Conventional tokenization methods for Chinese pretrained language models (PLMs) treat each character as an indivisible token (Devlin et al., 2019), which ignores the characteristics of the Chinese writing system. In this work, we comprehensively study the influences of three main factors on the Chinese tokenization for PLM: pronunciation, glyph (i.e., shape), and word boundary. Correspondingly, we propose three kinds of tokenizers: 1) SHUOWEN (meaning Talk Word), the pronunciation-based tokenizers; 2) JIEZI (meaning Solve Character), the glyph-based tokenizers; 3) Word segmented tokenizers, the tokenizers with Chinese word segmentation. To empirically compare the effectiveness of studied tokenizers, we pretrain BERT-style language models with them and evaluate the models on various downstream NLU tasks. We find that SHUOWEN and JIEZI tokenizers can generally outperform conventional single-character tokenizers, while Chinese word segmentation shows no benefit as a preprocessing step. Moreover, the proposed SHUOWEN and JIEZI tokenizers exhibit significantly better robustness in handling noisy texts. The code and pretrained models will be publicly released to facilitate linguistically informed Chinese NLP.
Recent explorations of large-scale pre-trained language models (PLMs) such as GPT-3 have revealed the power of PLMs with huge amounts of parameters, setting off a wave of training ever-larger PLMs. However, training a large-scale PLM requires tremendous amounts of computational resources, which is time-consuming and expensive. In addition, existing large-scale PLMs are mainly trained from scratch individually, ignoring the availability of many existing well-trained PLMs. To this end, we explore the question that how can previously trained PLMs benefit training larger PLMs in future. Specifically, we introduce a novel pre-training framework named "knowledge inheritance" (KI), which combines both self-learning and teacher-guided learning to efficiently train larger PLMs. Sufficient experimental results demonstrate the feasibility of our KI framework. 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 can well support lifelong learning and knowledge transfer.
Backdoor attacks are a kind of insidious security threat against machine learning models. After being injected with a backdoor in training, the victim model will produce adversary-specified outputs on the inputs embedded with predesigned triggers but behave properly on normal inputs during inference. As a sort of emergent attack, backdoor attacks in natural language processing (NLP) are investigated insufficiently. As far as we know, almost all existing textual backdoor attack methods insert additional contents into normal samples as triggers, which causes the trigger-embedded samples to be detected and the backdoor attacks to be blocked without much effort. In this paper, we propose to use syntactic structure as the trigger in textual backdoor attacks. We conduct extensive experiments to demonstrate that the syntactic trigger-based attack method can achieve comparable attack performance (almost 100\% success rate) to the insertion-based methods but possesses much higher invisibility and stronger resistance to defenses. These results also reveal the significant insidiousness and harmfulness of textual backdoor attacks. All the code and data of this paper can be obtained at https://github.com/thunlp/HiddenKiller.
Fine-tuning pre-trained language models (PLMs) has demonstrated its effectiveness on various downstream NLP tasks recently. However, in many low-resource scenarios, the conventional fine-tuning strategies cannot sufficiently capture the important semantic features for downstream tasks. To address this issue, we introduce a novel framework (named "CSS-LM") to improve the fine-tuning phase of PLMs via contrastive semi-supervised learning. Specifically, given a specific task, we retrieve positive and negative instances from large-scale unlabeled corpora according to their domain-level and class-level semantic relatedness to the task. We then perform contrastive semi-supervised learning on both the retrieved unlabeled and original labeled instances to help PLMs capture crucial task-related semantic features. The experimental results show that CSS-LM achieves better results than the conventional fine-tuning strategy on a series of downstream tasks with few-shot settings, and outperforms the latest supervised contrastive fine-tuning strategies. Our datasets and source code will be available to provide more details.
Due to the success of pre-trained models (PTMs), people usually fine-tune an existing PTM for downstream tasks. Most of PTMs are contributed and maintained by open sources and may suffer from backdoor attacks. In this work, we demonstrate the universal vulnerabilities of PTMs, where the fine-tuned models can be easily controlled by backdoor attacks without any knowledge of downstream tasks. Specifically, the attacker can add a simple pre-training task to restrict the output hidden states of the trigger instances to the pre-defined target embeddings, namely neuron-level backdoor attack (NeuBA). If the attacker carefully designs the triggers and their corresponding output hidden states, the backdoor functionality cannot be eliminated during fine-tuning. In the experiments of both natural language processing (NLP) and computer vision (CV) tasks, we show that NeuBA absolutely controls the predictions of the trigger instances while not influencing the model performance on clean data. Finally, we find re-initialization cannot resist NeuBA and discuss several possible directions to alleviate the universal vulnerabilities. Our findings sound a red alarm for the wide use of PTMs. Our source code and data can be accessed at \url{https://github.com/thunlp/NeuBA}.
Pre-trained language models (PLMs) fail miserably on adversarial attacks. To improve the robustness, adversarial data augmentation (ADA) has been widely adopted, which attempts to cover more search space of adversarial attacks by adding the adversarial examples during training. However, the number of adversarial examples added by ADA is extremely insufficient due to the enormously large search space. In this work, we propose a simple and effective method to cover much larger proportion of the attack search space, called Adversarial Data Augmentation with Mixup (MixADA). Specifically, MixADA linearly interpolates the representations of pairs of training examples to form new virtual samples, which are more abundant and diverse than the discrete adversarial examples used in conventional ADA. Moreover, to evaluate the robustness of different models fairly, we adopt a challenging setup, which dynamically generates new adversarial examples for each model. In the text classification experiments of BERT and RoBERTa, MixADA achieves significant robustness gains under two strong adversarial attacks and alleviates the performance degradation of ADA on the original data. Our source codes will be released to support further explorations.
Pre-trained Language Models (PLMs) have proven to be beneficial for various downstream NLP tasks. Recently, GPT-3, with 175 billion parameters and 570GB training data, drew a lot of attention due to the capacity of few-shot (even zero-shot) learning. However, applying GPT-3 to address Chinese NLP tasks is still challenging, as the training corpus of GPT-3 is primarily English, and the parameters are not publicly available. In this technical report, we release the Chinese Pre-trained Language Model (CPM) with generative pre-training on large-scale Chinese training data. To the best of our knowledge, CPM, with 2.6 billion parameters and 100GB Chinese training data, is the largest Chinese pre-trained language model, which could facilitate several downstream Chinese NLP tasks, such as conversation, essay generation, cloze test, and language understanding. Extensive experiments demonstrate that CPM achieves strong performance on many NLP tasks in the settings of few-shot (even zero-shot) learning. The code and parameters are available at https://github.com/TsinghuaAI/CPM-Generate.
Deep pre-trained Transformer models have achieved state-of-the-art results over a variety of natural language processing (NLP) tasks. By learning rich language knowledge with millions of parameters, these models are usually overparameterized and significantly increase the computational overhead in applications. It is intuitive to address this issue by model compression. In this work, we propose a method, called Single-Shot Meta-Pruning, to compress deep pre-trained Transformers before fine-tuning. Specifically, we focus on pruning unnecessary attention heads adaptively for different downstream tasks. To measure the informativeness of attention heads, we train our Single-Shot Meta-Pruner (SMP) with a meta-learning paradigm aiming to maintain the distribution of text representations after pruning. Compared with existing compression methods for pre-trained models, our method can reduce the overhead of both fine-tuning and inference. Experimental results show that our pruner can selectively prune 50% of attention heads with little impact on the performance on downstream tasks and even provide better text representations. The source code will be released in the future.
Several recent efforts have been devoted to enhancing pre-trained language models (PLMs) by utilizing extra heterogeneous knowledge in knowledge graphs (KGs), and achieved consistent improvements on various knowledge-driven NLP tasks. However, most of these knowledge-enhanced PLMs embed static sub-graphs of KGs ("knowledge context"), regardless of that the knowledge required by PLMs may change dynamically according to specific text ("textual context"). In this paper, we propose a novel framework named DKPLM to dynamically select and embed knowledge context according to textual context for PLMs, which can avoid the effect of redundant and ambiguous knowledge in KGs that cannot match the input text. Our experimental results show that DKPLM outperforms various baselines on typical knowledge-driven NLP tasks, indicating the effectiveness of utilizing dynamic knowledge context for language understanding. Besides the performance improvements, the dynamically selected knowledge in DKPLM can describe the semantics of text-related knowledge in a more interpretable form than the conventional PLMs. Our source code and datasets will be available to provide more details for DKPLM.