The recently proposed conformer architecture has been successfully used for end-to-end automatic speech recognition (ASR) architectures achieving state-of-the-art performance on different datasets. To our best knowledge, the impact of using conformer acoustic model for hybrid ASR is not investigated. In this paper, we present and evaluate a competitive conformer-based hybrid model training recipe. We study different training aspects and methods to improve word-error-rate as well as to increase training speed. We apply time downsampling methods for efficient training and use transposed convolutions to upsample the output sequence again. We conduct experiments on Switchboard 300h dataset and our conformer-based hybrid model achieves competitive results compared to other architectures. It generalizes very well on Hub5'01 test set and outperforms the BLSTM-based hybrid model significantly.
We introduce MTG, a new benchmark suite for training and evaluating multilingual text generation. It is the first and largest text generation benchmark with 120k human-annotated multi-way parallel data for three tasks (story generation, question generation, and title generation) across four languages (English, German, French, and Spanish). Based on it, we set various evaluation scenarios and make a deep analysis of several popular multilingual generation models from different aspects. Our benchmark suite will encourage the multilingualism for text generation community with more human-annotated parallel data and more diverse generation scenarios.
Sequence-to-sequence (seq2seq) problems such as machine translation are bidirectional, which naturally derive a pair of directional tasks and two directional learning signals. However, typical seq2seq neural networks are {\em simplex} that only model one unidirectional task, which cannot fully exploit the potential of bidirectional learning signals from parallel data. To address this issue, we propose a {\em duplex} seq2seq neural network, REDER (Reversible Duplex Transformer), and apply it to machine translation. The architecture of REDER has two ends, each of which specializes in a language so as to read and yield sequences in that language. As a result, REDER can simultaneously learn from the bidirectional signals, and enables {\em reversible machine translation} by simply flipping the input and output ends, Experiments on widely-used machine translation benchmarks verify that REDER achieves the first success of reversible machine translation, which helps obtain considerable gains over several strong baselines.
It is well accepted that the choice of token vocabulary largely affects the performance of machine translation. However, due to expensive trial costs, most studies only conduct simple trials with dominant approaches (e.g BPE) and commonly used vocabulary sizes. In this paper, we find an exciting relation between an information-theoretic feature and BLEU scores. With this observation, we formulate the quest of vocabularization -- finding the best token dictionary with a proper size -- as an optimal transport problem. We then propose VOLT, a simple and efficient vocabularization solution without the full and costly trial training. We evaluate our approach on multiple machine translation tasks, including WMT-14 English-German translation, TED bilingual translation, and TED multilingual translation. Empirical results show that VOLT beats widely-used vocabularies on diverse scenarios. For example, VOLT achieves 70% vocabulary size reduction and 0.6 BLEU gain on English-German translation. Also, one advantage of VOLT lies in its low resource consumption. Compared to naive BPE-search, VOLT reduces the search time from 288 GPU hours to 0.5 CPU hours.
Long text generation is an important but challenging task.The main problem lies in learning sentence-level semantic dependencies which traditional generative models often suffer from. To address this problem, we propose a Multi-hop Reasoning Generation (MRG) approach that incorporates multi-hop reasoning over a knowledge graph to learn semantic dependencies among sentences. MRG consists of twoparts, a graph-based multi-hop reasoning module and a path-aware sentence realization module. The reasoning module is responsible for searching skeleton paths from a knowledge graph to imitate the imagination process in the human writing for semantic transfer. Based on the inferred paths, the sentence realization module then generates a complete sentence. Unlike previous black-box models, MRG explicitly infers the skeleton path, which provides explanatory views tounderstand how the proposed model works. We conduct experiments on three representative tasks, including story generation, review generation, and product description generation. Automatic and manual evaluation show that our proposed method can generate more informative and coherentlong text than strong baselines, such as pre-trained models(e.g. GPT-2) and knowledge-enhanced models.
In sequence to sequence learning, the self-attention mechanism proves to be highly effective, and achieves significant improvements in many tasks. However, the self-attention mechanism is not without its own flaws. Although self-attention can model extremely long dependencies, the attention in deep layers tends to overconcentrate on a single token, leading to insufficient use of local information and difficultly in representing long sequences. In this work, we explore parallel multi-scale representation learning on sequence data, striving to capture both long-range and short-range language structures. To this end, we propose the Parallel MUlti-Scale attEntion (MUSE) and MUSE-simple. MUSE-simple contains the basic idea of parallel multi-scale sequence representation learning, and it encodes the sequence in parallel, in terms of different scales with the help from self-attention, and pointwise transformation. MUSE builds on MUSE-simple and explores combining convolution and self-attention for learning sequence representations from more different scales. We focus on machine translation and the proposed approach achieves substantial performance improvements over Transformer, especially on long sequences. More importantly, we find that although conceptually simple, its success in practice requires intricate considerations, and the multi-scale attention must build on unified semantic space. Under common setting, the proposed model achieves substantial performance and outperforms all previous models on three main machine translation tasks. In addition, MUSE has potential for accelerating inference due to its parallelism. Code will be available at https://github.com/lancopku/MUSE
Layer normalization (LayerNorm) is a technique to normalize the distributions of intermediate layers. It enables smoother gradients, faster training, and better generalization accuracy. However, it is still unclear where the effectiveness stems from. In this paper, our main contribution is to take a step further in understanding LayerNorm. Many of previous studies believe that the success of LayerNorm comes from forward normalization. Unlike them, we find that the derivatives of the mean and variance are more important than forward normalization by re-centering and re-scaling backward gradients. Furthermore, we find that the parameters of LayerNorm, including the bias and gain, increase the risk of over-fitting and do not work in most cases. Experiments show that a simple version of LayerNorm (LayerNorm-simple) without the bias and gain outperforms LayerNorm on four datasets. It obtains the state-of-the-art performance on En-Vi machine translation. To address the over-fitting problem, we propose a new normalization method, Adaptive Normalization (AdaNorm), by replacing the bias and gain with a new transformation function. Experiments show that AdaNorm demonstrates better results than LayerNorm on seven out of eight datasets.
We study fact-checking in this paper, which aims to verify a textual claim given textual evidence (e.g., retrieved sentences from Wikipedia). Existing studies typically either concatenate retrieved sentences as a single string or use feature fusion on the top of features of sentences, while ignoring semantic-level information including participants, location, and temporality of an event occurred in a sentence and relationships among multiple events. Such semantic-level information is crucial for understanding the relational structure of evidence and the deep reasoning procedure over that. In this paper, we address this issue by proposing a graph-based reasoning framework, called the Dynamic REAsoning Machine (DREAM) framework. We first construct a semantic-level graph, where nodes are extracted by semantic role labeling toolkits and are connected by inner- and inter- sentence edges. After having the automatically constructed graph, we use XLNet as the backbone of our approach and propose a graph-based contextual word representation learning module and a graph-based reasoning module to leverage the information of graphs. The first module is designed by considering a claim as a sequence, in which case we use the graph structure to re-define the relative distance of words. On top of this, we propose the second module by considering both the claim and the evidence as graphs and use a graph neural network to capture the semantic relationship at a more abstract level. We conduct experiments on FEVER, a large-scale benchmark dataset for fact-checking. Results show that both of the graph-based modules improve performance. Our system is the state-of-the-art system on the public leaderboard in terms of both accuracy and FEVER score.
Commonsense question answering aims to answer questions which require background knowledge that is not explicitly expressed in the question. The key challenge is how to obtain evidence from external knowledge and make predictions based on the evidence. Recent works either learn to generate evidence from human-annotated evidence which is expensive to collect, or extract evidence from either structured or unstructured knowledge bases which fails to take advantages of both sources. In this work, we propose to automatically extract evidence from heterogeneous knowledge sources, and answer questions based on the extracted evidence. Specifically, we extract evidence from both structured knowledge base (i.e. ConceptNet) and Wikipedia plain texts. We construct graphs for both sources to obtain the relational structures of evidence. Based on these graphs, we propose a graph-based approach consisting of a graph-based contextual word representation learning module and a graph-based inference module. The first module utilizes graph structural information to re-define the distance between words for learning better contextual word representations. The second module adopts graph convolutional network to encode neighbor information into the representations of nodes, and aggregates evidence with graph attention mechanism for predicting the final answer. Experimental results on CommonsenseQA dataset illustrate that our graph-based approach over both knowledge sources brings improvement over strong baselines. Our approach achieves the state-of-the-art accuracy (75.3%) on the CommonsenseQA leaderboard.