Towards Bootstrapping a Polarity Shifter Lexicon using Linguistic Features

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 624–633, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Towards Bootstrapping a Polarity Shifter Lexicon using Linguistic Features Marc Schulder*, Michael Wiegand*, Josef Ruppenhofer†, Benjamin Roth‡ * Spoken Language Systems, Saarland University † Institute for German Language, Mannheim ‡ Center for Information and Language Processing, LMU Munich marc.schulder@lsv.uni-saarland.de michael.wiegand@lsv.uni-saarland.de ruppenhofer@ids-mannheim.de beroth@cis.uni-muenchen.de Abstract We present a major step towards the cre- ation of the first high-coverage lexicon of polarity shifters. In this work, we boot- strap a lexicon of verbs by exploiting var- ious linguistic features. Polarity shifters, such as abandon, are similar to negations (e.g. not) in that they move the polarity of a phrase towards its inverse, as in abandon all hope. While there exist lists of nega- tion words, creating comprehensive lists of polarity shifters is far more challeng- ing due to their sheer number. On a sam- ple of manually annotated verbs we exam- ine a variety of linguistic features for this task. Then we build a supervised classifier to increase coverage. We show that this approach drastically reduces the annota- tion effort while ensuring a high-precision lexicon. We also show that our acquired knowledge of verbal polarity shifters im- proves phrase-level sentiment analysis. 1 Introduction We present an approach towards bootstrapping a lexicon of polarity shifters. Polarity shifters are content words that have semantic properties sim- ilar to negation. For example, the negated state- ment in (1) involving the negation word not can also be expressed by the verbal shifter fail in (2). (1) Peter did not pass the exam. (2) Peter failedshifter to pass the exam. Similarly, shifting is also caused by nouns (e.g. downfall) and adjectives (e.g. devoid). Polarity shifters are important for various tasks in NLP, such as relation extraction (Sanchez- Graillet and Poesio, 2007), recognition of textual entailment (Harabagiu et al., 2006) and particu- larly sentiment analysis (Wiegand et al., 2010). Similarly to negation words, they may cause the polarity of a statement to shift. Even though (3) contains the positive polar expression scholarship, the overall polarity of the sentence is negative. (4) conveys positive polarity despite the presence of the negative polar expression pain. (3) She was [deniedshifter the [scholarship]+]−. (4) The new treatment has [alleviatedshifter her [pain]−]+. Although there has been significant research on polarity shifting in sentiment analysis (Wiegand et al., 2010), this work has focused on the pres- ence of negation words. Negation words (no, not, never, etc.) are typically function words, so only a few exist. Polarity shifters are content words, of which there are a lot more. For instance, WordNet (Miller et al., 1990) contains over 10k verbal, 20K adjectival and 110K nominal lemmas. An exhaus- tive manual annotation would be far too costly. To reduce cost, we introduce a bootstrapping approach for the acquisition of polarity shifters. In this work we focus exclusively on verbs. As the main predicates of phrases they tend to have larger scopes than nouns and …

Content truncated to 500 words