Building Sentiment Lexicon with Representation Learning Based on Contrast and Label of Sentiment

摘要

Sentiment lexicon is an import part in many sentiment analysis or opinion-mining applications as its rich sentiment information plays an essential role in identifying the sentiment polarity of different text granularity. The latest work tries to automatically construct the sentiment lexicon by using the approaches of representation learning with word embedding. However, some study shows that similar word embedding may have opposite sentiment orientation, such ambiguous word embedding is difficult to distinguish them in sentiment classification model. To solve such problem, we integrate the sentiment label information of text and sentiment contrast information between target words and context word into word vector representation learning model. In this paper, the sentiment contrast information is synonyms and antonyms of the target word, which have positive LMI score with the context word and same sentiment orientation with the target word. We integrate sentiment label information of text and sentiment contrast information into Skip-gram to distinguish such ambiguous word embedding, and build the sentiment lexicon base on the improved word vector representation. The experiments demonstrated the effectiveness of our sentiment lexicon based on the improved word embedding, and its performance is better than the popular sentiment lexicons.