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 the STS benchmarks.

It is challenging to deploy deep neural nets on memory-constrained devices due to the explosion of numbers of parameters. Especially, the input embedding layer and Softmax layer usually dominate the memory usage in an RNN-based language model. For example, input embedding and Softmax matrices in IWSLT-2014 German-to-English data set account for more than 80% of the total model parameters. To compress these embedding layers, we propose MulCode, a novel multi-way multiplicative neural compressor. MulCode learns an adaptively created matrix and its multiplicative compositions. Together with a prior weighted loss, Multicode is more effective than the state-of-the-art compression methods. On the IWSLT-2014 machine translation data set, MulCode achieved 17 times compression rate for the embedding and Softmax matrices, and when combined with quantization technique, our method can achieve 41.38 times compression rate with very little loss in performance.

Named entity typing is the task of detecting the types of a named entity in context. For instance, given “Eric is giving a presentation”, our goal is to infer that ‘Eric’ is a speaker or a presenter and a person. Existing approaches to named entity typing cannot work with a growing type set and fails to recognize entity mentions of unseen types. In this paper, we present a label embedding method that incorporates prototypical and hierarchical information to learn pre-trained label embeddings. In addition, we adapt a zero-shot learning framework that can predict both seen and previously unseen entity types. We perform evaluation on three benchmark datasets with two settings: 1) few-shots recognition where all types are covered by the training set; and 2) zero-shot recognition where fine-grained types are assumed absent from training set. Results show that prior knowledge encoded using our label embedding methods can significantly boost the performance of classification for both cases.