Abstract: The concept of prominence has been used to decode pronoun resolution. Here, we would like to explore if LLM is able to mimic human prominence behaviour in pronoun resolution. We have focused on Portuguese; it allows speakers to drop subject pronouns, and human interpretations of null and overt pronouns vary. We used BERTimbau question-answer model to generate responses for stimuli used in an experiment by Fernandes et al. (2018). The results show that some aspects of prominence-based phenomena in pronoun resolution are not replicated by the LLM. However, examination of LLM confidence scores offers hope that the gap between human and LLM responses may be bridged by larger training corpora of language-specific data.
Paper Type: Short
Research Area: Discourse and Pragmatics
Research Area Keywords: discourse prominence, pronouns, pronoun resolution, LLM
Contribution Types: Data analysis
Languages Studied: European Portuguese
Submission Number: 8050
Loading