Killing Four Birds with Two Stones: Multi-Task Learning for Non-Literal Language Detection

Proceedings of the 27th International Conference on Computational Linguistics, pages 1558–1569 Santa Fe, New Mexico, USA, August 20-26, 2018. 1558 Killing Four Birds with Two Stones: Multi-Task Learning for Non-Literal Language Detection Erik-Lˆan Do Dinh, Steffen Eger, and Iryna Gurevych Ubiquitous Knowledge Processing Lab (UKP-TUDA) Department of Computer Science, Technische Universit¨at Darmstadt http://www.ukp.tu-darmstadt.de Abstract Non-literal language phenomena such as idioms or metaphors are commonly studied in isolation from each other in NLP. However, often similar definitions and features are being used for differ- ent phenomena, challenging the distinction. Instead, we propose to view the detection problem as a generalized non-literal language classification problem. In this paper, we investigate multi- task learning for related non-literal language phenomena. We show that in contrast to simply joining the data of multiple tasks, multi-task learning consistently improves upon four metaphor and idiom detection tasks in two languages, English and German. Comparing two state-of-the- art multi-task learning architectures, we also investigate when soft parameter sharing and learned information flow can be beneficial for our related tasks. We make our adapted code publicly available1. 1 Introduction As Lakoff and Johnson (1980) argued, metaphorical concept mappings, often from concrete to more abstract concepts, are ubiquitous in everyday life, thus they are ubiquitous in written texts. But the same is true for other kinds of so-called non-literal language, e.g., for idioms. Boundaries between these instances of non-literality are not always clear: for instance, “[...] a swathe of sunlight lay across the red-tiled floor.”2 can be classified as a metaphor because of its non-lexicalized figurative meaning. Few would dispute the idiomaticity of “kicking the bucket”, as the non-literal meaning in this multi-word expression is largely conventionalized. However, in the sentence “One approach would be to draw the line by reference [...]” the expression “draw the line” could be classified as either metaphorical (because it still evokes the literal senses of its constituents) or idiomatic (as it is a fixed expression with lexicalized figurative sense). Much effort has been spent on the detection of metaphors, idioms, and general non-literal language use (Shutova, 2015). However, because of the named vague and subjective nature of these phenomena, a multitude of datasets using differing definitions has emerged in the process. Even when datasets address the same aspect of non-literality, they may use diverging or underspecified definitions; compare, e.g., the guidelines for annotating metaphors of Tsvetkov et al. (2014) (“[...] all words that, in your opinion, are used non-literally [...]”) and Mohammad et al. (2016) (“more complex; more distant from our senses; more abstract; more vague; ...”). The fuzziness of non-literality has two natural consequences: (1) training data is sparse, because dif- ferent researchers may use diverging definitions, and hence may annotate different things rather than extend “the same story”; (2) high-quality training data is costly to obtain because there may be consid- erable disagreement among crowd-workers and even trained experts regarding the labels for different instances of (non-)literality. In this work, we address two research questions: This work is licensed under a Creative …

Content truncated to 500 words