Modelling the Interpretation of Discourse Connectives by Bayesian Pragmatics

Abstract

We propose a framework to model human comprehension of discourse connectives. Following the Bayesian pragmatic paradigm, we advocate that discourse connectives are interpreted based on a simulation of the production process by the speaker, who, in turn, considers the ease of interpretation for the listener when choosing connectives. Evaluation against the sense annotation of the Penn Discourse Treebank confirms the superiority of the model over literal comprehension. A further experiment demonstrates that the proposed model also improves automatic discourse parsing.

A Psycholinguistic Model for the Marking of Discourse Relations

“…In contrast to Yung et al (2016), who applied the listener model of RSA to model the human comprehension of discourse connectives, this work applies the speaker model of pragmatic speaker S 1 to model the production of discourse structure. Specifically, the proposed model predicts whether the speaker will use an explicit or implicit DC given the discourse relation to be conveyed.…”

“…We start by explaining the key elements of the RSA model. Then, in order to provide a full picture of the application of RSA to discourse processing, we also give a brief account of discourse relation interpretation using the listener model of RSA, as described in Yung et al (2016). This is followed by the details of our proposed marking model, which predicts the marking of a discourse relation produced by a speaker, and is based on the speaker model of RSA.…”

“…Pragmatic listener L 1 reasons about the pragmatic speaker S 1 , who in turn reasons about literal listener L 0 . Pragmatic listeners and speakers at higher levels (e.g., L 2 and S 2 ) reason with more iterations, but previous studies demonstrate that one level of reasoning is robust for modeling a human's interpretation (Goodman and Stuhlmüller, 2013;Lassiter and Goodman, 2013;Yung et al, 2016). Our proposed model for DC production is based on pragmatic speaker S 1 , who reasons about literal listener L 0 .…”

“…Instead, they reason about how likely it is that the speaker will use that utterance, in the current context, based on the informativeness and production effort of the utterance. To evaluate this claim, Yung et al (2016) compared the DC interpretation by the literal listener (based on the probability estimates of P L 0 , Equation 6), and the pragmatic listeners (based on the probability estimates of P L 1 and P L 2 , Equation 8). Their experiments based on the PDTB find that discourse sense predictions made by the pragmatic listeners outperform predictions by the literal listener, providing support to the RSA model.…”

“…Following Yung et al (2016), we define context C as the surrounding discourse relations, which are also used as features in discourse planning for natural language generation (Biran and McKeown, 2015). Specifically, the discourse contexts are: the full discourse sense annotated in PDTB (S), the fourway top level sense (TS), the form of discourse presentation (F) such as "explicit" or "implicit", the combination of sense and form (SF), and the combination of top sense and form (TSF).…”