Otomatik Duygu Sözlüğü Geliştirilmesi ve Haberlerin Duygu Analizi

Abstract

The level reached by mass media today, with respect to informing the society, raising awareness, affecting opinions, and even mobilizing masses, is impressive. Being mass communication tools, mainstream news media produces enormous amounts of content. Analysis of this content, which is mostly textual and treasurized from an academic perspective, is also crucial for people in a large spectrum, from decision makers to policy makers. Text analysis is studied under the discipline of Natural Language Processing (NLP), and Sentiment Analysis, which is a subdiscipline of NLP, is focused on enriching this analysis by evaluating the content with respect to sentimentality. The main requirement in content analysis is the existence of necessary language resources, whereas the correctness of the analysis depends on the coverage and sufficiency of these resources. Availability and state of the resources in the English language far surpasses the resources in many other languages, including Turkish. Most studies in the literature are, hence, language dependent.

In Turkish sentiment analysis studies, researchers mostly followed a route through translation. However, it is obvious that the correctness of the translation between the languages will deeply impact the performance of the analysis, with the effect being amplified in agglomerative languages, such as Turkish. Our base hypothesis in this work is that the language resources need to be produced within the language. With respect to this philosopohy, the first goal of this thesis is to produce a rich and correctly polarized General Purpose Turkish Sentiment Lexicon. Thus, we aim to provide an open resource to all disciplines working on the Turkish language. Our second goal, then, is to propose a Sentiment Map Model, which brings a fresh perspective to document sentiment analysis.

This thesis is mainly prepared in three phases. In the first phase, we aimed to produce a Turkish Sentiment Lexicon based on texts from main stream news media. To this end, a large corpus with known polarities was constructed. Once the tone and polarity values of the words from these texts were identified, they were merged with an existing Turkish Sentiment Lexicon, resulting in the first version of SWNetTR-PLUS, a 37K Turkish sentiment lexicon.

In the second phase, it was aimed to enrich SWNetTR-PLUS even further and stabilize the polarity and tone values of the words in the lexicon. To this end, synonyms and antonyms of the words in the lexicon were derived from different resources to extent the coverage of both positive and negative sentiments. At this point, it was chosen to model the whole lexicon with graphs to make it easier to explain and study the sentiment relations between the words. In order to compute missing polarities and tone values, and to stabilize the lexicon, we developed the concepts and methodologies of Tie Strength, Tone Propagation and Bias Balancing. At the end, we obtained SWNetTR++, which is a General Purpose Turkish Sentiment Lexicon with a capacity of 49K words.

In the third and the last phase, we proposed a new document sentiment analysis technique, namely Sentiment Map Model. Here, the motivation was the fact that in the literatüre document sentiment analysis is mostly based on assigning a positive/neutral/negative sentiment value to a document, and this omits the sentiment fluctuations within the text, crippling the richness of the text. As a solution to that, we proposed Sentiment Map Model, which allows detecting, exposing and interpreting sentiment shifts within an analysed document. The proposed model was extensively tested on multiple texts to outline its applicability and sufficiency.