Deception detection, new paper at EACL2021!

Glad to announce that the paper

“BERTective, or detective BERT: Language Models and Contextual Information for Deception Detection”

with Dirk Hovy, Federico Bianchi and Massimo Poesio has been accepted at EACL2021!

Linguistic cues of deception with Sampling and Occlusion (SOC) algorithm (Jin et al., 2019)

Overview

How do you spot a lie? It is a challenging task, with potential impact on security and private and public safety. Recent successful models look for different cues of deception, following multi-modal approaches when possible.

However, typically the focus is on the single communicative acts, overlooking the preceding parts of the dialogue. This is a limitation, as any communication takes place in a context, not in the vacuum.

Also, most studies rely on data collected in laboratory or online games/simulations, which reduces their findings' generalisability to high-stakes scenarios.

We study deception on a corpus of deceptive statements in natural environment, and for the first time

we train deep neural models that incorporate information from the texts' linguistic context.

We establish a new state-of-the-art in identifying deception, and discuss how the contextual information can be exploited by neural models.%, both those trained from scratch and those based on transfer learning strategies.

Postdoc

Direttore Tecnico Superiore della Polizia di Stato