Abstractive Thai opinion summarization

Abstract

With the advancement of the Internet technology, customers can easily share opinions about services and products in forms of reviews. There can be large amount of reviews for popular products. Manually summarizing those reviews for important issues is a daunting task. Automatic opinion summarization is a solution to the problem. The task is more complicated for reviews written in Thai. Thai words are written continuously without space, and there is no symbol to identify the end of a sentence. Many reviews are written informally, thus accurate word identification and linguistic annotation cannot be relied upon. Text summarization can be classified into two categories, which are extractive and abstractive summarization. In the extractive method, the summary is a set of actual sentences or phrases extracted from the reviews; on the other hand, abstractive summarization does not output original sentences from the reviews, but generates new sentences or phrases into a summary. The abstractive summarization approach is more difficult and thus less popular than the extractive approach. This research proposes a novel technique to generate abstractive summaries of customer reviews written in Thai. The proposed technique, which consists of local and global models, is evaluated by using actual reviews of fifty products, randomly selected from a popular cosmetic website. The results show that the local model outperforms the global model and the two baseline methods, both quantitatively and qualitatively.