Ditto: A Simple and Efficient Approach to Improve Sentence Embeddings
This addresses the bias in sentence representations for NLP practitioners, but it is incremental as it builds on existing methods without adding new parameters.
The paper tackles the anisotropy problem in sentence embeddings from pre-trained language models like BERT by proposing Ditto, an unsupervised method that uses diagonal attention pooling to weight words and compute embeddings, resulting in improved performance on semantic textual similarity tasks.
Prior studies diagnose the anisotropy problem in sentence representations from pre-trained language models, e.g., BERT, without fine-tuning. Our analysis reveals that the sentence embeddings from BERT suffer from a bias towards uninformative words, limiting the performance in semantic textual similarity (STS) tasks. To address this bias, we propose a simple and efficient unsupervised approach, Diagonal Attention Pooling (Ditto), which weights words with model-based importance estimations and computes the weighted average of word representations from pre-trained models as sentence embeddings. Ditto can be easily applied to any pre-trained language model as a postprocessing operation. Compared to prior sentence embedding approaches, Ditto does not add parameters nor requires any learning. Empirical evaluations demonstrate that our proposed Ditto can alleviate the anisotropy problem and improve various pre-trained models on STS tasks.