Theme Detection an Exploration of Opinion Subjectivity Amitava Das Jadavpur University Department of Computer Science and Engineering, Jadavpur University, Kolkata 700032, India Sivaji Bandyopadhyay Jadavpur University Department of Computer Science and Engineering, Jadavpur University, Kolkata 700032, India email@example.com firstname.lastname@example.org Abstract Work in opinion mining and classification often assumes the incoming documents to be opinionated. Opinion mining system makes false hits while attempting to compute polarity values for non-subjective or factual sentences or documents. It becomes imperative to decide whether a given document contains subjective information or not as well as to identify which portions of the document are subjective or factual. In this work a Theme Detection technique has been evolved for more generic domain independent subjectivity detection that classifies sentences with binary feature: opinionated or non-opinionated. Theme Detection technique examines sentence level opinion and finally accumulates the opinion clues to reach the discourse level subjectivity. The subjectivity detection system has been evaluated on the Multi Perspective Question Answering (MPQA) corpus as well as on Bengali corpus. The system evaluation has shown the precision and recall values of 76.08 and 83.33 for English and 72.16 and 76.00 for Bengali respectively. Keywords: Opinion Mining, Subjectivity Detection, Theme Detection. 1. Related Works An opinion could be defined as a private state that is not open to objective observation or verification . Opinion extraction, opinion summarization and opinion tracking are three important techniques for understanding opinions. Opinion-mining of product reviews, travel advice, consumer complaints, stock market predictions, real estate market predictions, email etc. are areas of interest for researchers since last few decades. Most research on opinion analysis has focused on sentiment analysis , subjectivity detection (,, ,), Methods on the extraction of opinionated sentences in a structured form can be found in . Some machine learning text labeling algorithms like Conditional Random Field (CRF) (,), Support Vector Machine (SVM)  have been used to cluster same type of opinions. Application of machine-learning techniques to any NLP task needs a large amount of 978-1-4244-4799-2/09/$25.00 ©2009 IEEE data. It is time-consuming and expensive to hand-label the large amounts of training data necessary for good performance. Hence, use of machine learning techniques to extract opinions in any new language may not be an acceptable solution. Opinion analysis of news document is an interesting area to explore. Newspapers generally attempt to present the news objectively, but textual affect analysis in news documents shows that many words carry positive or negative emotional charge . Some important works on opinion analysis in the newspaper domain are ,  and , but no such efforts have been taken up in Indian languages especially in Bengali. 2. Annotated Data Preparation The opinion subjectivity detection technique presented in this paper is rule based in nature and hence annotated data preparation is necessary for system testing and evaluation. The technique has been applied on both English and Bengali language texts. In case of English we choose MPQA corpus . Since in the MPQA corpus the private states in a sentence are annotated, sentence level opinion subjectivity annotation has been done as described in Section 3.1. No such corpus is available for Bengali. Hence, the editorial pages, i.e., Reader’s opinion section or Letters to the Editor Section, from the web archive of a popular Bengali newspaper are identified as the relevant corpus in Bengali. Detailed reports about this news corpus development in Bengali can be found in . The relevant Bengali news corpus is hand annotated as described in Section 3.2. 2.1. Subjectivity-Identification in MPQA The annotation scheme in the MPQA corpus identifies key components and properties of opinions, emotions, sentiments, speculations, evaluations, and other private states . The properties of a private state frame include the source of the private state (i.e., whose private state is being expressed), the target (i.e., what the private state is about), and various other details involving intensity, significance and type of attitude. The annotation within the MPQA corpus is not at sentence level but at word or phrase level. Private states in MPQA are classified into two basic categories i.e.
direct subjective frames and expressive subjective element. To distinguish opinion-oriented materials from other factual materials, objective speech event frames have been defined. Direct Subjective Frame: A private state containing direct subjective element is called a direct subjective frame. For example, in the sentence “The U.S. fears a spill-over,” said Xirao-Nima.” The word ‘fears’ represents a private state and is annotated as ‘Direct Subjective Frame’. Expressive subjective elements: A private state containing no direct opinion but only subjective references to opinion is called an expressive subjective element frame. For example, in the sentence “The report is full of absurdities,” Xirao-Nima said. The phrase “full of absurdities” represents a private state and is annotated as ‘Expressive Subjective Element’. Objective Speech Event: This is purely the factual part of any event. For example, the sentence does not carry any opinionated information but a description of a fact or event. “O’Leary said “the incident took place at 2:00pm.” During the subjective tagging of MPQA sentences, the hypothesis is that if a sentence has any Direct or Expressive Subjective phrases then the sentence would be subjective itself. Hence, the sentences containing either of the two private states (i.e. Direct Subjective and Expressive Subjective) are extracted as subjective sentences and the sentences containing only objective speech event or no annotated private states are discarded. There are features of an attitude frame like intensity in the MPQA annotation scheme. But for the present task other attributes of an attitude frame are not considered. 2.2. Subjectivity-Annotation in Bengali From the collected document set (Letters to the Editor Section), some documents have been chosen for the annotation task. Documents that appeared within an interval of four months are chosen on the hypothesis that these letters to the editors will be on related events. A simple annotation tool has been designed for annotating the subjective sentences. Three annotators participated in the present task. The documents with such annotated sentences are saved in XML format. The tool also highlights the sentiment words (SentiWordNet, see section 4.2.5 for details) by four different colors within a document according to their POS categories (Noun, Adjective, Adverb and Verb). This technique helps to increase the speed of annotation process. Finally 100 annotated documents have been produced. No inter-annotator agreement has been calculated for the present task. Some statistics about the Bengali news corpus is represented in the Table 1. Total number of documents in the corpus Total number of sentences in the corpus 100 2234 Avgerage number of sentences in a document 22 Total number of wordforms in the corpus 28807 Avgerage number of wordforms in a 288 document Total number of distinct wordforms in the 17176 corpus Table 1. Bengali News Corpus Statistics 3. Theme Detection Theme Detection is a rule based algorithm to identify subjective sentences in text documents. The algorithm takes the subjectivity decision on the basis of several features of the sentences in the text. These features are obtained using various machine learning algorithms. Various linguistic resources that are used to derive several binary features have been developed manually. Theme detection process works in two stages. Theme detection technique first captures discourse level opinion theme in terms of thematic expressions which best describes the opinionated theme of a document. In the next level the algorithm examines the presence of thematic expression as an opinion constituent (SubjectAspect-Evaluation) in any sentence. Subjectivity detection by lexicons like SentiWordNet or Subjectivity word list has been explored by researchers. Theme detection technique works on discourse level and takes care of syntactic structure of sentences. . The challenge is to identify the most concise feature set and construct the rules effectively for the two stage identification problem. Experiments are carried out with an initial list of features and finally some of the features are discarded as they are found to have no contribution towards increasing system performance. The Theme detection technique has been applied on both English and Bengali language texts. Motivation for the technique has been presented in Section 4.1. The various subjectivity clues or features and how these can be obtained have been discussed in Section 4.2. The evaluation results presented in the Section 6 show the effectiveness of the algorithm. 3.1. Motivation Many supervised and unsupervised techniques have been explored for subjectivity annotation task by many researchers over a long period of time. Several linguistic resources and tools like dependency parsing, Named Entity Recognition, Morphological Analyzer, Stemmer, SentiWordNet, WordNet etc have been used several times in the subjectivity detection task. But in the case of morphologically rich Indian languages like Bengali, such resources and tools are not readily available. Highly inspired by Janyce Wiebe et.al, 2005  the present work is initiated to develop a subjectivity classifier that will work on un-annotated text. Our aim is to design an automatic process that learns linguistically rich extraction patterns for subjective (opinionated) expressions and produces a rich ontological language-
specific (rather than domain dependent) knowledge. Subjective remarks come in a variety of forms, including opinions, rants, allegations, accusations, suspicions, humor and speculations. 3.2. Learning Subjective Clues Existing methods for opinion extraction tend to rely on relatively simple proximity-based or pattern-based techniques. However, these pattern-based techniques are not enough to extract opinions because these patterns can apply to the cases where all constituents of opinion appear in a sentence. It has been observed that most of the opinion constituents do not have a direct syntactic dependency relation within a sentence, mostly due to elliptical arguments. Based on a corpus study, it is proposed to define an opinion unit as a quadruple, i.e., the opinion holder, the subject being evaluated (Subject), the part or the attribute in which it is evaluated (Aspect), and the evaluation that expresses a positive or negative assessment (Evaluation). The present subjectivity detection algorithm has been applied to News corpus (both for English and Bengali) where name of the author of any article is rarely mentioned. Hence, the Opinion Holder information is not taken into consideration and only the SubjectAspect-Evaluation constituents are identified for analysis. These constituents can be further defined as: Subject: A named entity (Person or Location etc.) of a given particular class of interest (e.g. a leader name or a location name where an incident occurred). Aspect: An attribute of the subject with respect to which evaluation is made (size, color, date etc.). The aspect can define a characteristic of the subject or an integral part of the subject. Evaluation: An evaluative or subjective phrase used to express an evaluation or the opinion holder's mental/emotional attitude (good, poor, powerful, stylish etc.). Initially, a detailed analysis of the English MPQA and Bengali newspaper corpus has been done to understand the most concise and effective features for opinion (i.e., opinion constituents) identification and their characteristics in the corpus. In order to identify features we started with Part Of Speech (POS) categories and continued the exploration with the other features like chunk, functional word, ontology list, SentiWordNet, stemming cluster, frequency, positional aspect (e.g. title, first Paragraph, last two sentences, critical Issues) and average distribution. Each of the features and the methods for their identification are now being discussed. 3.2.1 Part Of Speech (POS) Hatzivassiloglou et. al., 2000 , Chesley et. al., 2006  etc. have proved that those words carrying opinion in sentences are mainly adjective, adverb, noun and verbs. Many opinion mining task, for example the one presented in , are mostly based on adjective words. This means that this part-of-speech (POS) tag is more important for Subjectivity Detection than others. The identified POS categories that carry opinion information are Adjective, Adverb, Verb and Noun while the opinion information for words of other POS categories is difficult to generalize. Thus, we concentrated on identifying the POS categories of the words, especially to see whether these words are adjectives, adverbs, noun and verbs. Stanford Parser 1 has been used for English text to get the word level POS category. The overall accuracy of this tool as reported in Klein et.al. 2003  is 86.7%. The POS Tagging Engine for Bengali text has been developed using the statistical Conditional Random Fields (CRF) 2 . The system makes use of the different contextual information of the words along with the variety of features that are helpful in predicting the various part of speech (POS) classes. The training set consists of 200K words and has been manually annotated with a POS tag-set 3 of twenty one tags developed by International Institute of Information Technology Hyderabad (IIIT-H). The system produces output in Shakti-Standard-Format 4 (SSF), also developed by IIIT-H for Indian languages. Experimental test results show the effectiveness of the CRF based POS tagging system with an overall average 87.23% accuracy. Feature selection plays a crucial role in CRF framework. Experiments were carried out to find out the most suitable features for POS in Bengali. The Experimental results are shown below in Table 2. Training-Set Test-Set Accuracy 16397 4587 87.23% Table 2. Experimental Result of POS Tagging. 3.2.2 Chunk Identification of subjective feature depends on opinion constituents. A detailed empirical study reveals that Subject, Aspect or Evaluation expressions may be defined in terms of chunk tags. Subject phrases are generally noun phrases with noun-noun, adjective-noun or noun-other combinations. The nouns in the subject phrases are generally named entities or low frequency noun words or out of vocabulary words. In case of multiple noun phrases in a sentence the head noun phrase is treated as a Subject phrase. Some empirical rules have been defined to identify the head noun phrase in a sentence: the distance of each NP (Noun phrase) from VP (Verb Phrase) is calculated in terms of characters. The NP which is situated at the farthest distance from VP is selected as the head NP. In case of multiple VPs, only the phrases tagged as Verb Finite Phrases are considered. 1 http://nlp.stanford.edu/software/lex-parser.shtml#Download http://crfpp.sourceforge.net 3 http://shiva.iiit.ac.in/SPSAL2007/iiit_tagset_guidelines.pdf 4 http://ltrc.iiit.net/MachineTrans/research/tb/shakti-analy-ssf.pdf 2
Aspect phrases are an attribute of the Subject. But in News domain aspect is not only sub part of the Subject but also a conceptual sub-event or an actor’s (i.e. Subject) activity on a particular event. More specifically the aspects are factual information which should be discarded by subjectivity detector. As an example: 1) Saddam hanged up yesterday night. 2) USA president says “He deserved it”. 3) USA people’s general opinion this is a form of real cruelty. In the sentence 2, the aspect is Saddam Hussain’s death is missing, which is an event reported in the sentence 1. In the sentence 3 subject is peoples of USA but the aspect is also an event that is president’s statement which is a subsequent event of the main event i.e., Saddam Hussain’s death is missing. Aspects are either missing or present as elliptical arguments in News corpus. In other cases when aspect present in a sentence, it has been observed that it is generally a noun phrase with some sentiment words. A small set of rules are defined for aspect identification in a sentence. Evaluation phrases are used to express an evaluation or the opinion holder's mental/emotional attitude. Detailed analysis of both the English and Bengali corpus has shown that Evaluation phrases are generally verb phrases with some evaluative sentiment phrases (adjective or adverbial). In case of missing finite verb phrases (Sentence 3 in the preceding example) in any sentence the evaluative phrases are generally adjective or adverbial phrases or noun phrases along with evaluative sentiment words (obtained from SentiWordNet). Rules have been defined accordingly to identify those phrases. The common feature among the three constituents of opinion is that they could be noun phrases. Positional information and language features have been used to disambiguate among the three constituents. The Stanford Parser was used for English chunking task. The parser produced a parse tree as the output that is converted into SSF format for further processing. The chunker for Bengali texts is trained on the feature templates for predicting the chunk boundary tags using CRF. The accuracy of the chunker is shown in Table 3. Training-Set Test-Set Accuracy 16397 4587 79.51% Table 3. Experimental Result of Chunker. 3.2.3 Functional Word Function words in a language are high frequency words and these words generally do not contribute to identify subjectivity; hence these words are dropped by system in the first stage. But function words help many times to understand syntactic pattern of an opinionated sentence, hence some rules are constructed based on functional words at the POS and Chunk level instead of the word itself. An example may illustrate the situation. 4) President and Army chief both congratulated NASA on the success of their expedition. In the sentence 4 the subject is “President and Army chief” and the clue is the function word “and”. Rules are defined to find out more than one consecutive NPs (Noun phrase) connected with any conjunct. A list of 253 entries is collected from the Bengali corpus. First a unique high frequency word list is generated where the assumed threshold frequency is considered as 20. The list is manually corrected keeping in mind that a word should not carry any opinionated or sentiment feature. For English the functional word list is collected from Website 5. In the English functional word list there are 300 entries. 3.2.4 Ontology List Four different ontology lists corresponding to the four POS categories (Noun, Verb, Adjective and Adverb) of words have been prepared by selecting the top 100 words in each category from POS tagged English and Bengali texts. During the second stage of subjectivity detection, i.e., Theme Sentence Identification, words with Noun POS categories were separated into Named Entities and Common nouns. 3.2.5 SentiWordNet in Bengali Words that are present in the SentiWordNet carry opinion information. SentiWordNet  is an automatically constructed lexical resource for English which assigns a positivity score, a negativity score, and a neutrality score to each WordNet synset. Release 1.1 of SentiWordNet for English was obtained from the authors. SentiWordNet Release 1.1 consists of 115,341 words marked with positive and negative orientation scores ranging from 0 to 1. A subset of 8,427 opinionated words was extracted from SentiWordNet, by selecting those whose orientation strength is above a threshold of 0.4. As there was no such SentiWordNet for Bengali, a task has been initiated to develop a similar resource for Bengali. For the task, Samsad 6, a widely used EnglishBengali dictionary available both in offline and online version, is selected. The Samsad English–Bengali dictionary has approximately 102119 entries. A word to word simple lexical-transfer technique is applied to each entry of SentiWordNet. Each dictionary search produces a set of Bengali words for a particular English word. Instead of making them into one entry we separate them into multiple one word entries for making the subsequent search process faster. The positive and negative opinion scores for the Bengali words are copied from their English equivalents. This process has resulted in 20,789 Bengali entries which is a useful resource in the Bengali Opinion Mining task. The words in the POS tagged English and Bengali texts with the following POS tags, namely Noun, Verb Adjective and Adverbs, were checked into the 5 http://www.dcs.gla.ac.uk/idom/ir_resources/linguistic_utils/stop_w ords 6 http://dsal.uchicago.edu/dictionaries/biswas_bengali/