Fine-Grained Contextual Predictions for Hard Sentiment Words

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1210–1215, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Fine-Grained Contextual Predictions for Hard Sentiment Words Sebastian Ebert and Hinrich Sch¨utze Center for Information and Language Processing University of Munich, Germany ebert@cis.lmu.de, inquiries@cislmu.org Abstract We put forward the hypothesis that high- accuracy sentiment analysis is only pos- sible if word senses with different polar- ity are accurately recognized. We pro- vide evidence for this hypothesis in a case study for the adjective “hard” and propose contextually enhanced sentiment lexicons that contain the information necessary for sentiment-relevant sense disambiguation. An experimental evaluation demonstrates that senses with different polarity can be distinguished well using a combination of standard and novel features. 1 Introduction This paper deals with fine-grained sentiment anal- ysis. We aim to make three contributions. First, based on a detailed linguistic analysis of contexts of the word “hard” (Section 3), we give evidence that highly accurate sentiment analysis is only pos- sible if senses with different polarity are accu- rately recognized. Second, based on this analysis, we propose to return to a lexicon-based approach to sentiment analysis that supports identifying sense distinc- tions relevant to sentiment. Currently available sentiment lexicons give the polarity for each word or each sense, but this is of limited utility if senses cannot be automatically identified in context. We extend the lexicon-based approach by introducing the concept of a contextually enhanced sentiment lexicon (CESL). The lexicon entry of a word w in CESL has three components: (i) the senses of w; (ii) a sentiment annotation of each sense; (iii) a data structure that, given a context in which w oc- curs, allows to identify the sense of w used in that context. As we will see in Section 3, the CESL sense inventory – (i) above – should be optimized for sentiment analysis: closely related senses with the same sentiment should be merged whereas subtle semantic distinctions that give rise to different po- larities should be distinguished. The data structure in (iii) is a statistical classi- fication model in the simplest case. We will give one other example for (iii) below: it can also be a set of centroids of context vector representations, with a mapping of these centroids to the senses. If sentiment-relevant sense disambiguation is the first step in sentiment analysis, then power- ful contextual features are necessary to support making fine-grained distinctions. Our third con- tribution is that we experiment with deep learn- ing as a source of such features. We look at two types of deep learning features: word em- beddings and neural network language model pre- dictions (Section 4). We show that deep learn- ing features significantly improve the accuracy of context-dependent polarity classification (Sec- tion 5). 2 Related work Initial work on sentiment analysis was either based on sentiment lexicons that listed words as posi- tive or negative sentiment indicators (e.g., Turney (2002), Yu and Hatzivassiloglou (2003)), on statis- tical classification approaches that …

Content truncated to 500 words