Probing for Bridging Inference in Transformer Language Models

Onkar Pandit, Yufang Hou

Abstract

Transformer-based language models have shown remarkable success in various NLP tasks, yet their internal understanding of complex discourse phenomena remains under-explored. This paper investigates whether pre-trained models like BERT and RoBERTa inherently capture bridging inference—the ability to link non-identical entities through conceptual relationships. Through attention head analysis and a novel "Of-Cloze" probing task, we demonstrate that these models possess significant zero-shot knowledge of bridging relations, particularly within higher layers and specialized attention heads.

Introduction

Bridging inference is a critical component of human discourse, allowing readers to connect an anaphor (e.g., "the economy") to an antecedent (e.g., "Poland") via implicit relationships. Unlike coreference resolution, which deals with identity, bridging requires understanding semantic and commonsense links. This study probes whether such knowledge is emergent in large-scale language models without explicit supervision.

Methodology

The researchers employed two primary probing techniques:

  1. Attention Head Analysis: Examining the internal attention mechanisms of BERT-base to identify heads that naturally attend to bridging pairs. The "bridging signal" is calculated as the ratio of attention weight between anaphors and their corresponding antecedents.
  2. "Of-Cloze" Probing: A zero-shot task where the model predicts a masked antecedent in a prompt like [Context] the economy of [MASK]. This evaluates the model's ability to leverage syntactic cues and semantic knowledge to resolve bridging links.

The experiments were conducted using the ISNotes corpus, which contains 663 manually annotated bridging anaphors across diverse relation types such as part-of and attribute-of.

Results and Discussion

The Hierarchy of Discourse Knowledge

Analysis reveals that bridging information is not uniformly distributed across the model. Lower layers primarily handle surface-level features, whereas higher layers exhibit increased activity related to discourse-level connections. This suggests that transformer models develop a hierarchical understanding of language, where complex referential relationships are resolved in the final stages of processing.

Heatmap analysis showing the distribution of bridging signals across layers (y-axis) and attention heads (x-axis).

Specialized Attention Heads

Specific heads act as "discourse specialists." For instance, head 1 in layer 5 and head 12 in layer 9 consistently focus on bridging relations. These heads demonstrate a natural affinity for linking anaphors to their antecedents even without task-specific fine-tuning.

Heatmap highlighting specialized attention heads, such as head 12 in layer 9 (indicated by the dark cell), which focus on bridging relations.

Zero-Shot Performance

Using the "Of-Cloze" test, RoBERTa-large achieved an accuracy of 28.05% on ISNotes. While lower than fine-tuned state-of-the-art models (~50%), this performance is significant for a zero-shot setting, outperforming traditional systems that rely on hand-crafted semantic features.

Linguistic Cues and Limitations

The preposition "of" serves as a vital trigger for bridging knowledge; removing it leads to a sharp decline in performance. Furthermore, models struggle with long-distance dependencies and complex commonsense reasoning (e.g., linking "OPEC" to "pool"), often defaulting to local semantic proximity rather than global discourse logic.

Conclusion

This research provides evidence that transformer models capture sophisticated discourse phenomena like bridging inference during pre-training. Identifying specialized "bridging heads" opens avenues for developing more discourse-aware NLP systems by leveraging these inherent signals.

Summary updated

Abstract

We probe pre-trained transformer language models for bridging inference. We first investigate individual attention heads in BERT and observe that attention heads at higher layers prominently focus on bridging relations in-comparison with the lower and middle layers, also, few specific attention heads concentrate consistently on bridging. More importantly, we consider language models as a whole in our second approach where bridging anaphora resolution is formulated as a masked token prediction task (Of-Cloze test). Our formulation produces optimistic results without any fine-tuning, which indicates that pre-trained language models substantially capture bridging inference. Our further investigation shows that the distance between anaphor-antecedent and the context provided to language models play an important role in the inference.
BibTeX
@inproceedings{pandit2021probing,
  author = {Onkar Pandit and Yufang Hou},
  title = {{Probing for Bridging Inference in Transformer Language Models}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2021},
  url = {https://openreview.net/forum?id=UHAGdX0ZWxDZ},
}
Figure