The impressive capabilities of recent language models can be largely attributed to the multi-trillion token pretraining datasets that they are trained on. However, model developers fail to disclose their construction methodology which has lead to a lack of open information on how to develop effective pretraining sets. To address this issue, we perform the first systematic study across the entire pipeline of pretraining set construction. First, we run ablations on existing techniques for pretraining set development to identify which methods translate to the largest gains in model accuracy on downstream evaluations. Then, we categorize the most widely used data source, web crawl snapshots, across the attributes of toxicity, quality, type of speech, and domain. Finally, we show how such attribute information can be used to further refine and improve the quality of a pretraining set. These findings constitute an actionable set of steps that practitioners can use to develop high quality pretraining sets.

An ideal dialogue system requires continuous skill acquisition and adaptation to new tasks while retaining prior knowledge. Dialogue State Tracking (DST), vital in these systems, often involves learning new services, confronting catastrophic forgetting and a critical capability loss termed the “Value Selection Quandary”. To address these challenges, we introduce the Reason-of-Select (RoS) distillation method by enhancing smaller models with a novel “meta-reasoning” capability. Meta-reasoning, employing an enhanced multi-domain perspective, combines fragments of meta-knowledge from domain-specific dialogues during continual learning, transcending traditional single-perspective reasoning. This domain bootstrapping process enhances the model’s ability to dissect intricate dialogues from multiple possible values, and its domain-agnostic property aligns data distribution across different domains, effectively mitigating forgetting. Besides, two novel improvements, “multi-value resolution” strategy and Semantic Contrastive Reasoning Selection method, significantly enhance RoS by generating DST-specific selection chains and mitigating hallucinations in teachers’ reasoning, ensuring effective and reliable knowledge transfer. Extensive experiments validate the exceptional performance and robust generalization capabilities of our method.

A practical dialogue system requires the capacity for ongoing skill acquisition and adaptability to new tasks while preserving prior knowledge. However, current methods for Continual Dialogue State Tracking (DST), a crucial function of dialogue systems, struggle with the catastrophic forgetting issue and knowledge transfer between tasks. We present TaSL, a novel framework for task skill localization and consolidation that enables effective knowledge transfer without relying on memory replay. TaSL uses a novel group-wise technique to pinpoint task-specific and task-shared areas. Additionally, a fine-grained skill consolidation strategy protects task-specific knowledge from being forgotten while updating shared knowledge for bi-directional knowledge transfer. As a result, TaSL strikes a balance between preserving previous knowledge and excelling at new tasks. Comprehensive experiments on various backbones highlight the significant performance improvements of TaSL, with a 7.6% absolute increase in Avg. JGA and an 11% absolute rise in BWT metrics over existing state-of-the-art methods. The source code is provided for reproducibility.

We present EasyGen, an efficient model designed to enhance multimodal understanding and generation by harnessing the capabilities of diffusion models and large language models (LLMs). Unlike existing multimodal models that predominately depend on encoders like CLIP or ImageBind and need ample amounts of training data to bridge modalities, EasyGen leverages BiDiffuser, a bidirectional conditional diffusion model, to foster more efficient modality interactions. EasyGen achieves text generation by training a projection layer linking BiDiffuser and an LLM, and facilities image generation by training an adapter to align the LLM’s text space with the BiDiffuser’s image space. Comprehensive quantitative and qualitative experiments show that EasyGen excels in data-efficient training, high-quality image generation, and extendibility, effectively addressing the challenges in multimodal generation.

Emotion Recognition in Conversation (ERC) has attracted increasing attention due to its wide applications in public opinion analysis, empathetic conversation generation, and so on. However, ERC research suffers from the problems of data imbalance and the presence of similar linguistic expressions for different emotions. These issues can result in limited learning for minority emotions, biased predictions for common emotions, and the misclassification of different emotions with similar linguistic expressions. To alleviate these problems, we propose a Contrast-Enhanced Prompt-Tuning (CEPT) framework for ERC. We transform the ERC task into a Masked Language Modeling (MLM) generation task and generate the emotion for each utterance in the conversation based on the prompt-tuning of the Pre-trained Language Model (PLM), where a novel mixed prompt template and a label mapping strategy are introduced for better context and emotion feature modeling. Moreover, Supervised Contrastive Learning (SCL) is employed to help the PLM mine more information from the labels and learn a more discriminative representation space for utterances with different emotions. We conduct extensive experiments and the results demonstrate that CEPT outperforms the state-of-the-art methods on all three benchmark datasets and excels in recognizing minority emotions.

Stance detection aims to determine the attitude expressed in text towards a given target. Zero-shot stance detection (ZSSD) has emerged to classify stances towards unseen targets during inference. Recent data augmentation techniques for ZSSD increase transferable knowledge between targets through text or target augmentation. However, these methods exhibit limitations. Target augmentation lacks logical connections between generated targets and source text, while text augmentation relies solely on training data, resulting in insufficient generalization. To address these issues, we propose an encoder-decoder data augmentation (EDDA) framework. The encoder leverages large language models and chain-of-thought prompting to summarize texts into target-specific if-then rationales, establishing logical relationships. The decoder generates new samples based on these expressions using a semantic correlation word replacement strategy to increase syntactic diversity. We also analyze the generated expressions to develop a rationale-enhanced network that fully utilizes the augmented data. Experiments on benchmark datasets demonstrate our approach substantially improves over state-of-the-art ZSSD techniques. The proposed EDDA framework increases semantic relevance and syntactic variety in augmented texts while enabling interpretable rationale-based learning.

Out-of-distribution (OOD) detection plays a vital role in enhancing the reliability of machine learning models. As large language models (LLMs) become more prevalent, the applicability of prior research on OOD detection that utilized smaller-scale Transformers such as BERT, RoBERTa, and GPT-2 may be challenged, due to the significant differences in the scale of these models, their pre-training objectives, and the paradigms used for inference. This paper initiates a pioneering empirical investigation into the OOD detection capabilities of LLMs, focusing on the LLaMA series ranging from 7B to 65B in size. We thoroughly evaluate commonly used OOD detectors, examining their performance in both zero-grad and fine-tuning scenarios. Notably, we alter previous discriminative in-distribution fine-tuning into generative fine-tuning, aligning the pre-training objective of LLMs with downstream tasks. Our findings unveil that a simple cosine distance OOD detector demonstrates superior efficacy, outperforming other OOD detectors. We provide an intriguing explanation for this phenomenon by highlighting the isotropic nature of the embedding spaces of LLMs, which distinctly contrasts with the anisotropic property observed in smaller BERT family models. The new insight enhances our understanding of how LLMs detect OOD data, thereby enhancing their adaptability and reliability in dynamic environments. We have released the source code at https://github.com/Awenbocc/LLM-OOD for other researchers to reproduce our results.

Out-of-distribution (OOD) detection plays a crucial role in ensuring the safety and reliability of deep neural networks in various applications. While there has been a growing focus on OOD detection in visual data, the field of textual OOD detection has received less attention. Only a few attempts have been made to directly apply general OOD detection methods to natural language processing (NLP) tasks, without adequately considering the characteristics of textual data. In this paper, we delve into textual OOD detection with Transformers. We first identify a key problem prevalent in existing OOD detection methods: the biased representation learned through the maximization of the conditional likelihood p(y|x) can potentially result in subpar performance. We then propose a novel variational inference framework for OOD detection (VI-OOD), which maximizes the likelihood of the joint distribution p(x, y) instead of p(y|x). VI-OOD is tailored for textual OOD detection by efficiently exploiting the representations of pre-trained Transformers. Through comprehensive experiments on various text classification tasks, VI-OOD demonstrates its effectiveness and wide applicability. Our code has been released at https://github.com/liam0949/LLM-OOD.

Teachers often guide students to improve their essays by adding engaging modifiers to polish the sentences. In this work, we present the first study on automatic sentence polishing by adding modifiers. Since there is no available dataset for the new task, we first automatically construct a large number of parallel data by removing modifiers in the engaging sentences collected from public resources. Then we fine-tune LongLM to reconstruct the original sentences from the corrupted ones. Considering that much overlap between inputs and outputs may bias the model to completely copy the inputs, we split each source sentence into sub-sentences and only require the model to generate the modified sub-sentences. Furthermore, we design a retrieval augmentation algorithm to prompt the model to add suitable modifiers. Automatic and manual evaluation on the auto-constructed test set and real human texts show that our model can generate more engaging sentences with suitable modifiers than strong baselines while keeping fluency. We deploy the model at http://coai.cs.tsinghua.edu.cn/static/polishSent/. A demo video is available at https://youtu.be/Y6gFHOgSv8Y.

Dialogue State Tracking (DST) is of paramount importance in ensuring accurate tracking of user goals and system actions within task-oriented dialogue systems. The emergence of large language models (LLMs) such as GPT3 and ChatGPT has sparked considerable interest in assessing their efficacy across diverse applications. In this study, we conduct an initial examination of ChatGPT’s capabilities in DST. Our evaluation uncovers the exceptional performance of ChatGPT in this task, offering valuable insights to researchers regarding its capabilities and providing useful directions for designing and enhancing dialogue systems. Despite its impressive performance, ChatGPT has significant limitations including its closed-source nature, request restrictions, raising data privacy concerns, and lacking local deployment capabilities. To address these concerns, we present LDST, an LLM-driven DST framework based on smaller, open-source foundation models. By utilizing a novel domain-slot instruction tuning method, LDST achieves performance on par with ChatGPT. Comprehensive evaluations across three distinct experimental settings, we find that LDST exhibits remarkable performance improvements in both zero-shot and few-shot setting compared to previous SOTA methods. The source code is provided for reproducibility.

In the dynamic field of eCommerce, the quality and comprehensiveness of product descriptions are pivotal for enhancing search visibility and customer engagement. Effective product descriptions can address the ‘cold start’ problem, align with market trends, and ultimately lead to increased click-through rates. Traditional methods for crafting these descriptions often involve significant human effort and may lack both consistency and scalability. This paper introduces a novel methodology for automating product description generation using the LLAMA 2.0 7B language model. We train the model on a dataset of authentic product descriptions from Walmart, one of the largest eCommerce platforms. The model is then fine-tuned for domain-specific language features and eCommerce nuances to enhance its utility in sales and user engagement. We employ multiple evaluation metrics—including NDCG, customer click-through rates, and human assessments—to validate the effectiveness of our approach. Our findings reveal that the system is not only scalable but also significantly reduces the human workload involved in creating product descriptions. This study underscores the considerable potential of large language models like LLAMA 2.0 7B in automating and optimizing various facets of eCommerce platforms, offering significant business impact, including improved search functionality and increased sales.

Commonsense question answering (QA) requires machines to utilize the QA content and external commonsense knowledge graph (KG) for reasoning when answering questions. Existing work uses two independent modules to model the QA contextual text representation and relationships between QA entities in KG, which prevents information sharing between modules for co-reasoning. In this paper, we propose a novel model, Co-Reasoning Network (CORN), which adopts a bidirectional multi-level connection structure based on Co-Attention Transformer. The structure builds bridges to connect each layer of the text encoder and graph encoder, which can introduce the QA entity relationship from KG to the text encoder and bring contextual text information to the graph encoder, so that these features can be deeply interactively fused to form comprehensive text and graph node representations. Meanwhile, we propose a QA-aware node based KG subgraph construction method. The QA-aware nodes aggregate the question entity nodes and the answer entity nodes, and further guide the expansion and construction process of the subgraph to enhance the connectivity and reduce the introduction of noise. We evaluate our model on QA benchmarks in the CommonsenseQA and OpenBookQA datasets, and CORN achieves state-of-the-art performance.

Noise Learning is important in the task of text classification which depends on massive labeled data that could be error-prone. However, we find that noise learning in text classification is relatively underdeveloped: 1. many methods that have been proven effective in the image domain are not explored in text classification, 2. it is difficult to conduct a fair comparison between previous studies as they do experiments in different noise settings. In this work, we adapt four state-of-the-art methods of noise learning from the image domain to text classification. Moreover, we conduct comprehensive experiments on our benchmark of noise learning with seven commonly-used methods, four datasets, and five noise modes. Additionally, most previous works are based on an implicit hypothesis that the commonly-used datasets such as TREC, Ag-News and Chnsenticorp contain no errors. However, these datasets indeed contain 0.61% to 15.77% noise labels which we define as intrinsic noise that can cause inaccurate evaluation. Therefore, we build a new dataset Golden-Chnsenticorp( G-Chnsenticorp) without intrinsic noise to more accurately compare the effects of different noise learning methods. To the best of our knowledge, this is the first benchmark of noise learning for text classification.

Meta-learning has emerged as an effective approach for few-shot text classification. However, current studies fail to realize the importance of the semantic interaction between sentence features and neglect to enhance the generalization ability of the model to new tasks. In this paper, we integrate an adversarial network architecture into the meta-learning system and leverage cost-effective modules to build a novel few-shot classification framework named SaAML. Significantly, our approach can exploit the temporal convolutional network to encourage more discriminative representation learning and explore the attention mechanism to promote more comprehensive feature expression, thus resulting in better adaptation for new classes. Through a series of experiments on four benchmark datasets, we demonstrate that our new framework acquires considerable superiority over state-of-the-art methods in all datasets, increasing the performance of 1-shot classification and 5-shot classification by 7.15% and 2.89%, respectively.

Out-of-scope intent detection is of practical importance in task-oriented dialogue systems. Since the distribution of outlier utterances is arbitrary and unknown in the training stage, existing methods commonly rely on strong assumptions on data distribution such as mixture of Gaussians to make inference, resulting in either complex multi-step training procedures or hand-crafted rules such as confidence threshold selection for outlier detection. In this paper, we propose a simple yet effective method to train an out-of-scope intent classifier in a fully end-to-end manner by simulating the test scenario in training, which requires no assumption on data distribution and no additional post-processing or threshold setting. Specifically, we construct a set of pseudo outliers in the training stage, by generating synthetic outliers using inliner features via self-supervision and sampling out-of-scope sentences from easily available open-domain datasets. The pseudo outliers are used to train a discriminative classifier that can be directly applied to and generalize well on the test task. We evaluate our method extensively on four benchmark dialogue datasets and observe significant improvements over state-of-the-art approaches. Our code has been released at https://github.com/liam0949/DCLOOS.

Multi-hop inference for explanation generation is to combine two or more facts to make an inference. The task focuses on generating explanations for elementary science questions. In the task, the relevance between the explanations and the QA pairs is of vital importance. To address the task, a three-step framework is proposed. Firstly, vector distance between two texts is utilized to recall the top-K relevant explanations for each question, reducing the calculation consumption. Then, a selection module is employed to choose those most relative facts in an autoregressive manner, giving a preliminary order for the retrieved facts. Thirdly, we adopt a re-ranking module to re-rank the retrieved candidate explanations with relevance between each fact and the QA pairs. Experimental results illustrate the effectiveness of the proposed framework with an improvement of 39.78% in NDCG over the official baseline.

Embedding-based entity alignment has been widely investigated in recent years, but most proposed methods still rely on an ideal supervised learning setting with a large number of unbiased seed mappings for training and validation, which significantly limits their usage. In this study, we evaluate those state-of-the-art methods in an industrial context, where the impact of seed mappings with different sizes and different biases is explored. Besides the popular benchmarks from DBpedia and Wikidata, we contribute and evaluate a new industrial benchmark that is extracted from two heterogeneous knowledge graphs (KGs) under deployment for medical applications. The experimental results enable the analysis of the advantages and disadvantages of these alignment methods and the further discussion of suitable strategies for their industrial deployment.

We present our 7th place solution to the Gendered Pronoun Resolution challenge, which uses BERT without fine-tuning and a novel augmentation strategy designed for contextual embedding token-level tasks. Our method anonymizes the referent by replacing candidate names with a set of common placeholder names. Besides the usual benefits of effectively increasing training data size, this approach diversifies idiosyncratic information embedded in names. Using same set of common first names can also help the model recognize names better, shorten token length, and remove gender and regional biases associated with names. The system scored 0.1947 log loss in stage 2, where the augmentation contributed to an improvements of 0.04. Post-competition analysis shows that, when using different embedding layers, the system scores 0.1799 which would be third place.

Paraphrase generation is important in various applications such as search, summarization, and question answering due to its ability to generate textual alternatives while keeping the overall meaning intact. Clinical paraphrase generation is especially vital in building patient-centric clinical decision support (CDS) applications where users are able to understand complex clinical jargons via easily comprehensible alternative paraphrases. This paper presents Neural Clinical Paraphrase Generation (NCPG), a novel approach that casts the task as a monolingual neural machine translation (NMT) problem. We propose an end-to-end neural network built on an attention-based bidirectional Recurrent Neural Network (RNN) architecture with an encoder-decoder framework to perform the task. Conventional bilingual NMT models mostly rely on word-level modeling and are often limited by out-of-vocabulary (OOV) issues. In contrast, we represent the source and target paraphrase pairs as character sequences to address this limitation. To the best of our knowledge, this is the first work that uses attention-based RNNs for clinical paraphrase generation and also proposes an end-to-end character-level modeling for this task. Extensive experiments on a large curated clinical paraphrase corpus show that the attention-based NCPG models achieve improvements of up to 5.2 BLEU points and 0.5 METEOR points over a non-attention based strong baseline for word-level modeling, whereas further gains of up to 6.1 BLEU points and 1.3 METEOR points are obtained by the character-level NCPG models over their word-level counterparts. Overall, our models demonstrate comparable performance relative to the state-of-the-art phrase-based non-neural models.

Essential grammatical information is conveyed in signed languages by clusters of events involving facial expressions and movements of the head and upper body. This poses a significant challenge for computer-based sign language recognition. Here, we present new methods for the recognition of nonmanual grammatical markers in American Sign Language (ASL) based on: (1) new 3D tracking methods for the estimation of 3D head pose and facial expressions to determine the relevant low-level features; (2) methods for higher-level analysis of component events (raised/lowered eyebrows, periodic head nods and head shakes) used in grammatical markings―with differentiation of temporal phases (onset, core, offset, where appropriate), analysis of their characteristic properties, and extraction of corresponding features; (3) a 2-level learning framework to combine low- and high-level features of differing spatio-temporal scales. This new approach achieves significantly better tracking and recognition results than our previous methods.

This paper addresses the problem of automatically recognizing linguistically significant nonmanual expressions in American Sign Language from video. We develop a fully automatic system that is able to track facial expressions and head movements, and detect and recognize facial events continuously from video. The main contributions of the proposed framework are the following: (1) We have built a stochastic and adaptive ensemble of face trackers to address factors resulting in lost face track; (2) We combine 2D and 3D deformable face models to warp input frames, thus correcting for any variation in facial appearance resulting from changes in 3D head pose; (3) We use a combination of geometric features and texture features extracted from a canonical frontal representation. The proposed new framework makes it possible to detect grammatically significant nonmanual expressions from continuous signing and to differentiate successfully among linguistically significant expressions that involve subtle differences in appearance. We present results that are based on the use of a dataset containing 330 sentences from videos that were collected and linguistically annotated at Boston University.