Is Incoherence Surprising? Targeted Evaluation of Coherence Prediction from Language Models

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4164–4173 June 6–11, 2021. ©2021 Association for Computational Linguistics 4164 Is Incoherence Surprising? Targeted Evaluation of Coherence Prediction from Language Models Anne Beyer Sharid Loáiciga Computational Linguistics, Department of Linguistics, University of Potsdam, Germany {anne.beyer, loaicigasanchez, david.schlangen}@uni-potsdam.de David Schlangen Abstract Coherent discourse is distinguished from a mere collection of utterances by the satisfac- tion of a diverse set of constraints, for example choice of expression, logical relation between denoted events, and implicit compatibility with world-knowledge. Do neural language models encode such constraints? We design an extend- able set of test suites addressing different as- pects of discourse and dialogue coherence. Un- like most previous coherence evaluation stud- ies, we address specific linguistic devices be- yond sentence order perturbations, allowing for a more fine-grained analysis of what constitutes coherence and what neural models trained on a language modelling objective do encode. Ex- tending the targeted evaluation paradigm for neural language models (Marvin and Linzen, 2018) to phenomena beyond syntax, we show that this paradigm is equally suited to evaluate linguistic qualities that contribute to the notion of coherence. 1 Introduction Statistical models trained on large amounts of data using the language modelling objective (predicting words in context) have shown to pick up an intrigu- ing amount of implicit knowledge about other tasks, for example syntactic knowledge (Warstadt et al., 2020; Hu et al., 2020) or world knowledge (Trinh and Le, 2019; Tamborrino et al., 2020). They have also been shown to exhibit, within these tasks, inter- esting divergences from expectation and sensitivity to confounding factors (e.g. McCoy et al. (2019)). Inspired by the recently released SyntaxGym (Gauthier et al., 2020), which enables specific and standardised evaluation of syntactic knowledge en- coded in such models, we explore whether similar methods can be applied to the study of discourse knowledge or coherence, i.e., constraints acting across sentence boundaries, as illustrated in (1) (where "#" marks the less acceptable variant). (1) a. #The lone ranger rode off into the sun- set. Then he jumped on his horse. b. The lone ranger jumped on his horse. Then he rode into the sunset. A common approach to coherence evaluation con- sists in shuffling the sentence order of a text, thereby creating incoherent text samples that need to be discriminated from the original (Barzilay and Lapata, 2008). While this approach to creating in- coherent test data is intuitive enough, recent studies suggest that it paints only a partial picture of what constitutes coherence (Lai and Tetreault, 2018; Mo- hammadi et al., 2020; Pishdad et al., 2020). It does not pinpoint the qualities that make the shuffled text incoherent, it does not tell us which linguistic devices are at fault, emphasising the need to move beyond this technique. This paper aims to add to the growing body of research stressing the need for more qualitative evaluations of text coherence (See et al., 2019; Mohammadi et al., 2020; Pishdad …

Content truncated to 500 words