Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Topic Aware Probing: From Sentence Length Prediction to Idiom Identification how reliant are Neural Language Models on Topic?

Vasudevan Nedumpozhimana, John D. Kelleher

TL;DR

The paper investigates how Transformer-based language models allocate information between topic-like word co-occurrence signals and word-order/syntactic cues. It introduces topic-aware probing and applies it to BERT and RoBERTa using tasks from simple (sentence length) to complex (idiom token identification), with a focus on idioms due to their reliance on multiple linguistic cues. Across probing tasks, the authors find that models encode both topic and non-topic information, with idiom identification showing strong sensitivity to topic signals and RoBERTa generally more topic-dependent than BERT. These findings suggest that incorporating more explicit syntactic/word-order information could further enhance downstream NLP tasks, and they highlight avenues for extending analyses to decoder-based models and multilingual data.

Abstract

Transformer-based Neural Language Models achieve state-of-the-art performance on various natural language processing tasks. However, an open question is the extent to which these models rely on word-order/syntactic or word co-occurrence/topic-based information when processing natural language. This work contributes to this debate by addressing the question of whether these models primarily use topic as a signal, by exploring the relationship between Transformer-based models' (BERT and RoBERTa's) performance on a range of probing tasks in English, from simple lexical tasks such as sentence length prediction to complex semantic tasks such as idiom token identification, and the sensitivity of these tasks to the topic information. To this end, we propose a novel probing method which we call topic-aware probing. Our initial results indicate that Transformer-based models encode both topic and non-topic information in their intermediate layers, but also that the facility of these models to distinguish idiomatic usage is primarily based on their ability to identify and encode topic. Furthermore, our analysis of these models' performance on other standard probing tasks suggests that tasks that are relatively insensitive to the topic information are also tasks that are relatively difficult for these models.

Topic Aware Probing: From Sentence Length Prediction to Idiom Identification how reliant are Neural Language Models on Topic?

TL;DR

The paper investigates how Transformer-based language models allocate information between topic-like word co-occurrence signals and word-order/syntactic cues. It introduces topic-aware probing and applies it to BERT and RoBERTa using tasks from simple (sentence length) to complex (idiom token identification), with a focus on idioms due to their reliance on multiple linguistic cues. Across probing tasks, the authors find that models encode both topic and non-topic information, with idiom identification showing strong sensitivity to topic signals and RoBERTa generally more topic-dependent than BERT. These findings suggest that incorporating more explicit syntactic/word-order information could further enhance downstream NLP tasks, and they highlight avenues for extending analyses to decoder-based models and multilingual data.

Abstract

Transformer-based Neural Language Models achieve state-of-the-art performance on various natural language processing tasks. However, an open question is the extent to which these models rely on word-order/syntactic or word co-occurrence/topic-based information when processing natural language. This work contributes to this debate by addressing the question of whether these models primarily use topic as a signal, by exploring the relationship between Transformer-based models' (BERT and RoBERTa's) performance on a range of probing tasks in English, from simple lexical tasks such as sentence length prediction to complex semantic tasks such as idiom token identification, and the sensitivity of these tasks to the topic information. To this end, we propose a novel probing method which we call topic-aware probing. Our initial results indicate that Transformer-based models encode both topic and non-topic information in their intermediate layers, but also that the facility of these models to distinguish idiomatic usage is primarily based on their ability to identify and encode topic. Furthermore, our analysis of these models' performance on other standard probing tasks suggests that tasks that are relatively insensitive to the topic information are also tasks that are relatively difficult for these models.
Paper Structure (12 sections, 9 figures, 6 tables)

This paper contains 12 sections, 9 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Transformer-based Neural Language Models
  4. Idiom Token Identification
  5. Topic-Aware Probing
  6. Experimentation
  7. Data Preparation
  8. Representations
  9. Experimental Design
  10. Results and Discussion
  11. Other Probing tasks
  12. Conclusions

Figures (9)

  • Figure 1: Topic-Aware probing method
  • Figure 2: Experimental design
  • Figure 3: Seen and Unseen AUC ROC scores from GloVe and different layers of BERT and RoBERTa on the Bigram Shift Task
  • Figure 4: Seen and Unseen AUC ROC scores from different layers of BERT and RoBERTa with GloVe baseline on General Idiom Token Identification Task
  • Figure 5: Difference between seen scores and unseen scores from different layers of BERT and RoBERTa on General Idiom Token Identification Task
  • ...and 4 more figures