Data-Efficient Concept Extraction from Pre-trained Language Models for Commonsense Explanation Generation

Yanbo Fang, Yongfeng ZhangEMNLP 2022 · OpenReview

Abstract: Predicting the key explanation concept is essential for generating commonsense explanations. This paper introduces a method to predict the concept from pre-trained language models for commonsense explanation generation. Our experiment found that adopting a language model as the concept extractor and fine-tuning it with 20% training data can improve the quality and accuracy of the generated explanations over multiple evaluation metrics. Compared with conventional methods that search concepts over knowledge graphs, our method does not require the preparation and training models to search through knowledge graphs. To better understand the results from pre-trained language models, we also designed a metric to evaluate the retrieved concepts. Through analysis and experiments, we show the correlation between this metric and the performance of the generators, and we also show the importance of attaching concepts for generating high-quality sentences.

Introduction

Artificial Intelligence often struggles with "commonsense"—the implicit knowledge humans use to navigate the world. For instance, understanding why "The school was open for summer" is typically false requires knowing that summer is vacation time. Current approaches to explaining such nonsensical statements generally fall into two categories: end-to-end generation, which lacks interpretability, and Knowledge Graph (KG) augmentation, which is computationally expensive and relies on structured external data.

This study addresses these challenges by proposing a method that leverages the latent knowledge within pre-trained language models (PLMs) to extract "bridge concepts" efficiently. The research utilizes the Commonsense Validation and Explanation (CoVE) dataset from SemEval 2020 Task 4, consisting of 10,000 counter-commonsense claims paired with human-written explanations.

Methodology

The proposed architecture consists of a two-stage pipeline: the Conceptor and the Generator.

  1. The Conceptor (Extraction): Instead of searching an external graph, a PLM (such as BERT or RoBERTa) acts as a "soft" knowledge base. The input statement is wrapped in a template: "Statement that [Statement] is unreasonable because of knowledge about [MASK]." The model predicts the top-$k$ words for the [MASK] token, which serve as bridge concepts.
  2. The Generator (Explanation): These concepts are concatenated with the original statement and fed into a T5 model to synthesize a coherent explanation.
Comparison of explanation generation strategies: (a) without concepts, (b) using external Knowledge Graphs, and (c) the proposed PLM-based concept extraction.

The authors found that fine-tuning the Conceptor on just 20% of the training data was sufficient to outperform baseline models, demonstrating significant data efficiency.

Key Results and Insights

The PLM-based extraction method outperformed both standalone T5 generators and KG-augmented versions. Using RoBERTa-large as the conceptor achieved a BLEU-4 score of 20.23, compared to 18.51 for the T5-only baseline and 19.33 for the KG-augmented method. Human evaluations corroborated these findings, rating the PLM-extracted explanations as more reasonable.

A critical finding was the existence of an optimal "Goldilocks zone" for the number of concepts provided to the generator. As shown in the performance metrics, the model's accuracy peaks when using $k=3$ concepts.

BLEU-4 scores as a function of the number of extracted concepts (k), showing a performance peak at k=3.

Adding too few concepts provides insufficient guidance, while adding too many introduces noise that degrades performance. Furthermore, the study demonstrated an "upper bound potential": using "perfect" concepts from human references yielded a BLEU-4 of 42.55, suggesting that improvements in concept retrieval will directly lead to higher-quality explanations.

Conclusion

This research demonstrates that large-scale pre-training imbues models with substantial world knowledge that can be effectively retrieved via prompt-based tuning. While challenges remain in improving concept precision and full generator interpretability, this approach offers a high-performing, data-efficient alternative to traditional knowledge-graph-based methods for commonsense reasoning.

Summary updated

Abstract

Predicting the key explanation concept is essential for generating commonsense explanations. This paper introduces a method to predict the concept from pre-trained language models for commonsense explanation generation. Our experiment found that adopting a language model as the concept extractor and fine-tuning it with 20% training data can improve the quality and accuracy of the generated explanations over multiple evaluation metrics. Compared with conventional methods that search concepts over knowledge graphs, our method does not require the preparation and training models to search through knowledge graphs. To better understand the results from pre-trained language models, we also designed a metric to evaluate the retrieved concepts. Through analysis and experiments, we show the correlation between this metric and the performance of the generators, and we also show the importance of attaching concepts for generating high-quality sentences.
BibTeX
@inproceedings{fang2022data,
  author = {Yanbo Fang and Yongfeng Zhang},
  title = {{Data-Efficient Concept Extraction from Pre-trained Language Models for Commonsense Explanation Generation}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2022},
  url = {https://openreview.net/forum?id=4LO4hlraLy},
}
Figure