[Paper Review] Does BERT Solve Commonsense Task via Commonsense Knowledge

Leyang Cui, Sijie Cheng+2 more|arXiv (Cornell University)|Aug 10, 2020

Topic Modeling30 references19 citations

TL;DR

This paper investigates whether BERT solves CommonsenseQA tasks through structured commonsense knowledge or shallow patterns. Using attention-based analysis, it shows BERT captures structured knowledge from ConceptNet, with fine-tuning enhancing its use of this knowledge on higher layers, demonstrating deep integration of commonsense reasoning in BERT.

ABSTRACT

The success of pre-trained contextualized language models such as BERT motivates a line of work that investigates linguistic knowledge inside such models in order to explain the huge improvement in downstream tasks. While previous work shows syntactic, semantic and word sense knowledge in BERT, little work has been done on investigating how BERT solves CommonsenseQA tasks. In particular, it is an interesting research question whether BERT relies on shallow syntactic patterns or deeper commonsense knowledge for disambiguation. We propose two attention-based methods to analyze commonsense knowledge inside BERT, and the contribution of such knowledge for the model prediction. We find that attention heads successfully capture the structured commonsense knowledge encoded in ConceptNet, which helps BERT solve commonsense tasks directly. Fine-tuning further makes BERT learn to use the commonsense knowledge on higher layers.

Motivation & Objective

• To determine whether BERT solves CommonsenseQA tasks via deep commonsense knowledge or shallow syntactic patterns.
• To analyze how attention heads in BERT encode and utilize structured commonsense knowledge from ConceptNet.
• To investigate the role of fine-tuning in enhancing BERT’s use of commonsense knowledge on higher layers.
• To quantify the contribution of commonsense knowledge to BERT’s prediction performance on commonsense tasks.

Proposed method

• Propose two attention-based methods to analyze the presence and contribution of commonsense knowledge within BERT’s attention heads.
• Leverage ConceptNet to identify structured commonsense knowledge relevant to CommonsenseQA tasks.
• Measure attention activation patterns to assess whether attention heads attend to relevant commonsense facts in ConceptNet.
• Compare attention patterns before and after fine-tuning to evaluate changes in knowledge utilization across layers.
• Use attention-based probing to isolate and quantify the contribution of commonsense knowledge to prediction decisions.

Experimental results

Research questions

• RQ1 Do BERT attention heads capture structured commonsense knowledge from ConceptNet during inference?
• RQ2 Is BERT's performance on CommonsenseQA driven by deep commonsense knowledge or superficial syntactic patterns?
• RQ3 How does fine-tuning affect BERT’s use of commonsense knowledge across different layers?
• RQ4 To what extent does commonsense knowledge contribute to BERT’s prediction confidence and accuracy?

Key findings

• BERT attention heads successfully capture and utilize structured commonsense knowledge encoded in ConceptNet.
• The model leverages this knowledge directly to solve commonsense reasoning tasks, indicating a deep integration of external knowledge.
• Fine-tuning enhances BERT’s ability to use commonsense knowledge, particularly on higher layers of the network.
• Attention patterns show that BERT dynamically attends to relevant commonsense facts during prediction, supporting knowledge-based reasoning.