Developing a concept-level polarity detection model through generation of a rule based semantic parser for Bengali sentences

Abstract

Public opinion over the Internet is getting importance with the rapid growth of online content every day. The sentiment of public opinion is considered a valuable piece of information in every interaction of human life. Concept-based approaches are the recent evolution in sentiment analysis, which is intended to infer the semantic and affective information associated with natural language opinion. Sentiment analysis at the concept level introduces a new opportunity for information retrieval like polarity detection, especially for a less privileged language like Bengali. In this work, a rule- based semantic parser is developed to generate the parse tree for a Bengali sentence. Concepts are extracted from the parse tree exploring the dependency among the constituents of the sentence. A domain specific classification model is proposed to detect the polarity of the concepts which in turn are used to find the sentence polarity through the parse tree traversal. Here, the AffectiveSpace is used as a knowledge base. Training data on targeted domain is generated from online contents using term frequency and inverse document frequency (tf-idf) where the concepts are labeled as positive, negative and neutral. The model uses the Linear Discriminant Analysis (LDA) to classify the training data where 81.8 percent of original grouped concepts correctly classified. The performance of the polarity detection method is evaluated using the precision and recall method. The overall accuracy for concept-level polarity detection is 70.24 percent. Whereas the accuracy at the sentence level is 65.63 percent for the simple sentence, and 73.77 percent for the complex or compound sentence, which can be considered an acceptable range for a less privileged language like Bengali. One of the limitations of the work is its failure to achieve the desired level of abstraction in forming the concept due to the language complexity of Bengali. Therefore, it is fully dependent on the terms available within the sentence and translates those to English for mapping in the AffectiveSpace. However, an independent dependency parser for Bengali can be generated by integrating the language morphology along with the language syntax to extract the concept with a high level of abstraction. Moreover, the generation of a Bengali affect space can be of great use in the field of NLP.