Inductive Natural Language Rules from Natural Language Facts

Home

Discover

Bookmarks

Keep Up With Latest Trending Papers. Computer Science, AI and Machine Learning and more.Subscribe

CC BY

Source arXiv

Artificial Intelligence

Computation and Language

Inductive Natural Language Rules from Natural Language Facts

Language Models as Inductive Reasoners

Our proposed framework (CoLM) for inductive reasoning with natural language representation task. Rule Proposer is a generative model based on input facts and desired rule template, aiming at generating (a large number of) rule candidates. Deductive consistency evaluator, indiscriminate confirmation handler, generalization checker, and triviality detector are classification models that filter improper rules according to four requirements of the induced rules in inductive reasoning.

Inductive reasoning is a core component of human intelligence.

However, logic language can cause systematic problems for inductive reasoning such as disability of handling raw input such as natural language, sensitiveness to mislabeled data, and incapacity to handle ambiguous input.

To this end, we propose a new task, which is to induce natural language rules from naturallanguage facts, and create a dataset termed deer containing 1.2k rule-fact pairs for the task, where rules and facts are written in natural language.

Moreover, we provide the first and comprehensive analysis of how well pretrained language models can induce natural language `` rules''from natural language facts.

We also propose a new framework drawing insights from philosophy literature for this task, which we show in the experiment section that surpasses baselines in both automatic and human evaluations.

Authors

Zonglin Yang, Li Dong, Xinya Du, Hao Cheng, Erik Cambria, Xiaodong Liu, Jianfeng Gao, Furu Wei

Related Topics

Introduction

Inductive reasoning is to reach to a hypothesis (usually a rule that explains an aspect of the law of nature) based on pieces of evidence (usually observed facts of the world) where the observations can not provide conclusive support to the hypothesis.

In this paper, we present a novel paradigm for inductive reasoning based entirely on natural language, i.e., inducing natural language rules from natural language facts.

In particular, we create a first-of-its-kind natural language inductive reasoning dataset named deer containing 1.2k rule-fact pairs.

With this dataset, we investigate a modern approach to inductive reasoning where both facts and rules are in natural language, and pretrained language models (plms) are used as the inductive reasoner to induce (generate) natural language rules (usually rules that explain an aspect of a law of nature) from (generate) natural language facts.

Specifically, human-written natural language rule sentences are first collected.

Based on the collected rules, we then ask human annotators to collect existing natural language texts as facts from the web where each fact can be possibly enough to induce the given rule.

Result

To overcome the systematic problems of using logic language for inductive reasoning, we propose a new paradigm of inducing natural language rules from natural language facts, and correspondingly propose a dataset deer and new evaluation metrics for this task.

We also propose a new framework drawing insights from philosophy literature, which show in the experiment section that surpasses baselines in both automatic and human evaluations.

Read the Paper

Read Paper

Bookmarks

Bookmarks

Keep Up With Latest Trending Papers. Computer Science, AI and Machine Learning and more.Subscribe

CC BY

Source arXiv

Artificial Intelligence

Computation and Language

Inductive Natural Language Rules from Natural Language Facts

Language Models as Inductive Reasoners

Inductive reasoning is a core component of human intelligence.

Moreover, we provide the first and comprehensive analysis of how well pretrained language models can induce natural language `` rules''from natural language facts.

We also propose a new framework drawing insights from philosophy literature for this task, which we show in the experiment section that surpasses baselines in both automatic and human evaluations.

Authors

Zonglin Yang, Li Dong, Xinya Du, Hao Cheng, Erik Cambria, Xiaodong Liu, Jianfeng Gao, Furu Wei

◐ Recommended Follows

◐ Latest Activity