Yuxian Meng

Shuhe Wang, Yuxian Meng, Rongbin Ouyang et al.

To better handle long-tail cases in the sequence labeling (SL) task, in this work, we introduce graph neural networks sequence labeling (GNN-SL), which augments the vanilla SL model output with similar tagging examples retrieved from the whole training set. Since not all the retrieved tagging examples benefit the model prediction, we construct a heterogeneous graph, and leverage graph neural networks (GNNs) to transfer information between the retrieved tagging examples and the input word sequence. The augmented node which aggregates information from neighbors is used to do prediction. This strategy enables the model to directly acquire similar tagging examples and improves the general quality of predictions. We conduct a variety of experiments on three typical sequence labeling tasks: Named Entity Recognition (NER), Part of Speech Tagging (POS), and Chinese Word Segmentation (CWS) to show the significant performance of our GNN-SL. Notably, GNN-SL achieves SOTA results of 96.9 (+0.2) on PKU, 98.3 (+0.4) on CITYU, 98.5 (+0.2) on MSR, and 96.9 (+0.2) on AS for the CWS task, and results comparable to SOTA performances on NER datasets, and POS datasets.

Shuhe Wang, Xiaoya Li, Yuxian Meng et al.

Inspired by recent advances in retrieval augmented methods in NLP~\citep{khandelwal2019generalization,khandelwal2020nearest,meng2021gnn}, in this paper, we introduce a $k$ nearest neighbor NER ($k$NN-NER) framework, which augments the distribution of entity labels by assigning $k$ nearest neighbors retrieved from the training set. This strategy makes the model more capable of handling long-tail cases, along with better few-shot learning abilities. $k$NN-NER requires no additional operation during the training phase, and by interpolating $k$ nearest neighbors search into the vanilla NER model, $k$NN-NER consistently outperforms its vanilla counterparts: we achieve a new state-of-the-art F1-score of 72.03 (+1.25) on the Chinese Weibo dataset and improved results on a variety of widely used NER benchmarks. Additionally, we show that $k$NN-NER can achieve comparable results to the vanilla NER model with 40\% less amount of training data. Code available at \url{https://github.com/ShannonAI/KNN-NER}.

Yuxian Meng, Shi Zong, Xiaoya Li et al.

Inspired by the notion that ``{\it to copy is easier than to memorize}``, in this work, we introduce GNN-LM, which extends the vanilla neural language model (LM) by allowing to reference similar contexts in the entire training corpus. We build a directed heterogeneous graph between an input context and its semantically related neighbors selected from the training corpus, where nodes are tokens in the input context and retrieved neighbor contexts, and edges represent connections between nodes. Graph neural networks (GNNs) are constructed upon the graph to aggregate information from similar contexts to decode the token. This learning paradigm provides direct access to the reference contexts and helps improve a model's generalization ability. We conduct comprehensive experiments to validate the effectiveness of the GNN-LM: GNN-LM achieves a new state-of-the-art perplexity of 14.8 on WikiText-103 (a 3.9 point improvement over its counterpart of the vanilla LM model), and shows substantial improvement on One Billion Word and Enwiki8 datasets against strong baselines. In-depth ablation studies are performed to understand the mechanics of GNN-LM. \footnote{The code can be found at https://github.com/ShannonAI/GNN-LM

Nan Wang, Jiwei Li, Yuxian Meng et al.

Semantic Role Labeling (SRL) aims at recognizing the predicate-argument structure of a sentence and can be decomposed into two subtasks: predicate disambiguation and argument labeling. Prior work deals with these two tasks independently, which ignores the semantic connection between the two tasks. In this paper, we propose to use the machine reading comprehension (MRC) framework to bridge this gap. We formalize predicate disambiguation as multiple-choice machine reading comprehension, where the descriptions of candidate senses of a given predicate are used as options to select the correct sense. The chosen predicate sense is then used to determine the semantic roles for that predicate, and these semantic roles are used to construct the query for another MRC model for argument labeling. In this way, we are able to leverage both the predicate semantics and the semantic role semantics for argument labeling. We also propose to select a subset of all the possible semantic roles for computational efficiency. Experiments show that the proposed framework achieves state-of-the-art or comparable results to previous work. Code is available at \url{https://github.com/ShannonAI/MRC-SRL}.

Xiaoya Li, Jiwei Li, Xiaofei Sun et al.

Out-of-Distribution (OOD) detection is an important problem in natural language processing (NLP). In this work, we propose a simple yet effective framework $k$Folden, which mimics the behaviors of OOD detection during training without the use of any external data. For a task with $k$ training labels, $k$Folden induces $k$ sub-models, each of which is trained on a subset with $k-1$ categories with the left category masked unknown to the sub-model. Exposing an unknown label to the sub-model during training, the model is encouraged to learn to equally attribute the probability to the seen $k-1$ labels for the unknown label, enabling this framework to simultaneously resolve in- and out-distribution examples in a natural way via OOD simulations. Taking text classification as an archetype, we develop benchmarks for OOD detection using existing text classification datasets. By conducting comprehensive comparisons and analyses on the developed benchmarks, we demonstrate the superiority of $k$Folden against current methods in terms of improving OOD detection performances while maintaining improved in-domain classification accuracy. The code and datasets can be found at: \url{https://github.com/ShannonAI/kfolden-ood-detection}.

Zijun Sun, Xiaoya Li, Xiaofei Sun et al.

Recent pretraining models in Chinese neglect two important aspects specific to the Chinese language: glyph and pinyin, which carry significant syntax and semantic information for language understanding. In this work, we propose ChineseBERT, which incorporates both the {\it glyph} and {\it pinyin} information of Chinese characters into language model pretraining. The glyph embedding is obtained based on different fonts of a Chinese character, being able to capture character semantics from the visual features, and the pinyin embedding characterizes the pronunciation of Chinese characters, which handles the highly prevalent heteronym phenomenon in Chinese (the same character has different pronunciations with different meanings). Pretrained on large-scale unlabeled Chinese corpus, the proposed ChineseBERT model yields significant performance boost over baseline models with fewer training steps. The porpsoed model achieves new SOTA performances on a wide range of Chinese NLP tasks, including machine reading comprehension, natural language inference, text classification, sentence pair matching, and competitive performances in named entity recognition. Code and pretrained models are publicly available at https://github.com/ShannonAI/ChineseBert.

Yuxian Meng, Xiaoya Li, Xiayu Zheng et al.

Though nearest neighbor Machine Translation ($k$NN-MT) \citep{khandelwal2020nearest} has proved to introduce significant performance boosts over standard neural MT systems, it is prohibitively slow since it uses the entire reference corpus as the datastore for the nearest neighbor search. This means each step for each beam in the beam search has to search over the entire reference corpus. $k$NN-MT is thus two-orders slower than vanilla MT models, making it hard to be applied to real-world applications, especially online services. In this work, we propose Fast $k$NN-MT to address this issue. Fast $k$NN-MT constructs a significantly smaller datastore for the nearest neighbor search: for each word in a source sentence, Fast $k$NN-MT first selects its nearest token-level neighbors, which is limited to tokens that are the same as the query token. Then at each decoding step, in contrast to using the entire corpus as the datastore, the search space is limited to target tokens corresponding to the previously selected reference source tokens. This strategy avoids search through the whole datastore for nearest neighbors and drastically improves decoding efficiency. Without loss of performance, Fast $k$NN-MT is two-orders faster than $k$NN-MT, and is only two times slower than the standard NMT model. Fast $k$NN-MT enables the practical use of $k$NN-MT systems in real-world MT applications. The code is available at \url{https://github.com/ShannonAI/fast-knn-nmt}

Shuhe Wang, Yuxian Meng, Xiaofei Sun et al.

Multi-modal dialog modeling is of growing interest. In this work, we propose frameworks to resolve a specific case of multi-modal dialog generation that better mimics multi-modal dialog generation in the real world, where each dialog turn is associated with the visual context in which it takes place. Specifically, we propose to model the mutual dependency between text-visual features, where the model not only needs to learn the probability of generating the next dialog utterance given preceding dialog utterances and visual contexts, but also the probability of predicting the visual features in which a dialog utterance takes place, leading the generated dialog utterance specific to the visual context. We observe significant performance boosts over vanilla models when the mutual dependency between text and visual features is modeled. Code is available at https://github.com/ShannonAI/OpenViDial.

Leilei Gan, Yuxian Meng, Kun Kuang et al.

Higher-order methods for dependency parsing can partially but not fully address the issue that edges in dependency trees should be constructed at the text span/subtree level rather than word level. In this paper, we propose a new method for dependency parsing to address this issue. The proposed method constructs dependency trees by directly modeling span-span (in other words, subtree-subtree) relations. It consists of two modules: the {\it text span proposal module} which proposes candidate text spans, each of which represents a subtree in the dependency tree denoted by (root, start, end); and the {\it span linking module}, which constructs links between proposed spans. We use the machine reading comprehension (MRC) framework as the backbone to formalize the span linking module, where one span is used as a query to extract the text span/subtree it should be linked to. The proposed method has the following merits: (1) it addresses the fundamental problem that edges in a dependency tree should be constructed between subtrees; (2) the MRC framework allows the method to retrieve missing spans in the span proposal stage, which leads to higher recall for eligible spans. Extensive experiments on the PTB, CTB and Universal Dependencies (UD) benchmarks demonstrate the effectiveness of the proposed method. The code is available at \url{https://github.com/ShannonAI/mrc-for-dependency-parsing}

Yuxiao Lin, Yuxian Meng, Xiaofei Sun et al.

In this work, we propose BertGCN, a model that combines large scale pretraining and transductive learning for text classification. BertGCN constructs a heterogeneous graph over the dataset and represents documents as nodes using BERT representations. By jointly training the BERT and GCN modules within BertGCN, the proposed model is able to leverage the advantages of both worlds: large-scale pretraining which takes the advantage of the massive amount of raw data and transductive learning which jointly learns representations for both training data and unlabeled test data by propagating label influence through graph convolution. Experiments show that BertGCN achieves SOTA performances on a wide range of text classification datasets. Code is available at https://github.com/ZeroRin/BertGCN.

Yuxian Meng, Wei Wu, Fei Wang et al.

It is intuitive that NLP tasks for logographic languages like Chinese should benefit from the use of the glyph information in those languages. However, due to the lack of rich pictographic evidence in glyphs and the weak generalization ability of standard computer vision models on character data, an effective way to utilize the glyph information remains to be found. In this paper, we address this gap by presenting Glyce, the glyph-vectors for Chinese character representations. We make three major innovations: (1) We use historical Chinese scripts (e.g., bronzeware script, seal script, traditional Chinese, etc) to enrich the pictographic evidence in characters; (2) We design CNN structures (called tianzege-CNN) tailored to Chinese character image processing; and (3) We use image-classification as an auxiliary task in a multi-task learning setup to increase the model's ability to generalize. We show that glyph-based models are able to consistently outperform word/char ID-based models in a wide range of Chinese NLP tasks. We are able to set new state-of-the-art results for a variety of Chinese NLP tasks, including tagging (NER, CWS, POS), sentence pair classification, single sentence classification tasks, dependency parsing, and semantic role labeling. For example, the proposed model achieves an F1 score of 80.6 on the OntoNotes dataset of NER, +1.5 over BERT; it achieves an almost perfect accuracy of 99.8\% on the Fudan corpus for text classification. Code found at https://github.com/ShannonAI/glyce.

Shuhe Wang, Jiwei Li, Yuxian Meng et al.

$k$NN based neural machine translation ($k$NN-MT) has achieved state-of-the-art results in a variety of MT tasks. One significant shortcoming of $k$NN-MT lies in its inefficiency in identifying the $k$ nearest neighbors of the query representation from the entire datastore, which is prohibitively time-intensive when the datastore size is large. In this work, we propose \textbf{Faster $k$NN-MT} to address this issue. The core idea of Faster $k$NN-MT is to use a hierarchical clustering strategy to approximate the distance between the query and a data point in the datastore, which is decomposed into two parts: the distance between the query and the center of the cluster that the data point belongs to, and the distance between the data point and the cluster center. We propose practical ways to compute these two parts in a significantly faster manner. Through extensive experiments on different MT benchmarks, we show that \textbf{Faster $k$NN-MT} is faster than Fast $k$NN-MT \citep{meng2021fast} and only slightly (1.2 times) slower than its vanilla counterpart while preserving model performance as $k$NN-MT. Faster $k$NN-MT enables the deployment of $k$NN-MT models on real-world MT services.

Leilei Gan, Jiwei Li, Tianwei Zhang et al.

Backdoor attacks pose a new threat to NLP models. A standard strategy to construct poisoned data in backdoor attacks is to insert triggers (e.g., rare words) into selected sentences and alter the original label to a target label. This strategy comes with a severe flaw of being easily detected from both the trigger and the label perspectives: the trigger injected, which is usually a rare word, leads to an abnormal natural language expression, and thus can be easily detected by a defense model; the changed target label leads the example to be mistakenly labeled and thus can be easily detected by manual inspections. To deal with this issue, in this paper, we propose a new strategy to perform textual backdoor attacks which do not require an external trigger, and the poisoned samples are correctly labeled. The core idea of the proposed strategy is to construct clean-labeled examples, whose labels are correct but can lead to test label changes when fused with the training set. To generate poisoned clean-labeled examples, we propose a sentence generation model based on the genetic algorithm to cater to the non-differentiable characteristic of text data. Extensive experiments demonstrate that the proposed attacking strategy is not only effective, but more importantly, hard to defend due to its triggerless and clean-labeled nature. Our work marks the first step towards developing triggerless attacking strategies in NLP.

Xiaofei Sun, Diyi Yang, Xiaoya Li et al.

Neural network models have achieved state-of-the-art performances in a wide range of natural language processing (NLP) tasks. However, a long-standing criticism against neural network models is the lack of interpretability, which not only reduces the reliability of neural NLP systems but also limits the scope of their applications in areas where interpretability is essential (e.g., health care applications). In response, the increasing interest in interpreting neural NLP models has spurred a diverse array of interpretation methods over recent years. In this survey, we provide a comprehensive review of various interpretation methods for neural models in NLP. We first stretch out a high-level taxonomy for interpretation methods in NLP, i.e., training-based approaches, test-based approaches, and hybrid approaches. Next, we describe sub-categories in each category in detail, e.g., influence-function based methods, KNN-based methods, attention-based models, saliency-based methods, perturbation-based methods, etc. We point out deficiencies of current methods and suggest some avenues for future research.

Kangjie Chen, Yuxian Meng, Xiaofei Sun et al.

Pre-trained Natural Language Processing (NLP) models can be easily adapted to a variety of downstream language tasks. This significantly accelerates the development of language models. However, NLP models have been shown to be vulnerable to backdoor attacks, where a pre-defined trigger word in the input text causes model misprediction. Previous NLP backdoor attacks mainly focus on some specific tasks. This makes those attacks less general and applicable to other kinds of NLP models and tasks. In this work, we propose \Name, the first task-agnostic backdoor attack against the pre-trained NLP models. The key feature of our attack is that the adversary does not need prior information about the downstream tasks when implanting the backdoor to the pre-trained model. When this malicious model is released, any downstream models transferred from it will also inherit the backdoor, even after the extensive transfer learning process. We further design a simple yet effective strategy to bypass a state-of-the-art defense. Experimental results indicate that our approach can compromise a wide range of downstream NLP tasks in an effective and stealthy way.

Shuhe Wang, Yuxian Meng, Xiaoya Li et al.

In order to better simulate the real human conversation process, models need to generate dialogue utterances based on not only preceding textual contexts but also visual contexts. However, with the development of multi-modal dialogue learning, the dataset scale gradually becomes a bottleneck. In this report, we release OpenViDial 2.0, a larger-scale open-domain multi-modal dialogue dataset compared to the previous version OpenViDial 1.0. OpenViDial 2.0 contains a total number of 5.6 million dialogue turns extracted from either movies or TV series from different resources, and each dialogue turn is paired with its corresponding visual context. We hope this large-scale dataset can help facilitate future researches on open-domain multi-modal dialog generation, e.g., multi-modal pretraining for dialogue generation.

Xiaofei Sun, Yufei Tian, Yuxian Meng et al.

In this paper, we propose a new paradigm for paraphrase generation by treating the task as unsupervised machine translation (UMT) based on the assumption that there must be pairs of sentences expressing the same meaning in a large-scale unlabeled monolingual corpus. The proposed paradigm first splits a large unlabeled corpus into multiple clusters, and trains multiple UMT models using pairs of these clusters. Then based on the paraphrase pairs produced by these UMT models, a unified surrogate model can be trained to serve as the final \sts model to generate paraphrases, which can be directly used for test in the unsupervised setup, or be finetuned on labeled datasets in the supervised setup. The proposed method offers merits over machine-translation-based paraphrase generation methods, as it avoids reliance on bilingual sentence pairs. It also allows human intervene with the model so that more diverse paraphrases can be generated using different filtering criteria. Extensive experiments on existing paraphrase dataset for both the supervised and unsupervised setups demonstrate the effectiveness the proposed paradigm.

Yuxian Meng, Xiang Ao, Qing He et al.

A long-standing issue with paraphrase generation is how to obtain reliable supervision signals. In this paper, we propose an unsupervised paradigm for paraphrase generation based on the assumption that the probabilities of generating two sentences with the same meaning given the same context should be the same. Inspired by this fundamental idea, we propose a pipelined system which consists of paraphrase candidate generation based on contextual language models, candidate filtering using scoring functions, and paraphrase model training based on the selected candidates. The proposed paradigm offers merits over existing paraphrase generation methods: (1) using the context regularizer on meanings, the model is able to generate massive amounts of high-quality paraphrase pairs; and (2) using human-interpretable scoring functions to select paraphrase pairs from candidates, the proposed framework provides a channel for developers to intervene with the data generation process, leading to a more controllable model. Experimental results across different tasks and datasets demonstrate that the effectiveness of the proposed model in both supervised and unsupervised setups.

Chun Fan, Jiwei Li, Xiang Ao et al.

The proposed pruning strategy offers merits over weight-based pruning techniques: (1) it avoids irregular memory access since representations and matrices can be squeezed into their smaller but dense counterparts, leading to greater speedup; (2) in a manner of top-down pruning, the proposed method operates from a more global perspective based on training signals in the top layer, and prunes each layer by propagating the effect of global signals through layers, leading to better performances at the same sparsity level. Extensive experiments show that at the same sparsity level, the proposed strategy offers both greater speedup and higher performances than weight-based pruning methods (e.g., magnitude pruning, movement pruning).

Xiaofei Sun, Xiaoya Li, Yuxian Meng et al.

The frustratingly fragile nature of neural network models make current natural language generation (NLG) systems prone to backdoor attacks and generate malicious sequences that could be sexist or offensive. Unfortunately, little effort has been invested to how backdoor attacks can affect current NLG models and how to defend against these attacks. In this work, by giving a formal definition of backdoor attack and defense, we investigate this problem on two important NLG tasks, machine translation and dialog generation. Tailored to the inherent nature of NLG models (e.g., producing a sequence of coherent words given contexts), we design defending strategies against attacks. We find that testing the backward probability of generating sources given targets yields effective defense performance against all different types of attacks, and is able to handle the {\it one-to-many} issue in many NLG tasks such as dialog generation. We hope that this work can raise the awareness of backdoor risks concealed in deep NLG systems and inspire more future work (both attack and defense) towards this direction.

Chun Fan, Yuxian Meng, Xiaofei Sun et al.

The standard way to estimate the parameters $Θ_\text{SEIR}$ (e.g., the transmission rate $β$) of an SEIR model is to use grid search, where simulations are performed on each set of parameters, and the parameter set leading to the least $L_2$ distance between predicted number of infections and observed infections is selected. This brute-force strategy is not only time consuming, as simulations are slow when the population is large, but also inaccurate, since it is impossible to enumerate all parameter combinations. To address these issues, in this paper, we propose to transform the non-differentiable problem of finding optimal $Θ_\text{SEIR}$ to a differentiable one, where we first train a recurrent net to fit a small number of simulation data. Next, based on this recurrent net that is able to generalize SEIR simulations, we are able to transform the objective to a differentiable one with respect to $Θ_\text{SEIR}$, and straightforwardly obtain its optimal value. The proposed strategy is both time efficient as it only relies on a small number of SEIR simulations, and accurate as we are able to find the optimal $Θ_\text{SEIR}$ based on the differentiable objective. On two COVID-19 datasets, we observe that the proposed strategy leads to significantly better parameter estimations with a smaller number of simulations.

Xiaofei Sun, Yuxian Meng, Xiang Ao et al.

Existing methods to measure sentence similarity are faced with two challenges: (1) labeled datasets are usually limited in size, making them insufficient to train supervised neural models; (2) there is a training-test gap for unsupervised language modeling (LM) based models to compute semantic scores between sentences, since sentence-level semantics are not explicitly modeled at training. This results in inferior performances in this task. In this work, we propose a new framework to address these two issues. The proposed framework is based on the core idea that the meaning of a sentence should be defined by its contexts, and that sentence similarity can be measured by comparing the probabilities of generating two sentences given the same context. The proposed framework is able to generate high-quality, large-scale dataset with semantic similarity scores between two sentences in an unsupervised manner, with which the train-test gap can be largely bridged. Extensive experiments show that the proposed framework achieves significant performance boosts over existing baselines under both the supervised and unsupervised settings across different datasets.

Yuxian Meng, Shuhe Wang, Qinghong Han et al.

When humans converse, what a speaker will say next significantly depends on what he sees. Unfortunately, existing dialogue models generate dialogue utterances only based on preceding textual contexts, and visual contexts are rarely considered. This is due to a lack of a large-scale multi-module dialogue dataset with utterances paired with visual contexts. In this paper, we release {\bf OpenViDial}, a large-scale multi-module dialogue dataset. The dialogue turns and visual contexts are extracted from movies and TV series, where each dialogue turn is paired with the corresponding visual context in which it takes place. OpenViDial contains a total number of 1.1 million dialogue turns, and thus 1.1 million visual contexts stored in images. Based on this dataset, we propose a family of encoder-decoder models leveraging both textual and visual contexts, from coarse-grained image features extracted from CNNs to fine-grained object features extracted from Faster R-CNNs. We observe that visual information significantly improves dialogue generation qualities, verifying the necessity of integrating multi-modal features for dialogue learning. Our work marks an important step towards large-scale multi-modal dialogue learning.

Zijun Sun, Chun Fan, Qinghong Han et al.

Existing approaches to explaining deep learning models in NLP usually suffer from two major drawbacks: (1) the main model and the explaining model are decoupled: an additional probing or surrogate model is used to interpret an existing model, and thus existing explaining tools are not self-explainable; (2) the probing model is only able to explain a model's predictions by operating on low-level features by computing saliency scores for individual words but are clumsy at high-level text units such as phrases, sentences, or paragraphs. To deal with these two issues, in this paper, we propose a simple yet general and effective self-explaining framework for deep learning models in NLP. The key point of the proposed framework is to put an additional layer, as is called by the interpretation layer, on top of any existing NLP model. This layer aggregates the information for each text span, which is then associated with a specific weight, and their weighted combination is fed to the softmax function for the final prediction. The proposed model comes with the following merits: (1) span weights make the model self-explainable and do not require an additional probing model for interpretation; (2) the proposed model is general and can be adapted to any existing deep learning structures in NLP; (3) the weight associated with each text span provides direct importance scores for higher-level text units such as phrases and sentences. We for the first time show that interpretability does not come at the cost of performance: a neural model of self-explaining features obtains better performances than its counterpart without the self-explaining nature, achieving a new SOTA performance of 59.1 on SST-5 and a new SOTA performance of 92.3 on SNLI.

Zijun Sun, Chun Fan, Xiaofei Sun et al.

The goal of semi-supervised learning is to utilize the unlabeled, in-domain dataset U to improve models trained on the labeled dataset D. Under the context of large-scale language-model (LM) pretraining, how we can make the best use of U is poorly understood: is semi-supervised learning still beneficial with the presence of large-scale pretraining? should U be used for in-domain LM pretraining or pseudo-label generation? how should the pseudo-label based semi-supervised model be actually implemented? how different semi-supervised strategies affect performances regarding D of different sizes, U of different sizes, etc. In this paper, we conduct comprehensive studies on semi-supervised learning in the task of text classification under the context of large-scale LM pretraining. Our studies shed important lights on the behavior of semi-supervised learning methods: (1) with the presence of in-domain pretraining LM on U, open-domain LM pretraining is unnecessary; (2) both the in-domain pretraining strategy and the pseudo-label based strategy introduce significant performance boosts, with the former performing better with larger U, the latter performing better with smaller U, and the combination leading to the largest performance boost; (3) self-training (pretraining first on pseudo labels D' and then fine-tuning on D) yields better performances when D is small, while joint training on the combination of pseudo labels D' and the original dataset D yields better performances when D is large. Using semi-supervised learning strategies, we are able to achieve a performance of around 93.8% accuracy with only 50 training data points on the IMDB dataset, and a competitive performance of 96.6% with the full IMDB dataset. Our work marks an initial step in understanding the behavior of semi-supervised learning models under the context of large-scale pretraining.

Xiaofei Sun, Zijun Sun, Yuxian Meng et al.

The difficulty of generating coherent long texts lies in the fact that existing models overwhelmingly focus on predicting local words, and cannot make high level plans on what to generate or capture the high-level discourse dependencies between chunks of texts. Inspired by human writing processes, where a list of bullet points or a catalog is first outlined, and then each bullet point is expanded to form the whole article, we propose {\it SOE}, a pipelined system that involves of summarizing, outlining and elaborating for long text generation: the model first outlines the summaries for different segments of long texts, and then elaborates on each bullet point to generate the corresponding segment. To avoid the labor-intensive process of summary soliciting, we propose the {\it reconstruction} strategy, which extracts segment summaries in an unsupervised manner by selecting its most informative part to reconstruct the segment. The proposed generation system comes with the following merits: (1) the summary provides high-level guidance for text generation and avoids the local minimum of individual word predictions; (2) the high-level discourse dependencies are captured in the conditional dependencies between summaries and are preserved during the summary expansion process and (3) additionally, we are able to consider significantly more contexts by representing contexts as concise summaries. Extensive experiments demonstrate that SOE produces long texts with significantly better quality, along with faster convergence speed.

Yuxian Meng, Chun Fan, Zijun Sun et al.

Any prediction from a model is made by a combination of learning history and test stimuli. This provides significant insights for improving model interpretability: {\it because of which part(s) of which training example(s), the model attends to which part(s) of a test example}. Unfortunately, existing methods to interpret a model's predictions are only able to capture a single aspect of either test stimuli or learning history, and evidences from both are never combined or integrated. In this paper, we propose an efficient and differentiable approach to make it feasible to interpret a model's prediction by jointly examining training history and test stimuli. Test stimuli is first identified by gradient-based methods, signifying {\it the part of a test example that the model attends to}. The gradient-based saliency scores are then propagated to training examples using influence functions to identify {\it which part(s) of which training example(s)} make the model attends to the test stimuli. The system is differentiable and time efficient: the adoption of saliency scores from gradient-based methods allows us to efficiently trace a model's prediction through test stimuli, and then back to training examples through influence functions. We demonstrate that the proposed methodology offers clear explanations about neural model decisions, along with being useful for performing error analysis, crafting adversarial examples and fixing erroneously classified examples.

Jiawei Wu, Xiaoya Li, Xiang Ao et al.

Supervised neural networks, which first map an input $x$ to a single representation $z$, and then map $z$ to the output label $y$, have achieved remarkable success in a wide range of natural language processing (NLP) tasks. Despite their success, neural models lack for both robustness and generality: small perturbations to inputs can result in absolutely different outputs; the performance of a model trained on one domain drops drastically when tested on another domain. In this paper, we present methods to improve robustness and generality of NLP models from the standpoint of disentangled representation learning. Instead of mapping $x$ to a single representation $z$, the proposed strategy maps $x$ to a set of representations $\{z_1,z_2,...,z_K\}$ while forcing them to be disentangled. These representations are then mapped to different logits $l$s, the ensemble of which is used to make the final prediction $y$. We propose different methods to incorporate this idea into currently widely-used models, including adding an $L$2 regularizer on $z$s or adding Total Correlation (TC) under the framework of variational information bottleneck (VIB). We show that models trained with the proposed criteria provide better robustness and domain adaptation ability in a wide range of supervised learning tasks.

Xiaoya Li, Yuxian Meng, Mingxin Zhou et al.

While the self-attention mechanism has been widely used in a wide variety of tasks, it has the unfortunate property of a quadratic cost with respect to the input length, which makes it difficult to deal with long inputs. In this paper, we present a method for accelerating and structuring self-attentions: Sparse Adaptive Connection (SAC). In SAC, we regard the input sequence as a graph and attention operations are performed between linked nodes. In contrast with previous self-attention models with pre-defined structures (edges), the model learns to construct attention edges to improve task-specific performances. In this way, the model is able to select the most salient nodes and reduce the quadratic complexity regardless of the sequence length. Based on SAC, we show that previous variants of self-attention models are its special cases. Through extensive experiments on neural machine translation, language modeling, graph representation learning and image classification, we demonstrate SAC is competitive with state-of-the-art models while significantly reducing memory cost.

Qinghong Han, Yuxian Meng, Fei Wu et al.

Maximum Mutual information (MMI), which models the bidirectional dependency between responses ($y$) and contexts ($x$), i.e., the forward probability $\log p(y|x)$ and the backward probability $\log p(x|y)$, has been widely used as the objective in the \sts model to address the dull-response issue in open-domain dialog generation. Unfortunately, under the framework of the \sts model, direct decoding from $\log p(y|x) + \log p(x|y)$ is infeasible since the second part (i.e., $p(x|y)$) requires the completion of target generation before it can be computed, and the search space for $y$ is enormous. Empirically, an N-best list is first generated given $p(y|x)$, and $p(x|y)$ is then used to rerank the N-best list, which inevitably results in non-globally-optimal solutions. In this paper, we propose to use non-autoregressive (non-AR) generation model to address this non-global optimality issue. Since target tokens are generated independently in non-AR generation, $p(x|y)$ for each target word can be computed as soon as it's generated, and does not have to wait for the completion of the whole sequence. This naturally resolves the non-global optimal issue in decoding. Experimental results demonstrate that the proposed non-AR strategy produces more diverse, coherent, and appropriate responses, yielding substantive gains in BLEU scores and in human evaluations.

Xiaoya Li, Yuxian Meng, Arianna Yuan et al.

Non-autoregressive translation (NAT) models generate multiple tokens in one forward pass and is highly efficient at inference stage compared with autoregressive translation (AT) methods. However, NAT models often suffer from the multimodality problem, i.e., generating duplicated tokens or missing tokens. In this paper, we propose two novel methods to address this issue, the Look-Around (LA) strategy and the Vocabulary Attention (VA) mechanism. The Look-Around strategy predicts the neighbor tokens in order to predict the current token, and the Vocabulary Attention models long-term token dependencies inside the decoder by attending the whole vocabulary for each position to acquire knowledge of which token is about to generate. %We also propose a dynamic bidirectional decoding approach to accelerate the inference process of the LAVA model while preserving the high-quality of the generated output. Our proposed model uses significantly less time during inference compared with autoregressive models and most other NAT models. Our experiments on four benchmarks (WMT14 En$\rightarrow$De, WMT14 De$\rightarrow$En, WMT16 Ro$\rightarrow$En and IWSLT14 De$\rightarrow$En) show that the proposed model achieves competitive performance compared with the state-of-the-art non-autoregressive and autoregressive models while significantly reducing the time cost in inference phase.

Xiaoya Li, Xiaofei Sun, Yuxian Meng et al.

Many NLP tasks such as tagging and machine reading comprehension are faced with the severe data imbalance issue: negative examples significantly outnumber positive examples, and the huge number of background examples (or easy-negative examples) overwhelms the training. The most commonly used cross entropy (CE) criteria is actually an accuracy-oriented objective, and thus creates a discrepancy between training and test: at training time, each training instance contributes equally to the objective function, while at test time F1 score concerns more about positive examples. In this paper, we propose to use dice loss in replacement of the standard cross-entropy objective for data-imbalanced NLP tasks. Dice loss is based on the Sorensen-Dice coefficient or Tversky index, which attaches similar importance to false positives and false negatives, and is more immune to the data-imbalance issue. To further alleviate the dominating influence from easy-negative examples in training, we propose to associate training examples with dynamically adjusted weights to deemphasize easy-negative examples.Theoretical analysis shows that this strategy narrows down the gap between the F1 score in evaluation and the dice loss in training. With the proposed training objective, we observe significant performance boost on a wide range of data imbalanced NLP tasks. Notably, we are able to achieve SOTA results on CTB5, CTB6 and UD1.4 for the part of speech tagging task; SOTA results on CoNLL03, OntoNotes5.0, MSRA and OntoNotes4.0 for the named entity recognition task; along with competitive results on the tasks of machine reading comprehension and paraphrase identification.

Xiaoya Li, Jingrong Feng, Yuxian Meng et al.

The task of named entity recognition (NER) is normally divided into nested NER and flat NER depending on whether named entities are nested or not. Models are usually separately developed for the two tasks, since sequence labeling models, the most widely used backbone for flat NER, are only able to assign a single label to a particular token, which is unsuitable for nested NER where a token may be assigned several labels. In this paper, we propose a unified framework that is capable of handling both flat and nested NER tasks. Instead of treating the task of NER as a sequence labeling problem, we propose to formulate it as a machine reading comprehension (MRC) task. For example, extracting entities with the \textsc{per} label is formalized as extracting answer spans to the question "{\it which person is mentioned in the text?}". This formulation naturally tackles the entity overlapping issue in nested NER: the extraction of two overlapping entities for different categories requires answering two independent questions. Additionally, since the query encodes informative prior knowledge, this strategy facilitates the process of entity extraction, leading to better performances for not only nested NER, but flat NER. We conduct experiments on both {\em nested} and {\em flat} NER datasets. Experimental results demonstrate the effectiveness of the proposed formulation. We are able to achieve vast amount of performance boost over current SOTA models on nested NER datasets, i.e., +1.28, +2.55, +5.44, +6.37, respectively on ACE04, ACE05, GENIA and KBP17, along with SOTA results on flat NER datasets, i.e.,+0.24, +1.95, +0.21, +1.49 respectively on English CoNLL 2003, English OntoNotes 5.0, Chinese MSRA, Chinese OntoNotes 4.0.

Yuxian Meng, Xiangyuan Ren, Zijun Sun et al.

In this paper, we investigate the problem of training neural machine translation (NMT) systems with a dataset of more than 40 billion bilingual sentence pairs, which is larger than the largest dataset to date by orders of magnitude. Unprecedented challenges emerge in this situation compared to previous NMT work, including severe noise in the data and prohibitively long training time. We propose practical solutions to handle these issues and demonstrate that large-scale pretraining significantly improves NMT performance. We are able to push the BLEU score of WMT17 Chinese-English dataset to 32.3, with a significant performance boost of +3.2 over existing state-of-the-art results.

Yuxian Meng, Xiaoya Li, Zijun Sun et al.

In this paper, we propose a new strategy for the task of named entity recognition (NER). We cast the task as a query-based machine reading comprehension task: e.g., the task of extracting entities with PER is formalized as answering the question of "which person is mentioned in the text ?". Such a strategy comes with the advantage that it solves the long-standing issue of handling overlapping or nested entities (the same token that participates in more than one entity categories) with sequence-labeling techniques for NER. Additionally, since the query encodes informative prior knowledge, this strategy facilitates the process of entity extraction, leading to better performances. We experiment the proposed model on five widely used NER datasets on English and Chinese, including MSRA, Resume, OntoNotes, ACE04 and ACE05. The proposed model sets new SOTA results on all of these datasets.

Yuxian Meng, Muyu Li, Xiaoya Li et al.

In this paper, we aim at tackling a general issue in NLP tasks where some of the negative examples are highly similar to the positive examples, i.e., hard-negative examples. We propose the distant supervision as a regularizer (DSReg) approach to tackle this issue. The original task is converted to a multi-task learning problem, in which distant supervision is used to retrieve hard-negative examples. The obtained hard-negative examples are then used as a regularizer. The original target objective of distinguishing positive examples from negative examples is jointly optimized with the auxiliary task objective of distinguishing softened positive (i.e., hard-negative examples plus positive examples) from easy-negative examples. In the neural context, this can be done by outputting the same representation from the last neural layer to different $softmax$ functions. Using this strategy, we can improve the performance of baseline models in a range of different NLP tasks, including text classification, sequence labeling and reading comprehension.

Xiaoya Li, Yuxian Meng, Xiaofei Sun et al.

Segmenting a chunk of text into words is usually the first step of processing Chinese text, but its necessity has rarely been explored. In this paper, we ask the fundamental question of whether Chinese word segmentation (CWS) is necessary for deep learning-based Chinese Natural Language Processing. We benchmark neural word-based models which rely on word segmentation against neural char-based models which do not involve word segmentation in four end-to-end NLP benchmark tasks: language modeling, machine translation, sentence matching/paraphrase and text classification. Through direct comparisons between these two types of models, we find that char-based models consistently outperform word-based models. Based on these observations, we conduct comprehensive experiments to study why word-based models underperform char-based models in these deep learning-based NLP tasks. We show that it is because word-based models are more vulnerable to data sparsity and the presence of out-of-vocabulary (OOV) words, and thus more prone to overfitting. We hope this paper could encourage researchers in the community to rethink the necessity of word segmentation in deep learning-based Chinese Natural Language Processing. \footnote{Yuxian Meng and Xiaoya Li contributed equally to this paper.}