Improving Question Intent Identification by Exploiting Its Synergy With User Age
No hay miniatura disponible
Archivos
Fecha
2023
Autores
Profesor/a Guía
Facultad/escuela
Idioma
en
Título de la revista
ISSN de la revista
Título del volumen
Editor
Institute of Electrical and Electronics Engineers Inc.
Nombre de Curso
Licencia CC
CC BY-NC-ND 4.0 ATTRIBUTION-NONCOMMERCIAL-NODERIVATIVES 4.0 INTERNATIONAL Deed
Licencia CC
https://creativecommons.org/licenses/by-nc-nd/4.0/
Resumen
At their heart, community Question-Answering (cQA) services are social networks that allow their members to prompt any kind of question expecting different answers produced by several community peers. Most of previous research on cQA has shown that questions can reflect two intents: learning information and starting a conversation. The purpose of this research is investigating the intrinsic relationship between models predicting question intent and user age. And if this relatedness can assist in overcoming one of the chief obstacles when constructing effective question intent recognizers: the scarcity of annotated data. The method adopted in this work involves addressing question intent recognition in a Multi-Task (MT) learning setting, where asker age identification is used as its auxiliary task. In other words, we exploit their task synergy by integrating both training signals with the aim of boosting the classification rate of question intent. Since MT learning is regarded as fruitful when a target task is improved wrt. single-task models, in our experiments, we compare four frontier frameworks with several state-of-the-art single-task neural network classifiers. In brief, our results show that a MT implementation of T5 yielded an increase of at least 10% over the best single-task models, when trained on full questions. Our experimental results also unveil that extra substantial improvements can be obtained by adjusting its parameters. All in all, we conclude that both variables are inherently related. Last but not least, we also make available a new question set labelled with the age of their askers and their intents with the hope of encouraging the research of MT learning into cQA tasks. © 2023 The Authors.
Notas
Indexación: Scopus.
Palabras clave
Multi-task Learning, Question Analysis, Question Answering, Shared Transformers
Citación
IEEE Access. Volume 11, Pages 112044 - 112059. 2023
DOI
10.1109/ACCESS.2023.3322457