Large language models (LLMs) have shown great potential to empower various domains and are often customized by fine-tuning for the requirements of different applications. However, the powerful learning ability of LLMs not only enables them to learn new tasks but also makes them vulnerable to learning undesired behaviors, such as harmfulness and hallucination, as the fine-tuning data often implicitly or explicitly contains such content. Can we fine-tune LLMs on harmful data without learning harmful behaviors? This paper proposes a controllable training framework to make undesired behaviors unlearnable during the fine-tuning process. Specifically, we introduce security vectors to control the model’s behavior and make it consistent with the undesired behavior. Security vectors are activated during fine-tuning, the consistent behavior makes the model believe that such behavior has already been learned and there is no need for further optimization, while inconsistent data can still be learned. After fine-tuning, security vectors are deactivated to restore the LLM’s normal behavior. Our experiments show that the security vectors can prevent LLM from learning harmful and hallucination behavior while preserving the ability to learn other information.

Recently, many studies have illustrated the robustness problem of Named Entity Recognition (NER) systems: the NER models often rely on superficial entity patterns for predictions, without considering evidence from the context. Consequently, even state-of-the-art NER models generalize poorly to out-of-domain scenarios when out-of-distribution (OOD) entity patterns are introduced. Previous research attributes the robustness problem to the existence of NER dataset bias, where simpler and regular entity patterns induce shortcut learning. In this work, we bring new insights into this problem by comprehensively investigating the NER dataset bias from a dataset difficulty view. We quantify the entity-context difficulty distribution in existing datasets and explain their relationship with model robustness. Based on our findings, we explore three potential ways to de-bias the NER datasets by altering entity-context distribution, and we validate the feasibility with intensive experiments. Finally, we show that the de-biased datasets can transfer to different models and even benefit existing model-based robustness-improving methods, indicating that building more robust datasets is fundamental for building more robust NER systems.

Recently, Few-shot Named Entity Recognition has received wide attention with the growing need for NER models to learn new classes with minimized annotation costs. However, one common yet understudied situation is to transfer a model trained with coarse-grained classes to recognize fine-grained classes, such as separating a product category into sub-classes. We find that existing few-shot NER solutions are not suitable for such a situation since they do not consider the sub-class discrimination during coarse training and various granularity of new classes during few-shot learning. In this work, we introduce the Coarse-to-fine Few-shot NER (C2FNER) task and propose an effective solution. Specifically, during coarse training, we propose a cluster-based prototype margin loss to learn group-wise discriminative representations, so as to benefit fine-grained learning. Targeting various granularity of new classes, we separate the coarse classes into extra-fine clusters and propose a novel prototype retrieval and bootstrapping algorithm to retrieve representative clusters for each fine class. We then adopt a mixture prototype loss to efficiently learn the representations of fine classes. We conduct experiments on both in-domain and cross-domain C2FNER settings with various target granularity, and the proposed method shows superior performance over the baseline methods.

In real-world applications, pre-trained language models are typically deployed on the cloud, allowing clients to upload data and perform compute-intensive inference remotely. To avoid sharing sensitive data directly with service providers, clients can upload numerical representations rather than plain text to the cloud. However, recent text reconstruction techniques have demonstrated that it is possible to transform representations into original words, suggesting that privacy risk remains. In this paper, we propose TextObfuscator, a novel framework for protecting inference privacy by applying random perturbations to clustered representations. The random perturbations make the representations indistinguishable from surrounding clustered representations, thus obscuring word information while retaining the original word functionality. To achieve this, we utilize prototypes to learn clustered representation, where tokens of similar functionality are encouraged to be closer to the same prototype during training. Additionally, we design different methods to find prototypes for token-level and sentence-level tasks, which can improve performance by incorporating semantic and task information. Experimental results on token and sentence classification tasks show that TextObfuscator achieves improvement over compared methods without increasing inference cost.

Pre-trained language models (PLMs) are often deployed as cloud services, enabling users to upload textual data and perform inference remotely. However, users’ personal text often contains sensitive information, and sharing such data directly with the service providers can lead to serious privacy leakage. To address this problem, we introduce a novel privacy-preserving inference framework called MixPi , which prevents plaintext leakage during the inference phase. Inspired by k-anonymity, MixPi aims to obfuscate a user’s private input by mixing it with multiple other inputs, thereby confounding potential privacy attackers. To achieve this, our approach involves: (1) proposing a novel encryption module, Privacy Mixer, which encrypts input from three distinct dimensions: mixing, representation, and position. (2) adopting a pre-trained Multi-input Multi-output network to handle mixed representations and obtain multiple predictions. (3) employing a Privacy Demixer to ensure only the user can decrypt the real output among the multiple predictions. Furthermore, we explore different ways to automatically generate synthetic inputs required for mixing. Experimental results on token and sentence classification tasks demonstrate that MixPi greatly surpasses existing privacy-preserving methods in both performance and privacy.

As the categories of named entities rapidly increase, the deployed NER models are required to keep updating toward recognizing more entity types, creating a demand for class-incremental learning for NER. Considering the privacy concerns and storage constraints, the standard paradigm for class-incremental NER updates the models with training data only annotated with the new classes, yet the entities from other entity classes are regarded as “Non-entity” (or “O”). In this work, we conduct an empirical study on the “Unlabeled Entity Problem” and find that it leads to severe confusion between “O” and entities, decreasing class discrimination of old classes and declining the model’s ability to learn new classes. To solve the Unlabeled Entity Problem, we propose a novel representation learning method to learn discriminative representations for the entity classes and “O”. Specifically, we propose an entity-aware contrastive learning method that adaptively detects entity clusters in “O”. Furthermore, we propose two effective distance-based relabeling strategies for better learning the old classes. We introduce a more realistic and challenging benchmark for class-incremental NER, and the proposed method achieves up to 10.62% improvement over the baseline methods.

Large pre-trained language models (PLMs) have demonstrated superior performance in industrial applications. Recent studies have explored parameter-efficient PLM tuning, which only updates a small amount of task-specific parameters while achieving both high efficiency and comparable performance against standard fine-tuning. However, all these methods ignore the inefficiency problem caused by the task-specific output layers, which is inflexible for us to re-use PLMs and introduces non-negligible parameters. In this work, we focus on the text classification task and propose plugin-tuning, a framework that further improves the efficiency of existing parameter-efficient methods with a unified classifier. Specifically, we re-formulate both token and sentence classification tasks into a unified language modeling task, and map label spaces of different tasks into the same vocabulary space. In this way, we can directly re-use the language modeling heads of PLMs, avoiding introducing extra parameters for different tasks. We conduct experiments on six classification benchmarks. The experimental results show that plugin-tuning can achieve comparable performance against fine-tuned PLMs, while further saving around 50% parameters on top of other parameter-efficient methods.

Prompt-based methods have been successfully applied in sentence-level few-shot learning tasks, mostly owing to the sophisticated design of templates and label words. However, when applied to token-level labeling tasks such as NER, it would be time-consuming to enumerate the template queries over all potential entity spans. In this work, we propose a more elegant method to reformulate NER tasks as LM problems without any templates. Specifically, we discard the template construction process while maintaining the word prediction paradigm of pre-training models to predict a class-related pivot word (or label word) at the entity position. Meanwhile, we also explore principled ways to automatically search for appropriate label words that the pre-trained models can easily adapt to. While avoiding the complicated template-based process, the proposed LM objective also reduces the gap between different objectives used in pre-training and fine-tuning, thus it can better benefit the few-shot performance. Experimental results demonstrate the effectiveness of the proposed method over bert-tagger and template-based method under few-shot settings. Moreover, the decoding speed of the proposed method is up to 1930.12 times faster than the template-based method.

Multilingual BERT (mBERT) has demonstrated considerable cross-lingual syntactic ability, whereby it enables effective zero-shot cross-lingual transfer of syntactic knowledge. The transfer is more successful between some languages, but it is not well understood what leads to this variation and whether it fairly reflects difference between languages. In this work, we investigate the distributions of grammatical relations induced from mBERT in the context of 24 typologically different languages. We demonstrate that the distance between the distributions of different languages is highly consistent with the syntactic difference in terms of linguistic formalisms. Such difference learnt via self-supervision plays a crucial role in the zero-shot transfer performance and can be predicted by variation in morphosyntactic properties between languages. These results suggest that mBERT properly encodes languages in a way consistent with linguistic diversity and provide insights into the mechanism of cross-lingual transfer.

Recently, more and more pre-trained language models are released as a cloud service. It allows users who lack computing resources to perform inference with a powerful model by uploading data to the cloud. The plain text may contain private information, as the result, users prefer to do partial computations locally and upload intermediate representations to the cloud for subsequent inference.However, recent studies have shown that intermediate representations can also be recovered to plain text with reasonable accuracy, thus the risk of privacy leakage still exists. To address this issue, we propose TextFusion, a novel method for preserving inference privacy.Specifically, we train a Fusion Predictor to dynamically fuse token representations, which hides multiple private token representations behind an unrecognizable one.Furthermore, an adversarial training regime is employed to privatize these representations. In this way, the cloud only receives incomplete and perturbed representations, making it difficult to accurately recover the complete plain text.The experimental results on diverse classification tasks show that our approach can effectively preserve inference privacy without significantly sacrificing performance in different scenarios.

Distant supervision for relation extraction provides uniform bag labels for each sentence inside the bag, while accurate sentence labels are important for downstream applications that need the exact relation type. Directly using bag labels for sentence-level training will introduce much noise, thus severely degrading performance. In this work, we propose the use of negative training (NT), in which a model is trained using complementary labels regarding that “the instance does not belong to these complementary labels”. Since the probability of selecting a true label as a complementary label is low, NT provides less noisy information. Furthermore, the model trained with NT is able to separate the noisy data from the training data. Based on NT, we propose a sentence-level framework, SENT, for distant relation extraction. SENT not only filters the noisy data to construct a cleaner dataset, but also performs a re-labeling process to transform the noisy data into useful training data, thus further benefiting the model’s performance. Experimental results show the significant improvement of the proposed method over previous methods on sentence-level evaluation and de-noise effect.

TextFlint is a multilingual robustness evaluation toolkit for NLP tasks that incorporates universal text transformation, task-specific transformation, adversarial attack, subpopulation, and their combinations to provide comprehensive robustness analyses. This enables practitioners to automatically evaluate their models from various aspects or to customize their evaluations as desired with just a few lines of code. TextFlint also generates complete analytical reports as well as targeted augmented data to address the shortcomings of the model in terms of its robustness. To guarantee acceptability, all the text transformations are linguistically based and all the transformed data selected (up to 100,000 texts) scored highly under human evaluation. To validate the utility, we performed large-scale empirical evaluations (over 67,000) on state-of-the-art deep learning models, classic supervised methods, and real-world systems. The toolkit is already available at https://github.com/textflint with all the evaluation results demonstrated at textflint.io.

Pre-Trained Models have been widely applied and recently proved vulnerable under backdoor attacks: the released pre-trained weights can be maliciously poisoned with certain triggers. When the triggers are activated, even the fine-tuned model will predict pre-defined labels, causing a security threat. These backdoors generated by the poisoning methods can be erased by changing hyper-parameters during fine-tuning or detected by finding the triggers. In this paper, we propose a stronger weight-poisoning attack method that introduces a layerwise weight poisoning strategy to plant deeper backdoors; we also introduce a combinatorial trigger that cannot be easily detected. The experiments on text classification tasks show that previous defense methods cannot resist our weight-poisoning method, which indicates that our method can be widely applied and may provide hints for future model robustness studies.

Recently, many works have tried to augment the performance of Chinese named entity recognition (NER) using word lexicons. As a representative, Lattice-LSTM has achieved new benchmark results on several public Chinese NER datasets. However, Lattice-LSTM has a complex model architecture. This limits its application in many industrial areas where real-time NER responses are needed. In this work, we propose a simple but effective method for incorporating the word lexicon into the character representations. This method avoids designing a complicated sequence modeling architecture, and for any neural NER model, it requires only subtle adjustment of the character representation layer to introduce the lexicon information. Experimental studies on four benchmark Chinese NER datasets show that our method achieves an inference speed up to 6.15 times faster than those of state-of-the-art methods, along with a better performance. The experimental results also show that the proposed method can be easily incorporated with pre-trained models like BERT.

Adversarial attacks for discrete data (such as texts) have been proved significantly more challenging than continuous data (such as images) since it is difficult to generate adversarial samples with gradient-based methods. Current successful attack methods for texts usually adopt heuristic replacement strategies on the character or word level, which remains challenging to find the optimal solution in the massive space of possible combinations of replacements while preserving semantic consistency and language fluency. In this paper, we propose BERT-Attack, a high-quality and effective method to generate adversarial samples using pre-trained masked language models exemplified by BERT. We turn BERT against its fine-tuned models and other deep neural models in downstream tasks so that we can successfully mislead the target models to predict incorrectly. Our method outperforms state-of-the-art attack strategies in both success rate and perturb percentage, while the generated adversarial samples are fluent and semantically preserved. Also, the cost of calculation is low, thus possible for large-scale generations. The code is available at https://github.com/LinyangLee/BERT-Attack.

In this work, we explore the way to quickly adjust an existing named entity recognition (NER) system to make it capable of recognizing entity types not defined in the system. As an illustrative example, consider the case that a NER system has been built to recognize person and organization names, and now it requires to additionally recognize job titles. Such a situation is common in the industrial areas, where the entity types required to recognize vary a lot in different products and keep changing. To avoid laborious data labeling and achieve fast adaptation, we propose to adjust the existing NER system using the previously labeled data and entity lexicons of the newly introduced entity types. We formulate such a task as a partially supervised learning problem and accordingly propose an effective algorithm to solve the problem. Comprehensive experimental studies on several public NER datasets validate the effectiveness of our method.