Transformers are powerful for sequence modeling. Nearly all state-of-the-art language models and pre-trained language models are based on the Transformer architecture. However, it distinguishes sequential tokens only with the token position index. We hypothesize that better contextual representations can be generated from the Transformer with richer positional information. To verify this, we propose a segment-aware Transformer (Segatron), by replacing the original token position encoding with a combined position encoding of paragraph, sentence, and token. We first introduce the segment-aware mechanism to Transformer-XL, which is a popular Transformer-based language model with memory extension and relative position encoding. We find that our method can further improve the Transformer-XL base model and large model, achieving 17.1 perplexity on the WikiText-103 dataset. We further investigate the pre-training masked language modeling task with Segatron. Experimental results show that BERT pre-trained with Segatron (SegaBERT) can outperform BERT with vanilla Transformer on various NLP tasks, and outperforms RoBERTa on zero-shot sentence representation learning. Our code is available on GitHub.

Yuqing Xie, Yi-An Lai, Yuanjun Xiong, Yi Zhang, Stefano Soatto. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

We tackle the problem of question answering directly on a large document collection, combining simple “bag of words” passage retrieval with a BERT-based reader for extracting answer spans. In the context of this architecture, we present a data augmentation technique using distant supervision to automatically annotate paragraphs as either positive or negative examples to supplement existing training data, which are then used together to fine-tune BERT. We explore a number of details that are critical to achieving high accuracy in this setup: the proper sequencing of different datasets during fine-tuning, the balance between “difficult” vs. “easy” examples, and different approaches to gathering negative examples. Experimental results show that, with the appropriate settings, we can achieve large gains in effectiveness on two English and two Chinese QA datasets. We are able to achieve results at or near the state of the art without any modeling advances, which once again affirms the cliché “there’s no data like more data”.