Neural Network Models For User Attribute Extraction From Dialogues

Abstract

In this paper, we present an approach to extracting user attributes represented in the form of Subject-Predicate-Object (SPO) triplets from utterances in dialogues. We treated user attribute extraction as a two-tasks process: 1) Predicate Classification and 2) Entity Generation. The predicate classification task is to determine predicates triggered in a sentence, and the entity generation task will generate related subjects and objects for each triggered predicate. Since multiple predicates may be triggered in one sentence, we formalized the predicate classification task as a multi-label classification problem. We fine-tuned a pre-trained BERT model to solve the problem. We formalized the entity generation task as a sequence-to-sequence problem and fine-tuned the T5 model on the task-specific data to create an entity generation model. Experimental results showed that our proposed method out-performs a state-of-the-art baseline model by a large margin.