PAED: Zero-Shot Persona Attribute Extraction in Dialogues

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 9771–9787 July 9-14, 2023 ©2023 Association for Computational Linguistics PAED: Zero-Shot Persona Attribute Extraction in Dialogues Luyao Zhu, Wei Li, Rui Mao, Vlad Pandelea and Erik Cambria Nanyang Technological University, Singapore {luyao001, wei008}@e.ntu.edu.sg, {rui.mao, vlad.pandelea, cambria}@ntu.edu.sg Abstract Persona attribute extraction is critical for per- sonalized human-computer interaction. Dia- logue is an important medium that communi- cates and delivers persona information. Al- though there is a public dataset for triplet-based persona attribute extraction from conversations, its automatically generated labels present many issues, including unspecific relations and incon- sistent annotations. We fix such issues by lever- aging more reliable text-label matching crite- ria to generate high-quality data for persona attribute extraction. We also propose a con- trastive learning- and generation-based model with a novel hard negative sampling strategy for generalized zero-shot persona attribute ex- traction. We benchmark our model with state- of-the-art baselines on our dataset and a public dataset, showing outstanding accuracy gains. Our sampling strategy also exceeds others by a large margin in persona attribute extraction. 1 Introduction Persona attribute extraction in dialogues (PAED) is a crucial task for persona-based dialogue sys- tems (Zheng et al., 2020; Cao et al., 2022). It can extract persona attribute information from conver- sations. Then, a dialogue system can use extracted persona attributes to generate personalized, user- preference-aware responses to user queries. Previ- ous works define the task as a sentence-level (Da- niulaityte et al., 2016) or utterance-level (Gu et al., 2021) classification task. A model learns to classify whether a text contains persona information or not. However, the identified persona-informed texts are still unstructured, resulting in the lower utility of the extracted information in downstream dialogue systems, e.g., irrelevant contexts and different rep- resentations towards the same persona attribute. Thus, we define persona attribute extraction as a triplet extraction task. A model should extract a subject, an object, and the persona-relevant relation linking the subject and object from utterances. The extracted attributes should be in the form of triplets (s, r, o), where the relation (r) indicates the persona attribute type of the subject (s) towards the object (o). Although the existing relation triplet extraction (RTE) task aims to extract triplets from documents (Li and Ji, 2014; Chia et al., 2022), its framework cannot be directly transferred to PAED task because the sentences in documents de- scribe the facts or knowledge in the real world and each pair of entities in triplets can be connected by very limited relations. For example, entities ‘Eiffel Tower’ and ‘France’ are very likely con- nected by relation located_in in traditional RTE datasets (Chen and Li, 2021). But the subject and object in dialogues can be linked by many rela- tions, causing hard sample problems, e.g., the rela- tion between ‘I’ and ‘my father’ may be live_with, raised_by, get_along, etc. Hence it is essential to formulate a framework for PAED, which is capable of processing hard samples. To the best of …

Content truncated to 500 words