Multilingual Irony Detection with Dependency Syntax and Neural Models

Proceedings of the 28th International Conference on Computational Linguistics, pages 1346–1358 Barcelona, Spain (Online), December 8-13, 2020 1346 Multilingual Irony Detection with Dependency Syntax and Neural Models Alessandra Teresa Cignarella1,2, Valerio Basile1, Manuela Sanguinetti3, Cristina Bosco1, Paolo Rosso2 and Farah Benamara4 1. Dipartimento di Informatica, Università degli Studi di Torino, Italy 2. PRHLT Research Center, Universitat Politècnica de València, Spain 3. Dipartimento di Matematica e Informatica, Università degli Studi di Cagliari, Italy 4. IRIT-CNRS, Université de Toulouse, France {cigna,basile,bosco}@di.unito.it, manuela.sanguinetti@unica.it, prosso@dsic.upv.es, farah.benamara@irit.fr Abstract This paper presents an in-depth investigation of the effectiveness of dependency-based syntactic features on the irony detection task in a multilingual perspective (English, Spanish, French and Italian). It focuses on the contribution from syntactic knowledge, exploiting linguistic resources where syntax is annotated according to the Universal Dependencies scheme. Three distinct ex- perimental settings are provided. In the first, a variety of syntactic dependency-based features combined with classical machine learning classifiers are explored. In the second scenario, two well-known types of word embeddings are trained on parsed data and tested against gold stan- dard datasets. In the third setting, dependency-based syntactic features are combined into the Multilingual BERT architecture. The results suggest that fine-grained dependency-based syntac- tic information is informative for the detection of irony. 1 Introduction Irony is a pragmatic phenomenon that can be defined as a sort of incongruity between the literal meaning of an utterance and its intended meaning (Grice, 1975; Sperber and Wilson, 1981; Attardo, 2000). In text processing, the challenging nature of irony is mainly due to the inherent features of the phenomenon itself, that may assume a large variety of merged forms and facets. Several semantic and syntactic devices can be used to express irony (e.g., analogy, oxymoron, paradox, euphemism and rhetorical questions), and a variety of different linguistic elements can cause the incongruity, determine the clash and play the role of irony triggers within a text. Most computational research in irony detection is applied on social media texts and focuses primarily on content-based processing of the linguistic information using semantic or pragmatic devices, neglecting syntax. For example, Joshi et al. (2015) propose a method for sarcasm detection in tweets as the task of detecting incongruity between an observed and an expected word, while Karoui et al. (2015) rely on external common knowledge to the utterance shared by author and reader. However, in addition to these pragmatic devices, some hints about the role of syntax can be found in linguistic literature where the deviation from syntactic norms has been reported as a possible trigger of the phenomenon. For instance, Michaelis and Feng (2015) explore the usage of split interrogatives and of topic-comment utterances in English, arguing they both have a double function (they express two distinct communicative acts) and that one of those is precisely that of conveying an ironic meaning (e.g. “Who are you, Willy Wonka?”).1 The interpretation of an ironic utterance involves not only the knowledge of the usage of the language, inventory and rules including grammar, lexis, syntax, structure …

Content truncated to 500 words