Researchers from the University of Washington, Salesforce Research and Allen Institute for Artificial Intelligence have introduced a graph-based method that retrieves reasoning paths to boost multi-hop open-domain question answering.
Open-domain Question Answering (QA) is the task of answering a question given a large collection of reference text documents such as Wikipedia. Current state-of-art open-domain QA approaches leverage non-parameterized models such as TF-IDF (term frequency-inverse document frequency) or the best-matching function BM25 to retrieve a fixed set of documents. Although these can achieve great performance on single-hop QA where questions can usually be answered based on a single paragraph of text, they struggle when answering questions that require multi-hop reasoning at web scale. This is generally because such pipeline approaches cannot retrieve the required evidence or properly capture relationships between evidence documents through the bridge entities required for multi-hop reasoning and multi-hop questions.
To tackle these problems, the researchers propose a novel recurrent graph-based retrieval method that can retrieve evidence documents as reasoning paths over the Wikipedia graph for answering multi-hop open-domain questions.
The proposed QA framework comprises a reasoning path retrieval and a reading and answering reasoning path. In the first stage, the retriever model learns to sequentially retrieve evidence paragraphs to form the reasoning path given the history of previously retrieved paragraphs and the graphical structure. The reader model subsequently re-ranks the reasoning paths and determines the final answer as the one extracted from the best reasoning path.
Experiment results show significant effectiveness and robustness, with state-of-the-art performance in the open-domain HotpotQA, SQuAD Open and Natural Questions Open datasets. The method achieves a remarkable improvement on HotpotQA, outperforming the previous best model by more than 14 points.
The research team says future work in this area will involve end-to-end training of the graph-based recurrent retriever and reader to further improve the two-stage training process.
The paper Learning to Retrieve Reasoning Paths over Wikipedia Graph for Question Answeringis on arXiv.
Author: Yuqing Li | Editor: Michael Sarazen
We know you don’t want to miss any story. Subscribe to our popular Synced Global AI Weekly to get weekly AI updates.