Large language models (LLMs) exhibit enhanced reasoning at larger scales, driving efforts to distill these capabilities into smaller models via teacher-student learning.Previous works simply fine-tune student models on teachers’ generated Chain-of-Thoughts (CoTs) data. Although these methods enhance in-domain (IND) reasoning performance, they struggle to generalize to out-of-domain (OOD) tasks.We believe that the widespread spurious correlations between questions and answers may lead the model to preset a specific answer which restricts the diversity and generalizability of its reasoning process.In this paper, we propose Cascading Decomposed CoTs Distillation (CasCoD) to address these issues by decomposing the traditional single-step learning process into two cascaded learning steps. Specifically, by restructuring the training objectives—removing the answer from outputs and concatenating the question with the rationale as input—CasCoD’s two-step learning process ensures that students focus on learning rationales without interference from the preset answers, thus improving reasoning generalizability. Extensive experiments demonstrate the effectiveness of CasCoD on both IND and OOD benchmark reasoning datasets

Generative language models are usually pre-trained on large text corpus via predicting the next token (i.e., sub-word/word/phrase) given the previous ones. Recent works have demonstrated the impressive performance of large generative language models on downstream tasks. However, existing generative language models generally neglect an inherent challenge in text corpus during training, i.e., the imbalance between frequent tokens and infrequent ones. It can lead a language model to be dominated by common and easy-to-learn tokens, thereby overlooking the infrequent and difficult-to-learn ones. To alleviate that, we propose a **MiLe Loss** function for **mi**tigating the bias of **le**arning difficulties with tokens. During training, it can dynamically assess the learning difficulty of a to-be-learned token, according to the information entropy of the corresponding predicted probability distribution over the vocabulary. Then it scales the training loss adaptively, trying to lead the model to focus more on the difficult-to-learn tokens. On the Pile dataset, we train generative language models at different scales of 468M, 1.2B, and 6.7B parameters. Experiments reveal that models incorporating the proposed MiLe Loss can gain consistent performance improvement on downstream benchmarks.

Capturing the unique knowledge demands for each dialogue context plays a crucial role in commonsense knowledge-grounded response generation. However, current CoT-based and RAG-based methods are still unsatisfactory in the era of LLMs because 1) CoT often overestimates the capabilities of LLMs and treats them as isolated knowledge Producers; thus, CoT only uses the inherent knowledge of LLM itself and then suffers from the hallucination and outdated knowledge, and 2) RAG underestimates LLMs because LLMs are the passive Receivers that can only use the knowledge retrieved by external retrievers. In contrast, this work regards LLMs as interactive Collaborators and proposes a novel DCRAG (Demands-Guided Collaborative RAG) to leverage the knowledge from both LLMs and the external knowledge graph. Specifically, DCRAG designs three Thought-then-Generate stages to collaboratively investigate knowledge demands, followed by a Demands-Guided Knowledge Retrieval to retrieve external knowledge by interacting with LLMs. Extensive experiments and in-depth analyses on English DailyDialog and Chinese Diamante datasets proved DCRAG can effectively capture knowledge demands and bring higher-quality responses.

Dialogue response selection aims to select an appropriate response from several candidates based on a given user and system utterance history. Most existing works primarily focus on post-training and fine-tuning tailored for cross-encoders. However, there are no post-training methods tailored for dense encoders in dialogue response selection. We argue that when the current language model, based on dense dialogue systems (such as BERT), is employed as a dense encoder, it separately encodes dialogue context and response, leading to a struggle to achieve the alignment of both representations. Thus, we propose Dial-MAE (Dialogue Contextual Masking Auto-Encoder), a straightforward yet effective post-training technique tailored for dense encoders in dialogue response selection. Dial-MAE uses an asymmetric encoder-decoder architecture to compress the dialogue semantics into dense vectors, which achieves better alignment between the features of the dialogue context and response. Our experiments have demonstrated that Dial-MAE is highly effective, achieving state-of-the-art performance on two commonly evaluated benchmarks.

Table Question Answering (TQA) aims at composing an answer to a question based on tabular data. While prior research has shown that TQA models lack robustness, understanding the underlying cause and nature of this issue remains predominantly unclear, posing a significant obstacle to the development of robust TQA systems. In this paper, we formalize three major desiderata for a fine-grained evaluation of robustness of TQA systems. They should (i) answer questions regardless of alterations in table structure, (ii) base their responses on the content of relevant cells rather than on biases, and (iii) demonstrate robust numerical reasoning capabilities. To investigate these aspects, we create and publish a novel TQA evaluation benchmark in English. Our extensive experimental analysis reveals that none of the examined state-of-the-art TQA systems consistently excels in these three aspects. Our benchmark is a crucial instrument for monitoring the behavior of TQA systems and paves the way for the development of robust TQA systems. We release our benchmark publicly.

This paper proposes an information-theoretic representation learning framework, named conditional information flow maximization, to extract noise-invariant sufficient representations for the input data and target task. It promotes the learned representations have good feature uniformity and sufficient predictive ability, which can enhance the generalization of pre-trained language models (PLMs) for the target task. Firstly, an information flow maximization principle is proposed to learn more sufficient representations for the input and target by simultaneously maximizing both input-representation and representation-label mutual information. Unlike the information bottleneck, we handle the input-representation information in an opposite way to avoid the over-compression issue of latent representations. Besides, to mitigate the negative effect of potential redundant features from the input, we design a conditional information minimization principle to eliminate negative redundant features while preserve noise-invariant features. Experiments on 13 language understanding benchmarks demonstrate that our method effectively improves the performance of PLMs for classification and regression. Extensive experiments show that the learned representations are more sufficient, robust and transferable.

Attribution scores indicate the importance of different input parts and can, thus, explain model behaviour. Currently, prompt-based models are gaining popularity, i.a., due to their easier adaptability in low-resource settings. However, the quality of attribution scores extracted from prompt-based models has not been investigated yet. In this work, we address this topic by analyzing attribution scores extracted from prompt-based models w.r.t. plausibility and faithfulness and comparing them with attribution scores extracted from fine-tuned models and large language models. In contrast to previous work, we introduce training size as another dimension into the analysis. We find that using the prompting paradigm (with either encoder-based or decoder-based models) yields more plausible explanations than fine-tuning the models in low-resource settings and Shapley Value Sampling consistently outperforms attention and Integrated Gradients in terms of leading to more plausible and faithful explanations.

Emotional support conversation (ESC) task aims to relieve the emotional distress of users who have high-intensity of negative emotions. However, due to the ignorance of emotion intensity modelling which is essential for ESC, previous methods fail to capture the transition of emotion intensity effectively. To this end, we propose a Multi-stream information Fusion Framework (MFF-ESC) to thoroughly fuse three streams (text semantics stream, emotion intensity stream, and feedback stream) for the modelling of emotion intensity, based on a designed multi-stream fusion unit. As the difficulty of modelling subtle transitions of emotion intensity and the strong emotion intensity-feedback correlations, we use the KL divergence between feedback distribution and emotion intensity distribution to further guide the learning of emotion intensities. Experimental results on automatic and human evaluations indicate the effectiveness of our method.

This paper describes our system designed for SemEval-2023 Task 12: Sentiment analysis for African languages. The challenge faced by this task is the scarcity of labeled data and linguistic resources in low-resource settings. To alleviate these, we propose a generalized multilingual system SACL-XLMR for sentiment analysis on low-resource languages. Specifically, we design a lexicon-based multilingual BERT to facilitate language adaptation and sentiment-aware representation learning. Besides, we apply a supervised adversarial contrastive learning technique to learn sentiment-spread structured representations and enhance model generalization. Our system achieved competitive results, largely outperforming baselines on both multilingual and zero-shot sentiment classification subtasks. Notably, the system obtained the 1st rank on the zero-shot classification subtask in the official ranking. Extensive experiments demonstrate the effectiveness of our system.

Extracting generalized and robust representations is a major challenge in emotion recognition in conversations (ERC). To address this, we propose a supervised adversarial contrastive learning (SACL) framework for learning class-spread structured representations in a supervised manner. SACL applies contrast-aware adversarial training to generate worst-case samples and uses joint class-spread contrastive learning to extract structured representations. It can effectively utilize label-level feature consistency and retain fine-grained intra-class features. To avoid the negative impact of adversarial perturbations on context-dependent data, we design a contextual adversarial training (CAT) strategy to learn more diverse features from context and enhance the model’s context robustness. Under the framework with CAT, we develop a sequence-based SACL-LSTM to learn label-consistent and context-robust features for ERC. Experiments on three datasets show that SACL-LSTM achieves state-of-the-art performance on ERC. Extended experiments prove the effectiveness of SACL and CAT.

Addressee recognition aims to identify addressees in multi-party conversations. While state-of-the-art addressee recognition models have achieved promising performance, they still suffer from the issue of robustness when applied in real-world scenes. When exposed to a noisy environment, these models regard the noise as input and identify the addressee in a pre-given addressee closed set, while the addressees of the noise do not belong to this closed set, thus leading to the wrong identification of addressee. To this end, we propose a Robust Addressee Recognition (RAR) method, which discrete the addressees into a character codebook, making it able to represent open set addressees and robust in a noisy environment. Experimental results show that the introduction of the addressee character codebook helps to represent the open set addressees and highly improves the robustness of addressee recognition even if the input is noise.

Generative language models have recently shown remarkable success in generating answers to questions in a given textual context. However, these answers may suffer from hallucination, wrongly cite evidence, and spread misleading information. In this work, we address this problem by employing ChatGPT, a state-of-the-art generative model, as a machine-reading system. We ask it to retrieve answers to lexically varied and open-ended questions from trustworthy instructive texts. We introduce WHERE (WikiHow Evidence REtrieval), a new high-quality evaluation benchmark of a set of WikiHow articles exhaustively annotated with evidence sentences to questions that comes with a special challenge: All questions are about the article’s topic, but not all can be answered using the provided context. We interestingly find that when using a regular question-answering prompt, ChatGPT neglects to detect the unanswerable cases. When provided with a few examples, it learns to better judge whether a text provides answer evidence or not. Alongside this important finding, our dataset defines a new benchmark for evidence retrieval in question answering, which we argue is one of the necessary next steps for making large language models more trustworthy.

Prior works have shown the promising results of commonsense knowledge-aware models in improving informativeness while reducing the hallucination issue. Nonetheless, prior works often can only use monolingual knowledge whose language is consistent with the dialogue context. Except for a few high-resource languages, such as English and Chinese, most languages suffer from insufficient knowledge issues, especially minority languages. To this end, this work proposes a new task, Multi-Lingual Commonsense Knowledge-Aware Response Generation (MCKRG), which tries to use commonsense knowledge in other languages to enhance the current dialogue generation. Then, we construct a MCKRG dataset MCK-Dialog of seven languages with multiple alignment methods. Finally, we verify the effectiveness of using multi-lingual commonsense knowledge with a proposed MCK-T5 model. Extensive experimental results demonstrate the great potential of using multi-lingual commonsense knowledge in high-resource and low-resource languages. To the best of our knowledge, this work is the first to explore Multi-Lingual Commonsense Knowledge-Aware Response Generation.

Neural network models are vulnerable to adversarial examples, and adversarial transferability further increases the risk of adversarial attacks. Current methods based on transferability often rely on substitute models, which can be impractical and costly in real-world scenarios due to the unavailability of training data and the victim model’s structural details. In this paper, we propose a novel approach that directly constructs adversarial examples by extracting transferable features across various tasks. Our key insight is that adversarial transferability can extend across different tasks. Specifically, we train a sequence-to-sequence generative model named CT-GAT (Cross-Task Generative Adversarial Attack) using adversarial sample data collected from multiple tasks to acquire universal adversarial features and generate adversarial examples for different tasks.We conduct experiments on ten distinct datasets, and the results demonstrate that our method achieves superior attack performance with small cost.

We study model extraction attacks in natural language processing (NLP) where attackers aim to steal victim models by repeatedly querying the open Application Programming Interfaces (APIs). Recent works focus on limited-query budget settings and adopt random sampling or active learning-based sampling strategies on publicly available, unannotated data sources. However, these methods often result in selected queries that lack task relevance and data diversity, leading to limited success in achieving satisfactory results with low query costs. In this paper, we propose MeaeQ (Model extraction attack with efficient Queries), a straightforward yet effective method to address these issues. Specifically, we initially utilize a zero-shot sequence inference classifier, combined with API service information, to filter task-relevant data from a public text corpus instead of a problem domain-specific dataset. Furthermore, we employ a clustering-based data reduction technique to obtain representative data as queries for the attack. Extensive experiments conducted on four benchmark datasets demonstrate that MeaeQ achieves higher functional similarity to the victim model than baselines while requiring fewer queries.

In this work we investigate the hypothesis that enriching contextualized models using fine-tuning tasks can improve theircapacity to detect lexical semantic change (LSC). We include tasks aimed to capture both low-level linguistic information like part-of-speech tagging, as well as higher level (semantic) information. Through a series of analyses we demonstrate that certain combinations of fine-tuning tasks, like sentiment, syntactic information, and logical inference, bring large improvements to standard LSC models that are based only on standard language modeling. We test on the binary classification and ranking tasks of SemEval-2020 Task 1 and evaluate using both permutation tests and under transfer-learningscenarios.

The emotion cause pair extraction (ECPE) task aims to extract emotions and causes as pairs from documents. We observe that the relative distance distribution of emotions and causes is extremely imbalanced in the typical ECPE dataset. Existing methods have set a fixed size window to capture relations between neighboring clauses. However, they neglect the effective semantic connections between distant clauses, leading to poor generalization ability towards position-insensitive data. To alleviate the problem, we propose a novel Multi-Granularity Semantic Aware Graph model (MGSAG) to incorporate fine-grained and coarse-grained semantic features jointly, without regard to distance limitation. In particular, we first explore semantic dependencies between clauses and keywords extracted from the document that convey fine-grained semantic features, obtaining keywords enhanced clause representations. Besides, a clause graph is also established to model coarse-grained semantic relations between clauses. Experimental results indicate that MGSAG surpasses the existing state-of-the-art ECPE models. Especially, MGSAG outperforms other models significantly in the condition of position-insensitive data.

The widespread of fake news has detrimental societal effects. Recent works model information propagation as graph structure and aggregate structural features from user interactions for fake news detection. However, they usually neglect a broader propagation uncertainty issue, caused by some missing and unreliable interactions during actual spreading, and suffer from learning accurate and diverse structural properties. In this paper, we propose a novel dual graph-based model, Uncertainty-aware Propagation Structure Reconstruction (UPSR) for improving fake news detection. Specifically, after the original propagation modeling, we introduce propagation structure reconstruction to fully explore latent interactions in the actual propagation. We design a novel Gaussian Propagation Estimation to refine the original deterministic node representation by multiple Gaussian distributions and arise latent interactions with KL divergence between distributions in a multi-facet manner. Extensive experiments on two real-world datasets demonstrate the effectiveness and superiority of our model.

Fake news’s quick propagation on social media brings severe social ramifications and economic damage. Previous fake news detection usually learn semantic and structural patterns within a single target propagation tree. However, they are usually limited in narrow signals since they do not consider latent information cross other propagation trees. Motivated by a common phenomenon that most fake news is published around a specific hot event/topic, this paper develops a new concept of propagation forest to naturally combine propagation trees in a semantic-aware clustering. We propose a novel Unified Propagation Forest-based framework (UniPF) to fully explore latent correlations between propagation trees to improve fake news detection. Besides, we design a root-induced training strategy, which encourages representations of propagation trees to be closer to their prototypical root nodes. Extensive experiments on four benchmarks consistently suggest the effectiveness and scalability of UniPF.

Detecting rumors on social media is a very critical task with significant implications to the economy, public health, etc. Previous works generally capture effective features from texts and the propagation structure. However, the uncertainty caused by unreliable relations in the propagation structure is common and inevitable due to wily rumor producers and the limited collection of spread data. Most approaches neglect it and may seriously limit the learning of features. Towards this issue, this paper makes the first attempt to explore propagation uncertainty for rumor detection. Specifically, we propose a novel Edge-enhanced Bayesian Graph Convolutional Network (EBGCN) to capture robust structural features. The model adaptively rethinks the reliability of latent relations by adopting a Bayesian approach. Besides, we design a new edge-wise consistency training framework to optimize the model by enforcing consistency on relations. Experiments on three public benchmark datasets demonstrate that the proposed model achieves better performance than baseline methods on both rumor detection and early rumor detection tasks.

Multi-label text classification is one of the fundamental tasks in natural language processing. Previous studies have difficulties to distinguish similar labels well because they learn the same document representations for different labels, that is they do not explicitly extract label-specific semantic components from documents. Moreover, they do not fully explore the high-order interactions among these semantic components, which is very helpful to predict tail labels. In this paper, we propose a novel label-specific dual graph neural network (LDGN), which incorporates category information to learn label-specific components from documents, and employs dual Graph Convolution Network (GCN) to model complete and adaptive interactions among these components based on the statistical label co-occurrence and dynamic reconstruction graph in a joint way. Experimental results on three benchmark datasets demonstrate that LDGN significantly outperforms the state-of-the-art models, and also achieves better performance with respect to tail labels.

Computational linguistic research on language change through distributional semantic (DS) models has inspired researchers from fields such as philosophy and literary studies, who use these methods for the exploration and comparison of comparatively small datasets traditionally analyzed by close reading. Research on methods for small data is still in early stages and it is not clear which methods achieve the best results. We investigate the possibilities and limitations of using distributional semantic models for analyzing philosophical data by means of a realistic use-case. We provide a ground truth for evaluation created by philosophy experts and a blueprint for using DS models in a sound methodological setup. We compare three methods for creating specialized models from small datasets. Though the models do not perform well enough to directly support philosophers yet, we find that models designed for small data yield promising directions for future work.

The dissemination of fake news significantly affects personal reputation and public trust. Recently, fake news detection has attracted tremendous attention, and previous studies mainly focused on finding clues from news content or diffusion path. However, the required features of previous models are often unavailable or insufficient in early detection scenarios, resulting in poor performance. Thus, early fake news detection remains a tough challenge. Intuitively, the news from trusted and authoritative sources or shared by many users with a good reputation is more reliable than other news. Using the credibility of publishers and users as prior weakly supervised information, we can quickly locate fake news in massive news and detect them in the early stages of dissemination. In this paper, we propose a novel structure-aware multi-head attention network (SMAN), which combines the news content, publishing, and reposting relations of publishers and users, to jointly optimize the fake news detection and credibility prediction tasks. In this way, we can explicitly exploit the credibility of publishers and users for early fake news detection. We conducted experiments on three real-world datasets, and the results show that SMAN can detect fake news in 4 hours with an accuracy of over 91%, which is much faster than the state-of-the-art models.

Multi-turn retrieval-based conversation is an important task for building intelligent dialogue systems. Existing works mainly focus on matching candidate responses with every context utterance on multiple levels of granularity, which ignore the side effect of using excessive context information. Context utterances provide abundant information for extracting more matching features, but it also brings noise signals and unnecessary information. In this paper, we will analyze the side effect of using too many context utterances and propose a multi-hop selector network (MSN) to alleviate the problem. Specifically, MSN firstly utilizes a multi-hop selector to select the relevant utterances as context. Then, the model matches the filtered context with the candidate response and obtains a matching score. Experimental results show that MSN outperforms some state-of-the-art methods on three public multi-turn dialogue datasets.

Vocabulary is one of the most important parts of language competence. Testing of vocabulary knowledge is central to research on reading and language. However, it usually costs a large amount of time and human labor to build an item bank and to test large number of students. In this paper, we propose a novel testing strategy by combining automatic item generation (AIG) and computerized adaptive testing (CAT) in vocabulary assessment for Chinese L2 learners. Firstly, we generate three types of vocabulary questions by modeling both the vocabulary knowledge and learners’ writing error data. After evaluation and calibration, we construct a balanced item pool with automatically generated items, and implement a three-parameter computerized adaptive test. We conduct manual item evaluation and online student tests in the experiments. The results show that the combination of AIG and CAT can construct test items efficiently and reduce test cost significantly. Also, the test result of CAT can provide valuable feedback to AIG algorithms.

In the era of big data, focused analysis for diverse topics with a short response time becomes an urgent demand. As a fundamental task, information filtering therefore becomes a critical necessity. In this paper, we propose a novel deep relevance model for zero-shot document filtering, named DAZER. DAZER estimates the relevance between a document and a category by taking a small set of seed words relevant to the category. With pre-trained word embeddings from a large external corpus, DAZER is devised to extract the relevance signals by modeling the hidden feature interactions in the word embedding space. The relevance signals are extracted through a gated convolutional process. The gate mechanism controls which convolution filters output the relevance signals in a category dependent manner. Experiments on two document collections of two different tasks (i.e., topic categorization and sentiment analysis) demonstrate that DAZER significantly outperforms the existing alternative solutions, including the state-of-the-art deep relevance ranking models.

Building multi-turn information-seeking conversation systems is an important and challenging research topic. Although several advanced neural text matching models have been proposed for this task, they are generally not efficient for industrial applications. Furthermore, they rely on a large amount of labeled data, which may not be available in real-world applications. To alleviate these problems, we study transfer learning for multi-turn information seeking conversations in this paper. We first propose an efficient and effective multi-turn conversation model based on convolutional neural networks. After that, we extend our model to adapt the knowledge learned from a resource-rich domain to enhance the performance. Finally, we deployed our model in an industrial chatbot called AliMe Assist and observed a significant improvement over the existing online model.