Efficient Large Scale Language Modeling with Mixtures of Experts

Mixture of Experts layers (MoEs) enable efficient scaling of language models through conditional computation. This paper presents a detailed empirical study of how autoregressive MoE language models scale in comparison with dense models in a wide range of settings: in- and out-of-domain language modeling, zero- and few-shot priming, and full fine-tuning. With the exception of fine-tuning, we find MoEs to be substantially more compute efficient. At more modest training budgets, MoEs can match the performance of dense models using $\sim$4 times less compute. This gap narrows at scale, but our largest MoE model (1.1T parameters) consistently outperforms a compute-equivalent dense model (6.7B parameters). Overall, this performance gap varies greatly across tasks and domains, suggesting that MoE and dense models generalize differently in ways that are worthy of future study. We make our code and models publicly available for research use.

General Purpose Text Embeddings from Pre-trained Language Models for Scalable Inference

The state of the art on many NLP tasks is currently achieved by large pre-trained language models, which require a considerable amount of computation. We explore a setting where many different predictions are made on a single piece of text. In that case, some of the computational cost during inference can be amortized over the different tasks using a shared text encoder. We compare approaches for training such an encoder and show that encoders pre-trained over multiple tasks generalize well to unseen tasks. We also compare ways of extracting fixed- and limited-size representations from this encoder, including different ways of pooling features extracted from multiple layers or positions. Our best approach compares favorably to knowledge distillation, achieving higher accuracy and lower computational cost once the system is handling around 7 tasks. Further, we show that through binary quantization, we can reduce the size of the extracted representations by a factor of 16 making it feasible to store them for later use. The resulting method offers a compelling solution for using large-scale pre-trained models at a fraction of the computational cost when multiple tasks are performed on the same text.

Keyphrase Extraction with Span-based Feature Representations

Keyphrases are capable of providing semantic metadata characterizing documents and producing an overview of the content of a document. Since keyphrase extraction is able to facilitate the management, categorization, and retrieval of information, it has received much attention in recent years. There are three approaches to address keyphrase extraction: (i) traditional two-step ranking method, (ii) sequence labeling and (iii) generation using neural networks. Two-step ranking approach is based on feature engineering, which is labor intensive and domain dependent. Sequence labeling is not able to tackle overlapping phrases. Generation methods (i.e., Sequence-to-sequence neural network models) overcome those shortcomings, so they have been widely studied and gain state-of-the-art performance. However, generation methods can not utilize context information effectively. In this paper, we propose a novelty Span Keyphrase Extraction model that extracts span-based feature representation of keyphrase directly from all the content tokens. In this way, our model obtains representation for each keyphrase and further learns to capture the interaction between keyphrases in one document to get better ranking results. In addition, with the help of tokens, our model is able to extract overlapped keyphrases. Experimental results on the benchmark datasets show that our proposed model outperforms the existing methods by a large margin.