Learning Interpretable Negation Rules via Weak Supervision at Document Level: A Reinforcement Learning Approach

Proceedings of NAACL-HLT 2019, pages 407–413 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 407 Learning Interpretable Negation Rules via Weak Supervision at Document Level: A Reinforcement Learning Approach Nicolas Pr¨ollochs Oxford-Man Institute University of Oxford nicolas.prollochs@eng.ox.ac.uk Stefan Feuerriegel ETH Zurich sfeuerriegel@ethz.ch Dirk Neumann University of Freiburg dirk.neumann@is.uni-freiburg.de Abstract Negation scope detection is widely performed as a supervised learning task which relies upon negation labels at word level. This suffers from two key drawbacks: (1) such granular an- notations are costly and (2) highly subjective, since, due to the absence of explicit linguistic resolution rules, human annotators often dis- agree in the perceived negation scopes. To the best of our knowledge, our work presents the first approach that eliminates the need for word-level negation labels, replacing it in- stead with document-level sentiment annota- tions. For this, we present a novel strategy for learning fully interpretable negation rules via weak supervision: we apply reinforcement learning to find a policy that reconstructs nega- tion rules from sentiment predictions at docu- ment level. Our experiments demonstrate that our approach for weak supervision can effec- tively learn negation rules. Furthermore, an out-of-sample evaluation via sentiment analy- sis reveals consistent improvements (of up to 4.66 %) over both a sentiment analysis with (i) no negation handling and (ii) the use of word-level annotations from humans. More- over, the inferred negation rules are fully in- terpretable. 1 Introduction Negations are a frequently utilized linguistic tool for expressing disapproval or framing negative content with positive words. Neglecting nega- tions can lead to false attributions (Morante et al., 2008) and, moreover, impair accuracy when an- alyzing natural language; e. g., in information re- trieval (Cruz D´ıaz et al., 2012; Rokach et al., 2008) and especially in sentiment analysis (Cruz et al., 2015; Wiegand et al., 2010). Hence, even simple heuristics for identifying negation scopes can yield substantial improvements in such cases (Jia et al., 2009). Negation scope detection is sometimes imple- mented as unsupervised learning (e. g., Pr¨ollochs et al., 2016), while a better performance is com- monly achieved via supervised learning (see our supplements for a detailed overview): the resulting models thus learn to identify negation scopes from word-level annotations (e. g., Li and Lu, 2018; Re- itan et al., 2015). We argue that this approach suffers from inherent drawbacks. (1) Such gran- ular annotations are costly and, especially at word level, a considerable number of them is needed. (2) Negation scope detection is highly subjec- tive (Councill et al., 2010). Due to the absence of explicit linguistic rules for resolutions, exist- ing corpora often come with annotation guidelines (Morante and Blanco, 2012; Morante and Daele- mans, 2012). Yet there are considerable differ- ences: some corpora were labeled in a way that negation scopes consist of single text spans, while others allowed disjoint spans (Fancellu et al., 2017). More importantly, given the absence of universal rules, human annotators largely disagree in their perception of what words should be la- beled as negated. …

Content truncated to 500 words