Multilingual Evaluation of Referential Disambiguation in Generative AI using Winograd Schema
DOI:
https://doi.org/10.4151/S0718-09342025011901274Keywords:
Referential disambiguation, Winograd schema, Computational Linguistics, Generative AIAbstract
The rise of chatbots, such as ChatGPT, opens an interesting path for analyzing Natural Language Generation. Therefore, this experimental study examines the referential disambiguation capability of several chatbots. The research is approached from the perspective of Computational Linguistics to analyze one of the most complex tasks for Natural Language Processing. Specifically, we evaluate the effectiveness of generative AI systems in the interpretation of sentences containing referential ambiguity in English, Spanish and Valencian. We use Winograd's scheme as a starting point to create multilingual datasets with examples of different levels of referential ambiguity. The relevance of this work consists of the analysis of referential disambiguation in depth and its multilingual perspective. The experimentation is carried out by two tests; in the first one, sentences containing a disambiguating word that guides the interpretation are used, and in the second one, this word is eliminated. The results show that without a disambiguating word none of the conversational agents exceeds 56% accuracy in any of the languages. Thus, it is concluded that the referential disambiguation task is difficult for generative AI to solve in cases where there is insufficient contextual information.
Downloads
Published
How to Cite
Issue
Section
License
Copyright (c) 2025 Revista Signos. Estudios de Lingüística

This work is licensed under a Creative Commons Attribution 4.0 International License.
Copyright agreement:
Authors who have a manuscript accepted for publication in this journal agree to the following terms:
Authors will retain their copyright and grant the journal the right of first publication of their work by means of this copyright agreement document, which is subject to the Creative Commons Acknowledgment License that allows third parties to share the work provided that its author and first publication in this journal are indicated.
Authors may adopt other non-exclusive license agreements for distribution of the published version of the work (e.g., depositing it in an institutional repository or publishing it in a monographic volume) as long as the initial publication in this journal is indicated.
Authors are allowed and encouraged to disseminate their work via the internet (e.g., in institutional publications or on their website) before and during the submission process, which can lead to interesting exchanges and increase citations of the published work (read more here).