Logic Against Bias: Textual Entailment Mitigates Stereotypical Sentence Reasoning

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 103–115 May 2-6, 2023 ©2023 Association for Computational Linguistics Sentiment as an Ordinal Latent Variable Niklas StoehrQ Ryan CotterellQ Aaron Schein@ QETH Zürich @The University of Chicago niklas.stoehr@inf.ethz.ch ryan.cotterell@inf.ethz.ch schein@uchicago.edu Abstract Sentiment analysis has become a central tool in various disciplines outside of natural language processing. In particular in applied and domain-specific settings with strong requirements for interpretable methods, dictionary-based approaches are still a popular choice. However, existing dictionaries are often limited in coverage, static once annota- tion is completed and sentiment scales differ widely; some are discrete others continuous. We propose a Bayesian generative model that learns a composite sentiment dictionary as an interpolation between six existing dictio- naries with different scales. We argue that sentiment is a latent concept with intrinsically ranking-based characteristics — the word “excellent” may be ranked more positive than “great” and “okay”, but it is hard to express how much more exactly. This prompts us to enforce an ordinal scale of ordered discrete sentiment values in our dictionary. We achieve this through an ordering transformation in the priors of our model. We evaluate the model intrinsically by imputing missing values in existing dictionaries. Moreover, we conduct extrinsic evaluations through sentiment classification tasks. Finally, we present two extension: first, we present a method to aug- ment dictionary-based approaches with word embeddings to construct sentiment scales along new semantic axes. Second, we demonstrate a Latent Dirichlet Allocation-inspired variant of our model that learns document topics that are ordered by sentiment. https://github.com/niklasstoehr/ ordinal-sentiment 1 Introduction Sentiment analysis is being applied in various do- mains from political science (Young and Soroka, 2012; Gründl, 2020; Widmann and Wich, 2022) to economics (Stephany et al., 2022) and computa- tional social science (West et al., 2014; Falck et al., 2020; Stoehr et al., 2021). In all of these applica- tions, there is a strong demand for domain-specific and interpretable methods (Hofman et al., 2021; Widmann and Wich, 2022) making dictionary- based sentiment analysis still a popular choice (Young and Soroka, 2012; Hoyle et al., 2019; Gründl, 2020; Friedrichs et al., 2022). Sentiment dictionaries describe a mapping be- tween word types and some form of sentiment val- ues. We consider the most general notion of senti- ment value referring to the polarity score along a positive-negative axis, instead of fine-grained emo- tion dimensions (Plutchik, 1980) or stance (Mo- hammad, 2016). Sentiment values are measured on scales of different support (§2): some dictionaries assign binary “positive” and “negative” values (Hu and Liu, 2004; Wilson et al., 2005; Stone et al., 2007). These discrete values are often falsely inter- preted as unordered, nominal categories. Other dic- tionaries have continuous scales that assign cardi- nal, floating point values (Hutto and Gilbert, 2014; Cambria et al., 2014). In this work, we propose a method for merg- ing sentiment dictionaries with different scales into a single, composite dictionary. Paying tribute to the subjective and ranking-based characteristics of sentiment, we design the …

Content truncated to 500 words