BERTective: Language Models and Contextual Information for Deception Detection

Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics, pages 2699–2708 April 19 - 23, 2021. ©2021 Association for Computational Linguistics 2699 BERTective: Language Models and Contextual Information for Deception Detection Tommaso Fornaciari Bocconi University fornaciari@unibocconi.it Massimo Poesio Queen Mary University of London m.poesio@qmul.ac.uk Federico Bianchi Bocconi University f.bianchi@unibocconi.it Dirk Hovy Bocconi University dirk.hovy@unibocconi.it Abstract Spotting a lie is challenging but has an enor- mous potential impact on security as well as private and public safety. Several NLP meth- ods have been proposed to classify texts as truthful or deceptive. In most cases, however, the target texts’ preceding context is not con- sidered. This is a severe limitation, as any com- munication takes place in context, not in a vac- uum, and context can help to detect deception. We study a corpus of Italian dialogues contain- ing deceptive statements and implement deep neural models that incorporate various linguis- tic contexts. We establish a new state-of-the- art identifying deception and find that not all context is equally useful to the task. Only the texts closest to the target, if from the same speaker (rather than questions by an interlocu- tor), boost performance. We also find that the semantic information in language models such as BERT contributes to the performance. However, BERT alone does not capture the im- plicit knowledge of deception cues: its contri- bution is conditional on the concurrent use of attention to learn cues from BERT’s represen- tations. 1 Introduction “The sky is bright green” is easily identified as false statement under normal circumstances. However, following “Look at this surreal painting,” the assess- ment changes. Spotting falsehoods and deception is useful in many personal, economic, legal, and political situations – but it is also extremely compli- cated. However, the reliability of communication is the basis of the social contract, with implications on personal, economic, legal, and political levels. There has been a growing interest in automatic de- ception detection from academia and industry in recent years (see section 9). One of the main research lines tries to increase the collection of deception cues in terms of number and variety. For example, several successful studies show how to exploit multi-modal signals, jointly an- alyzing verbal, video, and audio data (P´erez-Rosas et al., 2015). For the same reason, several early studies tried to identify deception cues through manual feature annotation, like irony or ambigu- ity (Fitzpatrick and Bachenko, 2012). While these approaches offer a broad and interpretable descrip- tion of the phenomenon, their main limitation lies in data collection and preprocessing difficulty. Surprisingly, so far, little attention has been paid to expanding the targets’ linguistic context, which is the easiest source of additional cues and data. Even in dialogues, which by definition are exchanges between different speakers/writers, the main focus is typically on the target text. None consider the preceding statements, be they issued by the same speaker of an interlocutor. We hypothesize that linguistic context can be useful for text classification. Based on a …

Content truncated to 500 words