Abstract

Complex questions often require combining multiple facts to correctly answer, particularly when generating detailed explanations for why those answers are correct. Combining multiple facts to answer questions is often modeled as a “multi-hop” graph traversal problem, where a given solver must find a series of interconnected facts in a knowledge graph that, taken together, answer the question and explain the reasoning behind that answer. Multi-hop inference currently suffers from semantic drift, or the tendency for chains of reasoning to “drift”’ to unrelated topics, and this semantic drift greatly limits the number of facts that can be combined in both free text or knowledge base inference. In this work we present our effort to mitigate semantic drift by extracting large high-confidence multi-hop inference patterns, generated by abstracting large-scale explanatory structure from a corpus of detailed explanations. We represent these inference patterns as sets of generalized constraints over sentences represented as rows in a knowledge base of semi-structured tables. We present a prototype tool for identifying common inference patterns from corpora of semi-structured explanations, and use it to successfully extract 67 inference patterns from a “matter” subset of standardized elementary science exam questions that span scientific and world knowledge.

Highlights

  • Combining separate pieces of knowledge to answer complex natural language questions is a central contemporary challenge in natural language inference

  • A single passage in a corpus or single fact in a knowledge base is often insufficient to arrive at an answer, and multiple sentences or facts must be combined through some inference process

  • This suggests that a subset of core facts are frequently reused, but that some form of abstraction or generalization of explanations would be required for those core facts to connect to the 69% of facts used in only a single explanation, or to knowledge imported from other knowledge bases that is not currently used in any explanation

Read more

Summary

Introduction

Combining separate pieces of knowledge to answer complex natural language questions is a central contemporary challenge in natural language inference. Das et al, 2017; Jansen et al, 2017; Ding et al, 2019) model inference as a progressive construction, iteratively adding nodes (facts) one at a time to a graph that represents the inference (and explanation) required to answer a question This approach suffers from the phenomenon of semantic drift (Fried et al, 2015), which is the observation that determining whether two facts can be meaningfully combined to answer a question is an extremely noisy process, and most often results in adding erroneous facts unrelated to answering a question that causes the inference to fail. Proceedings of the First Workshop on Commonsense Inference in Natural Language Processing, pages 53–65 Hongkong, China, November 3, 2019. c 2019 Association for Computational Linguistics

Common subgraphs in larger graph highlighted by human
Semi-Structured Explanation Corpus
Automatic Generation of Subgraphs
Merging and Curating Subgraphs
Preliminary Evaluation
Initial merging and curation
Extracting Inference Patterns
Executing constraint patterns
Conclusion and Future Work
Findings
Additional Tables and Figures
Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call