Through pretraining on a corpus with various sources, Large Language Models (LLMs) have gained impressive performance. However, the impact of each component of the pretraining corpus remains opaque. As a result, the organization of the pretraining corpus is still empirical and may deviate from the optimal. To address this issue, we systematically analyze the impact of 48 datasets from 5 major categories of pretraining data of LLMs and measure their impacts on LLMs using benchmarks about nine major categories of model capabilities. Our analyses provide empirical results about the contribution of multiple corpora on the performances of LLMs, along with their joint impact patterns, including complementary, orthogonal, and correlational relationships. We also identify a set of “high-impact data” such as Books that is significantly related to a set of model capabilities. These findings provide insights into the organization of data to support more efficient pretraining of LLMs.

Within the evolving landscape of deep learning, the dilemma of data quantity and quality has been a long-standing problem. The recent advent of Large Language Models (LLMs) offers a data-centric solution to alleviate the limitations of real-world data with synthetic data generation. However, current investigations into this field lack a unified framework and mostly stay on the surface. Therefore, this paper provides an organization of relevant studies based on a generic workflow of synthetic data generation. By doing so, we highlight the gaps within existing research and outline prospective avenues for future study. This work aims to shepherd the academic and industrial communities towards deeper, more methodical inquiries into the capabilities and applications of LLMs-driven synthetic data generation.

New intent discovery (NID) is an important problem for deploying practical dialogue systems, which trains intent classifiers on a semi-supervised corpus where unlabeled user utterances contain both known and novel intents. Most existing NID algorithms place hope on the sample similarity to cluster unlabeled corpus to known or new samples. Lacking supervision on new intents, we experimentally find the intent classifier fails to fully distinguish new intents since they tend to assemble into intertwined centers.To address this problem, we propose a novel GeoID framework that learns geometry-aware representations to maximally separate all intents. Specifically, we are motivated by the recent findings on Neural Collapse (NC) in classification tasks to derive optimal intent center structure. Meanwhile, we devise a dual pseudo-labeling strategy based on optimal transport assignments and semi-supervised clustering, ensuring proper utterances-to-center arrangement.Extensive results show that our GeoID method establishes a new state-of-the-art performance, achieving a +3.49% average accuracy improvement on three standardized benchmarking datasets. We also verify its usefulness in assisting large language models for improved in-context performance.

To improve the performance of large language models (LLMs), researchers have explored providing LLMs with textual task-solving experience via prompts. However, they rely on manual efforts to acquire and apply such experience for each task, which is not feasible for the growing demand for LLMs and the variety of user questions.To address this issue, we design a lifelong autonomous experiential learning framework based on LLMs to explore whether LLMs can imitate human ability for learning and utilizing experience. It autonomously learns and accumulates experience through experience transfer and induction, categorizing the types of input questions to select which accumulated experience to employ for them.Experimental results on six widely used NLP datasets show that our framework performs reliably in each intermediate step and effectively improves the performance of GPT-3.5 and GPT-4. This validates the feasibility of using LLMs to mimic human experiential learning and application capabilities, offering a new path worth further exploration for the evolution of machine intelligence. Additionally, we provide a detailed analysis of the behavior of our framework at each step.We will open source codes after the acceptance, fostering open research in the NLP community and beyond.

Although achieving promising performance, recent analyses show that current generative large language models (LLMs) may still capture dataset biases and utilize them for generation, leading to poor generalizability and harmfulness of LLMs. However, due to the diversity of dataset biases and the over-optimization problem, previous prior-knowledge-based debiasing methods and fine-tuning-based debiasing methods may not be suitable for current LLMs.To address this issue, we explore combining active learning with the causal mechanisms and propose a casual-guided active learning (CAL) framework, which utilizes LLMs itself to automatically and autonomously identify informative biased samples and induce the bias patterns. Then a cost-effective and efficient in-context learning based method is employed to prevent LLMs from utilizing dataset biases during generation.Experimental results show that CAL can effectively recognize typical biased instances and induce various bias patterns for debiasing LLMs.

Fact-checking is the task of verifying the factuality of a given claim by examining the available evidence. High-quality evidence plays a vital role in enhancing fact-checking systems and facilitating the generation of explanations that are understandable to humans. However, the provision of both sufficient and relevant evidence for explainable fact-checking systems poses a challenge. To tackle this challenge, we propose a method based on a Large Language Model to automatically retrieve and summarize evidence from the Web. Furthermore, we construct RU22Fact, a novel multilingual explainable fact-checking dataset on the Russia-Ukraine conflict in 2022 of 16K samples, each containing real-world claims, optimized evidence, and referenced explanation. To establish a baseline for our dataset, we also develop an end-to-end explainable fact-checking system to verify claims and generate explanations. Experimental results demonstrate the prospect of optimized evidence in increasing fact-checking performance and also indicate the possibility of further progress in the end-to-end claim verification and explanation generation tasks.

Large language models (LLMs) have achieved significant performance in various natural language reasoning tasks. However, they still struggle with performing first-order logic reasoning over formal logical theories expressed in natural language. This is because the previous LLMs-based reasoning systems have the theoretical incompleteness issue. As a result, it can only address a limited set of simple reasoning problems, which significantly decreases their generalization ability. To address this issue, we propose a novel framework, named Generalizable and Faithful Reasoner (GFaiR), which introduces the paradigm of resolution refutation. Resolution refutation has the capability to solve all first-order logic reasoning problems by extending reasoning rules and employing the principle of proof by contradiction, so our system’s completeness can be improved by introducing resolution refutation. Experimental results demonstrate that our system outperforms previous works by achieving state-of-the-art performances in complex scenarios while maintaining performances in simple scenarios. Besides, we observe that GFaiR is faithful to its reasoning process.

Although achieving promising performance, current Natural Language Understanding models tend to utilize dataset biases instead of learning the intended task, which always leads to performance degradation on out-of-distribution (OOD) samples. Toincrease the performance stability, previous debiasing methods empirically capture bias features from data to prevent the model from corresponding biases. However, our analyses show that the empirical debiasing methods may fail to capture part of the potential dataset biases and mistake semantic information of input text as biases, which limits the effectiveness of debiasing. To address these issues, we propose a debiasing framework IEGDB that comprehensively detects the dataset biases to induce a set of biased features, and then purifies the biased features with the guidance of information entropy. Experimental results show that IEGDB can consistently improve the stability of performance on OOD datasets for a set of widely adopted NLU models.

Large-scale datasets in the real world inevitably involve label noise. Deep models can gradually overfit noisy labels and thus degrade model generalization. To mitigate the effects of label noise, learning with noisy labels (LNL) methods are designed to achieve better generalization performance. Due to the lack of suitable datasets, previous studies have frequently employed synthetic label noise to mimic real-world label noise. However, synthetic noise is not instance-dependent, making this approximation not always effective in practice. Recent research has proposed benchmarks for learning with real-world noisy labels. However, the noise sources within may be single or fuzzy, making benchmarks different from data with heterogeneous label noises in the real world. To tackle these issues, we contribute NoisywikiHow, the largest NLP benchmark built with minimal supervision. Specifically, inspired by human cognition, we explicitly construct multiple sources of label noise to imitate human errors throughout the annotation, replicating real-world noise, whose corruption is affected by both ground-truth labels and instances. Moreover, we provide a variety of noise levels to support controlled experiments on noisy data, enabling us to evaluate LNL methods systematically and comprehensively. After that, we conduct extensive multi-dimensional experiments on a broad range of LNL methods, obtaining new and intriguing findings.

Large Language Models (LLMs) have shown impressive capabilities in various applications, but they still face various inconsistency issues. Existing works primarily focus on the inconsistency issues within a single LLM, while we complementarily explore the inter-consistency among multiple LLMs for collaboration. To examine whether LLMs can collaborate effectively to achieve a consensus for a shared goal, we focus on commonsense reasoning, and introduce a formal debate framework (FORD) to conduct a three-stage debate among LLMs with real-world scenarios alignment: fair debate, mismatched debate, and roundtable debate. Through extensive experiments on various datasets, LLMs can effectively collaborate to reach a consensus despite noticeable inter-inconsistencies, but imbalances in their abilities can lead to domination by superior LLMs. Leveraging a more advanced LLM like GPT-4 as an authoritative judge can boost collaboration performance. Our work contributes to understanding the inter-consistency among LLMs and lays the foundation for developing future collaboration methods. Codes and data are available at https://github.com/Waste-Wood/FORD.

Causal reasoning ability is crucial for numerous NLP applications. Despite the impressive emerging ability of ChatGPT in various NLP tasks, it is unclear how well ChatGPT performs in causal reasoning. In this paper, we conduct the first comprehensive evaluation of the ChatGPT’s causal reasoning capabilities. Experiments show that ChatGPT is not a good causal reasoner, but a good causal interpreter. Besides, ChatGPT has a serious hallucination on causal reasoning, possibly due to the reporting biases between causal and non-causal relationships in natural language, as well as ChatGPT’s upgrading processes, such as RLHF. The In-Context Learning (ICL) and Chain-of-Thought (COT) techniques can further exacerbate such causal hallucination. Additionally, the causal reasoning ability of ChatGPT is sensitive to the words used to express the causal concept in prompts, and close-ended prompts perform better than open-ended prompts. For events in sentences, ChatGPT excels at capturing explicit causality rather than implicit causality, and performs better in sentences with lower event density and smaller lexical distance between events.

Understanding causality has vital importance for various Natural Language Processing (NLP) applications. Beyond the labeled instances, conceptual explanations of the causality can provide deep understanding of the causal fact to facilitate the causal reasoning process. However, such explanation information still remains absent in existing causal reasoning resources. In this paper, we fill this gap by presenting a human-annotated explainable CAusal REasoning dataset (e-CARE), which contains over 20K causal reasoning questions, together with natural language formed explanations of the causal questions. Experimental results show that generating valid explanations for causal facts still remains especially challenging for the state-of-the-art models, and the explanation information can be helpful for promoting the accuracy and stability of causal reasoning models.

Causal chain reasoning (CCR) is an essential ability for many decision-making AI systems, which requires the model to build reliable causal chains by connecting causal pairs. However, CCR suffers from two main transitive problems: threshold effect and scene drift. In other words, the causal pairs to be spliced may have a conflicting threshold boundary or scenario.To address these issues, we propose a novel Reliable Causal chain reasoning framework (ReCo), which introduces exogenous variables to represent the threshold and scene factors of each causal pair within the causal chain, and estimates the threshold and scene contradictions across exogenous variables via structural causal recurrent neural networks (SRNN). Experiments show that ReCo outperforms a series of strong baselines on both Chinese and English CCR datasets. Moreover, by injecting reliable causal chain knowledge distilled by ReCo, BERT can achieve better performances on four downstream causal-related tasks than BERT models enhanced by other kinds of knowledge.

Noisy labels are ubiquitous in natural language processing (NLP) tasks. Existing work, namely learning with noisy labels in NLP, is often limited to dedicated tasks or specific training procedures, making it hard to be widely used. To address this issue, SGD noise has been explored to provide a more general way to alleviate the effect of noisy labels by involving benign noise in the process of stochastic gradient descent. However, previous studies exert identical perturbation for all samples, which may cause overfitting on incorrect ones or optimizing correct ones inadequately. To facilitate this, we propose a novel stochastic tailor-made gradient noise (STGN), mitigating the effect of inherent label noise by introducing tailor-made benign noise for each sample. Specifically, we investigate multiple principles to precisely and stably discriminate correct samples from incorrect ones and thus apply different intensities of perturbation to them. A detailed theoretical analysis shows that STGN has good properties, beneficial for model generalization. Experiments on three different NLP tasks demonstrate the effectiveness and versatility of STGN. Also, STGN can boost existing robust training methods.

Predicting the subsequent event for an existing event context is an important but challenging task, as it requires understanding the underlying relationship between events. Previous methods propose to retrieve relational features from event graph to enhance the modeling of event correlation. However, the sparsity of event graph may restrict the acquisition of relevant graph information, and hence influence the model performance. To address this issue, we consider automatically building of event graph using a BERT model. To this end, we incorporate an additional structured variable into BERT to learn to predict the event connections in the training process. Hence, in the test process, the connection relationship for unseen events can be predicted by the structured variable. Results on two event prediction tasks: script event prediction and story ending prediction, show that our approach can outperform state-of-the-art baseline methods.

We study the problem of integrating cognitive language processing signals (e.g., eye-tracking or EEG data) into pre-trained language models like BERT. Existing methods typically fine-tune pre-trained models on cognitive data, ignoring the semantic gap between the texts and cognitive signals. To fill the gap, we propose CogBERT, a framework that can induce fine-grained cognitive features from cognitive data and incorporate cognitive features into BERT by adaptively adjusting the weight of cognitive features for different NLP tasks. Extensive experiments show that: (1) Cognition-guided pre-trained models can consistently perform better than basic pre-trained models on ten NLP tasks. (2) Different cognitive features contribute differently to different NLP tasks. Based on this observation, we give a fine-grained explanation of why cognitive data is helpful for NLP. (3) Different transformer layers of pre-trained models should encode different cognitive features, with word-level cognitive features at the bottom and semantic-level cognitive features at the top. (4) Attention visualization demonstrates that CogBERT aligns with human gaze patterns and improves its natural language comprehension ability.

Prior work infers the causation between events mainly based on the knowledge induced from the annotated causal event pairs. However, additional evidence information intermediate to the cause and effect remains unexploited. By incorporating such information, the logical law behind the causality can be unveiled, and the interpretability and stability of the causal reasoning system can be improved. To facilitate this, we present an Event graph knowledge enhanced explainable CAusal Reasoning framework (ExCAR). ExCAR first acquires additional evidence information from a large-scale causal event graph as logical rules for causal reasoning. To learn the conditional probabilistic of logical rules, we propose the Conditional Markov Neural Logic Network (CMNLN) that combines the representation learning and structure learning of logical rules in an end-to-end differentiable manner. Experimental results demonstrate that ExCAR outperforms previous state-of-the-art methods. Adversarial evaluation shows the improved stability of ExCAR over baseline systems. Human evaluation shows that ExCAR can achieve a promising explainable performance.

Abductive reasoning aims at inferring the most plausible explanation for observed events, which would play critical roles in various NLP applications, such as reading comprehension and question answering. To facilitate this task, a narrative text based abductive reasoning task 𝛼NLI is proposed, together with explorations about building reasoning framework using pretrained language models. However, abundant event commonsense knowledge is not well exploited for this task. To fill this gap, we propose a variational autoencoder based model ege-RoBERTa, which employs a latent variable to capture the necessary commonsense knowledge from event graph for guiding the abductive reasoning task. Experimental results show that through learning the external event graph knowledge, our approach outperforms the baseline methods on the 𝛼NLI task.

Many open-domain question answering problems can be cast as a textual entailment task, where a question and candidate answers are concatenated to form hypotheses. A QA system then determines if the supporting knowledge bases, regarded as potential premises, entail the hypotheses. In this paper, we investigate a neural-symbolic QA approach that integrates natural logic reasoning within deep learning architectures, towards developing effective and yet explainable question answering models. The proposed model gradually bridges a hypothesis and candidate premises following natural logic inference steps to build proof paths. Entailment scores between the acquired intermediate hypotheses and candidate premises are measured to determine if a premise entails the hypothesis. As the natural logic reasoning process forms a tree-like, hierarchical structure, we embed hypotheses and premises in a Hyperbolic space rather than Euclidean space to acquire more precise representations. Empirically, our method outperforms prior work on answering multiple-choice science questions, achieving the best results on two publicly available datasets. The natural logic inference process inherently provides evidence to help explain the prediction process.

We describe our system for Task 5 of SemEval 2020: Modelling Causal Reasoning in Language: Detecting Counterfactuals. Despite deep learning has achieved significant success in many fields, it still hardly drives today’s AI to strong AI, as it lacks of causation, which is a fundamental concept in human thinking and reasoning. In this task, we dedicate to detecting causation, especially counterfactuals from texts. We explore multiple pre-trained models to learn basic features and then fine-tune models with counterfactual data and pseudo-labeling data. Our team HIT-SCIR wins the first place (1st) in Sub-task 1 — Detecting Counterfactual Statements and is ranked 4th in Sub-task 2 — Detecting Antecedent and Consequence. In this paper we provide a detailed description of the approach, as well as the results obtained in this task.

Understanding event and event-centered commonsense reasoning are crucial for natural language processing (NLP). Given an observed event, it is trivial for human to infer its intents and effects, while this type of If-Then reasoning still remains challenging for NLP systems. To facilitate this, a If-Then commonsense reasoning dataset Atomic is proposed, together with an RNN-based Seq2Seq model to conduct such reasoning. However, two fundamental problems still need to be addressed: first, the intents of an event may be multiple, while the generations of RNN-based Seq2Seq models are always semantically close; second, external knowledge of the event background may be necessary for understanding events and conducting the If-Then reasoning. To address these issues, we propose a novel context-aware variational autoencoder effectively learning event background information to guide the If-Then reasoning. Experimental results show that our approach improves the accuracy and diversity of inferences compared with state-of-the-art baseline methods.

Prior work has proposed effective methods to learn event representations that can capture syntactic and semantic information over text corpus, demonstrating their effectiveness for downstream tasks such as script event prediction. On the other hand, events extracted from raw texts lacks of commonsense knowledge, such as the intents and emotions of the event participants, which are useful for distinguishing event pairs when there are only subtle differences in their surface realizations. To address this issue, this paper proposes to leverage external commonsense knowledge about the intent and sentiment of the event. Experiments on three event-related tasks, i.e., event similarity, script event prediction and stock market prediction, show that our model obtains much better event embeddings for the tasks, achieving 78% improvements on hard similarity task, yielding more precise inferences on subsequent events under given contexts, and better accuracies in predicting the volatilities of the stock market.

Target-dependent classification tasks, such as aspect-level sentiment analysis, perform fine-grained classifications towards specific targets. Semantic compositions over tree structures are promising for such tasks, as they can potentially capture long-distance interactions between targets and their contexts. However, previous work that operates on tree structures resorts to syntactic parsers or Treebank annotations, which are either subject to noise in informal texts or highly expensive to obtain. To address above issues, we propose a reinforcement learning based approach, which automatically induces target-specific sentence representations over tree structures. The underlying model is a RNN encoder-decoder that explores possible binary tree structures and a reward mechanism that encourages structures that improve performances on downstream tasks. We evaluate our approach on two benchmark tasks: firm-specific cumulative abnormal return prediction (based on formal news texts) and aspect-level sentiment analysis (based on informal social media texts). Experimental results show that our model gives superior performances compared to previous work that operates on parsed trees. Moreover, our approach gives some intuitions on how target-specific sentence representations can be achieved from its word constituents.

Story generation is a challenging problem in artificial intelligence (AI) and has received a lot of interests in the natural language processing (NLP) community. Most previous work tried to solve this problem using Sequence to Sequence (Seq2Seq) model trained with Maximum Likelihood Estimation (MLE). However, the pure MLE training objective much limits the power of Seq2Seq model in generating high-quality storys. In this paper, we propose using adversarial training augmented Seq2Seq model to generate reasonable and diversified story endings given a story context. Our model includes a generator that defines the policy of generating a story ending, and a discriminator that labels story endings as human-generated or machine-generated. Carefully designed human and automatic evaluation metrics demonstrate that our adversarial training augmented Seq2Seq model can generate more reasonable and diversified story endings compared to purely MLE-trained Seq2Seq model. Moreover, our model achieves better performance on the task of Story Cloze Test with an accuracy of 62.6% compared with state-of-the-art baseline methods.

Texts from the Internet serve as important data sources for financial market modeling. Early statistical approaches rely on manually defined features to capture lexical, sentiment and event information, which suffers from feature sparsity. Recent work has considered learning dense representations for news titles and abstracts. Compared to news titles, full documents can contain more potentially helpful information, but also noise compared to events and sentences, which has been less investigated in previous work. To fill this gap, we propose a novel target-specific abstract-guided news document representation model. The model uses a target-sensitive representation of the news abstract to weigh sentences in the news content, so as to select and combine the most informative sentences for market modeling. Results show that document representations can give better performance for estimating cumulative abnormal returns of companies when compared to titles and abstracts. Our model is especially effective when it used to combine information from multiple document sources compared to the sentence-level baselines.

Representing structured events as vectors in continuous space offers a new way for defining dense features for natural language processing (NLP) applications. Prior work has proposed effective methods to learn event representations that can capture syntactic and semantic information over text corpus, demonstrating their effectiveness for downstream tasks such as event-driven stock prediction. On the other hand, events extracted from raw texts do not contain background knowledge on entities and relations that they are mentioned. To address this issue, this paper proposes to leverage extra information from knowledge graph, which provides ground truth such as attributes and properties of entities and encodes valuable relations between entities. Specifically, we propose a joint model to combine knowledge graph information into the objective function of an event embedding learning model. Experiments on event similarity and stock market prediction show that our model is more capable of obtaining better event embeddings and making more accurate prediction on stock market volatilities.