A small subset of dimensions within language Transformers’ representation spaces emerge as “outliers” during pretraining, encoding critical knowledge sparsely. We extend previous findings on emergent outliers to Encoder-Decoder Transformers and instruction-finetuned models, and tackle the problem of distilling a student Transformer from a larger teacher Transformer. Knowledge distillation reduces model size and cost by transferring knowledge from a larger teacher to a smaller student, necessitating a trade-off among representation dimensions. We show that emergent outlier dimensions contribute significantly more to zero-shot performance than non-outlier dimensions. Based on this, we propose the Emergent Outlier Focused Distillation (EOFD) method, which prioritizes critical outlier dimensions in distillation using a weighted MSE loss. We empirically demonstrate that EOFD outperforms state-of-the-art distillation methods and generalizes well across Encoder-only BERT, Decoder-only GPT-2, and Encoder-Decoder T5 architectures.

Joint entity and relation extraction has received increasing interests recently, due to the capability of utilizing the interactions between both steps. Among existing studies, the Multi-Head Selection (MHS) framework is efficient in extracting entities and relations simultaneously. However, the method is weak for its limited performance. In this paper, we propose several effective insights to address this problem. First, we propose an entity-specific Relative Position Representation (eRPR) to allow the model to fully leverage the distance information between entities and context tokens. Second, we introduce an auxiliary Global Relation Classification (GRC) to enhance the learning of local contextual features. Moreover, we improve the semantic representation by adopting a pre-trained language model BERT as the feature encoder. Finally, these new keypoints are closely integrated with the multi-head selection framework and optimized jointly. Extensive experiments on two benchmark datasets demonstrate that our approach overwhelmingly outperforms previous works in terms of all evaluation metrics, achieving significant improvements for relation F1 by +2.40% on CoNLL04 and +1.90% on ACE05, respectively.

Generation of natural language responses to the queries of structured language like SQL is very challenging as it requires generalization to new domains and the ability to answer ambiguous queries among other issues. We have participated in the CoSQL shared task organized in the IntEx-SemPar workshop at EMNLP 2020. We have trained a number of Neural Machine Translation (NMT) models to efficiently generate the natural language responses from SQL. Our shuffled back-translation model has led to a BLEU score of 7.47 on the unknown test dataset. In this paper, we will discuss our methodologies to approach the problem and future directions to improve the quality of the generated natural language responses.