Commonsense Knowledge Mining from Pretrained Models

TL;DR

Mining commonsense knowledge using pretrained models without updating weights, outperforming supervised methods.

cs.CL 🔴 Advanced 2019-09-02 4 views
Joshua Feldman Joe Davison Alexander M. Rush
NLP commonsense knowledge unsupervised learning language model knowledge mining

Key Findings

Methodology

This study proposes a method to generate commonsense knowledge using a large pretrained bidirectional language model. By transforming relational triples into masked sentences, the model estimates pointwise mutual information between entities to assess the validity of triples. The method does not update model weights, avoiding bias from any single commonsense knowledge base.

Key Results

  • Outperformed supervised methods on Wikipedia data with an average score of 3.00, exceeding existing methods by 0.50.
  • Achieved an F1 score of 78.8 on the ConceptNet test set, close to the supervised prototypical approach's 79.4.
  • Grammar and semantic accuracy in sentence generation significantly impact model performance.

Significance

This study demonstrates the potential of unsupervised methods in commonsense knowledge mining, especially when handling new data sources. It challenges the traditional view that machine learning models cannot implicitly learn commonsense and offers new insights for future knowledge mining techniques.

Technical Contribution

The fundamental difference from existing methods is the non-reliance on specific databases for training, leveraging the world knowledge of language models. This approach provides new theoretical guarantees and engineering possibilities, especially excelling in handling new data.

Novelty

This is the first to use pretrained bidirectional language models for commonsense knowledge mining, avoiding train-test leakage issues and offering a more generalizable knowledge mining method.

Limitations

  • Underperformed on the ConceptNet test set compared to supervised methods, possibly due to lack of optimization for specific datasets.
  • Generated sentences sometimes have grammatical errors or semantic inaccuracies, affecting model performance.

Future Work

Future work could explore the application of this method in non-commonsense knowledge mining and develop broader evaluation methods to validate the conclusions.

AI Executive Summary

Commonsense knowledge is a key challenge in natural language processing, with traditional supervised methods underperforming on novel data. This paper proposes a new method using pretrained bidirectional language models to generate commonsense knowledge. By transforming relational triples into masked sentences, the model estimates pointwise mutual information between entities to assess the validity of triples. Experiments show that this method outperformed supervised methods on Wikipedia data with an average score of 3.00. Although it underperformed on the ConceptNet test set, it excelled in handling new data. This study demonstrates the potential of unsupervised methods in commonsense knowledge mining, especially when handling new data sources. Future work could explore the application of this method in non-commonsense knowledge mining and develop broader evaluation methods to validate the conclusions.

Deep Analysis

Background

Commonsense knowledge consists of widely known facts about the world, rarely stated explicitly in natural language. Due to data sparsity, machine learning models struggle to implicitly learn commonsense. Traditional methods augment models with commonsense knowledge bases, but coverage is low. Commonsense knowledge base completion aims to improve the coverage of these resources.

Core Problem

Commonsense knowledge mining faces data sparsity and train-test leakage issues. Existing methods perform poorly on novel data, limiting their applicability. How to leverage the world knowledge of language models to directly identify commonsense facts is a key challenge.

Innovation

This paper proposes a novel unsupervised method to generate commonsense knowledge using pretrained bidirectional language models. The method avoids reliance on specific databases for training, addressing train-test leakage issues and offering a more generalizable knowledge mining approach.

Methodology

  • �� Use pretrained bidirectional language models to generate commonsense knowledge
  • �� Transform relational triples into masked sentences
  • �� Estimate pointwise mutual information between entities to assess triple validity
  • �� Avoid updating model weights to prevent bias from specific knowledge bases

Experiments

Experiments use Wikipedia and ConceptNet datasets to evaluate the performance of unsupervised methods versus supervised methods. BERT-large is used for masked language model evaluation, and GPT-2 for sentence generation. Clustering analysis is performed using the EM algorithm, adjusting PMI weights.

Results

Outperformed supervised methods on Wikipedia data with an average score of 3.00. Achieved an F1 score of 78.8 on the ConceptNet test set, close to the supervised prototypical approach's 79.4. Grammar and semantic accuracy in sentence generation significantly impact model performance.

Applications

This method can be used for commonsense knowledge base completion and knowledge mining from new data sources. It offers new insights for unsupervised learning applications in natural language processing, especially in addressing data sparsity and train-test leakage issues.

Limitations & Outlook

Underperformed on the ConceptNet test set compared to supervised methods, possibly due to lack of optimization for specific datasets. Generated sentences sometimes have grammatical errors or semantic inaccuracies, affecting model performance. Future work could explore the application of this method in non-commonsense knowledge mining.

Plain Language Accessible to non-experts

Imagine you're in a library looking for a book. Traditional methods are like relying on a librarian's recommendations, which might not suit your needs. This paper's method is like judging books by their titles and summaries yourself, allowing you to better decide if a new book is right for you. This method doesn't rely on specific recommendation systems but uses your own judgment to choose books.

ELI14 Explained like you're 14

Hey there! Imagine playing a game with lots of levels. Traditional methods are like trying to beat the game after playing only a few times, which is tough, right? This paper's method is like having a super guide to help you find the tricks to beat new levels. It doesn't rely on past experiences but helps you find solutions to new challenges. Isn't that cool?

Glossary

Pointwise Mutual Information

Measures statistical association between two events.

Used to estimate the validity of relations between entities.

Masked Language Model

A language model that predicts context by hiding certain words.

Used to evaluate word probabilities in sentences.

Commonsense Knowledge Base

A database containing widely known facts.

Used to enhance machine learning models' commonsense capabilities.

Pretrained Model

A model trained on large data, applicable to various tasks.

Used to generate commonsense knowledge.

Unsupervised Learning

Learning method without labeled data.

Used for commonsense knowledge mining.

Open Questions Unanswered questions from this research

  • 1 How to improve grammatical and semantic accuracy in sentence generation to enhance model performance.
  • 2 Explore the application of this method in non-commonsense knowledge mining, expanding its scope.

Applications

Immediate Applications

Commonsense Knowledge Base Completion

Enhance the coverage of knowledge bases, improving NLP models' commonsense capabilities.

Long-term Vision

Revolutionizing Knowledge Mining Techniques

Promote unsupervised learning applications in knowledge mining, addressing data sparsity issues.

Abstract

Inferring commonsense knowledge is a key challenge in natural language processing, but due to the sparsity of training data, previous work has shown that supervised methods for commonsense knowledge mining underperform when evaluated on novel data. In this work, we develop a method for generating commonsense knowledge using a large, pre-trained bidirectional language model. By transforming relational triples into masked sentences, we can use this model to rank a triple's validity by the estimated pointwise mutual information between the two entities. Since we do not update the weights of the bidirectional model, our approach is not biased by the coverage of any one commonsense knowledge base. Though this method performs worse on a test set than models explicitly trained on a corresponding training set, it outperforms these methods when mining commonsense knowledge from new sources, suggesting that unsupervised techniques may generalize better than current supervised approaches.

cs.CL cs.AI cs.LG