Zewen Chi, Shaohan Huang, Li Dong, Shuming Ma, Bo Zheng, Saksham Singhal, Payal Bajaj, Xia Song, Xian-Ling Mao, Heyan Huang, Furu Wei. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022.

The Mixture-of-Experts (MoE) technique can scale up the model size of Transformers with an affordable computational overhead.We point out that existing learning-to-route MoE methods suffer from the routing fluctuation issue, i.e., the target expert of the same input may change along with training, but only one expert will be activated for the input during inference.The routing fluctuation tends to harm sample efficiency because the same input updates different experts but only one is finally used.In this paper, we propose STABLEMOE with two training stages to address the routing fluctuation problem.In the first training stage, we learn a balanced and cohesive routing strategy and distill it into a lightweight router decoupled from the backbone model.In the second training stage, we utilize the distilled router to determine the token-to-expert assignment and freeze it for a stable routing strategy.We validate our method on language modeling and multilingual machine translation.The results show that STABLEMOE outperforms existing MoE methods in terms of both convergence speed and performance.

Zewen Chi, Li Dong, Bo Zheng, Shaohan Huang, Xian-Ling Mao, Heyan Huang, Furu Wei. 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.

Bo Zheng, Li Dong, Shaohan Huang, Wenhui Wang, Zewen Chi, Saksham Singhal, Wanxiang Che, Ting Liu, Xia Song, Furu Wei. 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.

Compared to monolingual models, crosslingual models usually require a more expressive vocabulary to represent all languages adequately.We find that many languages are under-represented in recent cross-lingual language models due to the limited vocabulary capacity.To this end, we propose an algorithm VOCAP to determine the desired vocabulary capacity of each language.However, increasing the vocabulary size significantly slows down the pre-training speed.In order to address the issues, we propose k-NN-based target sampling to accelerate the expensive softmax.Our experiments show that the multilingual vocabulary learned with VOCAP benefits cross-lingual language model pre-training.Moreover, k-NN-based target sampling mitigates the side-effects of increasing the vocabulary size while achieving comparable performance and faster pre-training speed.The code and the pretrained multilingual vocabularies are available at https://github. com/bozheng-hit/VoCapXLM.

Natural Questions is a new challenging machine reading comprehension benchmark with two-grained answers, which are a long answer (typically a paragraph) and a short answer (one or more entities inside the long answer).Despite the effectiveness of existing methods on this benchmark, they treat these two sub-tasks individually during training while ignoring their dependencies.To address this issue, we present a novel multi-grained machine reading comprehension framework that focuses on modeling documents at their hierarchical nature, which are different levels of granularity: documents, paragraphs, sentences, and tokens.We utilize graph attention networks to obtain different levels of representations so that they can be learned simultaneously.The long and short answers can be extracted from paragraphlevel representation and token-level representation, respectively.In this way, we can model the dependencies between the two-grained answers to provide evidence for each other.We jointly train the two sub-tasks, and our experiments show that our approach significantly outperforms previous systems at both long and short answer criteria.

Deep contextualized word embeddings (Embeddings from Language Model, short for ELMo), as an emerging and effective replacement for the static word embeddings, have achieved success on a bunch of syntactic and semantic NLP problems. However, little is known about what is responsible for the improvements. In this article, we focus on the effect of ELMo for a typical syntax problem—universal POS tagging and dependency parsing. We incorporate ELMo as additional word embeddings into the state-of-the-art POS tagger and dependency parser, and it leads to consistent performance improvements. Experimental results show the model using ELMo outperforms the state-of-the-art baseline by an average of 0.91 for POS tagging and 1.11 for dependency parsing. Further analysis reveals that the improvements mainly result from the ELMo’s better abstraction ability on the out-of-vocabulary (OOV) words, and the character-level word representation in ELMo contributes a lot to the abstraction. Based on ELMo’s advantage on OOV, experiments that simulate low-resource settings are conducted and the results show that deep contextualized word embeddings are effective for data-insufficient tasks where the OOV problem is severe.

In this paper, we propose a new rich resource enhanced AMR aligner which produces multiple alignments and a new transition system for AMR parsing along with its oracle parser.Our aligner is further tuned by our oracle parser via picking the alignment that leads to the highestscored achievable AMR graph.Experimental results show that our aligner outperforms the rule-based aligner in previous work by achieving higher alignment F1 score and consistently improving two open-sourced AMR parsers.Based on our aligner and transition system, we develop a transition-based AMR parser that parses a sentence into its AMR graph directly.An ensemble of our parsers with only words and POS tags as input leads to 68.4 Smatch F1 score, which outperforms the parser of Wang and Xue (2017).