pair2vec: Compositional Word-Pair Embeddings for Cross-Sentence Inference

TL;DR

pair2vec introduces compositional word-pair embeddings that encode background knowledge, improving cross-sentence inference tasks like question answering and natural language inference by enhancing model reasoning capabilities.

Contribution

The paper presents a novel method for learning word-pair embeddings via PMI maximization, which are integrated into existing models to improve inference performance.

Findings

2.7% accuracy gain on SQuAD2.0

1.3% accuracy gain on MultiNLI

Enhanced generalization on adversarial datasets

Abstract

Reasoning about implied relationships (e.g., paraphrastic, common sense, encyclopedic) between pairs of words is crucial for many cross-sentence inference problems. This paper proposes new methods for learning and using embeddings of word pairs that implicitly represent background knowledge about such relationships. Our pairwise embeddings are computed as a compositional function on word representations, which is learned by maximizing the pointwise mutual information (PMI) with the contexts in which the two words co-occur. We add these representations to the cross-sentence attention layer of existing inference models (e.g. BiDAF for QA, ESIM for NLI), instead of extending or replacing existing word embeddings. Experiments show a gain of 2.7% on the recently released SQuAD2.0 and 1.3% on MultiNLI. Our representations also aid in better generalization with gains of around 6-7% on adversarial SQuAD datasets, and 8.8% on the adversarial entailment test set by Glockner et al. (2018).