Khmer Sentiment Lexicon Based on PU Learning and Label Propagation Algorithm

Abstract

The sentiment lexicon is an important tool for natural language processing tasks. In addition to being able to determine the sentiment polarity of words or phrases, it can assist attribute-level, sentence-level, and text-level sentiment analysis tasks. In light of the fact that tagging data and corpora for the Khmer language are scarce, where most resources related to sentiment lexicons are for English, this paper proposes a method for constructing a sentiment lexicon for Khmer based on Positive-Unlabeled learning (PU Learning) and the label propagation algorithm. Sentiment words are first extracted from a corpus using the Spy technique of PU learning method. The main idea is to purify the set of N-class examples, train the MLP classifier, and continuously delete spy words and increase the number of P-class words in the iterative process. Following this, the sentiment polarity of the candidate words is determined. By considering the problem of determining the sentiment polarity of the candidate words as one of calculating its probability distribution, a small number of labeled sentiment words and candidate words are used to construct a graph model. The contextual information of the candidate words is used to construct a simple supplementary graph model of the set of sentiment words through word co-occurrence and triangulation, where this enhances the correlation between data items. The sentiment polarity of the candidate words is then determined through the label propagation algorithm. The results of experiments show that the proposed method can be used to construct a Khmer sentiment lexicon with a small number of labeled data and a small corpus without requiring excessive manual labeling.