Rui Liu

Rui Liu, Young Jin Kim, Alexandre Muzio et al. · microsoft-research

Sparsely activated transformers, such as Mixture of Experts (MoE), have received great interest due to their outrageous scaling capability which enables dramatical increases in model size without significant increases in computational cost. To achieve this, MoE models replace the feedforward sub-layer with Mixture-of-Experts sub-layer in transformers and use a gating network to route each token to its assigned experts. Since the common practice for efficient training of such models requires distributing experts and tokens across different machines, this routing strategy often incurs huge cross-machine communication cost because tokens and their assigned experts likely reside in different machines. In this paper, we propose \emph{Gating Dropout}, which allows tokens to ignore the gating network and stay at their local machines, thus reducing the cross-machine communication. Similar to traditional dropout, we also show that Gating Dropout has a regularization effect during training, resulting in improved generalization performance. We validate the effectiveness of Gating Dropout on multilingual machine translation tasks. Our results demonstrate that Gating Dropout improves a state-of-the-art MoE model with faster wall-clock time convergence rates and better BLEU scores for a variety of model sizes and datasets.

Rui Liu, Barzan Mozafari

Transformers achieve state-of-the-art performance for natural language processing tasks by pre-training on large-scale text corpora. They are extremely compute-intensive and have very high sample complexity. Memory replay is a mechanism that remembers and reuses past examples by saving to and replaying from a memory buffer. It has been successfully used in reinforcement learning and GANs due to better sample efficiency. In this paper, we propose \emph{Transformer with Memory Replay} (TMR), which integrates memory replay with transformer, making transformer more sample-efficient. Experiments on GLUE and SQuAD benchmark datasets show that Transformer with Memory Replay achieves at least $1\%$ point increase compared to the baseline transformer model when pretrained with the same number of examples. Further, by adopting a careful design that reduces the wall-clock time overhead of memory replay, we also empirically achieve a better runtime efficiency.

Wang Qi, Rui Liu, Yuan Zuo et al.

Creating an essay based on a few given topics is a challenging NLP task. Although several effective methods for this problem, topic-to-essay generation, have appeared recently, there is still much room for improvement, especially in terms of the coverage of the given topics and the coherence of the generated text. In this paper, we propose a novel approach called TegFormer which utilizes the Transformer architecture where the encoder is enriched with domain-specific contexts while the decoder is enhanced by a large-scale pre-trained language model. Specifically, a \emph{Topic-Extension} layer capturing the interaction between the given topics and their domain-specific contexts is plugged into the encoder. Since the given topics are usually concise and sparse, such an additional layer can bring more topic-related semantics in to facilitate the subsequent natural language generation. Moreover, an \emph{Embedding-Fusion} module that combines the domain-specific word embeddings learnt from the given corpus and the general-purpose word embeddings provided by a GPT-2 model pre-trained on massive text data is integrated into the decoder. Since GPT-2 is at a much larger scale, it contains a lot more implicit linguistic knowledge which would help the decoder to produce more grammatical and readable text. Extensive experiments have shown that the pieces of text generated by TegFormer have better topic coverage and higher text coherence than those from SOTA topic-to-essay techniques, according to automatic and human evaluations. As revealed by ablation studies, both the Topic-Extension layer and the Embedding-Fusion module contribute substantially to TegFormer's performance advantage.

Jianying Lin, Rui Liu, Quanye Jia

Due to the manifold ranking method has a significant effect on the ranking of unknown data based on known data by using a weighted network, many researchers use the manifold ranking method to solve the document summarization task. However, their models only consider the original features but ignore the semantic features of sentences when they construct the weighted networks for the manifold ranking method. To solve this problem, we proposed two improved models based on the manifold ranking method. One is combining the topic model and manifold ranking method (JTMMR) to solve the document summarization task. This model not only uses the original feature, but also uses the semantic feature to represent the document, which can improve the accuracy of the manifold ranking method. The other one is combining the lifelong topic model and manifold ranking method (JLTMMR). On the basis of the JTMMR, this model adds the constraint of knowledge to improve the quality of the topic. At the same time, we also add the constraint of the relationship between documents to dig out a better document semantic features. The JTMMR model can improve the effect of the manifold ranking method by using the better semantic feature. Experiments show that our models can achieve a better result than other baseline models for multi-document summarization task. At the same time, our models also have a good performance on the single document summarization task. After combining with a few basic surface features, our model significantly outperforms some model based on deep learning in recent years. After that, we also do an exploring work for lifelong machine learning by analyzing the effect of adding feedback. Experiments show that the effect of adding feedback to our model is significant.

Rui Liu, Tianyi Wu, Barzan Mozafari

There has been substantial research on sub-linear time approximate algorithms for Maximum Inner Product Search (MIPS). To achieve fast query time, state-of-the-art techniques require significant preprocessing, which can be a burden when the number of subsequent queries is not sufficiently large to amortize the cost. Furthermore, existing methods do not have the ability to directly control the suboptimality of their approximate results with theoretical guarantees. In this paper, we propose the first approximate algorithm for MIPS that does not require any preprocessing, and allows users to control and bound the suboptimality of the results. We cast MIPS as a Best Arm Identification problem, and introduce a new bandit setting that can fully exploit the special structure of MIPS. Our approach outperforms state-of-the-art methods on both synthetic and real-world datasets.

Rui Liu, Wei Wei, Weiguang Mao et al.

Attention models have been intensively studied to improve NLP tasks such as machine comprehension via both question-aware passage attention model and self-matching attention model. Our research proposes phase conductor (PhaseCond) for attention models in two meaningful ways. First, PhaseCond, an architecture of multi-layered attention models, consists of multiple phases each implementing a stack of attention layers producing passage representations and a stack of inner or outer fusion layers regulating the information flow. Second, we extend and improve the dot-product attention function for PhaseCond by simultaneously encoding multiple question and passage embedding layers from different perspectives. We demonstrate the effectiveness of our proposed model PhaseCond on the SQuAD dataset, showing that our model significantly outperforms both state-of-the-art single-layered and multiple-layered attention models. We deepen our results with new findings via both detailed qualitative analysis and visualized examples showing the dynamic changes through multi-layered attention models.

Rui Liu, Junjie Hu, Wei Wei et al.

Deep neural networks for machine comprehension typically utilizes only word or character embeddings without explicitly taking advantage of structured linguistic information such as constituency trees and dependency trees. In this paper, we propose structural embedding of syntactic trees (SEST), an algorithm framework to utilize structured information and encode them into vector representations that can boost the performance of algorithms for the machine comprehension. We evaluate our approach using a state-of-the-art neural attention model on the SQuAD dataset. Experimental results demonstrate that our model can accurately identify the syntactic boundaries of the sentences and extract answers that are syntactically coherent over the baseline methods.

Deqing Wang, Hui Zhang, Rui Liu et al.

Much work has been done on feature selection. Existing methods are based on document frequency, such as Chi-Square Statistic, Information Gain etc. However, these methods have two shortcomings: one is that they are not reliable for low-frequency terms, and the other is that they only count whether one term occurs in a document and ignore the term frequency. Actually, high-frequency terms within a specific category are often regards as discriminators. This paper focuses on how to construct the feature selection function based on term frequency, and proposes a new approach based on $t$-test, which is used to measure the diversity of the distributions of a term between the specific category and the entire corpus. Extensive comparative experiments on two text corpora using three classifiers show that our new approach is comparable to or or slightly better than the state-of-the-art feature selection methods (i.e., $χ^2$, and IG) in terms of macro-$F_1$ and micro-$F_1$.