Virtual Knowledge Graph Construction for Zero-Shot Domain-Specific Document Retrieval

Proceedings of the 29th International Conference on Computational Linguistics, pages 1169–1178 October 12–17, 2022. 1169 Virtual Knowledge Graph Construction for Zero-Shot Domain-Specific Document Retrieval Yeon Seonwoo†*, Seunghyun Yoon‡, Franck Dernoncourt‡, Trung Bui‡, Alice Oh† †KAIST, ‡Adobe Research, yeon.seonwoo@kaist.ac.kr {syoon, dernonco, bui}@adobe.com alice.oh@kaist.edu Abstract Domain-specific documents cover terminolo- gies and specialized knowledge. This has been the main challenge of domain-specific docu- ment retrieval systems. Previous approaches propose domain-adaptation and transfer learn- ing methods to alleviate this problem. How- ever, these approaches still follow the same document representation method in previous approaches; a document is embedded into a sin- gle vector. In this study, we propose VKGDR. VKGDR represents a given corpus into a graph of entities and their relations (known as a vir- tual knowledge graph) and computes the rele- vance between queries and documents based on the graph representation. We conduct three experiments 1) domain-specific document re- trieval, 2) comparison of our virtual knowledge graph construction method with previous ap- proaches, and 3) ablation study on each com- ponent of our virtual knowledge graph. From the results, we see that unsupervised VKGDR outperforms baselines in a zero-shot setting and even outperforms fully-supervised bi-encoder. We also verify that our virtual knowledge graph construction method results in better retrieval performance than previous approaches. 1 1 Introduction In domain-specific QA, building retrievers is chal- lenging since queries and documents in a specific domain cover terminologies and specialized knowl- edge, which are not well covered in general doc- uments. (Zhang et al., 2020; Ma et al., 2021; Yu et al., 2020, 2021). Another problem is the diffi- culty in building datasets for training retrievers. This problem comes from 1) the complexity of knowledge treated in the documents, and 2) costly dataset maintenance; recall that domain-specific documents are frequently updated (e.g., software *This work was done during an internship at Adobe Re- search. 1We provide the implementation of VKGDR at https: //github.com/yeonsw/VKGDR manuals are updated whenever there is a version update) (Castelli et al., 2020; Nandy et al., 2021; Voorhees et al., 2021; Maia et al., 2018). Recent domain-specific document retrieval stud- ies propose domain-adaptation and transfer learn- ing methods (Thakur et al., 2021; Ma et al., 2021; Beltagy et al., 2019; Gururangan et al., 2020; Chalkidis et al., 2020). However, these methods still use the conventional document representation method, embedding a document into a single vec- tor. This is problematic because a single vector is insufficient to cover complex knowledge in a domain-specific document. Semi-structured knowl- edge representation methods effectively address this problem, but they have only been applied to open-domain documents. (Dhingra et al., 2020; Sun et al., 2021; Zhang et al., 2018; Sun et al., 2018; Yasunaga et al., 2021; Talmor and Berant, 2018). In this paper, we propose an automatic virtual knowledge graph construction method for zero- shot domain-specific document retrieval. A virtual knowledge graph (VKG) is a graph representation of a corpus that consists of entities and their re- lations. In VKG, the relations are represented by relation vectors (Dhingra …

Content truncated to 500 words