Harnessing Privileged Information for Hyperbole Detection

Harnessing Privileged Information for Hyperbole Detection Rhys Biddle♠♣ Maciej Rybi´nski♣ Qian Li♠ C´ecile Paris♣ Guandong Xu♠ ♠Advanced Analytics Institute, University of Technology Sydney, Australia ♣CSIRO Data61, Sydney, Australia rhys.biddle@student.uts.edu.au {firstname.lastname}@uts.edu.au {firstname.lastname}@csiro.au Abstract The detection of hyperbole is an important stepping stone to understanding the intentions of a hyperbolic utterance. We propose a model that combines pre-trained language models with privileged information for the task of hy- perbole detection. We also introduce a suite of behavioural tests to probe the capabilities of hyperbole detection models across a range of hyperbole types. Our experiments show that our model improves upon baseline models on an existing hyperbole detection dataset. Prob- ing experiments combined with analysis using local linear approximations (LIME) show that our model excels at detecting one particular type of hyperbole. Further, we discover that our experiments highlight annotation artifacts introduced through the process of literal para- phrasing of hyperbole. These annotation arti- facts are likely to be a roadblock to further im- provements in hyperbole detection. 1 Introduction The analysis of figurative language by Natural Lan- guage Processing (NLP) systems is a challenge con- fronting researchers and practitioners (Reyes and Rosso, 2014; Rai and Chakraverty, 2020). Hyper- bole is a common type of figurative language that is defined by an intentionally excessive contrast be- tween utterance meaning and reality along a seman- tic scale to convey an evaluation (e.g., ‘my bedroom is the size of a postage stamp’) (McCarthy and Carter, 2004; Mora, 2009; Claridge, 2010; Carston and Wearing, 2015; Burgers et al., 2016). The de- tection of hyperbole has proven to be a challenging problem for NLP systems, much like the detec- tion of other figures of speech (Troiano et al., 2018; Kong et al., 2020; Abulaish et al., 2020). The evalu- ative nature of hyperbole motivates the importance of understanding hyperbole for affective computing applications (e.g., sentiment analysis). Learning under Privileged Information (LUPI) is a learning paradigm that involves providing ad- ditional information during training to help teach a model to learn a particular phenomenon (Pechy- ony and Vapnik, 2010). The source and type of privileged information (PI) varies depending on ap- plication, such as a list of ingredients present in an image to help teach a computer vision model to detect food in images (Meng et al., 2019), or the human ratings of various aesthetic categories of images for automated assessment of aesthetic photo quality (Shu et al., 2020). We propose to use literal paraphrases of hyperbole as a source of PI for hyperbole detection. We hypothesise that this information will help a model to learn the ex- cessive contrast within a particular hyperbole (e.g., ‘my head is exploding right now’ → ‘my head is hurting right now’). Our contributions in this paper are as follows; (1) We propose a method for hyperbole detection based on the injection of PI; (2) We introduce Hyper- Probe, a suite of behavioural tests for hyperbole detection models; (3) We reveal that annotation artifacts are a potential roadblock for progress on hyperbole …

Content truncated to 500 words