Learning Sentiment Composition from Sentiment Lexicons

Proceedings of the 27th International Conference on Computational Linguistics, pages 2230–2241 Santa Fe, New Mexico, USA, August 20-26, 2018. 2230 Learning Sentiment Composition from Sentiment Lexicons Orith Toledo-Ronen Roy Bar-Haim∗ Alon Halfon Charles Jochim Amir Menczel Ranit Aharonov Noam Slonim IBM Research {oritht,roybar,alonhal,amir.menczel,ranita,noams}@il.ibm.com charlesj@ie.ibm.com Abstract Sentiment composition is a fundamental sentiment analysis problem. Previous work relied on manual rules and manually-created lexical resources such as negator lists, or learned a com- position function from sentiment-annotated phrases or sentences. We propose a new approach for learning sentiment composition from a large, unlabeled corpus, which only requires a word- level sentiment lexicon for supervision. We automatically generate large sentiment lexicons of bigrams and unigrams, from which we induce a set of lexicons for a variety of sentiment com- position processes. The effectiveness of our approach is confirmed through manual annotation, as well as sentiment classification experiments with both phrase-level and sentence-level bench- marks. 1 Introduction Many sentiment analysis systems rely primarily on the sentiment of individual words. However, precise sentiment analysis often requires lexical-semantic knowledge that goes beyond word-level sentiment, even when dealing with short phrases such as bigrams. Phrases may be idiomatic, in which case their polarity cannot be inferred from the sentiment of their constituent words, e.g., black market, on track, and let down. The sentiment of most phrases, however, is compositional, namely, it can be determined from the interaction between their constituents. Sentiment composition is a fundamental problem in sentiment analysis. It involves a variety of se- mantic phenomena, the most studied of which are valence shifters (Polanyi and Zaenen, 2004) that re- verse sentiment polarity, e.g., not happy, hardly helpful, decrease unemployment, or change its intensity (deeply/somewhat disappointed). Sentiment composition also needs to resolve the polarity of phrases containing opposing polarities, such as cure cancer, sophisticated crime, and fake success. Another, less studied type of sentiment composition is expressions such as high income, long queue, and fast learner, that include gradable adjectives. The sentiment of these expressions cannot be derived from the senti- ment of their individual words (none of which have clear sentiment in the above examples). While they can be treated as fixed expressions, similar to idioms, it is beneficial to exploit their semantic composi- tionality. For example, for a given word w, the polarity of high w, higher w, and increasing w is likely to be the same, and opposite to the polarity of low/lower w. This allows more robust sentiment learning, as well as sentiment inference for such expressions. Previous approaches for sentiment composition relied on hand-written lists of negation words and composition rules. Such knowledge is hard to acquire and is typically incomplete. Other approaches aim to learn sentiment composition from sentiment-labeled texts. However, sentiment-labeled texts might not be available for certain domains or languages. In this work we propose a new approach for learning sentiment composition. We define a set of sentiment composition classes that cover a variety of sentiment composition processes, and propose a novel method for automatic acquisition of lexicons …

Content truncated to 500 words