A Research On Aspect Category Identification For Sentiment Analysis

Traditional sentiment analysis focuses on determine the sentiment polarity of a piece of given text. More meticulously, aspect based sentiment analysis (ABSA) looks at the opinion itself instead of looking at no matter which level of language constructs(documents, sentences). An ABSA system performs finer-grained analysis by firstly detecting what aspects people are appraising, then determining what their attitude is, towards these aspects.Aspect identification plays a fundamental role in ABSA. It has been widely inves-tigated and there exists a a variety of approaches to identify the aspects mentioned in a sentence. In previous works, nouns or noun phrases frequently mentioned in reviews are extracted to represent aspects. There are two main problems of these approaches. First of all, different words may have similar semantic. For instance, "screen" and "display" refers to same aspect of a laptop computer. If we regard these two words as different aspects, it will make people confused. Secondly, people sometimes comment without explicitly using an "aspect term". We talk about the price of a product by "too expensive" but not mentioning any nouns like "cost" or "price"To tackle these two problems, ABSA based on pre-defined aspect categories was proposed. In this approach, aspect identification is regarded a multil-label classification problem. However, a long sentence with complex structure usually express multiple opinions and that makes it difficult to distinguish which aspect is related to which part of the sentence. This thesis trys to tackle this problem from two perspectives. First, we propose a novel view of aspect based sentiment analysis. From this point of view, ABSA is regarded a translation from the natural language to a artifical language. Thus we can taking advantage the alignment models, which captures the alignments between text fragments and aspect categories.Secondly, we propose a multi-view multi-instance multi-label classification frame-work for aspect category identification. In this framework, texts are represented in different views and sentences are divided into clauses. A clause is the smallest gram-matical unit that can express a complet proposition. That means, a clause basically express one single opinion and it’s easier to analysis. We propose an algorithm called co-transfer training to solve this multi-view multi-instance multi-label problem.We demonstrate our methods empirically on two benchmark datasets in laptops and restaurants domains, and outperform the best results achieved in previous works.