QALinkPlus: Text Enrichment with Q&A data

Yandong Sun, Yixuan Tang, Anthony KH Tung

Abstract:

Text enrichment, the task of augmenting textual content by incorporating supplementary information to
bridge knowledge gaps and enhance reader engagement, is a critical aspect of information retrieval. This
study focuses on leveraging question answering datasets, such as Natural Questions and SQuAD, which
contain human-validated content from diverse domains as valuable knowledge sources. While QA datasets
hold promise for addressing informational needs, existing approaches, like employing dense retrieval for text
enrichment, often result in QA pairs that may lack relevance, diversity, or inherent interest. To address these
challenges, our paper proposes a novel graph-based method for text enrichment using QA pairs. We construct
an entity co-occurrence graph derived from QA datasets and derive context-QA-specific subgraphs. Through
rule-based path analysis, we develop an interpretable scoring system to assess the relevance and engagement
value of each QA pair. By intelligently re-ranking QA pairs with our scoring system, our method delivers
enriched text that fills knowledge gaps and captivates readers, thus improving the overall reading experience.
This framework is not only effective in text enrichment tasks, but it also offers advantages for personalization
and personal data management.