Pre-trained language model (PLM) can be stealthily misled to target outputs by backdoor attacks when encountering poisoned samples, without performance degradation on clean samples. The stealthiness of backdoor attacks is commonly attained through minimal cross-entropy loss fine-tuning on a union of poisoned and clean samples. Existing defense paradigms provide a workaround by detecting and removing poisoned samples at pre-training or inference time. On the contrary, we provide a new perspective where the backdoor attack is directly reversed. Specifically, maximum entropy loss is incorporated in training to neutralize the minimal cross-entropy loss fine-tuning on poisoned data. We defend against a range of backdoor attacks on classification tasks and significantly lower the attack success rate. In extension, we explore the relationship between intended backdoor attacks and unintended dataset bias, and demonstrate the feasibility of the maximum entropy principle in de-biasing.

We focus on dialogue reading comprehension (DRC) that extracts answers from dialogues. Compared to standard RC tasks, DRC has raised challenges because of the complex speaker information and noisy dialogue context. Essentially, the challenges come from the speaker-centric nature of dialogue utterances — an utterance is usually insufficient in its surface form, but requires to incorporate the role of its speaker and the dialogue context to fill the latent pragmatic and intention information. We propose to deal with these problems in two folds. First, we propose a new key-utterances-extracting method, which can realize more answer-contained utterances. Second, based on the extracted utterances, we then propose a Question-Interlocutor Scope Realized Graph (QuISG). QuISG involves the question and question-mentioning speaker as nodes. To realize interlocutor scopes, utterances are connected with corresponding speakers in the dialogue. Experiments on the benchmarks show that our method achieves state-of-the-art performance against previous works.

The success of ChatGPT validates the potential of large language models (LLMs) in artificial general intelligence (AGI). Subsequently, the release of LLMs has sparked the open-source community’s interest in instruction-tuning, which is deemed to accelerate ChatGPT’s replication process. However, research on instruction-tuning LLMs in Chinese, the world’s most spoken language, is still in its early stages. Therefore, this paper makes an in-depth empirical study of instruction-tuning LLMs in Chinese, which can serve as a cookbook that provides valuable findings for effectively customizing LLMs that can better respond to Chinese instructions. Specifically, we systematically explore the impact of LLM bases, parameter-efficient methods, instruction data types, which are the three most important elements for instruction-tuning. Besides, we also conduct experiment to study the impact of other factors, e.g., chain-of-thought data and human-value alignment. We hope that this empirical study can make a modest contribution to the open Chinese version of ChatGPT. This paper will release a powerful Chinese LLM that is comparable to ChatGLM. The code and data are available at https: //github.com/PhoebusSi/Alpaca-CoT.

Tabular mathematical reasoning task requires models to perform multi-step operations including information look-up and numerical calculation, based on heterogeneous data from tables and questions. Existing solutions tend to extend chain-of-thought (CoT) reasoning into powerful large language models (LLMs) to promote multi-hop mathematical reasoning. However, such LLM-based approaches are not a viable solution in the scenario of privatization deployment or limited resources. To address this problem, we revisit small-scale tabular language models (TaLMs) and extend chain-of-thought reasoning into TaLMs for the first time. Specifically, we propose a novel framework, TaCo, which coordinates two TaLMs responsible for CoT generation and answer inference, respectively. Besides, our framework can be combined with an external calculator to enhance accurate numerical calculation. On the TABMWP dataset, TaCo outperforms the state-of-the-art ChatGPT by 9.55% (82.60%→92.15% in accuracy) with much less parameters (0.8B). The code will be released along with the paper.

Despite the excellent performance of vision-language pre-trained models (VLPs) on conventional VQA task, they still suffer from two problems: First, VLPs tend to rely on language biases in datasets and fail to generalize to out-of-distribution (OOD) data. Second, they are inefficient in terms of memory footprint and computation. Although promising progress has been made in both problems, most existing works tackle them independently. To facilitate the application of VLP to VQA tasks, it is imperative to jointly study VLP compression and OOD robustness, which, however, has not yet been explored. This paper investigates whether a VLP can be compressed and debiased simultaneously by searching sparse and robust subnetworks. To this end, we systematically study the design of a training and compression pipeline to search the subnetworks, as well as the assignment of sparsity to different modality-specific modules. Our experiments involve 2 VLPs, 2 compression methods, 4 training methods, 2 datasets and a range of sparsity levels. Our results show that there indeed exist sparse and robust subnetworks, which are competitive with the debiased full VLP and clearly outperform the debiasing SoTAs with fewer parameters on OOD datasets VQA-CP v2 and VQA-VS. The codes can be found at https://github.com/PhoebusSi/Compress-Robust-VQA.

Knowledge-grounded dialogue generation aims to mitigate the issue of text degeneration by incorporating external knowledge to supplement the context. However, the model often fails to internalize this information into responses in a human-like manner. Instead, it simply inserts segments of the provided knowledge into generic responses. As a result, the generated responses tend to be tedious, incoherent, and in lack of interactivity which means the degeneration problem is still unsolved. In this work, we first find that such copying-style degeneration is primarily due to the weak likelihood objective, which allows the model to “cheat” the objective by merely duplicating knowledge segments in a superficial pattern matching based on overlap. To overcome this challenge, we then propose a Multi-level Adaptive Contrastive Learning (MACL) framework that dynamically samples negative examples and subsequently penalizes degeneration behaviors at both the token-level and sequence-level. Extensive experiments on the WoW dataset demonstrate the effectiveness of our approach across various pre-trained models and decoding strategies.

Parameter-Efficient Tuning (PET) has shown remarkable performance by fine-tuning only a small number of parameters of the pre-trained language models (PLMs) for the downstream tasks, while it is also possible to construct backdoor attacks due to the vulnerability of pre-trained weights. However, a large reduction in the number of attackable parameters in PET will cause the user’s fine-tuning to greatly affect the effectiveness of backdoor attacks, resulting in backdoor forgetting. We find that the backdoor injection process can be regarded as multi-task learning, which has a convergence imbalance problem between the training of clean and poisoned data. And this problem might result in forgetting the backdoor. Based on this finding, we propose a gradient control method to consolidate the attack effect, comprising two strategies. One controls the gradient magnitude distribution cross layers within one task and the other prevents the conflict of gradient directions between tasks. Compared with previous backdoor attack methods in the scenario of PET, our method improve the effect of the attack on sentiment classification and spam detection respectively, which shows that our method is widely applicable to different tasks.

Various datasets have been proposed to promote the development of Table Question Answering (TQA) technique. However, the problem setting of existing TQA benchmarks suffers from two limitations. First, they directly provide models with explicit table structures where row headers and column headers of the table are explicitly annotated and treated as model input during inference. Second, they only consider tables of limited types and ignore other tables especially complex tables with flexible header locations. Such simplified problem setting cannot cover practical scenarios where models need to process tables without header annotations in the inference phase or tables of different types. To address above issues, we construct a new TQA dataset with implicit and multi-type table structures, named IM-TQA, which not only requires the model to understand tables without directly available header annotations but also to handle multi-type tables including previously neglected complex tables. We investigate the performance of recent methods on our dataset and find that existing methods struggle in processing implicit and multi-type table structures. Correspondingly, we propose an RGCN-RCI framework outperforming recent baselines. We will release our dataset to facilitate future research.

Outside-knowledge visual question answering is a challenging task that requires both the acquisition and the use of open-ended real-world knowledge. Some existing solutions draw external knowledge into the cross-modality space which overlooks the much vaster textual knowledge in natural-language space, while others transform the image into a text which further fuses with the textual knowledge into the natural-language space and completely abandons the use of visual features. In this paper, we are inspired to constrain the cross-modality space into the same space of natural-language space which makes the visual features preserved directly, and the model still benefits from the vast knowledge in natural-language space. To this end, we propose a novel framework consisting of a multimodal encoder, a textual encoder and an answer decoder. Such structure allows us to introduce more types of knowledge including explicit and implicit multimodal and textual knowledge. Extensive experiments validate the superiority of the proposed method which outperforms the state-of-the-art by 6.17% accuracy. We also conduct comprehensive ablations of each component, and systematically study the roles of varying types of knowledge. Codes and knowledge data are to be released.

Visual Question Answering (VQA) models are prone to learn the shortcut solution formed by dataset biases rather than the intended solution. To evaluate the VQA models’ reasoning ability beyond shortcut learning, the VQA-CP v2 dataset introduces a distribution shift between the training and test set given a question type. In this way, the model cannot use the training set shortcut (from question type to answer) to perform well on the test set. However, VQA-CP v2 only considers one type of shortcut and thus still cannot guarantee that the model relies on the intended solution rather than a solution specific to this shortcut. To overcome this limitation, we propose a new dataset that considers varying types of shortcuts by constructing different distribution shifts in multiple OOD test sets. In addition, we overcome the three troubling practices in the use of VQA-CP v2, e.g., selecting models using OOD test sets, and further standardize OOD evaluation procedure. Our benchmark provides a more rigorous and comprehensive testbed for shortcut learning in VQA. We benchmark recent methods and find that methods specifically designed for particular shortcuts fail to simultaneously generalize to our varying OOD test sets. We also systematically study the varying shortcuts and provide several valuable findings, which may promote the exploration of shortcut learning in VQA.

Empathy, which is widely used in psychological counseling, is a key trait of everyday human conversations. Equipped with commonsense knowledge, current approaches to empathetic response generation focus on capturing implicit emotion within dialogue context, where the emotions are treated as a static variable throughout the conversations. However, emotions change dynamically between utterances, which makes previous works difficult to perceive the emotion flow and predict the correct emotion of the target response, leading to inappropriate response. Furthermore, simply importing commonsense knowledge without harmonization may trigger the conflicts between knowledge and emotion, which confuse the model to choose the correct information to guide the generation process. To address the above problems, we propose a Serial Encoding and Emotion-Knowledge interaction (SEEK) method for empathetic dialogue generation. We use a fine-grained encoding strategy which is more sensitive to the emotion dynamics (emotion flow) in the conversations to predict the emotion-intent characteristic of response. Besides, we design a novel framework to model the interaction between knowledge and emotion to solve the conflicts generate more sensible response. Extensive experiments on the utterance-level annotated EMPATHETICDIALOGUES demonstrate that SEEK outperforms the strong baseline in both automatic and manual evaluations.

Models for Visual Question Answering (VQA) often rely on the spurious correlations, i.e., the language priors, that appear in the biased samples of training set, which make them brittle against the out-of-distribution (OOD) test data. Recent methods have achieved promising progress in overcoming this problem by reducing the impact of biased samples on model training. However, these models reveal a trade-off that the improvements on OOD data severely sacrifice the performance on the in-distribution (ID) data (which is dominated by the biased samples). Therefore, we propose a novel contrastive learning approach, MMBS, for building robust VQA models by Making the Most of Biased Samples. Specifically, we construct positive samples for contrastive learning by eliminating the information related to spurious correlation from the original training samples and explore several strategies to use the constructed positive samples for training. Instead of undermining the importance of biased samples in model training, our approach precisely exploits the biased samples for unbiased information that contributes to reasoning. The proposed method is compatible with various VQA backbones. We validate our contributions by achieving competitive performance on the OOD dataset VQA-CP v2 while preserving robust performance on the ID dataset VQA v2.

Knowledge-grounded dialogue generation consists of two subtasks: knowledge selection and response generation. The knowledge selector generally constructs a query based on the dialogue context and selects the most appropriate knowledge to help response generation. Recent work finds that realizing who (the user or the agent) holds the initiative and utilizing the role-initiative information to instruct the query construction can help select knowledge. It depends on whether the knowledge connection between two adjacent rounds is smooth to assign the role. However, whereby the user takes the initiative only when there is a strong semantic transition between two rounds, probably leading to initiative misjudgment. Therefore, it is necessary to seek a more sensitive reason beyond the initiative role for knowledge selection. To address the above problem, we propose a Topic-shift Aware Knowledge sElector(TAKE). Specifically, we first annotate the topic shift and topic inheritance labels in multi-round dialogues with distant supervision. Then, we alleviate the noise problem in pseudo labels through curriculum learning and knowledge distillation. Extensive experiments on WoW show that TAKE performs better than strong baselines.

Stance detection aims to identify the attitude from an opinion towards a certain target. Despite the significant progress on this task, it is extremely time-consuming and budget-unfriendly to collect sufficient high-quality labeled data for every new target under fully-supervised learning, whereas unlabeled data can be collected easier. Therefore, this paper is devoted to few-shot stance detection and investigating how to achieve satisfactory results in semi-supervised settings. As a target-oriented task, the core idea of semi-supervised few-shot stance detection is to make better use of target-relevant information from labeled and unlabeled data. Therefore, we develop a novel target-aware semi-supervised framework. Specifically, we propose a target-aware contrastive learning objective to learn more distinguishable representations for different targets. Such an objective can be easily applied with or without unlabeled data. Furthermore, to thoroughly exploit the unlabeled data and facilitate the model to learn target-relevant stance features in the opinion content, we explore a simple but effective target-aware consistency regularization combined with a self-training strategy. The experimental results demonstrate that our approach can achieve state-of-the-art performance on multiple benchmark datasets in the few-shot setting.

Recent studies on the lottery ticket hypothesis (LTH) show that pre-trained language models (PLMs) like BERT contain matching subnetworks that have similar transfer learning performance as the original PLM. These subnetworks are found using magnitude-based pruning. In this paper, we find that the BERT subnetworks have even more potential than these studies have shown. Firstly, we discover that the success of magnitude pruning can be attributed to the preserved pre-training performance, which correlates with the downstream transferability. Inspired by this, we propose to directly optimize the subnetwork structure towards the pre-training objectives, which can better preserve the pre-training performance. Specifically, we train binary masks over model weights on the pre-training tasks, with the aim of preserving the universal transferability of the subnetwork, which is agnostic to any specific downstream tasks. We then fine-tune the subnetworks on the GLUE benchmark and the SQuAD dataset. The results show that, compared with magnitude pruning, mask training can effectively find BERT subnetworks with improved overall performance on downstream tasks. Moreover, our method is also more efficient in searching subnetworks and more advantageous when fine-tuning within a certain range of data scarcity. Our code is available at https://github.com/llyx97/TAMT.

Transformer-based pre-trained language models (PLMs) mostly suffer from excessive overhead despite their advanced capacity. For resource-constrained devices, there is an urgent need for a spatially and temporally efficient model which retains the major capacity of PLMs. However, existing statically compressed models are unaware of the diverse complexities between input instances, potentially resulting in redundancy and inadequacy for simple and complex inputs. Also, miniature models with early exiting encounter challenges in the trade-off between making predictions and serving the deeper layers. Motivated by such considerations, we propose a collaborative optimization for PLMs that integrates static model compression and dynamic inference acceleration. Specifically, the PLM is slenderized in width while the depth remains intact, complementing layer-wise early exiting to speed up inference dynamically. To address the trade-off of early exiting, we propose a joint training approach that calibrates slenderization and preserves contributive structures to each exit instead of only the final layer. Experiments are conducted on GLUE benchmark and the results verify the Pareto optimality of our approach at high compression and acceleration rate with 1/8 parameters and 1/19 FLOPs of BERT.

Conversational Emotion Recognition (CER) is a task to predict the emotion of an utterance in the context of a conversation. Although modeling the conversational context and interactions between speakers has been studied broadly, it is important to consider the speaker’s psychological state, which controls the action and intention of the speaker. The state-of-the-art method introduces CommonSense Knowledge (CSK) to model psychological states in a sequential way (forwards and backwards). However, it ignores the structural psychological interactions between utterances. In this paper, we propose a pSychological-Knowledge-Aware Interaction Graph (SKAIG). In the locally connected graph, the targeted utterance will be enhanced with the information of action inferred from the past context and intention implied by the future context. The utterance is self-connected to consider the present effect from itself. Furthermore, we utilize CSK to enrich edges with knowledge representations and process the SKAIG with a graph transformer. Our method achieves state-of-the-art and competitive performance on four popular CER datasets.

Recently, knowledge distillation (KD) has shown great success in BERT compression. Instead of only learning from the teacher’s soft label as in conventional KD, researchers find that the rich information contained in the hidden layers of BERT is conducive to the student’s performance. To better exploit the hidden knowledge, a common practice is to force the student to deeply mimic the teacher’s hidden states of all the tokens in a layer-wise manner. In this paper, however, we observe that although distilling the teacher’s hidden state knowledge (HSK) is helpful, the performance gain (marginal utility) diminishes quickly as more HSK is distilled. To understand this effect, we conduct a series of analysis. Specifically, we divide the HSK of BERT into three dimensions, namely depth, length and width. We first investigate a variety of strategies to extract crucial knowledge for each single dimension and then jointly compress the three dimensions. In this way, we show that 1) the student’s performance can be improved by extracting and distilling the crucial HSK, and 2) using a tiny fraction of HSK can achieve the same performance as extensive HSK distillation. Based on the second finding, we further propose an efficient KD paradigm to compress BERT, which does not require loading the teacher during the training of student. For two kinds of student models and computing devices, the proposed KD paradigm gives rise to training speedup of 2.7x 3.4x.

While sophisticated neural-based models have achieved remarkable success in Visual Question Answering (VQA), these models tend to answer questions only according to superficial correlations between question and answer. Several recent approaches have been developed to address this language priors problem. However, most of them predict the correct answer according to one best output without checking the authenticity of answers. Besides, they only explore the interaction between image and question, ignoring the semantics of candidate answers. In this paper, we propose a select-and-rerank (SAR) progressive framework based on Visual Entailment. Specifically, we first select the candidate answers relevant to the question or the image, then we rerank the candidate answers by a visual entailment task, which verifies whether the image semantically entails the synthetic statement of the question and each candidate answer. Experimental results show the effectiveness of our proposed framework, which establishes a new state-of-the-art accuracy on VQA-CP v2 with a 7.55% improvement.

Sarcasm is a pervasive phenomenon in today’s social media platforms such as Twitter and Reddit. These platforms allow users to create multi-modal messages, including texts, images, and videos. Existing multi-modal sarcasm detection methods either simply concatenate the features from multi modalities or fuse the multi modalities information in a designed manner. However, they ignore the incongruity character in sarcastic utterance, which is often manifested between modalities or within modalities. Inspired by this, we propose a BERT architecture-based model, which concentrates on both intra and inter-modality incongruity for multi-modal sarcasm detection. To be specific, we are inspired by the idea of self-attention mechanism and design inter-modality attention to capturing inter-modality incongruity. In addition, the co-attention mechanism is applied to model the contradiction within the text. The incongruity information is then used for prediction. The experimental results demonstrate that our model achieves state-of-the-art performance on a public multi-modal sarcasm detection dataset.

Open-domain question answering (OpenQA) aims to answer questions based on a number of unlabeled paragraphs. Existing approaches always follow the distantly supervised setup where some of the paragraphs are wrong-labeled (noisy), and mainly utilize the paragraph-question relevance to denoise. However, the paragraph-paragraph relevance, which may aggregate the evidence among relevant paragraphs, can also be utilized to discover more useful paragraphs. Moreover, current approaches mainly focus on the positive paragraphs which are known to contain the answer during training. This will affect the generalization ability of the model and make it be disturbed by the similar but irrelevant (distracting) paragraphs during testing. In this paper, we first introduce a ranking model leveraging the paragraph-question and the paragraph-paragraph relevance to compute a confidence score for each paragraph. Furthermore, based on the scores, we design a modified weighted sampling strategy for training to mitigate the influence of the noisy and distracting paragraphs. Experiments on three public datasets (Quasar-T, SearchQA and TriviaQA) show that our model advances the state of the art.