This work examines the presence of modularity in pre-trained Transformers, a feature commonly found in human brains and thought to be vital for general intelligence. In analogy to human brains, we consider two main characteristics of modularity: (1) functional specialization of neurons: we evaluate whether each neuron is mainly specialized in a certain function, and find that the answer is yes. (2) function-based neuron grouping: we explore to find a structure that groups neurons into modules by function, and each module works for its corresponding function. Given the enormous amount of possible structures, we focus on Mixture-of-Experts as a promising candidate, which partitions neurons into experts and usually activates different experts for different inputs. Experimental results show that there are functional experts, where clustered are the neurons specialized in a certain function. Moreover, perturbing the activations of functional experts significantly affects the corresponding function. Finally, we study how modularity emerges during pre-training, and find that the modular structure is stabilized at the early stage, which is faster than neuron stabilization. It suggests that Transformer first constructs the modular structure and then learns fine-grained neuron functions. Our code and data are available at https://github.com/THUNLP/modularity-analysis.

Parameter-efficient tuning methods (PETs) have achieved promising results in tuning large pre-trained language models (PLMs). By formalizing frozen PLMs and additional tunable parameters as systems and controls respectively, PETs can be theoretically grounded to optimal control and further viewed as optimizing the terminal cost and running cost in the optimal control literature. Despite the elegance of this theoretical grounding, in practice, existing PETs often ignore the running cost and only optimize the terminal cost, i.e., focus on optimizing the loss function of the output state, regardless of the running cost that depends on the intermediate states. Since it is non-trivial to directly model the intermediate states and design a running cost function, we propose to use latent stochastic bridges to regularize the intermediate states and use the regularization as the running cost of PETs. As the first work to propose regularized PETs that use stochastic bridges as the regularizers (running costs) for the intermediate states, we show the effectiveness and generality of this regularization across different tasks, PLMs and PETs. In view of the great potential and capacity, we believe more sophisticated regularizers can be designed for PETs and better performance can be achieved in the future.

Continual pre-training is the paradigm where pre-trained language models (PLMs) continually acquire fresh knowledge from growing data and gradually get upgraded. Before an upgraded PLM is released, we may have tuned the original PLM for various tasks and stored the adapted weights. However, when tuning the upgraded PLM, these outdated adapted weights will typically be ignored and discarded, causing a potential waste of resources. We bring this issue to the forefront and contend that proper algorithms for recycling outdated adapted weights should be developed. To this end, we formulate the task of recyclable tuning for continual pre-training. In pilot studies, we find that after continual pre-training, the upgraded PLM remains compatible with the outdated adapted weights to some extent. Motivated by this finding, we analyze the connection between continually pre-trained PLMs from two novel aspects, i.e., mode connectivity, and functional similarity. Based on the corresponding findings, we propose both an initialization-based method and a distillation-based method for our task. We demonstrate their feasibility in improving the convergence and performance for tuning the upgraded PLM. We also show that both methods can be combined to achieve better performance.

Pre-trained language models (PLM) have made impressive results in a wide range of NLP tasks and it has been revealed that one of the key factors to their success is the parameters of these models implicitly learn various types of knowledge in the pre-training corpus. However, encoding knowledge implicitly in the model parameters has two fundamental drawbacks. First, the knowledge is neither editable nor scalable once the model is trained, which is especially problematic in that knowledge is consistently evolving. Second, it lacks interpretability and prevents us from understanding what kind of knowledge PLM needs to solve a certain task. In this paper, we introduce {pasted macro ‘MODEL’}, a pre-training model with differentiable plug-in memory (DPM). The key intuition behind is to decouple the knowledge storage from model parameters with an editable and scalable key-value memory and leverage knowledge in an explainable manner by knowledge retrieval in the {pasted macro ‘MEMORY’}. We conduct extensive experiments under various settings to justify this design choice. In domain adaptation setting, {pasted macro ‘MODEL’} could be easily adapted to different domains with pluggable in-domain memory—obtaining 3.95 F1 improvements across four domains, without any in-domain training. {pasted macro ‘MODEL’} could also keep absorbing new knowledge after pre-training is done by knowledge updating operation in the {pasted macro ‘MEMORY’} without re-training. Finally, we show that by incorporating training samples into {pasted macro ‘MEMORY’} with knowledge prompting, {pasted macro ‘MODEL’} could further be improved by the instruction of in-task knowledge.

Large language models (LLMs) have achieved remarkable results on NLP tasks but at the expense of huge parameter sizes and the consequent computational costs. In this paper, we propose Variator, a parameter-efficient acceleration method that enhances computational efficiency through plug-and-play compression plugins. Compression plugins are designed to reduce the sequence length via compressing multiple hidden vectors into one and trained with original LLMs frozen. Different from traditional model acceleration methods, which compress LLMs to smaller sizes, Variator offers two distinct advantages: (1) In real-world applications, the plug-and-play nature of our compression plugins enables dynamic selection of different compression plugins with varying acceleration ratios based on the current workload. (2) The compression plugin comprises a few compact neural network layers with minimal parameters, significantly saving storage and memory overhead, particularly in scenarios with a growing number of tasks. We validate the effectiveness of Variator on seven datasets. Experimental results show that Variator can save 53% computational costs using only 0.9% additional parameters with a performance drop of less than 2%. Moreover, when the model scales to billions of parameters, Variator matches the strong performance of uncompressed LLMs. Our code and checkpoints will be released to facilitate future work.

Parameter-shared pre-trained language models (PLMs) have emerged as a successful approach in resource-constrained environments, enabling substantial reductions in model storage and memory costs without significant performance compromise. However, it is important to note that parameter sharing does not alleviate computational burdens associated with inference, thus impeding its practicality in situations characterized by limited stringent latency requirements or computational resources. Building upon neural ordinary differential equations (ODEs), we introduce a straightforward technique to enhance the inference efficiency of parameter-shared PLMs. Additionally, we propose a simple pre-training technique that leads to fully or partially shared models capable of achieving even greater inference acceleration. The experimental results demonstrate the effectiveness of our methods on both autoregressive and autoencoding PLMs, providing novel insights into more efficient utilization of parameter-shared models in resource-constrained settings.

Parameter-efficient tuning (PET) methods can effectively drive extremely large pre-trained language models (PLMs) by training only minimal parameters. Different PET methods utilize different manually designed tunable modules. In small PLMs, there are usually noticeable performance differences among PET methods. Nevertheless, as the model scale increases, the performance differences become marginal. Hence, we hypothesize that model scaling mitigates the impact of design differences on PET methods. To investigate this hypothesis, we introduce a more flexible PET method called Arbitrary PET (APET) method. The APET method is compatible with a tunable module, which consists of any number of parameters distributed in arbitrary positions. Then, we utilize it and conduct experiments on 11 NLP tasks across 3 representative PLMs. Our investigations reveal that model scaling (1) mitigates the effects of the positions of tunable parameters on performance, and (2) enables tuning methods to achieve performance comparable to full-parameter fine-tuning by optimizing fewer tunable parameters. Intriguingly, we also observe that tuning methods optimize the similar number of tunable parameters to exceed random guess performance on different tasks. We collectively discuss this phenomenon and the two aforementioned findings from an optimization perspective to understand the underlying mechanisms. These conclusions enhance our understanding of the impact of model scaling on PET and assist in designing more effective and efficient PET methods for PLMs of different scales. The source code can be obtained from this GitHub repository: https://github.com/yushengsu-thu/PET_Scaling.

Long-form question answering (LFQA) aims at answering complex, open-ended questions with detailed, paragraph-length responses. The de facto paradigm of LFQA necessitates two procedures: information retrieval, which searches for relevant supporting facts, and information synthesis, which integrates these facts into a coherent answer. In this paper, we introduce WebCPM, the first Chinese LFQA dataset. One unique feature of WebCPM is that its information retrieval is based on interactive web search, which engages with a search engine in real time. Following WebGPT, we develop a web search interface. We recruit annotators to search for relevant information using our interface and then answer questions. Meanwhile, the web search behaviors of our annotators would be recorded. In total, we collect 5,500 high-quality question-answer pairs, together with 15,372 supporting facts and 125,954 web search actions. We fine-tune pre-trained language models to imitate human behaviors for web search and to generate answers based on the collected facts. Our LFQA pipeline, built on these fine-tuned models, generates answers that are no worse than human-written ones in 32.5% and 47.5% of the cases on our dataset and DuReader, respectively. The interface, dataset, and codes are publicly available at https://github.com/thunlp/WebCPM.

Injecting external knowledge can improve the performance of pre-trained language models (PLMs) on various downstream NLP tasks. However, massive retraining is required to deploy new knowledge injection methods or knowledge bases for downstream tasks. In this work, we are the first to study how to improve the flexibility and efficiency of knowledge injection by reusing existing downstream models. To this end, we explore a new paradigm plug-and-play knowledge injection, where knowledge bases are injected into frozen existing downstream models by a knowledge plugin. Correspondingly, we propose a plug-and-play injection method map-tuning, which trains a mapping of knowledge embeddings to enrich model inputs with mapped embeddings while keeping model parameters frozen. Experimental results on three knowledge-driven NLP tasks show that existing injection methods are not suitable for the new paradigm, while map-tuning effectively improves the performance of downstream models. Moreover, we show that a frozen downstream model can be well adapted to different domains with different mapping networks of domain knowledge. Our code and models are available at https://github.com/THUNLP/Knowledge-Plugin.

Large-scale pre-trained models (PTMs) have been widely used in document-oriented NLP tasks, such as question answering. However, the encoding-task coupling requirement results in the repeated encoding of the same documents for different tasks and queries, which is highly computationally inefficient. To this end, we target to decouple document encoding from downstream tasks, and propose to represent each document as a plug-and-play document module, i.e., a document plugin, for PTMs (PlugD). By inserting document plugins into the backbone PTM for downstream tasks, we can encode a document one time to handle multiple tasks, which is more efficient than conventional encoding-task coupling methods that simultaneously encode documents and input queries using task-specific encoders. Extensive experiments on 8 datasets of 4 typical NLP tasks show that PlugD enables models to encode documents once and for all across different scenarios. Especially, PlugD can save 69% computational costs while achieving comparable performance to state-of-the-art encoding-task coupling methods. Additionally, we show that PlugD can serve as an effective post-processing way to inject knowledge into task-specific models, improving model performance without any additional model training. Our code and checkpoints can be found in https://github.com/thunlp/Document-Plugin.

Existing reference-free metrics have obvious limitations for evaluating controlled text generation models. Unsupervised metrics can only provide a task-agnostic evaluation result which correlates weakly with human judgments, whereas supervised ones may overfit task-specific data with poor generalization ability to other datasets. In this paper, we propose an unsupervised reference-free metric called CTRLEval, which evaluates controlled text generation from different aspects by formulating each aspect into multiple text infilling tasks. On top of these tasks, the metric assembles the generation probabilities from a pre-trained language model without any model training. Experimental results show that our metric has higher correlations with human judgments than other baselines, while obtaining better generalization of evaluating generated texts from different models and with different qualities.

Recent entity and relation extraction works focus on investigating how to obtain a better span representation from the pre-trained encoder. However, a major limitation of existing works is that they ignore the interrelation between spans (pairs). In this work, we propose a novel span representation approach, named Packed Levitated Markers (PL-Marker), to consider the interrelation between the spans (pairs) by strategically packing the markers in the encoder. In particular, we propose a neighborhood-oriented packing strategy, which considers the neighbor spans integrally to better model the entity boundary information. Furthermore, for those more complicated span pair classification tasks, we design a subject-oriented packing strategy, which packs each subject and all its objects to model the interrelation between the same-subject span pairs. The experimental results show that, with the enhanced marker feature, our model advances baselines on six NER benchmarks, and obtains a 4.1%-4.3% strict relation F1 improvement with higher speed over previous state-of-the-art models on ACE04 and ACE05. Our code and models are publicly available at https://github.com/thunlp/PL-Marker

Hyperbolic neural networks have shown great potential for modeling complex data. However, existing hyperbolic networks are not completely hyperbolic, as they encode features in the hyperbolic space yet formalize most of their operations in the tangent space (a Euclidean subspace) at the origin of the hyperbolic model. This hybrid method greatly limits the modeling ability of networks. In this paper, we propose a fully hyperbolic framework to build hyperbolic networks based on the Lorentz model by adapting the Lorentz transformations (including boost and rotation) to formalize essential operations of neural networks. Moreover, we also prove that linear transformation in tangent spaces used by existing hyperbolic networks is a relaxation of the Lorentz rotation and does not include the boost, implicitly limiting the capabilities of existing hyperbolic networks. The experimental results on four NLP tasks show that our method has better performance for building both shallow and deep networks. Our code will be released to facilitate follow-up research.

Pre-trained language models (PLMs) cannot well recall rich factual knowledge of entities exhibited in large-scale corpora, especially those rare entities. In this paper, we propose to build a simple but effective Pluggable Entity Lookup Table (PELT) on demand by aggregating the entity’s output representations of multiple occurrences in the corpora. PELT can be compatibly plugged as inputs to infuse supplemental entity knowledge into PLMs. Compared to previous knowledge-enhanced PLMs, PELT only requires 0.2%-5% pre-computation with capability of acquiring knowledge from out-of-domain corpora for domain adaptation scenario. The experiments on knowledge-related tasks demonstrate that our method, PELT, can flexibly and effectively transfer entity knowledge from related corpora into PLMs with different architectures. Our code and models are publicly available at https://github.com/thunlp/PELT

Recent work has shown that feed-forward networks (FFNs) in pre-trained Transformers are a key component, storing various linguistic and factual knowledge. However, the computational patterns of FFNs are still unclear. In this work, we study the computational patterns of FFNs and observe that most inputs only activate a tiny ratio of neurons of FFNs. This phenomenon is similar to the sparsity of the human brain, which drives research on functional partitions of the human brain. To verify whether functional partitions also emerge in FFNs, we propose to convert a model into its MoE version with the same parameters, namely MoEfication. Specifically, MoEfication consists of two phases: (1) splitting the parameters of FFNs into multiple functional partitions as experts, and (2) building expert routers to decide which experts will be used for each input. Experimental results show that MoEfication can conditionally use 10% to 30% of FFN parameters while maintaining over 95% original performance for different models on various downstream tasks. Besides, MoEfication brings two advantages: (1) it significantly reduces the FLOPS of inference, i.e., 2x speedup with 25% of FFN parameters, and (2) it provides a fine-grained perspective to study the inner mechanism of FFNs. The source code of this paper can be obtained from https://github.com/thunlp/MoEfication.

Current pre-trained language models (PLM) are typically trained with static data, ignoring that in real-world scenarios, streaming data of various sources may continuously grow. This requires PLMs to integrate the information from all the sources in a lifelong manner. Although this goal could be achieved by exhaustive pre-training on all the existing data, such a process is known to be computationally expensive. To this end, we propose ELLE, aiming at efficient lifelong pre-training for emerging data. Specifically, ELLE consists of (1) function preserved model expansion, which flexibly expands an existing PLM’s width and depth to improve the efficiency of knowledge acquisition; and (2) pre-trained domain prompts, which disentangle the versatile knowledge learned during pre-training and stimulate the proper knowledge for downstream tasks. We experiment ELLE with streaming data from 5 domains on BERT and GPT. The results show the superiority of ELLE over various lifelong learning baselines in both pre-training efficiency and downstream performances. The codes are publicly available at https://github.com/thunlp/ELLE.

In recent years, pre-trained language models (PLMs) have been shown to capture factual knowledge from massive texts, which encourages the proposal of PLM-based knowledge graph completion (KGC) models. However, these models are still quite behind the SOTA KGC models in terms of performance. In this work, we find two main reasons for the weak performance: (1) Inaccurate evaluation setting. The evaluation setting under the closed-world assumption (CWA) may underestimate the PLM-based KGC models since they introduce more external knowledge; (2) Inappropriate utilization of PLMs. Most PLM-based KGC models simply splice the labels of entities and relations as inputs, leading to incoherent sentences that do not take full advantage of the implicit knowledge in PLMs. To alleviate these problems, we highlight a more accurate evaluation setting under the open-world assumption (OWA), which manual checks the correctness of knowledge that is not in KGs. Moreover, motivated by prompt tuning, we propose a novel PLM-based KGC model named PKGC. The basic idea is to convert each triple and its support information into natural prompt sentences, which is further fed into PLMs for classification. Experiment results on two KGC datasets demonstrate OWA is more reliable for evaluating KGC, especially on the link prediction, and the effectiveness of our PKCG model on both CWA and OWA settings.

Despite the remarkable success deep models have achieved in Textual Matching (TM) tasks, it still remains unclear whether they truly understand language or measure the semantic similarity of texts by exploiting statistical bias in datasets. In this work, we provide a new perspective to study this issue — via the length divergence bias. We find the length divergence heuristic widely exists in prevalent TM datasets, providing direct cues for prediction. To determine whether TM models have adopted such heuristic, we introduce an adversarial evaluation scheme which invalidates the heuristic. In this adversarial setting, all TM models perform worse, indicating they have indeed adopted this heuristic. Through a well-designed probing experiment, we empirically validate that the bias of TM models can be attributed in part to extracting the text length information during training. To alleviate the length divergence bias, we propose an adversarial training method. The results demonstrate we successfully improve the robustness and generalization ability of models at the same time.

Delta tuning (DET, also known as parameter-efficient tuning) is deemed as the new paradigm for using pre-trained language models (PLMs). Up to now, various DETs with distinct design elements have been proposed, achieving performance on par with fine-tuning. However, the mechanisms behind the above success are still under-explored, especially the connections among various DETs. To fathom the mystery, we hypothesize that the adaptations of different DETs could all be reparameterized as low-dimensional optimizations in a unified optimization subspace, which could be found by jointly decomposing independent solutions of different DETs. Then we explore the connections among different DETs by conducting optimization within the subspace. In experiments, we find that, for a certain DET, conducting optimization simply in the subspace could achieve comparable performance to its original space, and the found solution in the subspace could be transferred to another DET and achieve non-trivial performance. We also visualize the performance landscape of the subspace, and find that, there exists a substantial region where different DETs all perform well. Finally, we extend our analysis and show the strong connections between fine-tuning and DETs. The codes are publicly available at https://github.com/thunlp/Unified-DeltaTuning.

Investigating better ways to reuse the released pre-trained language models (PLMs) can significantly reduce the computational cost and the potential environmental side-effects. This paper explores a novel PLM reuse paradigm, Knowledge Integration (KI). Without human annotations available, KI aims to merge the knowledge from different teacher-PLMs, each of which specializes in a different classification problem, into a versatile student model. To achieve this, we first derive the correlation between virtual golden supervision and teacher predictions. We then design a Model Uncertainty–aware Knowledge Integration (MUKI) framework to recover the golden supervision for the student. Specifically, MUKI adopts Monte-Carlo Dropout to estimate model uncertainty for the supervision integration. An instance-wise re-weighting mechanism based on the margin of uncertainty scores is further incorporated, to deal with the potential conflicting supervision from teachers.Experimental results demonstrate that MUKI achieves substantial improvements over baselines on benchmark datasets. Further analysis shows that MUKI can generalize well for merging teacher models with heterogeneous architectures, and even teachers major in cross-lingual datasets.

Recent explorations of large-scale pre-trained language models (PLMs) have revealed the power of PLMs with huge amounts of parameters, setting off a wave of training ever-larger PLMs. However, it requires tremendous computational resources to train a large-scale PLM, which may be practically unaffordable. In addition, existing large-scale PLMs are mainly trained from scratch individually, ignoring that many well-trained PLMs are available. To this end, we explore the question how could existing PLMs benefit training large-scale PLMs in future. Specifically, we introduce a pre-training framework named “knowledge inheritance” (KI) and explore how could knowledge distillation serve as auxiliary supervision during pre-training to efficiently learn larger PLMs. Experimental results demonstrate the superiority of KI in training efficiency. We also conduct empirical analyses to explore the effects of teacher PLMs’ pre-training settings, including model architecture, pre-training data, etc. Finally, we show that KI could be applied to domain adaptation and knowledge transfer.

Prompt tuning (PT) is a promising parameter-efficient method to utilize extremely large pre-trained language models (PLMs), which can achieve comparable performance to full-parameter fine-tuning by only tuning a few soft prompts. However, PT requires much more training time than fine-tuning. Intuitively, knowledge transfer can help to improve the efficiency. To explore whether we can improve PT via prompt transfer, we empirically investigate the transferability of soft prompts across different downstream tasks and PLMs in this work. We find that (1) in zero-shot setting, trained soft prompts can effectively transfer to similar tasks on the same PLM and also to other PLMs with a cross-model projector trained on similar tasks; (2) when used as initialization, trained soft prompts of similar tasks and projected prompts of other PLMs can significantly accelerate training and also improve the performance of PT. Moreover, to explore what decides prompt transferability, we investigate various transferability indicators and find that the overlapping rate of activated neurons strongly reflects the transferability, which suggests how the prompts stimulate PLMs is essential. Our findings show that prompt transfer is promising for improving PT, and further research shall focus more on prompts’ stimulation to PLMs. The source code can be obtained from https://github.com/thunlp/Prompt-Transferability.

The diverse relationships among real-world events, including coreference, temporal, causal, and subevent relations, are fundamental to understanding natural languages. However, two drawbacks of existing datasets limit event relation extraction (ERE) tasks: (1) Small scale. Due to the annotation complexity, the data scale of existing datasets is limited, which cannot well train and evaluate data-hungry models. (2) Absence of unified annotation. Different types of event relations naturally interact with each other, but existing datasets only cover limited relation types at once, which prevents models from taking full advantage of relation interactions. To address these issues, we construct a unified large-scale human-annotated ERE dataset MAVEN-ERE with improved annotation schemes. It contains 103,193 event coreference chains, 1,216,217 temporal relations, 57,992 causal relations, and 15,841 subevent relations, which is larger than existing datasets of all the ERE tasks by at least an order of magnitude. Experiments show that ERE on MAVEN-ERE is quite challenging, and considering relation interactions with joint learning can improve performances. The dataset and source codes can be obtained from https://github.com/THU-KEG/MAVEN-ERE.

Even though the large-scale language models have achieved excellent performances, they suffer from various adversarial attacks.A large body of defense methods has been proposed. However, they are still limited due to redundant attack search spaces and the inability to defend against various types of attacks.In this work, we present a novel fine-tuning approach called RObust SEletive fine-tuning (ROSE) to address this issue.ROSE conducts selective updates when adapting pre-trained models to downstream tasks, filtering out invaluable and unrobust updates of parameters.Specifically, we propose two strategies: the first-order and second-order ROSE for selecting target robust parameters.The experimental results show that ROSE achieves significant improvements in adversarial robustness on various downstream NLP tasks, and the ensemble method even surpasses both variants above.Furthermore, ROSE can be easily incorporated into existing fine-tuning methods to improve their adversarial robustness further.The empirical analysis confirms that ROSE eliminates unrobust spurious updates during fine-tuning, leading to solutions corresponding to flatter and wider optima than the conventional method.Code is available at https://github.com/jiangllan/ROSE.

Recent years have witnessed the prevalent application of pre-trained language models (PLMs) in NLP. From the perspective of parameter space, PLMs provide generic initialization, starting from which high-performance minima could be found. Although plenty of works have studied how to effectively and efficiently adapt PLMs to high-performance minima, little is known about the connection of various minima reached under different adaptation configurations. In this paper, we investigate the geometric connections of different minima through the lens of mode connectivity, which measures whether two minima can be connected with a low-loss path. We conduct empirical analyses to investigate three questions: (1) how could hyperparameters, specific tuning methods, and training data affect PLM’s mode connectivity? (2) How does mode connectivity change during pre-training? (3) How does the PLM’s task knowledge change along the path connecting two minima? In general, exploring the mode connectivity of PLMs conduces to understanding the geometric connection of different minima, which may help us fathom the inner workings of PLM downstream adaptation. The codes are publicly available at https://github.com/thunlp/Mode-Connectivity-PLM.

Prompting, which casts downstream applications as language modeling tasks, has shown to be sample efficient compared to standard fine-tuning with pre-trained models. However, one pitfall of prompting is the need of manually-designed patterns, whose outcome can be unintuitive and requires large validation sets to tune. To tackle the challenge, we propose AutoSeq, a fully automatic prompting method: (1) We adopt natural language prompts on sequence-to-sequence models, enabling free-form generation and larger label search space; (2) We propose label sequences – phrases with indefinite lengths to verbalize the labels – which eliminate the need of manual templates and are more expressive than single label words; (3) We use beam search to automatically generate a large amount of label sequence candidates and propose contrastive re-ranking to get the best combinations. AutoSeq significantly outperforms other no-manual-design methods, such as soft prompt tuning, adapter tuning, and automatic search on single label words; the generated label sequences are even better than curated manual ones on a variety of tasks. Our method reveals the potential of sequence-to-sequence models in few-shot learning and sheds light on a path to generic and automatic prompting. The source code of this paper can be obtained from https://github.com/thunlp/Seq2Seq-Prompt.

Existing pre-trained language models (PLMs) are often computationally expensive in inference, making them impractical in various resource-limited real-world applications. To address this issue, we propose a dynamic token reduction approach to accelerate PLMs’ inference, named TR-BERT, which could flexibly adapt the layer number of each token in inference to avoid redundant calculation. Specially, TR-BERT formulates the token reduction process as a multi-step token selection problem and automatically learns the selection strategy via reinforcement learning. The experimental results on several downstream NLP tasks show that TR-BERT is able to speed up BERT by 2-5 times to satisfy various performance demands. Moreover, TR-BERT can also achieve better performance with less computation in a suite of long-text tasks since its token-level layer number adaption greatly accelerates the self-attention operation in PLMs. The source code and experiment details of this paper can be obtained from https://github.com/thunlp/TR-BERT.

Knowledge distillation (KD) has been proved effective for compressing large-scale pre-trained language models. However, existing methods conduct KD statically, e.g., the student model aligns its output distribution to that of a selected teacher model on the pre-defined training dataset. In this paper, we explore whether a dynamic knowledge distillation that empowers the student to adjust the learning procedure according to its competency, regarding the student performance and learning efficiency. We explore the dynamical adjustments on three aspects: teacher model adoption, data selection, and KD objective adaptation. Experimental results show that (1) proper selection of teacher model can boost the performance of student model; (2) conducting KD with 10% informative instances achieves comparable performance while greatly accelerates the training; (3) the student performance can be boosted by adjusting the supervision contribution of different alignment objective. We find dynamic knowledge distillation is promising and provide discussions on potential future directions towards more efficient KD methods.

Existing relation extraction (RE) methods typically focus on extracting relational facts between entity pairs within single sentences or documents. However, a large quantity of relational facts in knowledge bases can only be inferred across documents in practice. In this work, we present the problem of cross-document RE, making an initial step towards knowledge acquisition in the wild. To facilitate the research, we construct the first human-annotated cross-document RE dataset CodRED. Compared to existing RE datasets, CodRED presents two key challenges: Given two entities, (1) it requires finding the relevant documents that can provide clues for identifying their relations; (2) it requires reasoning over multiple documents to extract the relational facts. We conduct comprehensive experiments to show that CodRED is challenging to existing RE methods including strong BERT-based models.

Backdoor attacks, which maliciously control a well-trained model’s outputs of the instances with specific triggers, are recently shown to be serious threats to the safety of reusing deep neural networks (DNNs). In this work, we propose an efficient online defense mechanism based on robustness-aware perturbations. Specifically, by analyzing the backdoor training process, we point out that there exists a big gap of robustness between poisoned and clean samples. Motivated by this observation, we construct a word-based robustness-aware perturbation to distinguish poisoned samples from clean samples to defend against the backdoor attacks on natural language processing (NLP) models. Moreover, we give a theoretical analysis about the feasibility of our robustness-aware perturbation-based defense method. Experimental results on sentiment analysis and toxic detection tasks show that our method achieves better defending performance and much lower computational costs than existing online defense methods. Our code is available at https://github.com/lancopku/RAP.

Pre-trained Language Models (PLMs) have shown superior performance on various downstream Natural Language Processing (NLP) tasks. However, conventional pre-training objectives do not explicitly model relational facts in text, which are crucial for textual understanding. To address this issue, we propose a novel contrastive learning framework ERICA to obtain a deep understanding of the entities and their relations in text. Specifically, we define two novel pre-training tasks to better understand entities and relations: (1) the entity discrimination task to distinguish which tail entity can be inferred by the given head entity and relation; (2) the relation discrimination task to distinguish whether two relations are close or not semantically, which involves complex relational reasoning. Experimental results demonstrate that ERICA can improve typical PLMs (BERT and RoBERTa) on several language understanding tasks, including relation extraction, entity typing and question answering, especially under low-resource settings.

Recent researches have shown that large natural language processing (NLP) models are vulnerable to a kind of security threat called the Backdoor Attack. Backdoor attacked models can achieve good performance on clean test sets but perform badly on those input sentences injected with designed trigger words. In this work, we point out a potential problem of current backdoor attacking research: its evaluation ignores the stealthiness of backdoor attacks, and most of existing backdoor attacking methods are not stealthy either to system deployers or to system users. To address this issue, we first propose two additional stealthiness-based metrics to make the backdoor attacking evaluation more credible. We further propose a novel word-based backdoor attacking method based on negative data augmentation and modifying word embeddings, making an important step towards achieving stealthy backdoor attacking. Experiments on sentiment analysis and toxic detection tasks show that our method is much stealthier while maintaining pretty good attacking performance. Our code is available at https://github.com/lancopku/SOS.

Event extraction (EE) has considerably benefited from pre-trained language models (PLMs) by fine-tuning. However, existing pre-training methods have not involved modeling event characteristics, resulting in the developed EE models cannot take full advantage of large-scale unsupervised data. To this end, we propose CLEVE, a contrastive pre-training framework for EE to better learn event knowledge from large unsupervised data and their semantic structures (e.g. AMR) obtained with automatic parsers. CLEVE contains a text encoder to learn event semantics and a graph encoder to learn event structures respectively. Specifically, the text encoder learns event semantic representations by self-supervised contrastive learning to represent the words of the same events closer than those unrelated words; the graph encoder learns event structure representations by graph contrastive pre-training on parsed event-related semantic structures. The two complementary representations then work together to improve both the conventional supervised EE and the unsupervised “liberal” EE, which requires jointly extracting events and discovering event schemata without any annotated data. Experiments on ACE 2005 and MAVEN datasets show that CLEVE achieves significant improvements, especially in the challenging unsupervised setting. The source code and pre-trained checkpoints can be obtained from https://github.com/THU-KEG/CLEVE.

Dynamic early exiting aims to accelerate the inference of pre-trained language models (PLMs) by emitting predictions in internal layers without passing through the entire model. In this paper, we empirically analyze the working mechanism of dynamic early exiting and find that it faces a performance bottleneck under high speed-up ratios. On one hand, the PLMs’ representations in shallow layers lack high-level semantic information and thus are not sufficient for accurate predictions. On the other hand, the exiting decisions made by internal classifiers are unreliable, leading to wrongly emitted early predictions. We instead propose a new framework for accelerating the inference of PLMs, CascadeBERT, which dynamically selects proper-sized and complete models in a cascading manner, providing comprehensive representations for predictions. We further devise a difficulty-aware objective, encouraging the model to output the class probability that reflects the real difficulty of each instance for a more reliable cascading mechanism. Experimental results show that CascadeBERT can achieve an overall 15% improvement under 4x speed-up compared with existing dynamic early exiting methods on six classification tasks, yielding more calibrated and accurate predictions.

Non-autoregressive neural machine translation (NAT) predicts the entire target sequence simultaneously and significantly accelerates inference process. However, NAT discards the dependency information in a sentence, and thus inevitably suffers from the multi-modality problem: the target tokens may be provided by different possible translations, often causing token repetitions or missing. To alleviate this problem, we propose a novel semi-autoregressive model RecoverSAT in this work, which generates a translation as a sequence of segments. The segments are generated simultaneously while each segment is predicted token-by-token. By dynamically determining segment length and deleting repetitive segments, RecoverSAT is capable of recovering from repetitive and missing token errors. Experimental results on three widely-used benchmark datasets show that our proposed model achieves more than 4 times speedup while maintaining comparable performance compared with the corresponding autoregressive model.

Continual relation learning aims to continually train a model on new data to learn incessantly emerging novel relations while avoiding catastrophically forgetting old relations. Some pioneering work has proved that storing a handful of historical relation examples in episodic memory and replaying them in subsequent training is an effective solution for such a challenging problem. However, these memory-based methods usually suffer from overfitting the few memorized examples of old relations, which may gradually cause inevitable confusion among existing relations. Inspired by the mechanism in human long-term memory formation, we introduce episodic memory activation and reconsolidation (EMAR) to continual relation learning. Every time neural models are activated to learn both new and memorized data, EMAR utilizes relation prototypes for memory reconsolidation exercise to keep a stable understanding of old relations. The experimental results show that EMAR could get rid of catastrophically forgetting old relations and outperform the state-of-the-art continual learning models.

Relational facts are an important component of human knowledge, which are hidden in vast amounts of text. In order to extract these facts from text, people have been working on relation extraction (RE) for years. From early pattern matching to current neural networks, existing RE methods have achieved significant progress. Yet with explosion of Web text and emergence of new relations, human knowledge is increasing drastically, and we thus require “more” from RE: a more powerful RE system that can robustly utilize more data, efficiently learn more relations, easily handle more complicated context, and flexibly generalize to more open domains. In this paper, we look back at existing RE methods, analyze key challenges we are facing nowadays, and show promising directions towards more powerful RE. We hope our view can advance this field and inspire more efforts in the community.

Event detection (ED), which means identifying event trigger words and classifying event types, is the first and most fundamental step for extracting event knowledge from plain text. Most existing datasets exhibit the following issues that limit further development of ED: (1) Data scarcity. Existing small-scale datasets are not sufficient for training and stably benchmarking increasingly sophisticated modern neural methods. (2) Low coverage. Limited event types of existing datasets cannot well cover general-domain events, which restricts the applications of ED models. To alleviate these problems, we present a MAssive eVENt detection dataset (MAVEN), which contains 4,480 Wikipedia documents, 118,732 event mention instances, and 168 event types. MAVEN alleviates the data scarcity problem and covers much more general event types. We reproduce the recent state-of-the-art ED models and conduct a thorough evaluation on MAVEN. The experimental results show that existing ED methods cannot achieve promising results on MAVEN as on the small datasets, which suggests that ED in the real world remains a challenging task and requires further research efforts. We also discuss further directions for general domain ED with empirical analyses. The source code and dataset can be obtained from https://github.com/THU-KEG/MAVEN-dataset.

Graph embedding (GE) methods embed nodes (and/or edges) in graph into a low-dimensional semantic space, and have shown its effectiveness in modeling multi-relational data. However, existing GE models are not practical in real-world applications since it overlooked the streaming nature of incoming data. To address this issue, we study the problem of continual graph representation learning which aims to continually train a GE model on new data to learn incessantly emerging multi-relational data while avoiding catastrophically forgetting old learned knowledge. Moreover, we propose a disentangle-based continual graph representation learning (DiCGRL) framework inspired by the human’s ability to learn procedural knowledge. The experimental results show that DiCGRL could effectively alleviate the catastrophic forgetting problem and outperform state-of-the-art continual learning models. The code and datasets are released on https://github.com/KXY-PUBLIC/DiCGRL.

Neural models have achieved remarkable success on relation extraction (RE) benchmarks. However, there is no clear understanding what information in text affects existing RE models to make decisions and how to further improve the performance of these models. To this end, we empirically study the effect of two main information sources in text: textual context and entity mentions (names). We find that (i) while context is the main source to support the predictions, RE models also heavily rely on the information from entity mentions, most of which is type information, and (ii) existing datasets may leak shallow heuristics via entity mentions and thus contribute to the high performance on RE benchmarks. Based on the analyses, we propose an entity-masked contrastive pre-training framework for RE to gain a deeper understanding on both textual context and type information while avoiding rote memorization of entities or use of superficial cues in mentions. We carry out extensive experiments to support our views, and show that our framework can improve the effectiveness and robustness of neural models in different RE scenarios. All the code and datasets are released at https://github.com/thunlp/RE-Context-or-Names.

Language representation models such as BERT could effectively capture contextual semantic information from plain text, and have been proved to achieve promising results in lots of downstream NLP tasks with appropriate fine-tuning. However, most existing language representation models cannot explicitly handle coreference, which is essential to the coherent understanding of the whole discourse. To address this issue, we present CorefBERT, a novel language representation model that can capture the coreferential relations in context. The experimental results show that, compared with existing baseline models, CorefBERT can achieve significant improvements consistently on various downstream NLP tasks that require coreferential reasoning, while maintaining comparable performance to previous models on other common NLP tasks. The source code and experiment details of this paper can be obtained from https://github.com/thunlp/CorefBERT.

Multiple entities in a document generally exhibit complex inter-sentence relations, and cannot be well handled by existing relation extraction (RE) methods that typically focus on extracting intra-sentence relations for single entity pairs. In order to accelerate the research on document-level RE, we introduce DocRED, a new dataset constructed from Wikipedia and Wikidata with three features: (1) DocRED annotates both named entities and relations, and is the largest human-annotated dataset for document-level RE from plain text; (2) DocRED requires reading multiple sentences in a document to extract entities and infer their relations by synthesizing all information of the document; (3) along with the human-annotated data, we also offer large-scale distantly supervised data, which enables DocRED to be adopted for both supervised and weakly supervised scenarios. In order to verify the challenges of document-level RE, we implement recent state-of-the-art methods for RE and conduct a thorough evaluation of these methods on DocRED. Empirical results show that DocRED is challenging for existing RE methods, which indicates that document-level RE remains an open problem and requires further efforts. Based on the detailed analysis on the experiments, we discuss multiple promising directions for future research. We make DocRED and the code for our baselines publicly available at https://github.com/thunlp/DocRED.

In this paper, we propose a novel graph neural network with generated parameters (GP-GNNs). The parameters in the propagation module, i.e. the transition matrices used in message passing procedure, are produced by a generator taking natural language sentences as inputs. We verify GP-GNNs in relation extraction from text, both on bag- and instance-settings. Experimental results on a human-annotated dataset and two distantly supervised datasets show that multi-hop reasoning mechanism yields significant improvements. We also perform a qualitative analysis to demonstrate that our model could discover more accurate relations by multi-hop relational reasoning.

Pattern-based labeling methods have achieved promising results in alleviating the inevitable labeling noises of distantly supervised neural relation extraction. However, these methods require significant expert labor to write relation-specific patterns, which makes them too sophisticated to generalize quickly. To ease the labor-intensive workload of pattern writing and enable the quick generalization to new relation types, we propose a neural pattern diagnosis framework, DIAG-NRE, that can automatically summarize and refine high-quality relational patterns from noise data with human experts in the loop. To demonstrate the effectiveness of DIAG-NRE, we apply it to two real-world datasets and present both significant and interpretable improvements over state-of-the-art methods.

Open-domain question answering (OpenQA) aims to answer questions through text retrieval and reading comprehension. Recently, lots of neural network-based models have been proposed and achieved promising results in OpenQA. However, the success of these models relies on a massive volume of training data (usually in English), which is not available in many other languages, especially for those low-resource languages. Therefore, it is essential to investigate cross-lingual OpenQA. In this paper, we construct a novel dataset XQA for cross-lingual OpenQA research. It consists of a training set in English as well as development and test sets in eight other languages. Besides, we provide several baseline systems for cross-lingual OpenQA, including two machine translation-based methods and one zero-shot cross-lingual method (multilingual BERT). Experimental results show that the multilingual BERT model achieves the best results in almost all target languages, while the performance of cross-lingual OpenQA is still much lower than that of English. Our analysis indicates that the performance of cross-lingual OpenQA is related to not only how similar the target language and English are, but also how difficult the question set of the target language is. The XQA dataset is publicly available at http://github.com/thunlp/XQA.

Numerical reasoning, such as addition, subtraction, sorting and counting is a critical skill in human’s reading comprehension, which has not been well considered in existing machine reading comprehension (MRC) systems. To address this issue, we propose a numerical MRC model named as NumNet, which utilizes a numerically-aware graph neural network to consider the comparing information and performs numerical reasoning over numbers in the question and passage. Our system achieves an EM-score of 64.56% on the DROP dataset, outperforming all existing machine reading comprehension models by considering the numerical relations among numbers.

During the past few decades, knowledge bases (KBs) have experienced rapid growth. Nevertheless, most KBs still suffer from serious incompletion. Researchers proposed many tasks such as knowledge base completion and relation prediction to help build the representation of KBs. However, there are some issues unsettled towards enriching the KBs. Knowledge base completion and relation prediction assume that we know two elements of the fact triples and we are going to predict the missing one. This assumption is too restricted in practice and prevents it from discovering new facts directly. To address this issue, we propose a new task, namely, fact discovery from knowledge base. This task only requires that we know the head entity and the goal is to discover facts associated with the head entity. To tackle this new problem, we propose a novel framework that decomposes the discovery problem into several facet discovery components. We also propose a novel auto-encoder based facet component to estimate some facets of the fact. Besides, we propose a feedback learning component to share the information between each facet. We evaluate our framework using a benchmark dataset and the experimental results show that our framework achieves promising results. We also conduct an extensive analysis of our framework in discovering different kinds of facts. The source code of this paper can be obtained from https://github.com/thunlp/FFD.

Sememes are defined as the minimum semantic units of human languages. As important knowledge sources, sememe-based linguistic knowledge bases have been widely used in many NLP tasks. However, most languages still do not have sememe-based linguistic knowledge bases. Thus we present a task of cross-lingual lexical sememe prediction, aiming to automatically predict sememes for words in other languages. We propose a novel framework to model correlations between sememes and multi-lingual words in low-dimensional semantic space for sememe prediction. Experimental results on real-world datasets show that our proposed model achieves consistent and significant improvements as compared to baseline methods in cross-lingual sememe prediction. The codes and data of this paper are available at https://github.com/thunlp/CL-SP.

We release an open toolkit for knowledge embedding (OpenKE), which provides a unified framework and various fundamental models to embed knowledge graphs into a continuous low-dimensional space. OpenKE prioritizes operational efficiency to support quick model validation and large-scale knowledge representation learning. Meanwhile, OpenKE maintains sufficient modularity and extensibility to easily incorporate new models into the framework. Besides the toolkit, the embeddings of some existing large-scale knowledge graphs pre-trained by OpenKE are also available, which can be directly applied for many applications including information retrieval, personalized recommendation and question answering. The toolkit, documentation, and pre-trained embeddings are all released on http://openke.thunlp.org/.

Distantly supervised open-domain question answering (DS-QA) aims to find answers in collections of unlabeled text. Existing DS-QA models usually retrieve related paragraphs from a large-scale corpus and apply reading comprehension technique to extract answers from the most relevant paragraph. They ignore the rich information contained in other paragraphs. Moreover, distant supervision data inevitably accompanies with the wrong labeling problem, and these noisy data will substantially degrade the performance of DS-QA. To address these issues, we propose a novel DS-QA model which employs a paragraph selector to filter out those noisy paragraphs and a paragraph reader to extract the correct answer from those denoised paragraphs. Experimental results on real-world datasets show that our model can capture useful information from noisy data and achieve significant improvements on DS-QA as compared to all baselines.

Multi-lingual relation extraction aims to find unknown relational facts from text in various languages. Existing models cannot well capture the consistency and diversity of relation patterns in different languages. To address these issues, we propose an adversarial multi-lingual neural relation extraction (AMNRE) model, which builds both consistent and individual representations for each sentence to consider the consistency and diversity among languages. Further, we adopt an adversarial training strategy to ensure those consistent sentence representations could effectively extract the language-consistent relation patterns. The experimental results on real-world datasets demonstrate that our AMNRE model significantly outperforms the state-of-the-art models. The source code of this paper can be obtained from https://github.com/thunlp/AMNRE.

Distantly supervised relation extraction has been widely used to find novel relational facts from plain text. To predict the relation between a pair of two target entities, existing methods solely rely on those direct sentences containing both entities. In fact, there are also many sentences containing only one of the target entities, which also provide rich useful information but not yet employed by relation extraction. To address this issue, we build inference chains between two target entities via intermediate entities, and propose a path-based neural relation extraction model to encode the relational semantics from both direct sentences and inference chains. Experimental results on real-world datasets show that, our model can make full use of those sentences containing only one target entity, and achieves significant and consistent improvements on relation extraction as compared with strong baselines. The source code of this paper can be obtained from https://github.com/thunlp/PathNRE.

Relation extraction has been widely used for finding unknown relational facts from plain text. Most existing methods focus on exploiting mono-lingual data for relation extraction, ignoring massive information from the texts in various languages. To address this issue, we introduce a multi-lingual neural relation extraction framework, which employs mono-lingual attention to utilize the information within mono-lingual texts and further proposes cross-lingual attention to consider the information consistency and complementarity among cross-lingual texts. Experimental results on real-world datasets show that, our model can take advantage of multi-lingual texts and consistently achieve significant improvements on relation extraction as compared with baselines.