactivedemonstrations. Association for Computational Linguistics, 2005, pp.34–35.[17] M. Taboada, J. Brooke, M. Tofiloski, K. Voll, and M. Stede, “Lexicon-based methods for sentiment analysis,” Computational linguistics,vol. 37, no. 2, pp. 267–307, 2011.[18] R. Mihalcea, C. Banea, and J. Wiebe, “Learning multilingual subjectivelanguage via cross-lingual projections,” in Proceedings of the 45thannual meeting of the association of computational linguistics, 2007,pp. 976–983.[19] V. Perez-Rosas, C. Banea, and R. Mihalcea, “Learning sentiment lexi-cons in spanish.” in LREC, vol. 12, 2012, p. 73.[20] J. W. Pennebaker, M. E. Francis, and R. J. Booth, “Linguistic inquiryand word count: Liwc 2001,” Mahway: Lawrence Erlbaum Associates,vol. 71, no. 2001, p. 2001, 2001.[21] J. W. Pennebaker, R. J. Booth, and M. E. Francis, “Linguistic inquiryand word count: Liwc [computer software],” Austin, TX: liwc. net, 2007.[22] X. Saralegi, I. San Vicente, and I. Ugarteburu, “Cross-lingual projectionsvs. corpora extracted subjectivity lexicons for less-resourced languages,”in International Conference on Intelligent Text Processing and Compu-tational Linguistics. Springer, 2013, pp. 96–108.[23] B. Liu, “Sentiment analysis and subjectivity.” Handbook of naturallanguage processing, vol. 2, pp. 627–666, 2010.[24] S. Baccianella, A. Esuli, and F. Sebastiani, “Sentiwordnet 3.0: anenhanced lexical resource for sentiment analysis and opinion mining.”in LREC, vol. 10, no. 2010, 2010, pp. 2200–2204.[25] K. A. Hasan, M. Rahman et al., “Sentiment detection from</s>
<s>banglatext using contextual valency analysis,” in Computer and InformationTechnology (ICCIT), 2014 17th International Conference on. IEEE,2014, pp. 292–295.[26] S. Chowdhury and W. Chowdhury, “Performing sentiment analysis inbangla microblog posts,” in Informatics, Electronics & Vision (ICIEV),2014 International Conference on. IEEE, 2014, pp. 1–6.[27] A. Das and S. Bandyopadhyay, “Topic-based bengali opinion sum-marization,” in Proceedings of the 23rd International Conference onComputational Linguistics: Posters. Association for ComputationalLinguistics, 2010, pp. 232–240.</s>
<s>Sentiment Analysis on Movie Review Data Using Machine Learning ApproachInternational Conference on Bangla Speech and Language Processing (ICBSLP), 27-28 September, 2019 978-1-7281-5242-4/19 ©2019 IEEE Sentiment Analysis on Movie Review Data Using Machine Learning Approach Atiqur Rahman, Md. Sharif Hossen Dept. of Information and Communication Technology Comilla University, Comilla, Bangladesh atikriyadict@gmail.com, sharif5613@gmail.com Abstract—At present Sentiment analysis is the most discussed topic which is purposed to assist one to get important information from a large dataset. It centers on the investigation and comprehension of the feelings from the text patterns. It automatically characterizes the expression of feelings, e.g., negative, positive or neutral about the existence of anything. Various sources like medical, social media, newspaper, and movie review can be used in data analysis. Here, we have collected movie review data as well as used five kinds of machine learning classifiers to analyze these data. Hence, the considered classifiers are Bernoulli Naïve Bayes (BNB), Decision Tree (DE), Support Vector Machine (SVM), Maximum Entropy (ME), as well as Multinomial Naïve Bayes (MNB). Our analysis outlines that MNB achieves better accuracy, precision and F-score while SVM shows higher recall compared to others. Besides it also show that BNB Classifier achieves better accuracy than previous experiment over this classifier. Keywords— precision, sentiment, classifiers, recall, opinion I. INTRODUCTION Internet makes people easier to connect each other. They express their opinion using internet through social media, blog post, movie review, product review site etc. Everyday huge amounts of data are generated by user. Movies are probably the best form of entertainment for mankind and it is common that people watch the movies and express their opinions either in social networking sites. By analyzing movie review data we can learn about the strong and weak point of a movie and tell us if the movie meets the expectation of the user. When a person wants to watch a movie, he first checks the review and rating of the movie. Sentiment analysis (SA) helps in obtaining the review of that movie. SA is the process of getting valuable fact from a big set of data. It automatically classifies the people opinion as a positive or negative view. According to [1], SA techniques are Sentence, Document, Aspect and User based. The first technique detects the sentiment of each sentence as positive or negative. The second technique detects the sentiment of full document as a single unit. The third technique focuses on all the properties of an entity. Last technique conducts the social interrelation having graph theory for different parities. Machine learning (ML) and Lexicon based techniques are the most common in SA where the first uses training and testing set to categorize data. While the second approaches are used as a dictionary which contain predefined positive and negative words [2]. Here, we have used ML technique to classify data. Using document based SA we find the accuracy of different classifiers where Multinomial Naïve Bayes (NB) has better accuracy (88.50%) than others while Bernoulli NB achieves 87.50%. The organization of the paper</s>
<s>is following: The related investigation has been discussed in section II. Section III includes the analysis procedure of SA. Section IV discusses about various types of machine learning algorithm. Section V describes the experimental results. Section VI includes the summary and future endeavor. II. RELATED WORK This paper evaluates the opinion on movie review data. SA is handled by natural language processing (NLP) using several levels. Various approaches are available for doing this. In [2] authors proposed a method using word embedding to create sentiment lexicon by word vector representation. Authors of [3] discussed about the usage of SVM in SA with different source of data. In [4], authors proposed a method based on a sentiment score vector for SVM. In [5] Hu and Liu research was churning out the product features and gave product based summary. Authors of [6] worked on feedback data from global support services survey. Sentiment analysis has many usages on movie review dataset using different techniques. Authors of [7] broadly categorized SA techniques into ML and lexical-based. Yonas Woldemariam has done SA based on ML and lexicon based approaches. He used apache hadoop framework with lexicon based model [8]. Authors of [9] discussed about entity-level sentiment analysis of issue comment. Authors of [10] discussed about the sentiment lexicon dictionary enrichment based on word2vec. They enlarge the opinion words by using SentiwordNet. In [11], authors deal the view-level SA on e-commerce data. According to [12], SA can be applied to detect the polarity of customer reviews in several dimensions. According to [13], SA can be done based on combined techniques. ML needs training and testing data set. There are two types of ML methods, namely, supervised method (SM) and unsupervised method (UM). SM is a method in which we at first teach the machine providing some data. It [14] has several algorithms to conduct the classification technique based on the trained data which are. Naïve Bayes (NB), SVM, ME, DE etc. There is no training data in unsupervised learning where data are unlabeled. Author of [15, 16] design advanced NB algorithm with parts-of-speech tagging. He also claim that NB algorithm is costly wastes a lot of resources. Authors of [13] discussed the NB and complement NB classifier algorithm on hadoop framework. According to [17] SVM classifier was also conduct to find the public user counsel on products. NB and SVM were used [18] on various medical forum data. Authors of [19] have done the sentiment analysis on Bengali data about Bangladesh cricket team using support vector machine. In [20], they done the sentiment analysis using KNN, Bagging, COCR, NB and 978-1-7281-5241-7/19/$31.00 ©2019 IEEEl C/ICAuthorized licensed use limited to: Auckland University of Technology. Downloaded on May 23,2020 at 04:37:18 UTC from IEEE Xplore. Restrictions apply. Decision tree classifier. Authors of [21] used deep learning recurrent neural network method and decision tree for movie review data. Two kinds of lexicons are available [22]. The 1st one is corpus based lexicon and second one is dictionary based lexicon. Corpus</s>
<s>based lexicon, like as SenticNet [23] can acquire more appropriate sentiment outcome as it is context oriented instead of accord of words oriented. Semantic oriented concept was applied based on a concept net lexicon. In [24] the authors show a statistical approach to find the sentiments. Authors of [25] show the 2 dictionaries. The 1st one is word dictionary and the second one is topic modeling. According to [14], a small set of counsel words are culled manually with acquainted orientations in dictionary based approach. This paper follows the machine learning technique for sentiment analysis. III. ANALYSIS PROCEDURE In this section, we discuss the analysis procedure. Figure 1 show the steps and techniques used in this paper for classification our text. Fig. 1: Steps for classification in SA First of all we created our dataset from the movie review data (Data source are mentioned in reference section [26]). We collected two thousands movie review posts where there exists equal number of positive and negative reviews. Then, we followed those steps for our sentiment classification. A. Text Preprocessing a) URL Removal: URL is a sharing link which is also known as HTML tag. Many texts contain URL or HTML tag. First of all we remove URL from the text. b) Bracket and Number Removal: Bracket and numbers have no meaning in sentiment classification. These are treated as noise. We need to clean bracket and number from the text. c) Tokenization: We divide our textual data into smaller components. This can be done by using tokenization. Tokenization provides us to turn text into sentences and sentences into words. d) Omitting Punctuation: Quotation mark, semi-colon, colon, etc. are omitted as there are no data. e) Case conversion: To remove the distinction between “Review” and “review” it is done. f) Omitting stop words such as “I”, “it”, “you”, “a”, “an”, “the” since they have no meaning in sentiment classification. We should remove those words from the input text. g) Stemming: It is the method to reform the inflected words and removes derivational affixes from a word. B. Feature Vector Creation Feature is a measurable characteristic of a phenomenon. We classify each review as positive or negative. Every review is considered as a simple document. In unigram model, every document is denoted by its primary words where positive and negative primary words are used for respective positive and negative document. We also added parts of speech (PoS) tags to get the more accurate sentiment. Emotions are determined by opinion words. Positive and negative sentiments are maintained by respective sentiment scores. For this reason, these words are instrumental feature to sentiment analysis. How many features are there is termed as dimensionality. Positive and negative keywords are signed as pos and neg respectively for every document. Then, PoS are included with sentence. For this reason every word has a feature. Depending on the more variety of dimensions classifiers shows worst outcome. IV. CONSIDERED CLASSIFIERS A. NB Classifier NB [27] is a SM which is the easiest and</s>
<s>most recognizable used classifier. It considers that every feature is distinct from one another. NB classifiers are a collection of classifications algorithm which is not a standalone algorithm but a family of algorithm. The mathematical expression is as follows: ( \| ) = ( \| ) ( )( ) Where ( \| ) = ( \| ) ( \| ) … ( \| ) Here, X is class variable and Y is a dependent feature vector. P(x) and P(x\|Y) denote the respective priori and posteriori probability of x and Y. We use two Bernoulli and Multinomial NB techniques where Multinomial NB is good for when features describe discrete frequency counts (e.g. word counts). If we want to estimate P(W \| Z), the decision rule for this classifier become ( \| ) = ( ) ( ) ( ) where W denotes sentiment. BNB is good for making predictions from binary features. Mathematically, Input data Pre-processing 1. URL and Punctuation removal 2. Bracket and Number removal 3. Tokenization 4. Stop-words removal 5. Case conversion 6. StemmingCreate Features Vector Classification Result Authorized licensed use limited to: Auckland University of Technology. Downloaded on May 23,2020 at 04:37:18 UTC from IEEE Xplore. Restrictions apply. ) = ( \| ) (1 ( \| ))(1 ) which differs from MNB approach. B. SVM SVM is another SM technique which is used to figure out each raw fact as a dot for fixed dimensions of features. Then, we choose a hyper-plane between two classes. There are several hyper-planes, but we choose one that maximizes the margin between the two classes. According to [14], text categorization is consummately appropriate for SVM due to the meager idea of content, where not many highlights are unimportant, yet they will in general be associated with each other and by and large planned into straightly particular classes. C. ME Classifier ME is an algorithm which uses probabilistic concept and does not guess that the characteristics are conditionally autonomous of each other. It is utilized when we cannot accept the restrictive autonomy of the features. Using the training data it makes a model which can prefer the trained data having higher entropy after a much time than other classifiers. D. DT Classifier DT differentiates those records having many features checking the property from the root and vertex in a tree. All terminal vertexes are assigned a class label pos or neg. The checker on property is the occurrence or non-occurrence of at least one word. Tree is partitioned until there exists least number of records. V. EXPERIMENTAL RESULT AND ANALYSIS A. Training and Testing Data After creating feature vector we apply ML techniques for classification. We take 1400 and 600 movie reviews for training and testing data set where first sets are trained by a classifier and the accuracy of that is calculated using the second. We use five types of main classifiers, namely, MNB, BNB, SVM, ME and DT). All classifiers are implemented by using python. The dataset contains 2000 movie review</s>
<s>where 1000 is negative and remaining is positive. We use the terms, namely, true positive (U), false positive (V), true negative (X), false negative (Y) for analysis. Here, first and second terms indicate that the review is really positive and negative respectively but both are featured as positive term. Third and fourth terms indicate that the review is really negative and positive respectively but both are featured as negative term. Accuracy, recall, precision, and F-score are determined from the above terms. Table 2 shows the analysis got from each classifier for data with label. Table 1. Classifiers with four terms for the 2000 review Method U V X Multinomial 250 19 281 50 Bernoulli 259 34 266 41 SVM 268 44 256 32 Maximum Entropy 254 190 110 46 Decision Tree 247 66 234 53 Table 2. Performance statistics of several classifiers Method Accuracy Precision Recall F-score Multinomial 88.50% 92.94% 83.33% 87.87% Bernoulli 87.50% 88.40% 86.33% 87.35% SVM 87.33% 85.90% 89.33% 87.58% Maximum Entropy 60.67% 57.21% 84.67% 68.28% Decision Tree 80.17% 78.91% 82.33% 80.58% Figure 2 shows the graphical representation of accuracy, precision and recall. We also show the recall versus precision graph for better comparison in Fig 3. From Fig. 2 and Table 2, we see that Multinomial NB has better accuracy compared to others. It obtains an accuracy of 88.50% while Bernoulli NB obtains 87.50%, SVM with 87.33%, Maximum Entropy with 60.67% and Decision tree with 80.17%. Multinomial NB also has high precision and F-score (Table 2), but the SVM has higher recall. Besides it also shows that Bernoulli Naïve Bayes Classifier achieves better accuracy than previous experiment over this classifier [27]. It has the good precision, recall and f-score Maximum Entropy classifier shows the low performance than other classifier. It has the low precision and f-score compare to the others. But, the precision is higher than Multinomial NB and Decision Tree. The above result shows the quality of features vector selected for movie review data. Fig. 2: Performance of difference classifiers Every classifier is sensitive to parameter optimization. Although the result shows that Multinomial Naïve Bayes classifier is better than SVM, this is only true for selected parameters because multinomial NB show the worse results when training dataset is small. Authorized licensed use limited to: Auckland University of Technology. Downloaded on May 23,2020 at 04:37:18 UTC from IEEE Xplore. Restrictions apply. Fig.3: Classifiers for precision vs recall VI. CONCLUSION AND FUTURE PLAN Sentiment analysis is very essential to understand the expression of feelings about anything like product, social media etc. It can be done by lexicon (LN) and machine learning (ML) approaches. LN can fail to calculate the score of expression if a word not found in the dictionary. While, ML is easier and more efficient but it requires labeled data. In this paper, we use ML approach for polarity classification on movie review data. This approach divides the dataset into two sets, i.e., train and test set. First of all a data set is collected from</s>
<s>the movie review site. Next, we perform pre-processing on data by using NLP tool. Then, after creating features vector the data set is trained using ML classifiers, namely, Multinomial NB, Bernoulli NB, SVM, Maximum Entropy and Decision Tree classifiers which are tested using test dataset. Finally, we show our experimental results which present that the accuracy (88.5%) of Multinomial NB is better than others. In near future, we would like to extend this work using deep learning approaches. REFERENCES [1] C. Tan, L. Lee, J. Tang, L. Jiang, M. Zhou, and P. Li, “User-level sentiment analysis incorporating social network,” In Proc. of ICKDDM, IEEE, pp. 1397-1405, 2011. [2] X. Fan, X. Li, F. Du, Xin Li, Mian Wei, “Apply word vectors for sentiment analysis of APP reviews,” In Proc. of ICSI, IEEE, 2016. [3] T. Mullen, N. Collier, “Sentiment analysis using support vector machines with diverse information sources”, In Proc. of ICEMNLP, pp. 412-418, 2004. [4] S. Naz, A. Sharan, N. Malik, “Sentiment classification on twitter data using support vector machine,” In Proc. of ICWI, 2018. [5] M. Hu and B. Liu, “Mining and summarizing customer remarks,” In Pro. of ICKDDM, IEEE, pp. 168–177, 2004. [6] M. Gammon, “Sentiment classification on customer feedback data: noisy data, large feature vectors, and the role of linguistic analysis,” In Proc. of ICCL, pp. 841-847, 2004. [7] B. Pang, L. Lee, and S. Vaithyanathan “Thumbs up?: sentiment classification using machine learning techniques,” In Proc. of ICEMNLP, pp. 79-86, 2002. [8] Y. Woldemariam, “Sentiment analysis in a cross-media analysis framework,” In Proc. of ICBDA, IEEE, 2016. [9] J. Ding, H. Sun, X. Wang, X. Liu, “Entity-level sentiment analysis of issue comments,” In Proc. of IWEASE, IEEE, 2018. [10] E. M. Alshari, A. Azman, S. Doraisamy, N. Mustapha, M. Alkeshr, “Effective method for sentiment lexical dictionary enrichment based on word2vec for sentiment analysis,” In Proc. of ICIRKM, Malaysia, 2018. [11] S. Vanaja, M. Belwal, “Aspect-level sentiment analysis on e-commerce data,” In Proc. of ICIRCA, 2018. [12] P. Porntrakoon, C. Moemeng, “ Thai sentiment analysis for consumer’s review in multiple dimension using sentiment compensation technique.,” In Proc. of ICEECTIT, 2018. [13] B. Seref, E. Bostanci, “Sentiment analysis using naïve bayes and complement naïve bayes classifier algorithms on handoop framework,” Int. Symp. on Multisciplinary Studies and Innovative Technologies, 2018. [14] W. Medhat, A. Hassan, “Sentiment analysis algorithms and applications: Asurvey,” Shams Engineering, vol. 5, pp. 1093–1113, 2014. [15] F. Xianghua, L. Guo, G. Yanyan, and W. Zhiqiang, “Multi-aspect sentiment analysis for chinese online social reviews based on topic modeling and hownet lexicon,” Knowledge-Based Sys., vol. 37, pp. 186-195, 2013 [16] Y. Wang, “Advanced naïve bayes algorithm design with part-of-speech tagger on sentiment analysis,” In Proc. of Int. Conf. on Computer System, Electronics and control, 2017. [17] H. Cho, S. Kim, J. Lee, and J.-S. Lee, “Data-driven integration of multiple sentiment dictionaries for lexicon-based sentiment classification of product reviews,” Knowledge-Based Sys., vol. 71, pp. 61–71, 2014. [18] T. Ali, D. Schramm, M. Sokolova, and D. Inkpen, “Can i hear</s>
<s>you? sentiment analysis on medical forums,” In Proc. of ICNLP, Asian Federation of Natural Language Processing, Nagoya, Japan, pp. 667–673, 2013. [19] S. A. Mahtab, N. Islam, M. Rahman, “Sentiment analysis on bangladesh cricket with support vector machine,” In Proc. of ICBSLP, 2018. [20] T. P. Sahu, Sanjeev Ahuja, “Sentiment analysis of movie review: A study on feature selection & classification algorithm,” In Proc. of ICMCC, 2016. [21] A. S. Zharmagambetov, A. A. Pak, “Sentiment analysis pf a document using deep learning approch and decision trees,” In Proc. of ICECC, 2015. [22] N. El-Fishawy, A. Hamouda, G. M. Attiya, and M. Atef, “Arabic summarization in twitter social network,” Ain Shams Eng., vol. 5, no. 2, pp. 411–420, 2014. [23] E. Cambria, A. Livingstone, and A. Hussain, “The hourglass of emotions,” Cognitive behavioural sys., Springer, pp. 144–157, 2012. [24] A. Hogenboom, F. Boon, and F. Frasincar, “A statistical approach to star rating classification of sentiment,” Management Int. Sys.. Springer, pp. 251–260, 2012. [25] M. Taboada, J. Brooke, M. Tofiloski, K. Voll, and M. Stede, “Lexiconbased methods for sentiment analysis,” Comp. linguistics, vol. 37, no. 2, pp. 267–307, 2011. [26] Movie Review 2000 dataset, https://github.com/riyadatik/Sentiment-Analysis-on-Movie-Review-Data/blob/master/Data%20set.xlsx, 2019. [27] V. Narayanan, I. Arora, A. Bhatia, “Fast and accurate sentiment classification using an enhanced Naïve Bayes model,” Int. Data Eng. and Aut. Learning, Lec. Notes. in Com. Sci., vol. 8206, pp 194-201, 2013. Authorized licensed use limited to: Auckland University of Technology. Downloaded on May 23,2020 at 04:37:18 UTC from IEEE Xplore. Restrictions apply. /ASCII85EncodePages false /AllowTransparency false /AutoPositionEPSFiles true /AutoRotatePages /None /Binding /Left /CalGrayProfile (Gray Gamma 2.2) /CalRGBProfile (sRGB IEC61966-2.1) /CalCMYKProfile (U.S. Web Coated \050SWOP\051 v2) /sRGBProfile (sRGB IEC61966-2.1) /CannotEmbedFontPolicy /Error /CompatibilityLevel 1.7 /CompressObjects /Off /CompressPages true /ConvertImagesToIndexed true /PassThroughJPEGImages true /CreateJobTicket false /DefaultRenderingIntent /Default /DetectBlends true /DetectCurves 0.0000 /ColorConversionStrategy /LeaveColorUnchanged /DoThumbnails false /EmbedAllFonts true /EmbedOpenType false /ParseICCProfilesInComments true /EmbedJobOptions true /DSCReportingLevel 0 /EmitDSCWarnings false /EndPage -1 /ImageMemory 1048576 /LockDistillerParams true /MaxSubsetPct 100 /Optimize true /OPM 0 /ParseDSCComments false /ParseDSCCommentsForDocInfo true /PreserveCopyPage true /PreserveDICMYKValues true /PreserveEPSInfo false /PreserveFlatness true /PreserveHalftoneInfo true /PreserveOPIComments false /PreserveOverprintSettings true /StartPage 1 /SubsetFonts true /TransferFunctionInfo /Remove /UCRandBGInfo /Preserve /UsePrologue false /ColorSettingsFile () /AlwaysEmbed [ true /AbadiMT-CondensedLight /ACaslon-Italic /ACaslon-Regular /ACaslon-Semibold /ACaslon-SemiboldItalic /AdobeArabic-Bold /AdobeArabic-BoldItalic /AdobeArabic-Italic /AdobeArabic-Regular /AdobeHebrew-Bold /AdobeHebrew-BoldItalic /AdobeHebrew-Italic /AdobeHebrew-Regular /AdobeHeitiStd-Regular /AdobeMingStd-Light /AdobeMyungjoStd-Medium /AdobePiStd /AdobeSongStd-Light /AdobeThai-Bold /AdobeThai-BoldItalic /AdobeThai-Italic /AdobeThai-Regular /AGaramond-Bold /AGaramond-BoldItalic /AGaramond-Italic /AGaramond-Regular /AGaramond-Semibold /AGaramond-SemiboldItalic /AgencyFB-Bold /AgencyFB-Reg /AGOldFace-Outline /AharoniBold /Algerian /Americana /Americana-ExtraBold /AndaleMono /AndaleMonoIPA /AngsanaNew /AngsanaNew-Bold /AngsanaNew-BoldItalic /AngsanaNew-Italic /AngsanaUPC /AngsanaUPC-Bold /AngsanaUPC-BoldItalic /AngsanaUPC-Italic /Anna /ArialAlternative /ArialAlternativeSymbol /Arial-Black /Arial-BlackItalic /Arial-BoldItalicMT /Arial-BoldMT /Arial-ItalicMT /ArialMT /ArialMT-Black /ArialNarrow /ArialNarrow-Bold /ArialNarrow-BoldItalic /ArialNarrow-Italic /ArialRoundedMTBold /ArialUnicodeMS /ArrusBT-Bold /ArrusBT-BoldItalic /ArrusBT-Italic /ArrusBT-Roman /AvantGarde-Book /AvantGarde-BookOblique /AvantGarde-Demi /AvantGarde-DemiOblique /AvantGardeITCbyBT-Book /AvantGardeITCbyBT-BookOblique /BakerSignet /BankGothicBT-Medium /Barmeno-Bold /Barmeno-ExtraBold /Barmeno-Medium /Barmeno-Regular /Baskerville /BaskervilleBE-Italic /BaskervilleBE-Medium /BaskervilleBE-MediumItalic /BaskervilleBE-Regular /Baskerville-Bold /Baskerville-BoldItalic /Baskerville-Italic /BaskOldFace /Batang /BatangChe /Bauhaus93 /Bellevue /BellMT /BellMTBold /BellMTItalic /BerlingAntiqua-Bold /BerlingAntiqua-BoldItalic /BerlingAntiqua-Italic /BerlingAntiqua-Roman /BerlinSansFB-Bold /BerlinSansFBDemi-Bold /BerlinSansFB-Reg /BernardMT-Condensed /BernhardModernBT-Bold /BernhardModernBT-BoldItalic /BernhardModernBT-Italic /BernhardModernBT-Roman /BiffoMT /BinnerD /BinnerGothic /BlackadderITC-Regular /Blackoak /blex /blsy /Bodoni /Bodoni-Bold /Bodoni-BoldItalic /Bodoni-Italic /BodoniMT /BodoniMTBlack /BodoniMTBlack-Italic /BodoniMT-Bold /BodoniMT-BoldItalic /BodoniMTCondensed /BodoniMTCondensed-Bold /BodoniMTCondensed-BoldItalic /BodoniMTCondensed-Italic /BodoniMT-Italic /BodoniMTPosterCompressed /Bodoni-Poster /Bodoni-PosterCompressed /BookAntiqua /BookAntiqua-Bold /BookAntiqua-BoldItalic /BookAntiqua-Italic /Bookman-Demi /Bookman-DemiItalic /Bookman-Light /Bookman-LightItalic /BookmanOldStyle /BookmanOldStyle-Bold /BookmanOldStyle-BoldItalic /BookmanOldStyle-Italic /BookshelfSymbolOne-Regular /BookshelfSymbolSeven /BookshelfSymbolThree-Regular</s>
<s>/BookshelfSymbolTwo-Regular /Botanical /Boton-Italic /Boton-Medium /Boton-MediumItalic /Boton-Regular /Boulevard /BradleyHandITC /Braggadocio /BritannicBold /Broadway /BrowalliaNew /BrowalliaNew-Bold /BrowalliaNew-BoldItalic /BrowalliaNew-Italic /BrowalliaUPC /BrowalliaUPC-Bold /BrowalliaUPC-BoldItalic /BrowalliaUPC-Italic /BrushScript /BrushScriptMT /CaflischScript-Bold /CaflischScript-Regular /Calibri /Calibri-Bold /Calibri-BoldItalic /Calibri-Italic /CalifornianFB-Bold /CalifornianFB-Italic /CalifornianFB-Reg /CalisMTBol /CalistoMT /CalistoMT-BoldItalic /CalistoMT-Italic /Cambria /Cambria-Bold /Cambria-BoldItalic /Cambria-Italic /CambriaMath /Candara /Candara-Bold /Candara-BoldItalic /Candara-Italic /Carta /CaslonOpenfaceBT-Regular /Castellar /CastellarMT /Centaur /Centaur-Italic /Century /CenturyGothic /CenturyGothic-Bold /CenturyGothic-BoldItalic /CenturyGothic-Italic /CenturySchL-Bold /CenturySchL-BoldItal /CenturySchL-Ital /CenturySchL-Roma /CenturySchoolbook /CenturySchoolbook-Bold /CenturySchoolbook-BoldItalic /CenturySchoolbook-Italic /CGTimes-Bold /CGTimes-BoldItalic /CGTimes-Italic /CGTimes-Regular /CharterBT-Bold /CharterBT-BoldItalic /CharterBT-Italic /CharterBT-Roman /CheltenhamITCbyBT-Bold /CheltenhamITCbyBT-BoldItalic /CheltenhamITCbyBT-Book /CheltenhamITCbyBT-BookItalic /Chiller-Regular /Cmb10 /CMB10 /Cmbsy10 /CMBSY10 /CMBSY5 /CMBSY6 /CMBSY7 /CMBSY8 /CMBSY9 /Cmbx10 /CMBX10 /Cmbx12 /CMBX12 /Cmbx5 /CMBX5 /Cmbx6 /CMBX6 /Cmbx7 /CMBX7 /Cmbx8 /CMBX8 /Cmbx9 /CMBX9 /Cmbxsl10 /CMBXSL10 /Cmbxti10 /CMBXTI10 /Cmcsc10 /CMCSC10 /Cmcsc8 /CMCSC8 /Cmcsc9 /CMCSC9 /Cmdunh10 /CMDUNH10 /Cmex10 /CMEX10 /CMEX7 /CMEX8 /CMEX9 /Cmff10 /CMFF10 /Cmfi10 /CMFI10 /Cmfib8 /CMFIB8 /Cminch /CMINCH /Cmitt10 /CMITT10 /Cmmi10 /CMMI10 /Cmmi12 /CMMI12 /Cmmi5 /CMMI5 /Cmmi6 /CMMI6 /Cmmi7 /CMMI7 /Cmmi8 /CMMI8 /Cmmi9 /CMMI9 /Cmmib10 /CMMIB10 /CMMIB5 /CMMIB6 /CMMIB7 /CMMIB8 /CMMIB9 /Cmr10 /CMR10 /Cmr12 /CMR12 /Cmr17 /CMR17 /Cmr5 /CMR5 /Cmr6 /CMR6 /Cmr7 /CMR7 /Cmr8 /CMR8 /Cmr9 /CMR9 /Cmsl10 /CMSL10 /Cmsl12 /CMSL12 /Cmsl8 /CMSL8 /Cmsl9 /CMSL9 /Cmsltt10 /CMSLTT10 /Cmss10 /CMSS10 /Cmss12 /CMSS12 /Cmss17 /CMSS17 /Cmss8 /CMSS8 /Cmss9 /CMSS9 /Cmssbx10 /CMSSBX10 /Cmssdc10 /CMSSDC10 /Cmssi10 /CMSSI10 /Cmssi12 /CMSSI12 /Cmssi17 /CMSSI17 /Cmssi8 /CMSSI8 /Cmssi9 /CMSSI9 /Cmssq8 /CMSSQ8 /Cmssqi8 /CMSSQI8 /Cmsy10 /CMSY10 /Cmsy5 /CMSY5 /Cmsy6 /CMSY6 /Cmsy7 /CMSY7 /Cmsy8 /CMSY8 /Cmsy9 /CMSY9 /Cmtcsc10 /CMTCSC10 /Cmtex10 /CMTEX10 /Cmtex8 /CMTEX8 /Cmtex9 /CMTEX9 /Cmti10 /CMTI10 /Cmti12 /CMTI12 /Cmti7 /CMTI7 /Cmti8 /CMTI8 /Cmti9 /CMTI9 /Cmtt10 /CMTT10 /Cmtt12 /CMTT12 /Cmtt8 /CMTT8 /Cmtt9 /CMTT9 /Cmu10 /CMU10 /Cmvtt10 /CMVTT10 /ColonnaMT /Colossalis-Bold /ComicSansMS /ComicSansMS-Bold /Consolas /Consolas-Bold /Consolas-BoldItalic /Consolas-Italic /Constantia /Constantia-Bold /Constantia-BoldItalic /Constantia-Italic /CooperBlack /CopperplateGothic-Bold /CopperplateGothic-Light /Copperplate-ThirtyThreeBC /Corbel /Corbel-Bold /Corbel-BoldItalic /Corbel-Italic /CordiaNew /CordiaNew-Bold /CordiaNew-BoldItalic /CordiaNew-Italic /CordiaUPC /CordiaUPC-Bold /CordiaUPC-BoldItalic /CordiaUPC-Italic /Courier /Courier-Bold /Courier-BoldOblique /CourierNewPS-BoldItalicMT /CourierNewPS-BoldMT /CourierNewPS-ItalicMT /CourierNewPSMT /Courier-Oblique /CourierStd /CourierStd-Bold /CourierStd-BoldOblique /CourierStd-Oblique /CourierX-Bold /CourierX-BoldOblique /CourierX-Oblique /CourierX-Regular /CreepyRegular /CurlzMT /David-Bold /David-Reg /DavidTransparent /Dcb10 /Dcbx10 /Dcbxsl10 /Dcbxti10 /Dccsc10 /Dcitt10 /Dcr10 /Desdemona /DilleniaUPC /DilleniaUPCBold /DilleniaUPCBoldItalic /DilleniaUPCItalic /Dingbats /DomCasual /Dotum /DotumChe /DoulosSIL /EdwardianScriptITC /Elephant-Italic /Elephant-Regular /EngraversGothicBT-Regular /EngraversMT /EraserDust /ErasITC-Bold /ErasITC-Demi /ErasITC-Light /ErasITC-Medium /ErieBlackPSMT /ErieLightPSMT /EriePSMT /EstrangeloEdessa /Euclid /Euclid-Bold /Euclid-BoldItalic /EuclidExtra /EuclidExtra-Bold /EuclidFraktur /EuclidFraktur-Bold /Euclid-Italic /EuclidMathOne /EuclidMathOne-Bold /EuclidMathTwo /EuclidMathTwo-Bold /EuclidSymbol /EuclidSymbol-Bold /EuclidSymbol-BoldItalic /EuclidSymbol-Italic /EucrosiaUPC /EucrosiaUPCBold /EucrosiaUPCBoldItalic /EucrosiaUPCItalic /EUEX10 /EUEX7 /EUEX8 /EUEX9 /EUFB10 /EUFB5 /EUFB7 /EUFM10 /EUFM5 /EUFM7 /EURB10 /EURB5 /EURB7 /EURM10 /EURM5 /EURM7 /EuroMono-Bold /EuroMono-BoldItalic /EuroMono-Italic /EuroMono-Regular /EuroSans-Bold /EuroSans-BoldItalic /EuroSans-Italic /EuroSans-Regular /EuroSerif-Bold /EuroSerif-BoldItalic /EuroSerif-Italic /EuroSerif-Regular /EUSB10 /EUSB5 /EUSB7 /EUSM10 /EUSM5 /EUSM7 /FelixTitlingMT /Fences /FencesPlain /FigaroMT /FixedMiriamTransparent /FootlightMTLight /Formata-Italic /Formata-Medium /Formata-MediumItalic /Formata-Regular /ForteMT /FranklinGothic-Book /FranklinGothic-BookItalic /FranklinGothic-Demi /FranklinGothic-DemiCond /FranklinGothic-DemiItalic /FranklinGothic-Heavy /FranklinGothic-HeavyItalic /FranklinGothicITCbyBT-Book /FranklinGothicITCbyBT-BookItal /FranklinGothicITCbyBT-Demi /FranklinGothicITCbyBT-DemiItal /FranklinGothic-Medium /FranklinGothic-MediumCond /FranklinGothic-MediumItalic /FrankRuehl /FreesiaUPC /FreesiaUPCBold /FreesiaUPCBoldItalic /FreesiaUPCItalic /FreestyleScript-Regular /FrenchScriptMT /Frutiger-Black /Frutiger-BlackCn /Frutiger-BlackItalic /Frutiger-Bold /Frutiger-BoldCn /Frutiger-BoldItalic /Frutiger-Cn /Frutiger-ExtraBlackCn /Frutiger-Italic /Frutiger-Light /Frutiger-LightCn /Frutiger-LightItalic /Frutiger-Roman /Frutiger-UltraBlack /Futura-Bold /Futura-BoldOblique /Futura-Book /Futura-BookOblique /FuturaBT-Bold /FuturaBT-BoldItalic /FuturaBT-Book /FuturaBT-BookItalic /FuturaBT-Medium /FuturaBT-MediumItalic /Futura-Light /Futura-LightOblique /GalliardITCbyBT-Bold /GalliardITCbyBT-BoldItalic /GalliardITCbyBT-Italic /GalliardITCbyBT-Roman /Garamond /Garamond-Bold /Garamond-BoldCondensed /Garamond-BoldCondensedItalic /Garamond-BoldItalic /Garamond-BookCondensed /Garamond-BookCondensedItalic /Garamond-Italic /Garamond-LightCondensed /Garamond-LightCondensedItalic /Gautami /GeometricSlab703BT-Light /GeometricSlab703BT-LightItalic /Georgia /Georgia-Bold /Georgia-BoldItalic /Georgia-Italic /GeorgiaRef /Giddyup /Giddyup-Thangs /Gigi-Regular /GillSans /GillSans-Bold /GillSans-BoldItalic /GillSans-Condensed /GillSans-CondensedBold /GillSans-Italic /GillSans-Light /GillSans-LightItalic /GillSansMT /GillSansMT-Bold /GillSansMT-BoldItalic /GillSansMT-Condensed /GillSansMT-ExtraCondensedBold /GillSansMT-Italic /GillSans-UltraBold /GillSans-UltraBoldCondensed /GloucesterMT-ExtraCondensed /Gothic-Thirteen /GoudyOldStyleBT-Bold /GoudyOldStyleBT-BoldItalic /GoudyOldStyleBT-Italic /GoudyOldStyleBT-Roman /GoudyOldStyleT-Bold /GoudyOldStyleT-Italic /GoudyOldStyleT-Regular /GoudyStout /GoudyTextMT-LombardicCapitals /GSIDefaultSymbols /Gulim /GulimChe /Gungsuh /GungsuhChe /Haettenschweiler /HarlowSolid /Harrington /Helvetica /Helvetica-Black /Helvetica-BlackOblique /Helvetica-Bold /Helvetica-BoldOblique /Helvetica-Condensed</s>
<s>/Helvetica-Condensed-Black /Helvetica-Condensed-BlackObl /Helvetica-Condensed-Bold /Helvetica-Condensed-BoldObl /Helvetica-Condensed-Light /Helvetica-Condensed-LightObl /Helvetica-Condensed-Oblique /Helvetica-Fraction /Helvetica-Narrow /Helvetica-Narrow-Bold /Helvetica-Narrow-BoldOblique /Helvetica-Narrow-Oblique /Helvetica-Oblique /HighTowerText-Italic /HighTowerText-Reg /Humanist521BT-BoldCondensed /Humanist521BT-Light /Humanist521BT-LightItalic /Humanist521BT-RomanCondensed /Imago-ExtraBold /Impact /ImprintMT-Shadow /InformalRoman-Regular /IrisUPC /IrisUPCBold /IrisUPCBoldItalic /IrisUPCItalic /Ironwood /ItcEras-Medium /ItcKabel-Bold /ItcKabel-Book /ItcKabel-Demi /ItcKabel-Medium /ItcKabel-Ultra /JasmineUPC /JasmineUPC-Bold /JasmineUPC-BoldItalic /JasmineUPC-Italic /JoannaMT /JoannaMT-Italic /Jokerman-Regular /JuiceITC-Regular /Kartika /Kaufmann /KaufmannBT-Bold /KaufmannBT-Regular /KidTYPEPaint /KinoMT /KodchiangUPC /KodchiangUPC-Bold /KodchiangUPC-BoldItalic /KodchiangUPC-Italic /KorinnaITCbyBT-Regular /KristenITC-Regular /KrutiDev040Bold /KrutiDev040BoldItalic /KrutiDev040Condensed /KrutiDev040Italic /KrutiDev040Thin /KrutiDev040Wide /KrutiDev060 /KrutiDev060Bold /KrutiDev060BoldItalic /KrutiDev060Condensed /KrutiDev060Italic /KrutiDev060Thin /KrutiDev060Wide /KrutiDev070 /KrutiDev070Condensed /KrutiDev070Italic /KrutiDev070Thin /KrutiDev070Wide /KrutiDev080 /KrutiDev080Condensed /KrutiDev080Italic /KrutiDev080Wide /KrutiDev090 /KrutiDev090Bold /KrutiDev090BoldItalic /KrutiDev090Condensed /KrutiDev090Italic /KrutiDev090Thin /KrutiDev090Wide /KrutiDev100 /KrutiDev100Bold /KrutiDev100BoldItalic /KrutiDev100Condensed /KrutiDev100Italic /KrutiDev100Thin /KrutiDev100Wide /KrutiDev120 /KrutiDev120Condensed /KrutiDev120Thin /KrutiDev120Wide /KrutiDev130 /KrutiDev130Condensed /KrutiDev130Thin /KrutiDev130Wide /KunstlerScript /Latha /LatinWide /LetterGothic /LetterGothic-Bold /LetterGothic-BoldOblique /LetterGothic-BoldSlanted /LetterGothicMT /LetterGothicMT-Bold /LetterGothicMT-BoldOblique /LetterGothicMT-Oblique /LetterGothic-Slanted /LevenimMT /LevenimMTBold /LilyUPC /LilyUPCBold /LilyUPCBoldItalic /LilyUPCItalic /Lithos-Black /Lithos-Regular /LotusWPBox-Roman /LotusWPIcon-Roman /LotusWPIntA-Roman /LotusWPIntB-Roman /LotusWPType-Roman /LucidaBright /LucidaBright-Demi /LucidaBright-DemiItalic /LucidaBright-Italic /LucidaCalligraphy-Italic /LucidaConsole /LucidaFax /LucidaFax-Demi /LucidaFax-DemiItalic /LucidaFax-Italic /LucidaHandwriting-Italic /LucidaSans /LucidaSans-Demi /LucidaSans-DemiItalic /LucidaSans-Italic /LucidaSans-Typewriter /LucidaSans-TypewriterBold /LucidaSans-TypewriterBoldOblique /LucidaSans-TypewriterOblique /LucidaSansUnicode /Lydian /Magneto-Bold /MaiandraGD-Regular /Mangal-Regular /Map-Symbols /MathA /MathB /MathC /Mathematica1 /Mathematica1-Bold /Mathematica1Mono /Mathematica1Mono-Bold /Mathematica2 /Mathematica2-Bold /Mathematica2Mono /Mathematica2Mono-Bold /Mathematica3 /Mathematica3-Bold /Mathematica3Mono /Mathematica3Mono-Bold /Mathematica4 /Mathematica4-Bold /Mathematica4Mono /Mathematica4Mono-Bold /Mathematica5 /Mathematica5-Bold /Mathematica5Mono /Mathematica5Mono-Bold /Mathematica6 /Mathematica6Bold /Mathematica6Mono /Mathematica6MonoBold /Mathematica7 /Mathematica7Bold /Mathematica7Mono /Mathematica7MonoBold /MatisseITC-Regular /MaturaMTScriptCapitals /Mesquite /Mezz-Black /Mezz-Regular /MICR /MicrosoftSansSerif /MingLiU /Minion-BoldCondensed /Minion-BoldCondensedItalic /Minion-Condensed /Minion-CondensedItalic /Minion-Ornaments /MinionPro-Bold /MinionPro-BoldIt /MinionPro-It /MinionPro-Regular /Miriam /MiriamFixed /MiriamTransparent /Mistral /Modern-Regular /MonotypeCorsiva /MonotypeSorts /MSAM10 /MSAM5 /MSAM6 /MSAM7 /MSAM8 /MSAM9 /MSBM10 /MSBM5 /MSBM6 /MSBM7 /MSBM8 /MSBM9 /MS-Gothic /MSHei /MSLineDrawPSMT /MS-Mincho /MSOutlook /MS-PGothic /MS-PMincho /MSReference1 /MSReference2 /MSReferenceSansSerif /MSReferenceSansSerif-Bold /MSReferenceSansSerif-BoldItalic /MSReferenceSansSerif-Italic /MSReferenceSerif /MSReferenceSerif-Bold /MSReferenceSerif-BoldItalic /MSReferenceSerif-Italic /MSReferenceSpecialty /MSSong /MS-UIGothic /MT-Extra /MTExtraTiger /MT-Symbol /MT-Symbol-Italic /MVBoli /Myriad-Bold /Myriad-BoldItalic /Myriad-Italic /Myriad-Roman /Narkisim /NewCenturySchlbk-Bold /NewCenturySchlbk-BoldItalic /NewCenturySchlbk-Italic /NewCenturySchlbk-Roman /NewMilleniumSchlbk-BoldItalicSH /NewsGothic /NewsGothic-Bold /NewsGothicBT-Bold /NewsGothicBT-BoldItalic /NewsGothicBT-Italic /NewsGothicBT-Roman /NewsGothic-Condensed /NewsGothic-Italic /NewsGothicMT /NewsGothicMT-Bold /NewsGothicMT-Italic /NiagaraEngraved-Reg /NiagaraSolid-Reg /NimbusMonL-Bold /NimbusMonL-BoldObli /NimbusMonL-Regu /NimbusMonL-ReguObli /NimbusRomNo9L-Medi /NimbusRomNo9L-MediItal /NimbusRomNo9L-Regu /NimbusRomNo9L-ReguItal /NimbusSanL-Bold /NimbusSanL-BoldCond /NimbusSanL-BoldCondItal /NimbusSanL-BoldItal /NimbusSanL-Regu /NimbusSanL-ReguCond /NimbusSanL-ReguCondItal /NimbusSanL-ReguItal /Nimrod /Nimrod-Bold /Nimrod-BoldItalic /Nimrod-Italic /NSimSun /Nueva-BoldExtended /Nueva-BoldExtendedItalic /Nueva-Italic /Nueva-Roman /NuptialScript /OCRA /OCRA-Alternate /OCRAExtended /OCRB /OCRB-Alternate /OfficinaSans-Bold /OfficinaSans-BoldItalic /OfficinaSans-Book /OfficinaSans-BookItalic /OfficinaSerif-Bold /OfficinaSerif-BoldItalic /OfficinaSerif-Book /OfficinaSerif-BookItalic /OldEnglishTextMT /Onyx /OnyxBT-Regular /OzHandicraftBT-Roman /PalaceScriptMT /Palatino-Bold /Palatino-BoldItalic /Palatino-Italic /PalatinoLinotype-Bold /PalatinoLinotype-BoldItalic /PalatinoLinotype-Italic /PalatinoLinotype-Roman /Palatino-Roman /PapyrusPlain /Papyrus-Regular /Parchment-Regular /Parisian /ParkAvenue /Penumbra-SemiboldFlare /Penumbra-SemiboldSans /Penumbra-SemiboldSerif /PepitaMT /Perpetua /Perpetua-Bold /Perpetua-BoldItalic /Perpetua-Italic /PerpetuaTitlingMT-Bold /PerpetuaTitlingMT-Light /PhotinaCasualBlack /Playbill /PMingLiU /Poetica-SuppOrnaments /PoorRichard-Regular /PopplLaudatio-Italic /PopplLaudatio-Medium /PopplLaudatio-MediumItalic /PopplLaudatio-Regular /PrestigeElite /Pristina-Regular /PTBarnumBT-Regular /Raavi /RageItalic /Ravie /RefSpecialty /Ribbon131BT-Bold /Rockwell /Rockwell-Bold /Rockwell-BoldItalic /Rockwell-Condensed /Rockwell-CondensedBold /Rockwell-ExtraBold /Rockwell-Italic /Rockwell-Light /Rockwell-LightItalic /Rod /RodTransparent /RunicMT-Condensed /Sanvito-Light /Sanvito-Roman /ScriptC /ScriptMTBold /SegoeUI /SegoeUI-Bold /SegoeUI-BoldItalic /SegoeUI-Italic /Serpentine-BoldOblique /ShelleyVolanteBT-Regular /ShowcardGothic-Reg /Shruti /SILDoulosIPA /SimHei /SimSun /SimSun-PUA /SnapITC-Regular /StandardSymL /Stencil /StoneSans /StoneSans-Bold /StoneSans-BoldItalic /StoneSans-Italic /StoneSans-Semibold /StoneSans-SemiboldItalic /Stop /Swiss721BT-BlackExtended /Sylfaen /Symbol /SymbolMT /SymbolTiger /SymbolTigerExpert /Tahoma /Tahoma-Bold /Tci1 /Tci1Bold /Tci1BoldItalic /Tci1Italic /Tci2 /Tci2Bold /Tci2BoldItalic /Tci2Italic /Tci3 /Tci3Bold /Tci3BoldItalic /Tci3Italic /Tci4 /Tci4Bold /Tci4BoldItalic /Tci4Italic /TechnicalItalic /TechnicalPlain /Tekton /Tekton-Bold /TektonMM /Tempo-HeavyCondensed /Tempo-HeavyCondensedItalic /TempusSansITC /Tiger /TigerExpert /Times-Bold /Times-BoldItalic /Times-BoldItalicOsF /Times-BoldSC /Times-ExtraBold /Times-Italic /Times-ItalicOsF /TimesNewRomanMT-ExtraBold /TimesNewRomanPS-BoldItalicMT /TimesNewRomanPS-BoldMT /TimesNewRomanPS-ItalicMT /TimesNewRomanPSMT /Times-Roman /Times-RomanSC /Trajan-Bold /Trebuchet-BoldItalic /TrebuchetMS /TrebuchetMS-Bold /TrebuchetMS-Italic /Tunga-Regular /TwCenMT-Bold /TwCenMT-BoldItalic /TwCenMT-Condensed /TwCenMT-CondensedBold /TwCenMT-CondensedExtraBold /TwCenMT-CondensedMedium /TwCenMT-Italic /TwCenMT-Regular /Univers-Bold /Univers-BoldItalic /UniversCondensed-Bold /UniversCondensed-BoldItalic /UniversCondensed-Medium /UniversCondensed-MediumItalic /Univers-Medium /Univers-MediumItalic /URWBookmanL-DemiBold /URWBookmanL-DemiBoldItal /URWBookmanL-Ligh /URWBookmanL-LighItal /URWChanceryL-MediItal /URWGothicL-Book /URWGothicL-BookObli /URWGothicL-Demi /URWGothicL-DemiObli /URWPalladioL-Bold /URWPalladioL-BoldItal /URWPalladioL-Ital /URWPalladioL-Roma /USPSBarCode /VAGRounded-Black /VAGRounded-Bold /VAGRounded-Light /VAGRounded-Thin /Verdana /Verdana-Bold /Verdana-BoldItalic /Verdana-Italic /VerdanaRef /VinerHandITC /Viva-BoldExtraExtended /Vivaldii /Viva-LightCondensed /Viva-Regular /VladimirScript /Vrinda /Webdings /Westminster /Willow /Wingdings2 /Wingdings3 /Wingdings-Regular /WNCYB10 /WNCYI10 /WNCYR10 /WNCYSC10 /WNCYSS10 /WoodtypeOrnaments-One /WoodtypeOrnaments-Two /WP-ArabicScriptSihafa /WP-ArabicSihafa</s>
<s>/WP-BoxDrawing /WP-CyrillicA /WP-CyrillicB /WP-GreekCentury /WP-GreekCourier /WP-GreekHelve /WP-HebrewDavid /WP-IconicSymbolsA /WP-IconicSymbolsB /WP-Japanese /WP-MathA /WP-MathB /WP-MathExtendedA /WP-MathExtendedB /WP-MultinationalAHelve /WP-MultinationalARoman /WP-MultinationalBCourier /WP-MultinationalBHelve /WP-MultinationalBRoman /WP-MultinationalCourier /WP-Phonetic /WPTypographicSymbols /XYATIP10 /XYBSQL10 /XYBTIP10 /XYCIRC10 /XYCMAT10 /XYCMBT10 /XYDASH10 /XYEUAT10 /XYEUBT10 /ZapfChancery-MediumItalic /ZapfDingbats /ZapfHumanist601BT-Bold /ZapfHumanist601BT-BoldItalic /ZapfHumanist601BT-Demi /ZapfHumanist601BT-DemiItalic /ZapfHumanist601BT-Italic /ZapfHumanist601BT-Roman /ZWAdobeF /NeverEmbed [ true /AntiAliasColorImages false /CropColorImages true /ColorImageMinResolution 150 /ColorImageMinResolutionPolicy /OK /DownsampleColorImages true /ColorImageDownsampleType /Bicubic /ColorImageResolution 300 /ColorImageDepth -1 /ColorImageMinDownsampleDepth 1 /ColorImageDownsampleThreshold 2.00333 /EncodeColorImages true /ColorImageFilter /DCTEncode /AutoFilterColorImages true /ColorImageAutoFilterStrategy /JPEG /ColorACSImageDict << /QFactor 0.76 /HSamples [2 1 1 2] /VSamples [2 1 1 2] /ColorImageDict << /QFactor 0.76 /HSamples [2 1 1 2] /VSamples [2 1 1 2] /JPEG2000ColorACSImageDict << /TileWidth 256 /TileHeight 256 /Quality 15 /JPEG2000ColorImageDict << /TileWidth 256 /TileHeight 256 /Quality 15 /AntiAliasGrayImages false /CropGrayImages true /GrayImageMinResolution 150 /GrayImageMinResolutionPolicy /OK /DownsampleGrayImages true /GrayImageDownsampleType /Bicubic /GrayImageResolution 300 /GrayImageDepth -1 /GrayImageMinDownsampleDepth 2 /GrayImageDownsampleThreshold 2.00333 /EncodeGrayImages true /GrayImageFilter /DCTEncode /AutoFilterGrayImages true /GrayImageAutoFilterStrategy /JPEG /GrayACSImageDict << /QFactor 0.76 /HSamples [2 1 1 2] /VSamples [2 1 1 2] /GrayImageDict << /QFactor 0.76 /HSamples [2 1 1 2] /VSamples [2 1 1 2] /JPEG2000GrayACSImageDict << /TileWidth 256 /TileHeight 256 /Quality 15 /JPEG2000GrayImageDict << /TileWidth 256 /TileHeight 256 /Quality 15 /AntiAliasMonoImages false /CropMonoImages true /MonoImageMinResolution 1200 /MonoImageMinResolutionPolicy /OK /DownsampleMonoImages true /MonoImageDownsampleType /Bicubic /MonoImageResolution 600 /MonoImageDepth -1 /MonoImageDownsampleThreshold 1.00167 /EncodeMonoImages true /MonoImageFilter /CCITTFaxEncode /MonoImageDict << /K -1 /AllowPSXObjects false /CheckCompliance [ /None /PDFX1aCheck false /PDFX3Check false /PDFXCompliantPDFOnly false /PDFXNoTrimBoxError true /PDFXTrimBoxToMediaBoxOffset [ 0.00000 0.00000 0.00000 0.00000 /PDFXSetBleedBoxToMediaBox true /PDFXBleedBoxToTrimBoxOffset [ 0.00000 0.00000 0.00000 0.00000 /PDFXOutputIntentProfile (None) /PDFXOutputConditionIdentifier () /PDFXOutputCondition () /PDFXRegistryName () /PDFXTrapped /False /CreateJDFFile false /Description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f7f75288fd94e9b8bbe5b9a521b5efa7684002000410064006f006200650020005000440046002065876863900275284e8e55464e1a65876863768467e5770b548c62535370300260a853ef4ee54f7f75280020004100630072006f0062006100740020548c002000410064006f00620065002000520065006100640065007200200035002e003000204ee553ca66f49ad87248672c676562535f00521b5efa768400200050004400460020658768633002> /CHT <FEFF4f7f752890194e9b8a2d7f6e5efa7acb7684002000410064006f006200650020005000440046002065874ef69069752865bc666e901a554652d965874ef6768467e5770b548c52175370300260a853ef4ee54f7f75280020004100630072006f0062006100740020548c002000410064006f00620065002000520065006100640065007200200035002e003000204ee553ca66f49ad87248672c4f86958b555f5df25efa7acb76840020005000440046002065874ef63002> /CZE <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> /DAN <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> /DEU <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> /ESP <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> /FRA <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> /GRE <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a stvaranje Adobe PDF dokumenata pogodnih za pouzdani prikaz i ispis poslovnih dokumenata koristite ove postavke. Stvoreni PDF dokumenti mogu se otvoriti Acrobat i Adobe Reader 5.0 i kasnijim verzijama.) /HUN <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> /ITA (Utilizzare queste impostazioni per creare documenti Adobe PDF adatti per visualizzare e stampare documenti aziendali in modo affidabile. I documenti PDF creati possono essere aperti con Acrobat e Adobe Reader 5.0 e versioni successive.) /JPN <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> /KOR <FEFFc7740020c124c815c7440020c0acc6a9d558c5ec0020be44c988b2c8c2a40020bb38c11cb97c0020c548c815c801c73cb85c0020bcf4ace00020c778c1c4d558b2940020b3700020ac00c7a50020c801d569d55c002000410064006f0062006500200050004400460020bb38c11cb97c0020c791c131d569b2c8b2e4002e0020c774b807ac8c0020c791c131b41c00200050004400460020bb38c11cb2940020004100630072006f0062006100740020bc0f002000410064006f00620065002000520065006100640065007200200035002e00300020c774c0c1c5d0c11c0020c5f40020c2180020c788c2b5b2c8b2e4002e> /NLD (Gebruik deze instellingen om Adobe PDF-documenten te maken waarmee zakelijke documenten betrouwbaar kunnen worden weergegeven en afgedrukt. De gemaakte PDF-documenten kunnen worden geopend met Acrobat en Adobe Reader 5.0 en hoger.) /NOR <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> /POL <FEFF0055007300740061007700690065006e0069006100200064006f002000740077006f0072007a0065006e0069006100200064006f006b0075006d0065006e007400f300770020005000440046002000700072007a0065007a006e00610063007a006f006e00790063006800200064006f0020006e00690065007a00610077006f0064006e00650067006f002000770079015b0077006900650074006c0061006e00690061002000690020006400720075006b006f00770061006e0069006100200064006f006b0075006d0065006e007400f300770020006600690072006d006f0077007900630068002e002000200044006f006b0075006d0065006e0074007900200050004400460020006d006f017c006e00610020006f007400770069006500720061010700200077002000700072006f006700720061006d006900650020004100630072006f00620061007400200069002000410064006f00620065002000520065006100640065007200200035002e0030002000690020006e006f00770073007a0079006d002e> /PTB <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> /RUM <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> /RUS <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> /SLV <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> /SUO <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> /SVE <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> /TUR <FEFF005400690063006100720069002000620065006c00670065006c006500720069006e0020006700fc00760065006e0069006c0069007200200062006900720020015f0065006b0069006c006400650020006700f6007200fc006e007400fc006c0065006e006d006500730069002000760065002000790061007a0064013100720131006c006d006100730131006e006100200075007900670075006e002000410064006f006200650020005000440046002000620065006c00670065006c0065007200690020006f006c0075015f007400750072006d0061006b0020006900e70069006e00200062007500200061007900610072006c0061007201310020006b0075006c006c0061006e0131006e002e00200020004f006c0075015f0074007500720075006c0061006e0020005000440046002000620065006c00670065006c0065007200690020004100630072006f006200610074002000760065002000410064006f00620065002000520065006100640065007200200035002e003000200076006500200073006f006e0072006100730131006e00640061006b00690020007300fc007200fc006d006c00650072006c00650020006100e70131006c006100620069006c00690072002e> /ENU (Use these settings to create Adobe PDF documents suitable for reliable viewing and printing of business documents. Created PDF documents can be opened with Acrobat and Adobe Reader 5.0 and later.)>> setdistillerparams /HWResolution [600 600] /PageSize [612.000 792.000]>> setpagedevice</s>
<s>Sentiment Extraction From Text Using Emotion Tagged CorpusSentiment Extraction From Text Using EmotionTagged CorpusThasina TabashumComputer Science and EngineeringAmerican International University Bangladeshthasinatabashumabonti@gmail.comAbdul Mutalab ShaykatComputer Science and EngineeringAmerican International University Bangladeshshaykat2057@gmail.comSheikh AbujarComputer Science and EngineeringDaffodil International Universitysheikh.cse@diu.edu.bdMd MohibullahDepartment of CSEComilla Universitymohib.cse.bd@gmail.comShekhor ChandaComputer Science and EngineeringAmerican International University-Bangladeshshekhor.chanda333@gmail.comAbstract—Emotion tagging aims to tag words with appropriateemotion that a word expresses in a document or a sentence.In this paper, authors introduced an approach to tag wordsautomatically and then analyzed in sentence and document level.The evaluation of tagged words described and the performancebased on changing the tone of a document has been explained.The experiment showed that the approach could be effective if thetag words are not constant or static. Application of this approachin a contextual level analysis is briefly concluded and proposedtentative emotional tree by using emotion tagging in word level.Keywords— Sentiment Analysis, Natural Language Pro-cessing, Emotion TaggingI. INTRODUCTIONEmotion is a feeling from one’s circumstances which is notconstant and varies from place to place, time to time or personto person so determining emotion of a document or an articleis not straightforward process. Sentiment classification of atext is quite difficult task. This imponderable research haslots of applicable and practical uses in the field of NaturalLanguage Processing. Analysis of opinion, events, news arti-cles, topics, posts on social media provide us a vast problemspace. Sentiment analysis is divided into three distinct levels: -document level, sentence level and aspect level[1]. Documentlevel classification is classifying overall text for example: -review of service or product. Sentence level classification isclassifying a sentence opinion. In this study word tagging withemotion polarity will be approached. Assigning emotion tagsto the words with Ekaman’s (1993) six basic emotions- joy,sadness, anger, fear, surprise, disgust along with a distancemeasurement. The goal with this approach is to allow oneto easily and accurately analyze these sentiments by taggingwords with emotions. In this study, a tagged dataset is created.However, all the words could not be included therefore asystem has been created to tag the untagged words. A dynamicsystem has been proposed to label the unlabeled words witha correct emotion that word expresses. The objective of thisstudy is to tag words according to the emotion it expresses witha distance measurement from actual standard emotion level.By distance measurement the intensity of that emotion can beconcluded. In section 2, literature reviews has been discussed.Resource collection and formation is described in section 3,followed by process of data analysis and evaluation process inSection 4 and 5 respectively. Emotional tree is proposed withemotion tagged word in section 6. Finally, the last sectionincludes a discussion of the future scope of this paper.II. LITERATURE REVIEWIn [2] retrieved English sentiWordNet and translated it toBangla for emotion tagging word at sentence level. For trainingtenacity here used CRF framework and some topographieshave been nominated for six emotions classification. In [3] theauthor introduced co-training for the sentiment classificationof multiple languages. The author used English dataset toperform classification in Chinese. Explanation of importanceof adjectives and verbs for opinion mining has been introducedin [4]. In recent past, the domain of NLP for academics hasbeen broadening. Online</s>
<s>product recommendation system [5],determining the subjective adjectives [6], topic detection andidentifying the pilot of the study [7] are focused on text pro-cessing. As the importance of opinion mining is increasing dayby day in many sectors of modern life, new methods are beingintroduced and researchers are improving the existing meth-ods. Researchers are mostly focused on general solution for allkind of sentences in a document. Classifications of sentencesbefore go for sentiment analysis can improve performanceand accuracy [8]. Sentiment analysis is not like other textmining researches, it depends on many parameters like context,background of the context, sentence pattern etc. and that’swhy accurate results from sentiment analysis is infrequent.To extract attribute from document NaÃŕve Bayes classifierhas been used in [9]. Authors simplified a data set modelfor sentiment analysis in Bangla text and Randomized Banglatext in [10] , therefore many works have been completed onphrase level sentiment analysis. In NLP it is very importantto identify the subjective parts of the document in order toidentify the context, which is essential to determine. In [11]the author conferred the identification of subjective sentencesof a document to determine its subject using SVM classifier.IEEE - 4567010th ICCCNT 2019 July 6-8, 2019, IIT - Kanpur Kanpur, IndiaA chat scheme [12] is offered that customs enthusiastic textwith expressive material. Here is an Arrangement to assembleinformation using some physiological sensors which are at-tached to the user’s body. Context is major term in sentimentanalysis and to detect contextual polarity in [13] it comes withan approach which has the capability to inevitably recognizecontextual polarity. Manual footnote of contextual polarityand an inter-annotator covenant is premeditated in this paper[14]. Contextual valence is tricky thing in sentiment analysisespecially in Bangla text. It is used here for sentiment analysisfrom Bangla text. WordNet is a strongly structured structurethat contains words according to their relations with each otherand contains explanations and examples in it. It is easy toget the sense of each word conferring to its parts of speechand SentiWordNet has been castoff to recognize polarity.Summarize collection of similar opinion has been presentedand Graph-based technique is used for summarization and thestatement is combined using sentiment analysis [15]. Data isevaluated on the collection of 250 English news text leveledwith sentiment from sentiment voting system. There are someavailable part of speech tagger Stanford NLP parser is one ofthem and it is castoff to gain the part of speech tagging [16].III. RESOURCE COLLECTION AND FORMATIONResources are collected and separated in to six basic emo-tions. Six emotion categories- joy, sadness, anger, fear, disgustand surprise have been labeled. In this study, wordnet effectemotion list for tagged data has been used and all the taggedwords are included in one dataset according to the tags. Totalnumber of tagged words in dataset is 1135. Labeled wordsinclude 400 "Joy" , 148 "Fear" , 202 "sadness", 261 "anger" ,53 " disgust" and "71" surprise In the table 1 tag names andtable 2 some tagged data from dataset have been shown:TABLE IEMOTION TAGSEmotion TagJoy JYSadness SDAnger ANFear FRDisgust DGSurprise SPTABLE IITAGGED DATAWords Tag Words TagAngry AG trepidation FRAngered AG timidity FREnraged AG</s>
<s>satisfaction JYFurious AG rejoicing JYDistasteful DG regard JYDisgustful DG amaze SPUnassertiveness FR astonish SPIV. DATA ANALYSISIn this section, our objective is to label the unlabeledwords. To achieve this, the words were tagged on both anautomatic and manual level. Therefore this method of emotiontagging ultimately gives the accurate result for all the level ofsentiment analysis. The aim is to accurately tag words withthe correct emotions using this method for opinion mining.Fig. 1. Work flow of proposed emotional word tagging systemA. Automatic Word TaggingMachine has given an article about Hospital incident [17].Comparing with the tagged dataset only 7 words are taggedof an article of length 6531 characters, and consisting of 1246words. Those 7 words were already existed in the taggeddataset.[(’heart’, ’JY’), (’bad’, ’SD’), (’suffering’, ’SD’), (’sur-prise’, ’SP’), (’get’, ’SP’), (’bad’, ’SD’), (’contentment’,’JY’)]. Since it is not possible to obtain the sentiment of thatdocument more emotion tagged words are essential to perceivethe intended result. After compiling the article through themodule, it separated the untagged words. These unlabeledwords needed to be labeled however they must be furtheranalyzed in order to be accurately tagged with the rightemotions. All the untagged words are compared with thetagged words. For dynamical approach Wu-Palmer metric(WUP) has been applied, WUP weights the edges based ondistance in the hierarchy for the unlabeled word from all otherlabeled words. In this paper polarity has been assigned for eachemotion. The polarity calculation is not constant. It varies withthe tagged dataset. The larger the dataset, the more accuratethe result is. Polarity equation: -P (x =′ JY ′,′ SP ′,′ SD′,′ FR′,′ DG′,′ AG′) =Max(L)L(x)Max(L) = maximum length , L(x)=length of x. Here,maximum length of tagged emotion is "JY’ which is 400. If,x=’FR’; Length of x = 148; P (x=’FR’) = 400/148 =2.703In this manner all the emotion polarities have been assigned.Six emotion percentage of a word calculated with equation (2).P.E(w, x) =i S(i, w) ∗ pix∑∑ji S(i, w) ∗ pj∗ 100[s >= 0.5]...(2)In this equation, ’j’ is from joy, fear, disgust, sadness toanger. ’P’ is polarity. Next,’s’ is the WUP similarity betweenIEEE - 4567010th ICCCNT 2019 July 6-8, 2019, IIT - Kanpur Kanpur, IndiaTABLE IIIEMOTION TAGS WITH POLARITYJoy 1.0Surprise 5.633Sad 1.980Fear 2.703Disgust 7.547Anger 1.532untagged word and the tagged data and s is greater than 0.5.n= number of total tagged data (1135) and x can be any ofthe six emotions. P.E of some words is shown:P.E ("punishment", ’JY’) = 22.026P.E ("punishment", ’SP’) = 12.428P.E ("punishment", ’SD’) = 4.114P.E ("punishment", ’FR’) = 15.117P.E ("punishment", ’DG’) = 35.700P.E ("punishment", ’AG’) = 10.616P.E ("punishment", ’JY’) = 22.026P.E ("hospital", ’SP’) = 34.5478P.E ("hospital", ’SD’) = 11.4684P.E ("hospital", ’FR’) = 0.0P.E ("hospital", ’DG’) = 0.0P.E ("hospital", ’AG’) = 35.78067P.E ("punishment", ’JY’) = 22.026P.E ("punishment", ’SP’) = 12.428P.E ("punishment", ’SD’) = 4.114P.E ("punishment", ’FR’) = 15.117P.E ("punishment", ’DG’) = 35.700P.E ("punishment", ’AG’) = 10.616P.E ("hospital", ’JY’) = 18.2031P.E ("hospital", ’SP’) = 34.5478P.E ("hospital", ’SD’) = 11.4684P.E ("hospital", ’FR’) = 0.0P.E ("hospital", ’DG’) = 0.0P.E ("hospital", ’AG’) = 35.78067Through this process, the P.E. of 240 words was calculated.These 240 words were each tagged</s>
<s>with the maximum P.Evalue that was found for a single word. Taking maximum P.E"punishment" has been tagged as "DG". The total number ofwords tagged summed up to 247.TABLE IVNUMBER OF WORDS WITH EMOTION TAGEmotion Tag JY SP SD FR DG AGNumber of Words 16 33 77 12 64 43B. Non-Automatic Word TaggingThe same article is provided to readers to manually tag thewords. For this purpose a system has been developed whichwill provide the words and the users will select the emotion.Around 17-19 readers gave the feedback. Average values aretaken then maximum valued emotion is assigned to the word.Total number of manual tagged words is 311, which ismore than dynamically tagged data because the latter searchesTABLE VTAGGED WORDS BY READERSWords Joy Surprise Sad Fear Disgust Angercareless 0 0 25.5 18.5 15 41childcare 38.34 1.05 2.66 42 10 5.95parents 29.9 20 22 13.1 0 15.1baby 27 22.5 3 2 0 7girl 18.33 0.67 15.01 8.9 5 21.3Ctg 13.3 0 0 0 0 0hospital 8.2 23.9 15.8 33.8 0.3 18TABLE VIMANUALLY TAGGED EMOTIONEmotion Tag JY SP SD FR DG AGNumber of words 28 65 76 26 63 53for similarity above 0.5 and derives its own emotion. As thedataset was not substantial, it may not find many words whichare above the aforementioned mark thus many words might notbe taken into the calculation.C. Merging Dynamic and Manual TagsThe automatic and manual tags were then compared. Thewords that had the same tag in both cases are inserted intothe main dataset. The main objective of this is to createan enriched emotion tagged dataset. Manual tagging will nolonger be necessary once the dataset is large enough and thesystem will be capable of automatically analyze accurately.Since many words were not being tagged initially, manualtagging in the small scale was important. Moreover, themanually tagged words are used for cross check the resultof the dynamic tagging process. Taking set of automaticallyand manually tagged words, 96 words were tagged the same.These tagged data was added to the main dataset.Fig. 2. Merge emotional words from Auto tagged dataset and manually taggeddatasetV. EVALUATIONEvaluation has done in word level and sentence level. Auto-matic tagging is compared with manual tagging for accuracy.IEEE - 4567010th ICCCNT 2019 July 6-8, 2019, IIT - Kanpur Kanpur, IndiaIn this section, it has been discussed about the issues for thechange in accuracy.A. Word Level EvaluationIn order to measure the accuracy, a confusion matrix wasdeveloped comparing manual and auto tags. The accuracy wasfound to be at 82.75/100. Updated tagged dataset includes -406.0 "JY", 157.0 "SP", 224.0 "SD", 276.0 "FR", 76.0 "DG",89.0 "AG". The table 7 is a confusion matrix of an articleabout child bus accident which gave very good results. But ifthis learning approach is provided with a positive article aboutthe ’Prime minister’ [18] the accuracy level was 56.67/100.Correctly tagged words were only 34 among 60. Since thesystem had initially been provided with a negatively tonedarticle, it learned to tag words with negative emotions morethan positive.TABLE VIICONFUSION MATRIXJY SP SD FR DG AGJY 8 2 1 0 2 0SP 4 9 2</s>
<s>0 2 1SD 0 0 23 0 1 0FR 0 0 1 15 2 0DG 0 0 0 0 23 0AG 1 0 0 0 2 18Therefore, when given an article with a positive toneafterward, it was unable to accurately determine the overallsentiment of the article. This proved that word level evaluationis not sufficient, and the analysis must be done on a contextuallevel as well.B. Sentence Level EvaluationSentence level evaluation was considered as the ultimateaim of word tagging is to determine the opinion or tone of asentence or an article, eventually. Three tagged sentences areshown below:"With a heavy heart, Rokhsana Akter, 21, began makingpreparations for the funeral of her five-day-old girl, also herfirst child." (s.1)TABLE VIIINUMBER OF WORDS TAGGED UNDER EACH TAG IN S1Tag JY SP SD FR DG AGWords "heart" – "funeral","girl","child"– – –NumberWords1 0 3 0 0 0"The authorities allegedly warned the mother and her familymembers not to unwrap the shroud as the baby’s face was inbad shape due to excessive bleeding." (s.2)"The authorities helped the mother with the situation, herbaby is fine now" (s.3)This approach will not always yield accurate results. It wasused to accurately tag s.1 and s.2, but was unable to do theTABLE IXNUMBER OF WORDS TAGGED UNDER EACH TAG IN S2Tag JY SP SD FR DG AGWords – "Unwrap","shroud""baby","shape","bad"– "Authorities","family""face"NumberWords0 2 3 0 2 1TABLE XNUMBER OF WORDS TAGGED UNDER EACH TAG IN S3Tag JY SP SD FR DG AGWords "fine" – "baby" – "Authorities" "situation"NumberWords1 0 1 0 1 1same for s.3. This learning method gave accurate result fornegative toned article because the machine learned the wordsfrom a bitter article titled "Careless Childcare" from Daily star.Thus, when happy sentences are provided to the machine, itsprediction is not as accurate. Statically tagging the words withemotion cannot give the actually emotion that word expressesin a particular sentence or article, therefore a context level isneeded to have an exact result.VI. CONTEXTUAL LEVEL APPLICATIONIn this section, a contextual level analysis using emotiontagged word is proposed. Contextual methodology for futurestudy will involve the method used in this paper to createa dynamic, rather than static, tree corpus of tagged words,where the weight of the words and the distance between themis taken into consideration. Rather than static emotion tags, thedataset will learn according to the nodes around it. TaggingS.2 , "authorities" tagged as "DG", but tagging manual s.3authorities should no longer be tagged as "DG", but rather"JY". Creating sentence level tree will be very helpful toanalyze sentence sentiment in the future. As an example aftercreating two trees, if the machine is provided with anothersentence:Fig. 3. Tagged words in the sentenceHere, tagging mother as "FR" is inaccurate. It should bechanged. It doesn’t necessarily express any emotion in thissentence. Thus there is abundant number of neutral words.If more sentences are provided about mother, for example: -mother loves her child, mother is crying, the tag for motherIEEE - 4567010th ICCCNT 2019 July 6-8, 2019, IIT - Kanpur Kanpur, Indiais changing every time. These kinds of words, which arefrequently changing shall be tagged as</s>
<s>’Neutral’ The proposedmethod would create a tree where only parts of speech tags-noun, adjective, verb, adverb shall be considered. If pos =Nounthat won’t create another child branch under it. Like in figure2 under "Unwrap", there are "shroud", "face", "baby". Sincethey are nouns, a child is not created until the word "was"is found.Proposed methodology is, initially creating a corpusof sentence level tree and emotion tagged words which willconsider its branches, weight, distance and polarity in orderto calculate the emotion tag. Below is a demonstration of theproposed algorithm for dynamic emotion word tagging:parts_of_speech_words = pos_tagallowed_types = [”JJ”, ”JJR”, ”JJS”, ”NN”,”NNS”, ”RB”, RBR”, ”RBS”,”V BD”, ”V BN”, ”V BS”]3) Creating tree:a) Root(first word of the sentence)b) Parent=rootc) For p in allowed_types:i) If p is not noun and is not parent:create edge from parent to pparent=pii) else if (parent-1) is not root then create edgefrom all(parent-1) to p4) Search the created tree on the emotional tree5) If found then tag the words6) Else if found some branches of the tree, then computeequation (1) and (2) for untagged words and the sis WUP similarity between the words of surroundedbranches and the polarity will also be calculated withthose branches. Weight will be assigned with the eachof neighboring words considering the distance betweenthem. So equation (2)P.E(w, x) =i S(i, w) ∗ pix ∗ dis(i, w)∑∑ji S(i, w) ∗ pj ∗ dis(i, w)∗ 100[s >= 0.5] ...........(3)7) Else if no match found then approach will be the oneused in this paper8) Create set of each word store in emotion trackingdataset:a) wordEmotionTrack[key]=[[eos_tag][P.E(key,eos_tag)]]b) totalOccurence[key]+=totalOccurence[key]9) for w in all_wordsWth =len(set(wordEmotionTracj[w]))totalOccurence[w]10) if Wth>0.5:word_tag[w]= "neu"In step 10 , Wth is word threshold which is measuring thevariance of each word. If any word tag is changing frequentlythat word will added under neutral tag. This paper study actualpurpose is for noise free sentiment analysis by tagging wordswith emotion which is only be possible if the analysis is incontextual level. Contextual level analysis is more accuratecompared to word frequency analysis.A. Evaluation word level tagging with dynamic tree leveltaggingInitially tagged sentence "The authorities allegedly warnedthe mother and her family members not to unwrap the shroudas the baby’s face was in bad Shape due to excessive bleeding."Another sentence "The authorities helped the mother withthe situation, her baby is fine now", for word level taggingusing equation (1) and (2) the result has shown table 9 wherebaby tagged as ’SD’ which is wrong after close investigation.Contextual tree has shown in figure 4 which is built on thebasis of word level tagging. But in this new sentence babyis not conveying sadness. In the new sentence "baby" shouldbe tagged as a positive toned sentiment rather than negativetoned sentiment. On the other hand, for the previous sentencesadness it is accurate for the "baby" word.Fig. 4. Initial/Data Tree Using Word Level TaggingApplying the contextual tree algorithm for word taggingwith emotion gives more satisfying result. Using equation 4to measure the distance to "baby" node to surround nodes hasbeen shown in figure 5. In this tree, it will travel choosing thepath of immediate parent nodes. After</s>
<s>reaching the startingnode "Authorities", it travels the existing data tree withoutvisiting same node again. Distance is measured with belowequation:dis =...(4), En = numberofedgetraveled.IEEE - 4567010th ICCCNT 2019 July 6-8, 2019, IIT - Kanpur Kanpur, IndiaFig. 5. Distance form a node to other nodes in contextual treeUsing Equation 3, the word "baby" tag has been updated to"SP".P.E(Baby,SP) = 82.5263 , P.E(Baby,FR)=44.23The contextual level analysis given the better result which isdynamic, emotion tagging should be dependent on the contextrather than individual words. Article [18] words extracted tagswhich were previously tagged incorrect after using contextualanalysis tags are updated shown in table 11 :TABLE XICOMPARISON BETWEEN WORD LEVEL TAGGING AND CONTEXTUAL LEVELTAGGINGmeeting JY DG SDPrince SP SD SDVictory JY DG SPcriticism SP SD SPCurrents tags are comparatively more accurate than theprevious ones. Considering as a word how it is contributing tothe sentence to expressing an emotion is giving valid result.Applying the contextual tree level learning on [18] currentaccurate tags number is 39 among 60. But other inaccuratetags as example "victory" tagged as "SP", which make moresense than "DG". This learning method will learn more errorfreely if its data tree is large enough to estimate the actualemotion of a word in a sentence or passage.VII. CONCLUSIONThe limitation of this paper was in the depth of the analysis.The word tagging also should be made dynamic instead ofstatic so the system is better able to conclude the contextualmeaning rather than go by previous learning. This way, thedatabase will be better enriched and would not depend on itsinitial learning curve. Since the system was too dependent onthe meaning of the initially tagged words we were unable toaccurately determine the overall sentiment of an oppositelytoned article supplied to it afterwards. In conclusion, as thedocuments were evaluated according to word level sentimentbut this alone can’t give actual sentiment analysis. Thuscontextual tree level approach is needed along with the taggedwords to better evaluate this. Using this approach of labelingemotion on word level will give better and faster result.ACKNOWLEDGMENTWe thank Daffodil International University Machine Learn-ing and NLP lab for their support through-out the research.REFERENCES[1] Liu, B. (2014). Sentiment Analysis and Opinion Mining. Disserta-tion Abstracts International, B: Sciences and Engineering (Vol. 70).https://doi.org/10.1162/COLI[2] Das, D. (2009). Word to Sentence Level Emotion Tagging for BengaliBlogs. American Psychologist, (August), 149âĂŞ152.[3] Wan, X. (2009). Co-Training for Cross-Lingual Sentiment Classification,(August), 235âĂŞ243.[4] Chesley, P., Vincent, B., Xu, L., Srihari, R. (2006). Using verbs andadjectives to automatically classify blog sentiment. Training, 9(4), 8.https://doi.org/10.1016/0020-0271(73)90092-2[5] Chai, J., Horvath, V., Nicolov, N., Stys, M., Kambhatla, N., Zadrozny,W., Melville, P. (2002). Natural Language Assistant: A Dialog Sys-tem for Online Product Recommendation. AI Magazine, 23(2), 63.https://doi.org/10.1609/aimag.v23i2.1641[6] Baroni, M., Vegnaduzzo, S. (2004). Identifying subjective adjec-tives through web-based mutual information. Proceedings of KON-VENS, (2000), 17âĂŞ24. Retrieved from http://sslmit.unibo.it/ ba-roni/publications/konvens2004/wmiKONV.pdf[7] Allan, J., Carbonell, J., Doddington, G., Yamron, J., Yang, Y. (1998).Topic Detection and Tracking Pilot Study: Final Report. Proceedings ofthe DARPA Broadcast News Transcription and Understanding Workshop,194âĂŞ218.[8] Chen, T., Xu, R., He, Y., Wang, X. (2017). Improving senti-ment analysis via sentence type classification using BiLSTM-CRFand CNN. Expert Systems</s>
<s>with Applications, 72, 221âĂŞ230.https://doi.org/10.1016/j.eswa.2016.10.065[9] Hasan, K. M. A., Sabuj, M. S., and Afrin, Z. (2015). Opinion Min-ing using Naive Bayes. 2015 IEEE International WIE Conference onElectrical and Computer Engineering (WIECON-ECE), 511âĂŞ514.https://doi.org/10.1109/WIECON-ECE.2015.7443981[10] Wilson, T., Wiebe, J., and Hoffmann, P. (2005). Recognizing con-textual polarity in phrase-level sentiment analysis. Proceedings of theConference on Human Language Technology and Empirical Methodsin Natural Language Processing - HLT ’05, (October), 347âĂŞ354.https://doi.org/10.3115/1220575.1220619[11] Das, a., and Bandyopadhyay, S. (2010). Opinion-Polarity Identi-fication in Bengali. In ICCPOL 2010, (October). Retrieved fromhttp://www.amitavadas.com/Pub/ICCPOL 2010.pdf[12] Hasan, K. M. A., Rahman, M., and Badiuzzaman. (2014). Sentimentdetection from Bangla text using contextual valency analysis. 2014 17thInternational Conference on Computer and Information Technology, IC-CIT 2014, 292âĂŞ295. https://doi.org/10.1109/ICCITechn.2014.7073151[13] Asif Hasan, Mohammad Rashedul Amin, Abul Kalam Al Azad, NabeelMohammad, Sentiment Analysis on Bangla and Romanized Bangla Textusing Deep Recurrent Models. 2016 International Workshop on Compu-tational Intelligence (IWCI)12-13 December 2016, Dhaka, Bangladesh[14] Wang, H., Prendinger, H., Igarashi, T. (2004). Communicating emotionsin online chat using physiological sensors and animated text. ExtendedAbstracts of the 2004 Conference on Human Factors and ComputingSystems - CHI ’04, 1171. https://doi.org/10.1145/985921.986016[15] Bhargava, R., Sharma, Y., Sharma, G. (2016). ATSSI: Abstractive TextSummarization Using Sentiment Infusion. Procedia Computer Science, 89,404âĂŞ411. https://doi.org/10.1016/j.procs.2016.06.088[16] Hui, J. L. O., Hoon, G. K., Zainon, W. M. N. W. (2017).Effects of Word Class and Text Position in Sentiment-basedNews Classification. Procedia Computer Science, 124, 77âĂŞ85.https://doi.org/10.1016/j.procs.2017.12.132[17] https://www.thedailystar.net/frontpage/careless-childcare-1564540[18] https://www.thedailystar.net/backpage/hasina-makes-times-list-100-most-influential-people-1565233IEEE - 4567010th ICCCNT 2019 July 6-8, 2019, IIT - Kanpur Kanpur, India /ASCII85EncodePages false /AllowTransparency false /AutoPositionEPSFiles false /AutoRotatePages /None /Binding /Left /CalGrayProfile (Gray Gamma 2.2) /CalRGBProfile (sRGB IEC61966-2.1) /CalCMYKProfile (U.S. Web Coated \050SWOP\051 v2) /sRGBProfile (sRGB IEC61966-2.1) /CannotEmbedFontPolicy /Warning /CompatibilityLevel 1.4 /CompressObjects /Off /CompressPages true /ConvertImagesToIndexed true /PassThroughJPEGImages true /CreateJobTicket false /DefaultRenderingIntent /Default /DetectBlends true /DetectCurves 0.0000 /ColorConversionStrategy /LeaveColorUnchanged /DoThumbnails false /EmbedAllFonts true /EmbedOpenType false /ParseICCProfilesInComments true /EmbedJobOptions true /DSCReportingLevel 0 /EmitDSCWarnings false /EndPage -1 /ImageMemory 1048576 /LockDistillerParams true /MaxSubsetPct 100 /Optimize false /OPM 0 /ParseDSCComments false /ParseDSCCommentsForDocInfo false /PreserveCopyPage true /PreserveDICMYKValues true /PreserveEPSInfo false /PreserveFlatness true /PreserveHalftoneInfo true /PreserveOPIComments false /PreserveOverprintSettings true /StartPage 1 /SubsetFonts false /TransferFunctionInfo /Remove /UCRandBGInfo /Preserve /UsePrologue false /ColorSettingsFile () /AlwaysEmbed [ true /Arial-Black /Arial-BoldItalicMT /Arial-BoldMT /Arial-ItalicMT /ArialMT /ArialNarrow /ArialNarrow-Bold /ArialNarrow-BoldItalic /ArialNarrow-Italic /ArialUnicodeMS /BookAntiqua /BookAntiqua-Bold /BookAntiqua-BoldItalic /BookAntiqua-Italic /BookmanOldStyle /BookmanOldStyle-Bold /BookmanOldStyle-BoldItalic /BookmanOldStyle-Italic /BookshelfSymbolSeven /Century /CenturyGothic /CenturyGothic-Bold /CenturyGothic-BoldItalic /CenturyGothic-Italic /CenturySchoolbook /CenturySchoolbook-Bold /CenturySchoolbook-BoldItalic /CenturySchoolbook-Italic /ComicSansMS /ComicSansMS-Bold /CourierNewPS-BoldItalicMT /CourierNewPS-BoldMT /CourierNewPS-ItalicMT /CourierNewPSMT /EstrangeloEdessa /FranklinGothic-Medium /FranklinGothic-MediumItalic /Garamond /Garamond-Bold /Garamond-Italic /Gautami /Georgia /Georgia-Bold /Georgia-BoldItalic /Georgia-Italic /Haettenschweiler /Impact /Kartika /Latha /LetterGothicMT /LetterGothicMT-Bold /LetterGothicMT-BoldOblique /LetterGothicMT-Oblique /LucidaConsole /LucidaSans /LucidaSans-Demi /LucidaSans-DemiItalic /LucidaSans-Italic /LucidaSansUnicode /Mangal-Regular /MicrosoftSansSerif /MonotypeCorsiva /MSReferenceSansSerif /MSReferenceSpecialty /MVBoli /PalatinoLinotype-Bold /PalatinoLinotype-BoldItalic /PalatinoLinotype-Italic /PalatinoLinotype-Roman /Raavi /Shruti /Sylfaen /SymbolMT /Tahoma /Tahoma-Bold /TimesNewRomanMT-ExtraBold /TimesNewRomanPS-BoldItalicMT /TimesNewRomanPS-BoldMT /TimesNewRomanPS-ItalicMT /TimesNewRomanPSMT /Trebuchet-BoldItalic /TrebuchetMS /TrebuchetMS-Bold /TrebuchetMS-Italic /Tunga-Regular /Verdana /Verdana-Bold /Verdana-BoldItalic /Verdana-Italic /Vrinda /Webdings /Wingdings2 /Wingdings3 /Wingdings-Regular /ZWAdobeF /NeverEmbed [ true /AntiAliasColorImages false /CropColorImages true /ColorImageMinResolution 200 /ColorImageMinResolutionPolicy /OK /DownsampleColorImages true /ColorImageDownsampleType /Bicubic /ColorImageResolution 300 /ColorImageDepth -1 /ColorImageMinDownsampleDepth 1 /ColorImageDownsampleThreshold 1.50000 /EncodeColorImages true /ColorImageFilter /DCTEncode /AutoFilterColorImages false /ColorImageAutoFilterStrategy /JPEG /ColorACSImageDict << /QFactor 0.76 /HSamples [2 1 1 2] /VSamples [2 1 1 2] /ColorImageDict << /QFactor 0.76 /HSamples [2 1 1 2] /VSamples [2 1 1 2] /JPEG2000ColorACSImageDict << /TileWidth 256 /TileHeight 256</s>
<s>/Quality 15 /JPEG2000ColorImageDict << /TileWidth 256 /TileHeight 256 /Quality 15 /AntiAliasGrayImages false /CropGrayImages true /GrayImageMinResolution 200 /GrayImageMinResolutionPolicy /OK /DownsampleGrayImages true /GrayImageDownsampleType /Bicubic /GrayImageResolution 300 /GrayImageDepth -1 /GrayImageMinDownsampleDepth 2 /GrayImageDownsampleThreshold 1.50000 /EncodeGrayImages true /GrayImageFilter /DCTEncode /AutoFilterGrayImages false /GrayImageAutoFilterStrategy /JPEG /GrayACSImageDict << /QFactor 0.76 /HSamples [2 1 1 2] /VSamples [2 1 1 2] /GrayImageDict << /QFactor 0.76 /HSamples [2 1 1 2] /VSamples [2 1 1 2] /JPEG2000GrayACSImageDict << /TileWidth 256 /TileHeight 256 /Quality 15 /JPEG2000GrayImageDict << /TileWidth 256 /TileHeight 256 /Quality 15 /AntiAliasMonoImages false /CropMonoImages true /MonoImageMinResolution 400 /MonoImageMinResolutionPolicy /OK /DownsampleMonoImages true /MonoImageDownsampleType /Bicubic /MonoImageResolution 600 /MonoImageDepth -1 /MonoImageDownsampleThreshold 1.50000 /EncodeMonoImages true /MonoImageFilter /CCITTFaxEncode /MonoImageDict << /K -1 /AllowPSXObjects false /CheckCompliance [ /None /PDFX1aCheck false /PDFX3Check false /PDFXCompliantPDFOnly false /PDFXNoTrimBoxError true /PDFXTrimBoxToMediaBoxOffset [ 0.00000 0.00000 0.00000 0.00000 /PDFXSetBleedBoxToMediaBox true /PDFXBleedBoxToTrimBoxOffset [ 0.00000 0.00000 0.00000 0.00000 /PDFXOutputIntentProfile (None) /PDFXOutputConditionIdentifier () /PDFXOutputCondition () /PDFXRegistryName () /PDFXTrapped /False /CreateJDFFile false /Description << /CHS <FEFF4f7f75288fd94e9b8bbe5b9a521b5efa7684002000410064006f006200650020005000440046002065876863900275284e8e55464e1a65876863768467e5770b548c62535370300260a853ef4ee54f7f75280020004100630072006f0062006100740020548c002000410064006f00620065002000520065006100640065007200200035002e003000204ee553ca66f49ad87248672c676562535f00521b5efa768400200050004400460020658768633002> /CHT <FEFF4f7f752890194e9b8a2d7f6e5efa7acb7684002000410064006f006200650020005000440046002065874ef69069752865bc666e901a554652d965874ef6768467e5770b548c52175370300260a853ef4ee54f7f75280020004100630072006f0062006100740020548c002000410064006f00620065002000520065006100640065007200200035002e003000204ee553ca66f49ad87248672c4f86958b555f5df25efa7acb76840020005000440046002065874ef63002> /DAN <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> /DEU <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> /ESP <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> /FRA <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> /ITA (Utilizzare queste impostazioni per creare documenti Adobe PDF adatti per visualizzare e stampare documenti aziendali in modo affidabile. I documenti PDF creati possono essere aperti con Acrobat e Adobe Reader 5.0 e versioni successive.) /JPN <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> /KOR <FEFFc7740020c124c815c7440020c0acc6a9d558c5ec0020be44c988b2c8c2a40020bb38c11cb97c0020c548c815c801c73cb85c0020bcf4ace00020c778c1c4d558b2940020b3700020ac00c7a50020c801d569d55c002000410064006f0062006500200050004400460020bb38c11cb97c0020c791c131d569b2c8b2e4002e0020c774b807ac8c0020c791c131b41c00200050004400460020bb38c11cb2940020004100630072006f0062006100740020bc0f002000410064006f00620065002000520065006100640065007200200035002e00300020c774c0c1c5d0c11c0020c5f40020c2180020c788c2b5b2c8b2e4002e> /NLD (Gebruik deze instellingen om Adobe PDF-documenten te maken waarmee zakelijke documenten betrouwbaar kunnen worden weergegeven en afgedrukt. De gemaakte PDF-documenten kunnen worden geopend met Acrobat en Adobe Reader 5.0 en hoger.) /NOR <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> /PTB <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> /SUO <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> /SVE <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> /ENU (Use these settings to create PDFs that match the "Required" settings for PDF Specification 4.01)>> setdistillerparams /HWResolution [600 600] /PageSize [612.000 792.000]>> setpagedevice</s>
<s>Sentiment Analysis on Bengali Text using Lexicon Based Approach978-1-7281-5842-6/19/$31.00 ©2019 IEEE 2019 22nd International Conference on Computer and Information Technology (ICCIT), 18-20 December, 2019 Sentiment Analysis on Bengali Text using Lexicon Based Approach Abstract—In this modern era, we daily involve in the internet strongly. We express our opinion about products, services, books, movies, songs, politics, sports, organizations, etc. through the internet in social media, blogs, micro-blogging websites or any media. Public opinion with Bengali text in internet media is increasing very rapidly. Due to a few works in Bengali text sentiment analysis, it has become an important issue of extracting opinions, emotions from Bengali textual data through Sentiment Analysis (SA) for better knowledge extraction. Sentiment Analysis (SA) is effectively used for classifying the opinion expressed in a text according to its polarity (e.g., positive, negative or neutral). This paper represents a lexicon dictionary-based approach for polarity detection of Bengali text data. We compared our proposed model with machine learning classifiers such as Decision Tree (DT), Naive Bayes (NB) and Support Vector Machine (SVM) classifiers and it works as a much better accurate model for Bengali text polarity detection. Keywords—Natural Language Processing, Sentiment Analysis, Lexicon Based Approach, Polarity Detection, Bengali Text I. INTRODUCTION Internet users around the world are increasing very quickly. In their day to day life, they express their reviews/opinions through the internet. Manual analysis of textual reviews/opinions is required too much time. For a huge amount of textual data analysis, it is quite impossible. For proper knowledge extraction from this huge amount of textual data, we need to build a system that can be done through Sentiment Analysis (SA), which plays a great role as a Natural Language Processing (NLP) tool to detect the polarity of the textual data. Sentiment Analysis (SA) or Opinion Mining is a way of detecting the polarity that is expressed in written text or through textual data [1]. In another way, Sentiment analysis indicates to analyze the public’s real opinions, emotions, sentiments, attitudes, and evaluations about various issues and services [2]. For commercial organizations, consumer’s feedback about particular products, and services are very much important. Business organizations can improve their product quality [3] and services depending on customer opinions/reviews about their products and services. In Twitter, Facebook, Blogs or any internet sites where a large number of users use it to give their opinions/reviews to a specific topic. According to opinions/reviews, sentiments can be estimated through analysis. Two popular approaches that help very much in sentiment analysis. The first process is lexicon-based weighted words of textual data and the second process is based on machine learning approaches. Lexicon based approach uses a dictionary of sentiment words with a sentiment score and for finding the polarity of a text line, token words are matched with this dictionary. Bengali is a very popular language around the world. Every day more than 261 million people of the world are using Bengali and it is the primary language in Bangladesh [4] where it has huge importance and</s>
<s>influence for businesses as well as Government offices. To avoid unwanted circumstances, it is important for the Government to understand the public sentiments as well as be updated about political agendas. In this paper, we have developed a polarity detection system on Bengali textual opinions/reviews like as products, services, books, movies, songs, sports, politics, organizations using Python and lexicon dictionary-based approach where we have collected 5200 opinions/reviews by crawling and manually from some internet sites and social media. Data is preprocessed for getting clean and in a suitable format. Tokenization, punctuation removing, stop words removing & stemming is used during preprocessing. With the help of the Bengali sentiment words dictionary, boost word check, negation check & normalization of the score, we finally achieved sentiment results. We also have evaluated the system efficiency and performance by comparing with well-known machine learning classifiers. The evaluation of our developed system has provided with the acceptability of 92% accuracy as well as better precision, recall & F1-score. The remaining of this paper is organized in the following way: related works in Sentiment Analysis in Section II. Our methodology for Bengali text Sentiment Analysis is given in Section III. In Section IV, presents Results and discussion. Finally, Section V is for the conclusion of this paper. II. RELATED WORK Due to a lack of tools in Sentiment Analysis (SA) on Bengali Text, this field has become a growing issue for research. Most of the sentiment tools developed for English. Few researchers work to develop Sentiment Analysis (SA) tools for Bengali Text polarity detection. Document-level Sentiment Analysis has been performed in [5]. For polarity detection in Bengali language texts using WordNet [6] and SentiWordNet [7], the authors proposed valency analysis. In [8] authors translate 16000 Amazon watch reviews from English into Bangla using Google Translator for polarity detection on product reviews for both Bengali and English. Polarities are divided into weak, steady Orvila Sarker School of Computer Science The University of Adelaide Adelaide, Australia orvila.sarker@gmail.com Rajib Chandra Dey Dept. of Information & Communication Technology Comilla University Comilla - 3506, Bangladesh rajibict14@gmail.com Authorized licensed use limited to: University of Durham. Downloaded on May 16,2020 at 09:39:11 UTC from IEEE Xplore. Restrictions apply. and strong where they achieve 85% accuracy for Bengali. Paper [9] proposes a human sentiment analysis model using the lexicon-based approach on any given data set. A lexicon-based approach has been proposed for sentiment analysis of news articles [10]. For sentiment analysis, a dataset from BBC, comprising of news articles between the years 2004 and 2005 has been used. They used the WordNet dictionary for positive, negative or neutral classification. The paper [11] demonstrates a sentiment analysis model trained using TF-IDF and they also used lexicon-based features to analyze the sentiments expressed by students in their textual feedback. A semi-supervised bootstrapping approach for polarity detection is used in [12]. They separate text line polarity as positive and negative. Sentiment analysis was done at the sentence level [13]. A dynamic dictionary with predefined positive and negative</s>
<s>words was used to find the sentiment polarity of the sentence. 91% accurate results were achieved for the classification of news articles. In [14] authors proposed for the unigram presence method with negation handling and stemming. This proposed system obtained average accuracy but showed low performance. A polarity detection system is shown in [15] on Bengali movie reviews. They mainly used two machine learning approaches Support Vector Machines and Naive Bayes. Using a lexicon-based approach in [16] they show how cricket fan’s sentiments vary over the period of time and they collected the tweets over Cricket tournament for sentiment analysis. Sentiment analysis has been performed in [17] for product review using the lexicon method. They have used product reviews from Twitter using twitter API. III. METHODOLOGY The steps of our proposed Bengali text polarity detection system are shown in Fig. 1. We have developed a polarity detection system that can detect the polarity of the Bengali Text. A lexicon Bengali words dictionary is used in our system. Our dictionary contains sentiment words with a sentiment score. We have performed tokenization, punctuation remove, stop words remove and stemming. We calculate each word polarity of the text with the help of our lexicon dictionary and also checked boost word & negation. After that, the total sentiment score of the text is calculated and finally, we got sentiment results through normalization of the total sentiment score. The whole process of our developed system is described in the following way: A. Data Collection We have collected opinions/reviews of Bengali Text by web crawling and manually from some sites and social media such as Facebook groups, Twitter, Blogs, online newspaper sites. This data set contains opinions/reviews about products, services, books, songs, organizations, sports, movies and politics. We have collected 5200 sentences where 2600 sentences are positive and 2600 sentences are negative reviews/opinions. B. Preprocessing of Text Collected text data from several sites and social media may be contained some unwanted or noisy information. For getting clean and suitable data, we have to process the collected web data. It is a very important step in sentiment analysis for obtaining accurate sentiment results. Fig. 1. Step by step process of proposed model. Our steps of preprocessing are in the following way: 1) Tokenization: Tokenization is the process of splitting a text line into individual components such as words, phrases or symbols. Each component is known as a token. We can see the following example: A book’s review: eকাtেরর িদন িল বiিট খুব ভােলা eবং aেনক পছেnর !!! ভােলা লাগেল আপনারা aনলাiেন িকনেত পােরন। After tokenization, this review text appeared as ‘eকাtেরর’, ‘িদন িল’, ‘বiিট’, ‘খুব’, ‘ভােলা’, ‘eবং’, ‘aেনক’, ‘পছেnর’, ‘!!!’, ‘ভােলা’, ‘লাগেল’, ‘আপিনo’, ‘aনলাiেন’, ‘িকনেত’, ‘পােরন’, ‘।’ 2) Punctuation Remove: In Bengali text, it contains several punctuations and it has very little influence in sentiment analysis. To get clear text, we can simply remove those punctuations. After removing punctuations, review text appeared as: ‘eকাtেরর’, ‘িদন িল’, ‘বiিট’, ‘খুব’, ‘ভােলা’, ‘eবং’, ‘aেনক’, ‘পছেnর’, ‘ভােলা’, ‘লাগেল’, ‘আপিনo’, ‘aনলাiেন’, ‘িকনেত’,</s>
<s>‘পােরন’ 3) Stop words Remove: Stop words are used to complete a sentence but it has no importance in sentiment analysis. Normally in sentiment analysis, stop words are removed [18]. Boost Word CheckNegation Check Total Sentiment Score of Text Calculation Bengali Text (Reviews/Opinions) Bengali Sentiment Words Dictionary Polarity of Sentiment Words Calculation Sentiment Results Preprocessing of Text Tokenization Punctuation Remove Stop Words Remove Stemming Normalization Authorized licensed use limited to: University of Durham. Downloaded on May 16,2020 at 09:39:11 UTC from IEEE Xplore. Restrictions apply. In Bengali Text, there are many stop words such as ‘আমার’, ‘আিম’, ‘আপিন’, ‘আপিনo’, ‘হয়’, ‘হেব’, ‘পাির’, ‘পােরন’, ‘কের’, ‘কেরন’, ‘থােক’, ‘থােকন’, ‘eবং’ etc. We have created a Bengali stop words list. During the stop words remove process, we have checked our token with the stop words list. If a match is found then we removed it from the token list of review/opinions. After removing stop words, the above token list appeared as ‘eকাtেরর’, ‘িদন িল’, ‘বiিট’, ‘খুব’, ‘ভােলা’, ‘aেনক’, ‘পছেnর’, ‘ভােলা’, ‘লাগেল’, ‘aনলাiেন’, ‘িকনেত’ 4) Stemming: Stemming is used in natural language processing to change a word into its original form or stem. Then each word can easily be recognized from the lexicon dictionary. We have checked each word for stemming. If a word contains such as (‘◌া’, ‘ি◌’, ‘ে◌’, ‘◌ী’, ‘i’, ‘য়’, ‘িট’, ‘টা’) at the end, then the word is used for stemming. We checked stem word with lexicon dictionary where our lexicon dictionary contains only sentiment words with a sentiment score. If a match occurs, then we change the word from the token list with the stemming word. Otherwise, the word remains unchanged. After stemming, the above token list appeared as ‘eকাtেরর’, ‘িদন িল’, ‘বiিট’, ‘খুব’, ‘ভাল’, ‘aেনক’, ‘পছn’, ‘ভাল’, ‘লাগেল’, ‘aনলাiেন’, ‘িকনেত’ and we can say it our final token list. Here, [‘ভােলা’, ‘পছেnর’] they are sentiment words. Then, they are changed with stemming words [‘ভাল’, ‘পছn’]. But [‘eকাtেরর’, ‘িদন িল’, ‘বiিট’, ‘লাগেল’, ‘aনলাiেন’, ‘িকনেত’] they are not sentiment words and that’s why they remain unchanged. C. Polarity of Sentiment Words Calculation For calculating polarity of sentiment words, we have to be familiar with the following terms: 1) Bengali Sentiment Words Dictionary: We have developed a lexicon dictionary1 of Bangla sentiment words. Our lexicon dictionary contains more than 5100 words. We assigned a sentiment score to each word. This sentiment score is given in the range of -4 to +4 where +4 is for strong positivity and -4 is for strong negativity. Sentiment score with zero (0) indicates neutral. 2) Boost Word Check: In our proposed system, we specified some words which are used in Bengali text to boost its next words in a text line. According to the presence of boost word in a text, polarity of text will be increased or decreased. We have created a list of boost words such as ‘al’, ‘aিধক’, ‘েবিশ’, ‘সবেচেয়’, ‘aেনক’, ‘খুব’ etc. For example, ‘আiেফান aেনক ভােলা েমাবাiল ।’. Here, ‘aেনক’ is a boosted word according to our system. It boosts</s>
<s>the next word ‘ভােলা’ and ‘ভােলা’ is a positive sentiment word. Then ‘aেনক’ increases the positivity of the text line. 3) Negation Check: Naturally, in Bengali text, negation words are placed at the end of the text line. We have developed a list of Bengali negation words such as ‘নয়’, ‘েনi’, ‘না’, ‘নাi’, ‘িন’. If any negation word is found in a text line, the score of the sentiment word is multiplied by -1 which means that negative sentiment will be positive or positive sentiment will be negative. For example, ‘শাoিম েমাবাiল খারাপ না’. Though ‘খারাপ’ is a negative sentiment word, this sentence is predicted as positive by our system due to the presence of negation word ‘না'. This sentence is an actual positive sentence. If negation & boost word both are found in the text line, then the score of sentiment word is increased or decreased according to the boosted word. After the preprocessing of text, it is a very important task to calculate the polarity of each sentiment word from the final token list. We compared each final token word with our Bengali sentiment words dictionary. For each token word, a separate sentiment score is assigned according to our lexicon dictionary. If any token did not contain any sentiment word and then its token sentiment score value is assigned to zero (0). The final token list is also checked for boost word & negation word. If a match has occurred, the sentiment of the token list will be increased or decreased according to the boost word & negation word. Finally, we got a separate sentiment score for each token of the final token list. D. Total Sentiment Score of Text Calculation Total sentiment score of a text is calculated combining all final token sentiment score of the text line. In our lexicon Bengali dictionary, sentiment word score ranges from -4 to 4. It needs to normalize the total sentiment score for a text line. Using normalization2, we constructed our sentiment results within range -1 to +1 where -1 is for strong negativity and +1 is for strong positivity. 1) Normalization: If the total sentiment score of a text line is denoted as T-Score and normalization parameter is denoted as alpha where alpha = 15, Then normalize score equation can be given as follows: Normalize Score = ( ) (1) Normalize the total sentiment score between -1 and 1 and alpha approximates the max expected value. E. Sentiment Results Our main purpose is to analyze reviews/opinions about such as products, services, sports, books, songs, organizations, and politics. We want to analyze which reviews/opinions are positive and which reviews/opinions are negative. After normalization, we can easily detect which are positive reviews/opinions and which are negative reviews/opinions. For a text line, if normalization score is greater than zero and less than or equal to 1 (0 < Normalize score <= 1), then the text line is predicted as positive. Else if normalization score is less than zero and greater than or</s>
<s>equal to -1 (-1 <= Normalize score < 0), then the text line is predicted as negative. Otherwise, the text is neutral. Some predicted reviews/opinions by our proposed system are shown in Table I. 1https://github.com/Fighter-1/Programming-/tree/master/Dictionary 2https://www.nltk.org/_modules/nltk/sentiment/vader.html#normalize Authorized licensed use limited to: University of Durham. Downloaded on May 16,2020 at 09:39:11 UTC from IEEE Xplore. Restrictions apply. TABLE I. SOME PREDICTED REVIEWS/OPINIONS BY OUR PROPOSED SYSTEM Reviews/Opinions Normalized Score Predicted Result eকাtেরর িদন িল বiিট খুব ভােলা eবং aেনক পছেnর !!! ভােলা লাগেল আপিনo aনলাiেন িকনেত পােরন । 0.92 Positive শাoিম েমাবাiল খুব খারাপ ...!! -0.73 Negative শাoমী েমাবাiেল kিতকর িবিকরণ eকটু েবিশi হয় । -0.53 Negative বাংলােদশ খুব ভােলা িkেকট েখেলেছ ......... 0.60 Positive জেলর গান বয্াn aেনক ভােলা গান কের 0.59 Positive IV. RESULTS AND DISCUSSION The following performance measurement techniques are used to evaluate our proposed polarity detection system: Accuracy, Precision, Recall, and F1-score. Our data set contains 2600 positive reviews/opinions and 2600 negative reviews/opinions. We also have used popular machine learning classifiers such as Decision Tree (DT), Naive Bayes (NB) and Support Vector Machine (SVM) classifiers for our data set. We used these machine learning classifiers using Scikit-learn3. Scikit-learn is an open-source and very popular machine learning Python library. We compared our proposed system with machine learning classifiers in terms of Accuracy, Precision, Recall, F1-score and achieved better acceptable Accuracy, Precision, Recall & F1-score than machine learning classifiers. Accuracy indicates the percentage of text lines or sentences in the test set that the classifier correctly labels. True Positive, True Negative, False Positive and False Negative is denoted as TP, TN, FP, and FN. Accuracy = ∗ 100% (2) Precision indicates the number of text lines or sentences in the test set that is correctly labeled by the classifier from the total text lines or sentences in the test set that is classified by the classifier for a particular class. Precision = (3) Recall indicates the number of text lines or sentences in the test set that is correctly labeled by the classifier from the total lines or sentences in the test set that are actually labeled for a particular class. Recall = (4) F1-score is the weighted harmonic mean of Precision and Recall for a particular class. F1-score = ∗ ∗ (5) A scatter plot is shown in Fig. 2. indicates the predicted positive and negative sentences of data set by our proposed system. The blue portion is for positive sentences and the red portion is for negative sentences. All positive sentences are remains in the range 0 to 1 (without including 0). All negative sentences are remains in the range -1 to 0 (without including 0). In the figure, we can see that a large amount of positive data are remains in range (0.30 to 1) and a large amount of negative data are remains in range (-1 to -.33). Fig. 2. Scatter plot for positive and negative predicted sentences by our proposed system. TABLE II. ACCURACY COMPARISON OF OUR PROPOSED SYSTEM AND MACHINE LEARNING CLASSIFIERS.</s>
<s>Fig. 3. Visualization of accuracy comparison in bar diagram. TABLE III. PERFORMANCE COMPARISON OF OUR PROPOSED SYSTEM AND MACHINE LEARNING CLASSIFIERS. 3https://scikit-learn.org Method Accuracy (%) Proposed System 92% Decision Tree Classifier 87% Naive Bayes Classifier 86% Support Vector Machine Classifier 88% Method Precision Recall F1-score Proposed System 0.91 0.94 0.92 Decision Tree Classifier 0.83 0.94 0.88 Naive Bayes Classifier 0.82 0.93 0.87 Support Vector Machine Classifier 0.84 0.94 0.89 Authorized licensed use limited to: University of Durham. Downloaded on May 16,2020 at 09:39:11 UTC from IEEE Xplore. Restrictions apply. Fig. 4. Visualization of performance comparison in bar diagram. We observed that our proposed model for polarity detection in Bengali text is better than supervised machine learning classifiers and accuracy, precision, recall & f1-score of our proposed technique is much superior. Our Bengali text polarity detection system is mostly depending on the Bengali lexicon sentiment word dictionary. A lack of sentiment words in the dictionary may change the results. Increasing proper sentiment words in the lexicon dictionary, very much accurate results can be achieved by our polarity detection system. V. CONCLUSION Our main purpose of this work is to develop a Bengali text polarity detection system for sentiment analysis. We have developed a polarity detection system using a lexicon dictionary-based approach. Our system is mostly depending on the Bengali sentiment words dictionary and we developed this dictionary much properly where it contains more than 5100 sentiment words. Due to the lack of standard data set in Bangla, we have collected Bengali reviews/opinions from sites and social media. Data is preprocessed for getting clean and in a suitable format. Tokenization, punctuation removing, stop words removing & stemming is used during preprocessing. With the help of the Bengali sentiment words dictionary, boost word check, negation check & normalization of the score, we finally achieved sentiment results. We also evaluated our system efficiency and performances and also compared with machine learning classifiers. We observed that our Bengali text polarity detection system is much better with 92% accuracy as well as better precision, recall & F1-score. In the future, we will enrich our sentiment Bengali words dictionary with more sentiment words for achieving much better accuracy. REFERENCES [1] J. Reis, P. Olmo, F. Benevenuto, H. Kwak, R. Prates, and J. An, “Breaking the news: first impressions matter on online news,” in International AAAI Conference on Web and Social Media(ICWSM), 2015. [2] O. Kolchyna, T.T. P. Souza and P. C. Treleaven and T. Aste, “Twitter Sentiment Analysis: Lexicon Method, Machine Learning Method and Their Combination,” arXiv:1507.00955v3 [cs.CL], Sep 2015. [3] B. Liu, “Sentiment analysis and opinion mining,” Synthesis lectures on human language technologies, vol. 5, 2012, pp. 1–167. [4] N. Tabassum and M. I. Khan, “Design an Empirical Framework for Sentiment Analysis from Bangla Text using Machine Learning,” in Proc. IEEE International Conference on Electrical, Computer and Communication Engineering (ECCE), 2019. [5] K. M. A. Hasan, M. Rahman and Badiuzzaman, “Sentiment detection from Bangla text using contextual valency analysis,” in Proc. IEEE 17th International Conference on</s>
<s>Computer and Information Technology (ICCIT), 2014, pp. 292–295. [6] T. S. Utomo, R. Sarno and Suhariyanto, “Emotion Label from ANEW Dataset for Searching Best Definition from Wordnet,” IEEE International Seminar on Application for Technology of Information and Communication (ISEMANTIC), 2018, pp. 249-252. [7] A. Agarwal, V. Sharma, G. Sikka and R. Dhir, “Opinion mining of news headlines using SentiWordNet,” in Proc. IEEE Symposium on Colossal Data Analysis and Networking (CDAN), 2016. [8] K. M. A. Hasan, M. S. Sabuj and Z. Afrin, “Opinion mining using Naive Bayes,” in Proc. IEEE International WIE Conference on Electrical and Computer Engineering (WIECON-ECE), 2015, pp. 511–514. [9] V.Singh, G. Singh, P. Rastogi and D. Deswal, “Sentiment Analysis Using Lexicon Based Approach,” IEEE International Conference on Parallel, Distributed and Grid Computing (PDGC), 2018, pp. 13-18. [10] S. Taj, B. B. Shaikh and A. F. Meghji, “Sentiment Analysis of News Articles: A Lexicon based Approach,” International Conference on Computing, Mathematics and Engineering Technologies (iCoMET), 2019. [11] Z. Nasim, Q. Rajput and S. Haider, “Sentiment Analysis of student feedback using machine learning and lexicon based approaches,” IEEE International Conference on Research and Innovation in Information Systems (ICRIIS), 2017. [12] S. Chowdhury and W. Chowdhury, “Performing sentiment analysis in Bangla microblog posts,” International Conference on Informatics, Electronics & Vision (ICIEV), 2014. [13] M. U. Islam, F. B. Ashraf, A. I. Abir and M. A. Mottalib, “Polarity detection of online news articles based on sentence structure and dynamic dictionary,” in Proc. IEEE 20th International Conference on Computer and Information Technology (ICCIT), Dhaka, 2017. [14] A. Kaur and V. Gupta, “Proposed algorithm of sentiment analysis for Punjabi text,” Journal of Emerging Technologies in Web Intelligence, vol. 6, no. 2, 2014, pp. 180-183. [15] N. Banik and H. H. Rahman, “Evaluation of Naïve Bayes and Support Vector Machines on Bangla textual movie reviews” IEEE International Conference on Bangla Speech and Language Processing(ICBSLP), 2018. [16] A. Agarwal and D. Toshniwal, “Application of Lexicon Based Approach in Sentiment Analysis for short Tweets,” in Proc. IEEE International Conference on Advances in Computing and Communication Engineering (ICACCE), 2018, pp. 189-193. [17] P. Ray and A. Chakrabarti, “Twitter sentiment analysis for product review using lexicon method,” International Conference on Data Management, Analytics and Innovation (ICDMAI), 2017, pp. 211-216. [18] J. Leskovec, A. Rajaraman, and J. D. Ullman, “Mining of massive datasets,” Cambridge University Press, 2014. Authorized licensed use limited to: University of Durham. Downloaded on May 16,2020 at 09:39:11 UTC from IEEE Xplore. Restrictions apply. /ASCII85EncodePages false /AllowTransparency false /AutoPositionEPSFiles true /AutoRotatePages /None /Binding /Left /CalGrayProfile (Gray Gamma 2.2) /CalRGBProfile (sRGB IEC61966-2.1) /CalCMYKProfile (U.S. Web Coated \050SWOP\051 v2) /sRGBProfile (sRGB IEC61966-2.1) /CannotEmbedFontPolicy /Error /CompatibilityLevel 1.7 /CompressObjects /Off /CompressPages true /ConvertImagesToIndexed true /PassThroughJPEGImages true /CreateJobTicket false /DefaultRenderingIntent /Default /DetectBlends true /DetectCurves 0.0000 /ColorConversionStrategy /LeaveColorUnchanged /DoThumbnails false /EmbedAllFonts true /EmbedOpenType false /ParseICCProfilesInComments true /EmbedJobOptions true /DSCReportingLevel 0 /EmitDSCWarnings false /EndPage -1 /ImageMemory 1048576 /LockDistillerParams true /MaxSubsetPct 100 /Optimize true /OPM 0 /ParseDSCComments false /ParseDSCCommentsForDocInfo true /PreserveCopyPage true /PreserveDICMYKValues true /PreserveEPSInfo false /PreserveFlatness true /PreserveHalftoneInfo true /PreserveOPIComments false /PreserveOverprintSettings</s>
<s>true /StartPage 1 /SubsetFonts true /TransferFunctionInfo /Remove /UCRandBGInfo /Preserve /UsePrologue false /ColorSettingsFile () /AlwaysEmbed [ true /AbadiMT-CondensedLight /ACaslon-Italic /ACaslon-Regular /ACaslon-Semibold /ACaslon-SemiboldItalic /AdobeArabic-Bold /AdobeArabic-BoldItalic /AdobeArabic-Italic /AdobeArabic-Regular /AdobeHebrew-Bold /AdobeHebrew-BoldItalic /AdobeHebrew-Italic /AdobeHebrew-Regular /AdobeHeitiStd-Regular /AdobeMingStd-Light /AdobeMyungjoStd-Medium /AdobePiStd /AdobeSongStd-Light /AdobeThai-Bold /AdobeThai-BoldItalic /AdobeThai-Italic /AdobeThai-Regular /AGaramond-Bold /AGaramond-BoldItalic /AGaramond-Italic /AGaramond-Regular /AGaramond-Semibold /AGaramond-SemiboldItalic /AgencyFB-Bold /AgencyFB-Reg /AGOldFace-Outline /AharoniBold /Algerian /Americana /Americana-ExtraBold /AndaleMono /AndaleMonoIPA /AngsanaNew /AngsanaNew-Bold /AngsanaNew-BoldItalic /AngsanaNew-Italic /AngsanaUPC /AngsanaUPC-Bold /AngsanaUPC-BoldItalic /AngsanaUPC-Italic /Anna /ArialAlternative /ArialAlternativeSymbol /Arial-Black /Arial-BlackItalic /Arial-BoldItalicMT /Arial-BoldMT /Arial-ItalicMT /ArialMT /ArialMT-Black /ArialNarrow /ArialNarrow-Bold /ArialNarrow-BoldItalic /ArialNarrow-Italic /ArialRoundedMTBold /ArialUnicodeMS /ArrusBT-Bold /ArrusBT-BoldItalic /ArrusBT-Italic /ArrusBT-Roman /AvantGarde-Book /AvantGarde-BookOblique /AvantGarde-Demi /AvantGarde-DemiOblique /AvantGardeITCbyBT-Book /AvantGardeITCbyBT-BookOblique /BakerSignet /BankGothicBT-Medium /Barmeno-Bold /Barmeno-ExtraBold /Barmeno-Medium /Barmeno-Regular /Baskerville /BaskervilleBE-Italic /BaskervilleBE-Medium /BaskervilleBE-MediumItalic /BaskervilleBE-Regular /Baskerville-Bold /Baskerville-BoldItalic /Baskerville-Italic /BaskOldFace /Batang /BatangChe /Bauhaus93 /Bellevue /BellMT /BellMTBold /BellMTItalic /BerlingAntiqua-Bold /BerlingAntiqua-BoldItalic /BerlingAntiqua-Italic /BerlingAntiqua-Roman /BerlinSansFB-Bold /BerlinSansFBDemi-Bold /BerlinSansFB-Reg /BernardMT-Condensed /BernhardModernBT-Bold /BernhardModernBT-BoldItalic /BernhardModernBT-Italic /BernhardModernBT-Roman /BiffoMT /BinnerD /BinnerGothic /BlackadderITC-Regular /Blackoak /blex /blsy /Bodoni /Bodoni-Bold /Bodoni-BoldItalic /Bodoni-Italic /BodoniMT /BodoniMTBlack /BodoniMTBlack-Italic /BodoniMT-Bold /BodoniMT-BoldItalic /BodoniMTCondensed /BodoniMTCondensed-Bold /BodoniMTCondensed-BoldItalic /BodoniMTCondensed-Italic /BodoniMT-Italic /BodoniMTPosterCompressed /Bodoni-Poster /Bodoni-PosterCompressed /BookAntiqua /BookAntiqua-Bold /BookAntiqua-BoldItalic /BookAntiqua-Italic /Bookman-Demi /Bookman-DemiItalic /Bookman-Light /Bookman-LightItalic /BookmanOldStyle /BookmanOldStyle-Bold /BookmanOldStyle-BoldItalic /BookmanOldStyle-Italic /BookshelfSymbolOne-Regular /BookshelfSymbolSeven /BookshelfSymbolThree-Regular /BookshelfSymbolTwo-Regular /Botanical /Boton-Italic /Boton-Medium /Boton-MediumItalic /Boton-Regular /Boulevard /BradleyHandITC /Braggadocio /BritannicBold /Broadway /BrowalliaNew /BrowalliaNew-Bold /BrowalliaNew-BoldItalic /BrowalliaNew-Italic /BrowalliaUPC /BrowalliaUPC-Bold /BrowalliaUPC-BoldItalic /BrowalliaUPC-Italic /BrushScript /BrushScriptMT /CaflischScript-Bold /CaflischScript-Regular /Calibri /Calibri-Bold /Calibri-BoldItalic /Calibri-Italic /CalifornianFB-Bold /CalifornianFB-Italic /CalifornianFB-Reg /CalisMTBol /CalistoMT /CalistoMT-BoldItalic /CalistoMT-Italic /Cambria /Cambria-Bold /Cambria-BoldItalic /Cambria-Italic /CambriaMath /Candara /Candara-Bold /Candara-BoldItalic /Candara-Italic /Carta /CaslonOpenfaceBT-Regular /Castellar /CastellarMT /Centaur /Centaur-Italic /Century /CenturyGothic /CenturyGothic-Bold /CenturyGothic-BoldItalic /CenturyGothic-Italic /CenturySchL-Bold /CenturySchL-BoldItal /CenturySchL-Ital /CenturySchL-Roma /CenturySchoolbook /CenturySchoolbook-Bold /CenturySchoolbook-BoldItalic /CenturySchoolbook-Italic /CGTimes-Bold /CGTimes-BoldItalic /CGTimes-Italic /CGTimes-Regular /CharterBT-Bold /CharterBT-BoldItalic /CharterBT-Italic /CharterBT-Roman /CheltenhamITCbyBT-Bold /CheltenhamITCbyBT-BoldItalic /CheltenhamITCbyBT-Book /CheltenhamITCbyBT-BookItalic /Chiller-Regular /Cmb10 /CMB10 /Cmbsy10 /CMBSY10 /CMBSY5 /CMBSY6 /CMBSY7 /CMBSY8 /CMBSY9 /Cmbx10 /CMBX10 /Cmbx12 /CMBX12 /Cmbx5 /CMBX5 /Cmbx6 /CMBX6 /Cmbx7 /CMBX7 /Cmbx8 /CMBX8 /Cmbx9 /CMBX9 /Cmbxsl10 /CMBXSL10 /Cmbxti10 /CMBXTI10 /Cmcsc10 /CMCSC10 /Cmcsc8 /CMCSC8 /Cmcsc9 /CMCSC9 /Cmdunh10 /CMDUNH10 /Cmex10 /CMEX10 /CMEX7 /CMEX8 /CMEX9 /Cmff10 /CMFF10 /Cmfi10 /CMFI10 /Cmfib8 /CMFIB8 /Cminch /CMINCH /Cmitt10 /CMITT10 /Cmmi10 /CMMI10 /Cmmi12 /CMMI12 /Cmmi5 /CMMI5 /Cmmi6 /CMMI6 /Cmmi7 /CMMI7 /Cmmi8 /CMMI8 /Cmmi9 /CMMI9 /Cmmib10 /CMMIB10 /CMMIB5 /CMMIB6 /CMMIB7 /CMMIB8 /CMMIB9 /Cmr10 /CMR10 /Cmr12 /CMR12 /Cmr17 /CMR17 /Cmr5 /CMR5 /Cmr6 /CMR6 /Cmr7 /CMR7 /Cmr8 /CMR8 /Cmr9 /CMR9 /Cmsl10 /CMSL10 /Cmsl12 /CMSL12 /Cmsl8 /CMSL8 /Cmsl9 /CMSL9 /Cmsltt10 /CMSLTT10 /Cmss10 /CMSS10 /Cmss12 /CMSS12 /Cmss17 /CMSS17 /Cmss8 /CMSS8 /Cmss9 /CMSS9 /Cmssbx10 /CMSSBX10 /Cmssdc10 /CMSSDC10 /Cmssi10 /CMSSI10 /Cmssi12 /CMSSI12 /Cmssi17 /CMSSI17 /Cmssi8 /CMSSI8 /Cmssi9 /CMSSI9 /Cmssq8 /CMSSQ8 /Cmssqi8 /CMSSQI8 /Cmsy10 /CMSY10 /Cmsy5 /CMSY5 /Cmsy6 /CMSY6 /Cmsy7 /CMSY7 /Cmsy8 /CMSY8 /Cmsy9 /CMSY9 /Cmtcsc10 /CMTCSC10 /Cmtex10 /CMTEX10 /Cmtex8 /CMTEX8 /Cmtex9 /CMTEX9 /Cmti10 /CMTI10 /Cmti12 /CMTI12 /Cmti7 /CMTI7 /Cmti8 /CMTI8 /Cmti9 /CMTI9 /Cmtt10 /CMTT10 /Cmtt12 /CMTT12 /Cmtt8 /CMTT8 /Cmtt9 /CMTT9 /Cmu10 /CMU10 /Cmvtt10 /CMVTT10 /ColonnaMT /Colossalis-Bold /ComicSansMS /ComicSansMS-Bold /Consolas /Consolas-Bold /Consolas-BoldItalic /Consolas-Italic /Constantia /Constantia-Bold /Constantia-BoldItalic /Constantia-Italic /CooperBlack /CopperplateGothic-Bold /CopperplateGothic-Light /Copperplate-ThirtyThreeBC /Corbel /Corbel-Bold /Corbel-BoldItalic /Corbel-Italic /CordiaNew /CordiaNew-Bold /CordiaNew-BoldItalic /CordiaNew-Italic /CordiaUPC /CordiaUPC-Bold /CordiaUPC-BoldItalic /CordiaUPC-Italic /Courier /Courier-Bold /Courier-BoldOblique /CourierNewPS-BoldItalicMT /CourierNewPS-BoldMT /CourierNewPS-ItalicMT /CourierNewPSMT /Courier-Oblique /CourierStd /CourierStd-Bold /CourierStd-BoldOblique /CourierStd-Oblique /CourierX-Bold /CourierX-BoldOblique /CourierX-Oblique /CourierX-Regular /CreepyRegular /CurlzMT /David-Bold /David-Reg /DavidTransparent /Dcb10 /Dcbx10 /Dcbxsl10 /Dcbxti10 /Dccsc10 /Dcitt10 /Dcr10 /Desdemona /DilleniaUPC /DilleniaUPCBold /DilleniaUPCBoldItalic /DilleniaUPCItalic /Dingbats /DomCasual /Dotum /DotumChe /DoulosSIL /EdwardianScriptITC /Elephant-Italic /Elephant-Regular /EngraversGothicBT-Regular /EngraversMT /EraserDust /ErasITC-Bold /ErasITC-Demi /ErasITC-Light /ErasITC-Medium /ErieBlackPSMT /ErieLightPSMT /EriePSMT /EstrangeloEdessa /Euclid /Euclid-Bold /Euclid-BoldItalic /EuclidExtra /EuclidExtra-Bold /EuclidFraktur /EuclidFraktur-Bold /Euclid-Italic /EuclidMathOne /EuclidMathOne-Bold /EuclidMathTwo /EuclidMathTwo-Bold /EuclidSymbol /EuclidSymbol-Bold /EuclidSymbol-BoldItalic /EuclidSymbol-Italic /EucrosiaUPC /EucrosiaUPCBold</s>
<s>/EucrosiaUPCBoldItalic /EucrosiaUPCItalic /EUEX10 /EUEX7 /EUEX8 /EUEX9 /EUFB10 /EUFB5 /EUFB7 /EUFM10 /EUFM5 /EUFM7 /EURB10 /EURB5 /EURB7 /EURM10 /EURM5 /EURM7 /EuroMono-Bold /EuroMono-BoldItalic /EuroMono-Italic /EuroMono-Regular /EuroSans-Bold /EuroSans-BoldItalic /EuroSans-Italic /EuroSans-Regular /EuroSerif-Bold /EuroSerif-BoldItalic /EuroSerif-Italic /EuroSerif-Regular /EUSB10 /EUSB5 /EUSB7 /EUSM10 /EUSM5 /EUSM7 /FelixTitlingMT /Fences /FencesPlain /FigaroMT /FixedMiriamTransparent /FootlightMTLight /Formata-Italic /Formata-Medium /Formata-MediumItalic /Formata-Regular /ForteMT /FranklinGothic-Book /FranklinGothic-BookItalic /FranklinGothic-Demi /FranklinGothic-DemiCond /FranklinGothic-DemiItalic /FranklinGothic-Heavy /FranklinGothic-HeavyItalic /FranklinGothicITCbyBT-Book /FranklinGothicITCbyBT-BookItal /FranklinGothicITCbyBT-Demi /FranklinGothicITCbyBT-DemiItal /FranklinGothic-Medium /FranklinGothic-MediumCond /FranklinGothic-MediumItalic /FrankRuehl /FreesiaUPC /FreesiaUPCBold /FreesiaUPCBoldItalic /FreesiaUPCItalic /FreestyleScript-Regular /FrenchScriptMT /Frutiger-Black /Frutiger-BlackCn /Frutiger-BlackItalic /Frutiger-Bold /Frutiger-BoldCn /Frutiger-BoldItalic /Frutiger-Cn /Frutiger-ExtraBlackCn /Frutiger-Italic /Frutiger-Light /Frutiger-LightCn /Frutiger-LightItalic /Frutiger-Roman /Frutiger-UltraBlack /Futura-Bold /Futura-BoldOblique /Futura-Book /Futura-BookOblique /FuturaBT-Bold /FuturaBT-BoldItalic /FuturaBT-Book /FuturaBT-BookItalic /FuturaBT-Medium /FuturaBT-MediumItalic /Futura-Light /Futura-LightOblique /GalliardITCbyBT-Bold /GalliardITCbyBT-BoldItalic /GalliardITCbyBT-Italic /GalliardITCbyBT-Roman /Garamond /Garamond-Bold /Garamond-BoldCondensed /Garamond-BoldCondensedItalic /Garamond-BoldItalic /Garamond-BookCondensed /Garamond-BookCondensedItalic /Garamond-Italic /Garamond-LightCondensed /Garamond-LightCondensedItalic /Gautami /GeometricSlab703BT-Light /GeometricSlab703BT-LightItalic /Georgia /Georgia-Bold /Georgia-BoldItalic /Georgia-Italic /GeorgiaRef /Giddyup /Giddyup-Thangs /Gigi-Regular /GillSans /GillSans-Bold /GillSans-BoldItalic /GillSans-Condensed /GillSans-CondensedBold /GillSans-Italic /GillSans-Light /GillSans-LightItalic /GillSansMT /GillSansMT-Bold /GillSansMT-BoldItalic /GillSansMT-Condensed /GillSansMT-ExtraCondensedBold /GillSansMT-Italic /GillSans-UltraBold /GillSans-UltraBoldCondensed /GloucesterMT-ExtraCondensed /Gothic-Thirteen /GoudyOldStyleBT-Bold /GoudyOldStyleBT-BoldItalic /GoudyOldStyleBT-Italic /GoudyOldStyleBT-Roman /GoudyOldStyleT-Bold /GoudyOldStyleT-Italic /GoudyOldStyleT-Regular /GoudyStout /GoudyTextMT-LombardicCapitals /GSIDefaultSymbols /Gulim /GulimChe /Gungsuh /GungsuhChe /Haettenschweiler /HarlowSolid /Harrington /Helvetica /Helvetica-Black /Helvetica-BlackOblique /Helvetica-Bold /Helvetica-BoldOblique /Helvetica-Condensed /Helvetica-Condensed-Black /Helvetica-Condensed-BlackObl /Helvetica-Condensed-Bold /Helvetica-Condensed-BoldObl /Helvetica-Condensed-Light /Helvetica-Condensed-LightObl /Helvetica-Condensed-Oblique /Helvetica-Fraction /Helvetica-Narrow /Helvetica-Narrow-Bold /Helvetica-Narrow-BoldOblique /Helvetica-Narrow-Oblique /Helvetica-Oblique /HighTowerText-Italic /HighTowerText-Reg /Humanist521BT-BoldCondensed /Humanist521BT-Light /Humanist521BT-LightItalic /Humanist521BT-RomanCondensed /Imago-ExtraBold /Impact /ImprintMT-Shadow /InformalRoman-Regular /IrisUPC /IrisUPCBold /IrisUPCBoldItalic /IrisUPCItalic /Ironwood /ItcEras-Medium /ItcKabel-Bold /ItcKabel-Book /ItcKabel-Demi /ItcKabel-Medium /ItcKabel-Ultra /JasmineUPC /JasmineUPC-Bold /JasmineUPC-BoldItalic /JasmineUPC-Italic /JoannaMT /JoannaMT-Italic /Jokerman-Regular /JuiceITC-Regular /Kartika /Kaufmann /KaufmannBT-Bold /KaufmannBT-Regular /KidTYPEPaint /KinoMT /KodchiangUPC /KodchiangUPC-Bold /KodchiangUPC-BoldItalic /KodchiangUPC-Italic /KorinnaITCbyBT-Regular /KristenITC-Regular /KrutiDev040Bold /KrutiDev040BoldItalic /KrutiDev040Condensed /KrutiDev040Italic /KrutiDev040Thin /KrutiDev040Wide /KrutiDev060 /KrutiDev060Bold /KrutiDev060BoldItalic /KrutiDev060Condensed /KrutiDev060Italic /KrutiDev060Thin /KrutiDev060Wide /KrutiDev070 /KrutiDev070Condensed /KrutiDev070Italic /KrutiDev070Thin /KrutiDev070Wide /KrutiDev080 /KrutiDev080Condensed /KrutiDev080Italic /KrutiDev080Wide /KrutiDev090 /KrutiDev090Bold /KrutiDev090BoldItalic /KrutiDev090Condensed /KrutiDev090Italic /KrutiDev090Thin /KrutiDev090Wide /KrutiDev100 /KrutiDev100Bold /KrutiDev100BoldItalic /KrutiDev100Condensed /KrutiDev100Italic /KrutiDev100Thin /KrutiDev100Wide /KrutiDev120 /KrutiDev120Condensed /KrutiDev120Thin /KrutiDev120Wide /KrutiDev130 /KrutiDev130Condensed /KrutiDev130Thin /KrutiDev130Wide /KunstlerScript /Latha /LatinWide /LetterGothic /LetterGothic-Bold /LetterGothic-BoldOblique /LetterGothic-BoldSlanted /LetterGothicMT /LetterGothicMT-Bold /LetterGothicMT-BoldOblique /LetterGothicMT-Oblique /LetterGothic-Slanted /LevenimMT /LevenimMTBold /LilyUPC /LilyUPCBold /LilyUPCBoldItalic /LilyUPCItalic /Lithos-Black /Lithos-Regular /LotusWPBox-Roman /LotusWPIcon-Roman /LotusWPIntA-Roman /LotusWPIntB-Roman /LotusWPType-Roman /LucidaBright /LucidaBright-Demi /LucidaBright-DemiItalic /LucidaBright-Italic /LucidaCalligraphy-Italic /LucidaConsole /LucidaFax /LucidaFax-Demi /LucidaFax-DemiItalic /LucidaFax-Italic /LucidaHandwriting-Italic /LucidaSans /LucidaSans-Demi /LucidaSans-DemiItalic /LucidaSans-Italic /LucidaSans-Typewriter /LucidaSans-TypewriterBold /LucidaSans-TypewriterBoldOblique /LucidaSans-TypewriterOblique /LucidaSansUnicode /Lydian /Magneto-Bold /MaiandraGD-Regular /Mangal-Regular /Map-Symbols /MathA /MathB /MathC /Mathematica1 /Mathematica1-Bold /Mathematica1Mono /Mathematica1Mono-Bold /Mathematica2 /Mathematica2-Bold /Mathematica2Mono /Mathematica2Mono-Bold /Mathematica3 /Mathematica3-Bold /Mathematica3Mono /Mathematica3Mono-Bold /Mathematica4 /Mathematica4-Bold /Mathematica4Mono /Mathematica4Mono-Bold /Mathematica5 /Mathematica5-Bold /Mathematica5Mono /Mathematica5Mono-Bold /Mathematica6 /Mathematica6Bold /Mathematica6Mono /Mathematica6MonoBold /Mathematica7 /Mathematica7Bold /Mathematica7Mono /Mathematica7MonoBold /MatisseITC-Regular /MaturaMTScriptCapitals /Mesquite /Mezz-Black /Mezz-Regular /MICR /MicrosoftSansSerif /MingLiU /Minion-BoldCondensed /Minion-BoldCondensedItalic /Minion-Condensed /Minion-CondensedItalic /Minion-Ornaments /MinionPro-Bold /MinionPro-BoldIt /MinionPro-It /MinionPro-Regular /Miriam /MiriamFixed /MiriamTransparent /Mistral /Modern-Regular /MonotypeCorsiva /MonotypeSorts /MSAM10 /MSAM5 /MSAM6 /MSAM7 /MSAM8 /MSAM9 /MSBM10 /MSBM5 /MSBM6 /MSBM7 /MSBM8 /MSBM9 /MS-Gothic /MSHei /MSLineDrawPSMT /MS-Mincho /MSOutlook /MS-PGothic /MS-PMincho /MSReference1 /MSReference2 /MSReferenceSansSerif /MSReferenceSansSerif-Bold /MSReferenceSansSerif-BoldItalic /MSReferenceSansSerif-Italic /MSReferenceSerif /MSReferenceSerif-Bold /MSReferenceSerif-BoldItalic /MSReferenceSerif-Italic /MSReferenceSpecialty /MSSong /MS-UIGothic /MT-Extra /MTExtraTiger /MT-Symbol /MT-Symbol-Italic /MVBoli /Myriad-Bold /Myriad-BoldItalic /Myriad-Italic /Myriad-Roman /Narkisim /NewCenturySchlbk-Bold /NewCenturySchlbk-BoldItalic /NewCenturySchlbk-Italic /NewCenturySchlbk-Roman /NewMilleniumSchlbk-BoldItalicSH /NewsGothic /NewsGothic-Bold /NewsGothicBT-Bold /NewsGothicBT-BoldItalic /NewsGothicBT-Italic /NewsGothicBT-Roman /NewsGothic-Condensed /NewsGothic-Italic /NewsGothicMT /NewsGothicMT-Bold /NewsGothicMT-Italic /NiagaraEngraved-Reg /NiagaraSolid-Reg /NimbusMonL-Bold /NimbusMonL-BoldObli /NimbusMonL-Regu /NimbusMonL-ReguObli /NimbusRomNo9L-Medi /NimbusRomNo9L-MediItal /NimbusRomNo9L-Regu /NimbusRomNo9L-ReguItal /NimbusSanL-Bold /NimbusSanL-BoldCond /NimbusSanL-BoldCondItal /NimbusSanL-BoldItal /NimbusSanL-Regu /NimbusSanL-ReguCond /NimbusSanL-ReguCondItal /NimbusSanL-ReguItal /Nimrod /Nimrod-Bold /Nimrod-BoldItalic /Nimrod-Italic /NSimSun /Nueva-BoldExtended /Nueva-BoldExtendedItalic /Nueva-Italic /Nueva-Roman /NuptialScript /OCRA /OCRA-Alternate /OCRAExtended /OCRB /OCRB-Alternate /OfficinaSans-Bold /OfficinaSans-BoldItalic /OfficinaSans-Book /OfficinaSans-BookItalic /OfficinaSerif-Bold /OfficinaSerif-BoldItalic /OfficinaSerif-Book /OfficinaSerif-BookItalic /OldEnglishTextMT /Onyx /OnyxBT-Regular /OzHandicraftBT-Roman /PalaceScriptMT /Palatino-Bold /Palatino-BoldItalic /Palatino-Italic /PalatinoLinotype-Bold /PalatinoLinotype-BoldItalic /PalatinoLinotype-Italic /PalatinoLinotype-Roman /Palatino-Roman /PapyrusPlain /Papyrus-Regular /Parchment-Regular /Parisian /ParkAvenue /Penumbra-SemiboldFlare /Penumbra-SemiboldSans /Penumbra-SemiboldSerif /PepitaMT /Perpetua /Perpetua-Bold /Perpetua-BoldItalic /Perpetua-Italic /PerpetuaTitlingMT-Bold /PerpetuaTitlingMT-Light /PhotinaCasualBlack /Playbill /PMingLiU /Poetica-SuppOrnaments /PoorRichard-Regular /PopplLaudatio-Italic /PopplLaudatio-Medium /PopplLaudatio-MediumItalic /PopplLaudatio-Regular /PrestigeElite</s>

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.