Towards a noise-tolerant, representation-independent mechanism for argument interpretation

Abstract

We describe a mechanism for the interpretation of arguments, which can cope with noisy conditions in terms of wording, beliefs and argument structure. This is achieved through the application of the Minimum Message Length Principle to evaluate candidate interpretations. Our system receives as input a quasi-Natural Language argument, where propositions are presented in English, and generates an interpretation of the argument in the form of a Bayesian network (BN). Performance was evaluated by distorting the system's arguments (generated from a BN) and feeding them to the system for interpretation. In 75% of the cases, the interpretations produced by the system matched precisely or almost-precisely the representation of the original arguments.

Incorporating a User Model into an Information Theoretic Framework for Argument Interpretation

“…The previous version of the system was evaluated by making it generate synthetic arguments, and then produce interpretations of its own arguments [4]. The results of this evaluation were encouraging, with the system generating plausible interpretations of its own arguments in 75% of the 5400 tried cases.…”

“…As indicated in Section 1, our research builds on work described in [3,4]. In this paper, we apply a principled approach based on the MML criterion [1] to select an interpretation for unrestricted arguments, instead of the heuristics used in [3] to select an interpretation for single-proposition rejoinders.…”

“…This may require adding propositions and relations to the argument structure proposed by the user, deleting relations from the user's argument, or postulating degrees of belief in the propositions in the user's argument which differ from those stated by the user. The procedure for proposing interpretations is described in [4]. Here we provide a brief overview, focusing on the structure of the interpretations (the beliefs in the propositions are obtained by performing Bayesian propagation).…”

“…In this section, we first review the MML encoding of an NL argument (a detailed description of this encoding appears in [4]). We then discuss the incorporation of a user model into this formalism.…”

“…Here BIAS interprets user arguments of arbitrary complexity, which may differ from BIAS' beliefs and inference patterns. The basic discourse-interpretation mechanism relies on the Minimum Message Length (MML) Principle [1] to evaluate candidate discourse interpretations [4]. The main contribution of this paper is in its principled incorporation of a user model into the discourse-interpretation mechanism.…”