Explicit Pairwise Word Interaction Modeling Improves Pretrained Transformers for English Semantic Similarity Tasks
This work addresses a specific bottleneck in semantic similarity tasks for NLP practitioners, offering an incremental enhancement to existing models.
The authors tackled the problem of improving pretrained transformers for English semantic similarity tasks by introducing an explicit pairwise word interaction mechanism, which led to consistent quality improvements on four tasks including semantic textual similarity and answer sentence selection.
In English semantic similarity tasks, classic word embedding-based approaches explicitly model pairwise "interactions" between the word representations of a sentence pair. Transformer-based pretrained language models disregard this notion, instead modeling pairwise word interactions globally and implicitly through their self-attention mechanism. In this paper, we hypothesize that introducing an explicit, constrained pairwise word interaction mechanism to pretrained language models improves their effectiveness on semantic similarity tasks. We validate our hypothesis using BERT on four tasks in semantic textual similarity and answer sentence selection. We demonstrate consistent improvements in quality by adding an explicit pairwise word interaction module to BERT.