Wenjie Yang, Xiaojuan Ma.

Consumers assess and select products and services based on a combination of objec- tive factual attributes (e.g., price) and subjective experiential factors. For example, when choosing a restaurant, users often focus on the food quality and ambiance. State-of-the-art search services provide powerful interfaces for filtering objective properties but struggle to support users through the process of considering experiential factors. One of the key reasons for this discrepancy is that the objective properties are clearly represented by a database schema, but there is no such equivalent for experiential properties, which are vaguer by nature. This paper introduces CoNex, a pipeline for building knowledge graphs (KGs) that describe concepts concerning consumers’ experiences in a given domain and the relationships between them. CoNex begins by harvesting experience-related concepts on a domain-specific corpus and then discovering experiential connections between them. CoNex further expands its knowledge coverage by a pre-trained language model fine-tuned via data from hybrid sources. Our experiments demonstrate that the KGs constructed by CoNex accurately reflect the experiential relationships between concepts as judged by humans. Finally, we show the effectiveness of using these KGs as tools to improve the performance of an experience-oriented search task.

Citation

@inproceedings{
yang2022building,
title={Building Knowledge Graphs of Experientially Related Concepts},
author={Wenjie Yang and Xiaojuan Ma},
booktitle={4th Conference on Automated Knowledge Base Construction},
year={2022}
}