We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:


Current browse context:


Change to browse by:

References & Citations

DBLP - CS Bibliography


(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Computation and Language

Title: Best of Both Worlds: A Hybrid Approach for Multi-Hop Explanation with Declarative Facts

Abstract: Language-enabled AI systems can answer complex, multi-hop questions to high accuracy, but supporting answers with evidence is a more challenging task which is important for the transparency and trustworthiness to users. Prior work in this area typically makes a trade-off between efficiency and accuracy; state-of-the-art deep neural network systems are too cumbersome to be useful in large-scale applications, while the fastest systems lack reliability. In this work, we integrate fast syntactic methods with powerful semantic methods for multi-hop explanation generation based on declarative facts. Our best system, which learns a lightweight operation to simulate multi-hop reasoning over pieces of evidence and fine-tunes language models to re-rank generated explanation chains, outperforms a purely syntactic baseline from prior work by up to 7% in gold explanation retrieval rate.
Comments: Accepted to CLeaR Workshop @ AAAI 2022
Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
Cite as: arXiv:2201.02740 [cs.CL]
  (or arXiv:2201.02740v1 [cs.CL] for this version)

Submission history

From: Shane Storks [view email]
[v1] Fri, 17 Dec 2021 20:27:48 GMT (4498kb,D)

Link back to: arXiv, form interface, contact.