UniK-QA: Unified Representations of Structured and Unstructured Knowledge for Open-Domain Question Answering

Abstract

We study open-domain question answering with structured, unstructured and semistructured knowledge sources, including text, tables, lists and knowledge bases. Departing from prior work, we propose a unifying approach that homogenizes all sources by reducing them to text and applies the retrieverreader model which has so far been limited to text sources only. Our approach greatly improves the results on knowledge-base QA tasks by 11 points, compared to latest graphbased methods. More importantly, we demonstrate that our unified knowledge (UniK-QA 1 ) model is a simple and yet effective way to combine heterogeneous sources of knowledge, advancing the state-of-the-art results on two popular question answering benchmarks, Nat-uralQuestions and WebQuestions, by 3.5 and 2.6 points, respectively.

TABi: Type-Aware Bi-Encoders for Open-Domain Entity Retrieval

“…7 We also normalize embeddings before computing ψ. Following recent work (Gillick et al, 2019;Karpukhin et al, 2020), we use hard negative sampling to add the top nearest incorrect entities for each query to the batch. 8 We follow Botha et al (2020) to balance the hard negatives by fixing the ratio of positive to negative examples allowed for each entity, reducing the proportion of hard negatives that are rare entities (see Appendix A.4).…”

“…Prior work has shown that a hybrid model that combines sparse retrievers (e.g. TF-IDF) and dense retrievers can improve performance (Karpukhin et al, 2020;Luan et al, 2021) and that entity popularity can help disambiguation (Ganea and Hofmann, 2017).…”

“…There has been extensive work on dense retrieval for open-domain NLP tasks (e.g. Karpukhin et al, 2020;Oguz et al, 2020). However, most prior work has assumed unstructured text as the only input.…”

“…Entity retrieval (ER) is the process of finding the most relevant entities in a knowledge base for a natural language query. 2 ER is crucial for open-domain NLP tasks, where systems are provided with a query without the information needed to answer the query (Karpukhin et al, 2020). For instance, to answer the query, "What team does George Washington play for?"…”

“…In this work, we introduce TABi, a method for training entity retrievers on knowledge graph types and unstructured text. TABi builds on the bi-encoder model for dense retrieval (e.g., Wu et al, 2020;Karpukhin et al, 2020) (Figure 1). Bi-encoders learn embeddings of queries and entities contrastively: query embeddings are pulled close to their ground truth entity embedding and pushed away from other entity embeddings.…”