Paraphrasing vs Coreferring: Two Sides of the Same Coin

Yehudit Meged, Avi Caciularu, Vered Shwartz, Ido Dagan

Abstract

This paper explores the symbiotic relationship between two fundamental NLP tasks: predicate paraphrasing and event coreference resolution. While predicate paraphrasing identifies generic lexical similarity, event coreference resolution determines if two specific mentions in text refer to the same event instance. The authors propose a two-way synergy where event coreference data is used to improve paraphrase ranking (Chirps), and these improved scores are subsequently integrated back into a coreference model. Experimental results demonstrate that this mutual reinforcement significantly boosts performance in both tasks, with Chirps achieving an 80.0 Average Precision in paraphrase ranking and contributing to state-of-the-art results in cross-document event coreference.

Introduction

Lexical variability remains a significant challenge in Natural Language Processing. For instance, understanding that a celebrity "checked into" a clinic and "entered" a rehab center refers to the same event requires a system to recognize these as paraphrases within a specific context. Standard resources like WordNet often lack the coverage for such context-specific nuances, while distributional methods like word embeddings frequently struggle to distinguish between synonyms and antonyms.

This research posits that event coreference (specific instance matching) and predicate paraphrasing (generic meaning matching) are "two sides of the same coin." If two mentions corefer, their predicates are likely paraphrases; conversely, known paraphrases are more likely to corefer in a given text. By exploiting this link, the authors aim to overcome the limitations of existing lexical resources.

Methods

The researchers developed a two-stage approach to exploit this synergy. First, they improved the ranking of Chirps, a large-scale resource of 5 million predicate pairs extracted from Twitter, creating Chirps*. This supervised scorer utilizes distant supervision from the ECB+ dataset, incorporating features such as Named Entity Coverage, graph-based connectivity (cliques and connected components), and outputs from existing coreference models.

Second, these Chirps features were integrated into a state-of-the-art cross-document event coreference model (Barhom et al., 2019). As illustrated in Figure 1, a Multi-Layer Perceptron (MLP) transforms the Chirps features into the same embedding space as the internal representations of the coreference model. These representations are then concatenated and passed through a scorer to produce a final pairwise coreference score.

Architecture of the pairwise coreference scorer integrating Chirps* features.

Results

The Chirps scorer demonstrated substantial improvements in paraphrase ranking tasks. As detailed in Figure 5, Chirps achieved an Average Precision (AP) of 80.0 on the test set, representing an 18-point gain over the original heuristic Chirps ranking (62.5 AP) and nearly 30 points over GloVe-based baselines (50.7 AP).

Performance comparison of Chirps* against GloVe and original Chirps on the ch-ECB+ test set.

In the coreference resolution task, the addition of Chirps* features led to a CoNLL F1 score of 80.0, a statistically significant improvement over the previous state-of-the-art of 79.5. Qualitative analysis indicates that the system became more effective at linking mentions (improving recall) while simultaneously avoiding erroneous links (maintaining precision) by leveraging generic knowledge of valid paraphrases.

Discussion and Limitations

The study highlights that coreference is essentially a context-sensitive version of paraphrasing. The use of distant supervision from news events allows for learning high-quality paraphrases without extensive manual labeling. However, several limitations remain. The current Chirps resource focuses exclusively on verbal predicates, omitting nominalizations (e.g., "the investigation" vs. "they investigated") which are prevalent in event descriptions. Furthermore, Chirps covered only approximately 30% of the mention pairs in the ECB+ dataset. Expanding the resource's domain and part-of-speech coverage is expected to yield further gains. Future work may explore joint modeling of both tasks to fully realize their inherent synergy.

Summary updated

BibTeX
@inproceedings{meged2020paraphrasing,
  author = {Yehudit Meged and Avi Caciularu and Vered Shwartz and Ido Dagan},
  title = {{Paraphrasing vs Coreferring: Two Sides of the Same Coin}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2020},
  url = {https://openreview.net/forum?id=mSLj4144YX1},
}
Figure