Using Mined Coreference Chains as a Resource for a Semantic Task

Heike Adel, Hinrich SchützeEMNLP 2014 · OpenReview

Abstract

We propose to use coreference chains extracted from a large corpus as a resource for semantic tasks. We extract three million coreference chains and train word embeddings on them. Then, we compare these embeddings to word vectors derived from raw text data and show that coreference-based word embeddings improve F1 on the task of antonym classification by up to .09.

Introduction

In Natural Language Processing (NLP), word embeddings are typically built on the principle of distributional similarity: words appearing in similar contexts are assumed to have similar meanings. While this approach is highly effective for identifying synonyms, it often fails to distinguish between antonyms. For example, the words "hot" and "cold" frequently appear in identical contexts (e.g., "the water is ____"), leading standard models like word2vec to treat them as semantically similar. This research addresses this "guilt by association" by leveraging coreference chains—sequences of words that refer to the exact same entity—to capture true semantic identity rather than mere contextual relatedness.

Methodology

The researchers developed a novel approach to training word representations by comparing two distinct embedding types:

  1. Text-based Embeddings: A standard skip-gram model (word2vec) trained on raw text from the English Gigaword corpus.
  2. Coreference-based Embeddings: Using the Stanford CoreNLP coreference resolution system, the authors extracted approximately 3.1 million "chains" of words referring to the same entity (e.g., "the CEO," "she," and "the executive"). Each chain was treated as a pseudo-sentence for training word2vec.

The underlying hypothesis is that while "cows" and "milk" share a context, only "cows" and "cattle" would appear in the same coreference chain. To evaluate these embeddings, the authors built an Antonym Classifier using a Support Vector Machine (SVM), utilizing features such as cosine similarity and neighborhood rank.

Qualitative Analysis

A qualitative comparison of the two embedding types reveals that coreference-based vectors are significantly better at preserving specific attributes like gender and identity.

Comparison of nearest neighbors for "his" and "woman" between text-based and coreference-based models.

As seen in the table above, the text-based model identifies "man" as a top neighbor for "woman" due to their shared contexts. In contrast, the coreference-based model identifies more specific entities and synonyms (like "girl" or specific names), reflecting the model's focus on referential identity.

Results and Discussion

The system was evaluated on a balanced dataset of 4,674 word pairs derived from the Merriam-Webster online dictionary. The coreference-based approach showed a marked advantage in noun classification. While text-based features achieved an F1 score of .60, the addition of coreference-based features boosted the score to .69—a substantial .09 improvement.

t-SNE visualization comparing text-based embeddings (left) and coreference-based embeddings (right).

Visualization via t-SNE (above) demonstrates that coreference embeddings successfully "push" antonyms further apart in vector space. In the text-based plot, pairs like "literacy" and "illiteracy" are clustered closely, whereas the coreference-based plot shows greater spatial separation, facilitating easier classification by the SVM.

Limitations and Conclusion

While highly effective for nouns, the method faces challenges with other parts of speech, particularly adjectives. Antonymous adjectives like "black" and "white" are sometimes mistakenly grouped by coreference systems when they describe a set (e.g., "a collection of black and white photos"). Despite this, the study proves that mined coreference chains are a powerful resource for distinguishing semantic identity from contextual association. Future work may involve combining these embeddings with pattern-based methods to refine antonym detection across all word classes.

Summary updated

Abstract

We propose to use coreference chains extracted from a large corpus as a resource for semantic tasks. We extract three million coreference chains and train word embeddings on them. Then, we compare these embeddings to word vectors derived from raw text data and show that coreference-based word embeddings improve F1 on the task of antonym classification by up to .09.
BibTeX
@inproceedings{adel2014using,
  author = {Heike Adel and Hinrich Schütze},
  title = {{Using Mined Coreference Chains as a Resource for a Semantic Task}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2014},
  url = {https://openreview.net/forum?id=rJ-oufz_br},
}
Figure