Extending the Comparative Argumentative Machine: Multilingualism and Stance Detection

Abstract

AbstractThe comparative argumentative machine CAM can retrieve arguments that answer comparative questions—questions that ask which of several to-be-compared options should be favored in some scenario. In this paper, we describe how we equipped CAM with a better answer stance detection (i.e., a better detection of which option “wins” a comparison) and with system variants to support non-English requests. As for the improved answer stance detection, we develop RoBERTa-based approaches and experimentally show them to be more effective than previous feature-based and LLM-based stance detectors. As for the multilingualism, in a proof of concept, we compare two approaches to support Russian requests and answers: (1) translating the original English CAM data and (2) using an existing replica of CAM on native Russian data. Comparing the translation-based and the replica-based CAM variants in a user study shows that combining their answers seems to be the most promising. For individual questions, the retrieved arguments of the two variants are often different and of quite diverse relevance and quality. As a demonstrator, we deploy a first multilingual CAM version that combines translation-based and replica-based outputs for English and Russian and that can easily be extended to further languages.