Recently, the evaluation of Large Language Models has emerged as a popular area of research. The three crucial questions for LLM evaluation are ``what, where, and how to evaluate''. However, the existing research mainly focuses on the first two questions, which are basically what tasks to give the LLM during testing and what kind of knowledge it should deal with. As for the third question, which is about what standards to use, the types of evaluators, how to score, and how to rank, there hasn't been much discussion. In this paper, we analyze evaluation methods by comparing various criteria with both manual and automatic evaluation, utilizing onsite, crowd-sourcing, public annotators and GPT-4, with different scoring methods and ranking systems. We propose a new dataset, LLMEval and conduct evaluations on 20 LLMs. A total of 2,186 individuals participated, leading to the generation of 243,337 manual annotations and 57,511 automatic evaluation results. We perform comparisons and analyses of different settings and conduct 10 conclusions that can provide some insights for evaluating LLM in the future. The dataset and the results are publicly available at https://github.com/llmeval. The version with the appendix are publicly available at https://arxiv.org/abs/2312.07398.
论文检索
输入标题、作者或关键词，从 2,866 篇学术成果中精准定位
Accurately typing entity mentions from text segments is a fundamental task for various natural language processing applications. Many previous approaches rely on massive human-annotated data to perform entity typing. Nevertheless, collecting such data in highly specialized science and engineering domains (e.g., software engineering and security) can be time-consuming and costly, without mentioning the domain gaps between training and inference data if the model needs to be applied to confidential datasets. In this paper, we study the task of seed-guided fine-grained entity typing in science and engineering domains, which takes the name and a few seed entities for each entity type as the only supervision and aims to classify new entity mentions into both seen and unseen types (i.e., those without seed entities). To solve this problem, we propose SEType which first enriches the weak supervision by finding more entities for each seen type from an unlabeled corpus using the contextualized representations of pre-trained language models. It then matches the enriched entities to unlabeled text to get pseudo-labeled samples and trains a textual entailment model that can make inferences for both seen and unseen types. Extensive experiments on two datasets covering four domains demonstrate the effectiveness of SEType in comparison with various baselines. Code and data are available at: https://github.com/yuzhimanhua/SEType.
Style transfer for out-of-domain (OOD) singing voice synthesis (SVS) focuses on generating high-quality singing voices with unseen styles (such as timbre, emotion, pronunciation, and articulation skills) derived from reference singing voice samples. However, the endeavor to model the intricate nuances of singing voice styles is an arduous task, as singing voices possess a remarkable degree of expressiveness. Moreover, existing SVS methods encounter a decline in the quality of synthesized singing voices in OOD scenarios, as they rest upon the assumption that the target vocal attributes are discernible during the training phase. To overcome these challenges, we propose StyleSinger, the first singing voice synthesis model for zero-shot style transfer of out-of-domain reference singing voice samples. StyleSinger incorporates two critical approaches for enhanced effectiveness: 1) the Residual Style Adaptor (RSA) which employs a residual quantization module to capture diverse style characteristics in singing voices, and 2) the Uncertainty Modeling Layer Normalization (UMLN) to perturb the style attributes within the content representation during the training phase and thus improve the model generalization. Our extensive evaluations in zero-shot style transfer undeniably establish that StyleSinger outperforms baseline models in both audio quality and similarity to the reference singing voice samples. Access to singing voice samples can be found at https://stylesinger.github.io/.
Effectively and efficiently adapting a pre-trained language model (PLM) for human-centered text understanding (HCTU) is challenging since user tokens are million-level in most personalized applications and do not have concrete explicit semantics. A standard and parameter-efficient approach (e.g., LoRA) necessitates memorizing numerous suits of adapters for each user. In this work, we introduce a personalized LoRA (PLoRA) with a plug-and-play (PnP) framework for the HCTU task. PLoRA is effective, parameter-efficient, and dynamically deploying in PLMs. Moreover, a personalized dropout and a mutual information maximizing strategies are adopted and hence the proposed PLoRA can be well adapted to few/zero-shot learning scenarios for the cold-start issue. Experiments conducted on four benchmark datasets show that the proposed method outperforms existing methods in full/few/zero-shot learning scenarios for the HCTU task, even though it has fewer trainable parameters. For reproducibility, the code for this paper is available at: https://github.com/yoyo-yun/PLoRA.
Multi-goal conversational recommender system (MG-CRS) which is more in line with realistic scenarios has attracted a lot of attention. MG-CRS can dynamically capture the demands of users in conversation, continuously engage their interests, and make recommendations. The key of accomplishing these tasks is to plan a reasonable goal sequence which can naturally guide the user to accept the recommended goal. Previous works have demonstrated that mining the correlations of goals from the goal sequences in the dialogue corpus is helpful for recommending the goal that the user is interested in. However, they independently model correlations for each level of goal (i.e., goal type or entity) and neglect the order of goals appear in the dialogue. In this paper, we propose a goal interaction graph planning framework which constructs a directed heterogeneous graph to flexibly model the correlations between any level of goals and retain the order of goals. We design a goal interaction graph learning module to model the goal correlations and propagate goal representations via directed edges, then use an encoder and a dual-way fusion decoder to extract the most relevant information with the current goal from the conversation and domain knowledge, making the next-goal prediction fully exploit the prior goal correlations and user feedback. Finally we generate engaging responses based on the predicted goal sequence to complete the recommendation task. Experiments on two benchmark datasets show that our method achieves significant improvements in both the goal planning and response generation tasks.
The rapid evolution of speech synthesis and voice conversion has raised substantial concerns due to the potential misuse of such technology, prompting a pressing need for effective audio deepfake detection mechanisms. Existing detection models have shown remarkable success in discriminating known deepfake audio, but struggle when encountering new attack types. To address this challenge, one of the emergent effective approaches is continual learning. In this paper, we propose a continual learning approach called Radian Weight Modification (RWM) for audio deepfake detection. The fundamental concept underlying RWM involves categorizing all classes into two groups: those with compact feature distributions across tasks, such as genuine audio, and those with more spread-out distributions, like various types of fake audio. These distinctions are quantified by means of the in-class cosine distance, which subsequently serves as the basis for RWM to introduce a trainable gradient modification direction for distinct data types. Experimental evaluations against mainstream continual learning methods reveal the superiority of RWM in terms of knowledge acquisition and mitigating forgetting in audio deepfake detection. Furthermore, RWM's applicability extends beyond audio deepfake detection, demonstrating its potential significance in diverse machine learning domains such as image recognition.
Tree-of-Reasoning Question Decomposition for Complex Question Answering with Large Language Models
PDF ↗Large language models (LLMs) have recently demonstrated remarkable performance across various Natual Language Processing tasks. In the field of multi-hop reasoning, the Chain-of-thought (CoT) prompt method has emerged as a paradigm, using curated stepwise reasoning demonstrations to enhance LLM's ability to reason and produce coherent rational pathways. To ensure the accuracy, reliability, and traceability of the generated answers, many studies have incorporated information retrieval (IR) to provide LLMs with external knowledge. However, existing CoT with IR methods decomposes questions into sub-questions based on a single compositionality type, which limits their effectiveness for questions involving multiple compositionality types. Additionally, these methods suffer from inefficient retrieval, as complex questions often contain abundant information, leading to the retrieval of irrelevant information inconsistent with the query's intent. In this work, we propose a novel question decomposition framework called TRQA for multi-hop question answering, which addresses these limitations. Our framework introduces a reasoning tree (RT) to represent the structure of complex questions. It consists of four components: the Reasoning Tree Constructor (RTC), the Question Generator (QG), the Retrieval and LLM Interaction Module (RAIL), and the Answer Aggregation Module (AAM). Specifically, the RTC predicts diverse sub-question structures to construct the reasoning tree, allowing a more comprehensive representation of complex questions. The QG generates sub-questions for leaf-node in the reasoning tree, and we explore two methods for QG: prompt-based and T5-based approaches. The IR module retrieves documents aligned with sub-questions, while the LLM formulates answers based on the retrieved information. Finally, the AAM aggregates answers along the reason tree, producing a definitive response from bottom to top.
Word Sense Disambiguation (WSD) aims to determine the meaning of the target word according to the given context. Currently, a single representation enhanced by glosses from different dictionaries or languages is used to characterize each word sense. By analyzing the similarity between glosses of the same word sense, we find semantic biases among them, revealing that the glosses have their own descriptive perspectives. Therefore, the traditional approach of integrating all glosses by a single representation results in failing to present the unique semantics revealed by the individual glosses. In this paper, a quantum superposition state is employed to formalize the representations of multiple glosses of the same word sense to reveal their distributions. Furthermore, the quantum interference model is leveraged to calculate the probability that the target word belongs to this superposition state. The advantage is that the interference term can be regarded as a confidence level to guide word sense recognition. Finally, experiments are performed under standard WSD evaluation framework and the latest cross-lingual datasets, and the results verify the effectiveness of our model.
Due to the detrimental impact of noise on the conventional audio speech recognition (ASR) task, audio-visual speech recognition~(AVSR) has been proposed by incorporating both audio and visual video signals. Although existing methods have demonstrated that the aligned visual input of lip movements can enhance the robustness of AVSR systems against noise, the paired videos are not always available during inference, leading to the problem of the missing visual modality, which restricts their practicality in real-world scenarios. To tackle this problem, we propose a Discrete Feature based Visual Generative Model (DFVGM) which exploits semantic correspondences between the audio and visual modalities during training, generating visual hallucinations in lieu of real videos during inference. To achieve that, the primary challenge is to generate the visual hallucination given the noisy audio while preserving semantic correspondences with the clean speech. To tackle this challenge, we start with training the audio encoder in the Audio-Only (AO) setting, which generates continuous semantic features closely associated with the linguistic information. Simultaneously, the visual encoder is trained in the Visual-Only (VO) setting, producing visual features that are phonetically related. Next, we employ K-means to discretize the continuous audio and visual feature spaces. The discretization step allows DFVGM to capture high-level semantic structures that are more resilient to noise and generate visual hallucinations with high quality. To evaluate the effectiveness and robustness of our approach, we conduct extensive experiments on two publicly available datasets. The results demonstrate that our method achieves a remarkable 53% relative reduction (30.5%->12.9%) in Word Error Rate (WER) on average compared to the current state-of-the-art Audio-Only (AO) baselines while maintaining comparable results (< 5% difference) under the Audio-Visual (AV) setting even without video as input.
Multi-hop fact verification aims to detect the veracity of the given claim by integrating and reasoning over multiple pieces of evidence. Conventional multi-hop fact verification models are prone to rely on spurious correlations from the annotation artifacts, leading to an obvious performance decline on unbiased datasets. Among the various debiasing works, the causal inference-based methods become popular by performing theoretically guaranteed debiasing such as casual intervention or counterfactual reasoning. However, existing causal inference-based debiasing methods, which mainly formulate fact verification as a single-hop reasoning task to tackle shallow bias patterns, cannot deal with the complicated bias patterns hidden in multiple hops of evidence. To address the challenge, we propose Causal Walk, a novel method for debiasing multi-hop fact verification from a causal perspective with front-door adjustment. Specifically, in the structural causal model, the reasoning path between the treatment (the input claim-evidence graph) and the outcome (the veracity label) is introduced as the mediator to block the confounder. With the front-door adjustment, the causal effect between the treatment and the outcome is decomposed into the causal effect between the treatment and the mediator, which is estimated by applying the idea of random walk, and the causal effect between the mediator and the outcome, which is estimated with normalized weighted geometric mean approximation. To investigate the effectiveness of the proposed method, an adversarial multi-hop fact verification dataset and a symmetric multi-hop fact verification dataset are proposed with the help of the large language model. Experimental results show that Causal Walk outperforms some previous debiasing methods on both existing datasets and the newly constructed datasets. Code and data will be released at https://github.com/zcccccz/CausalWalk.
As an effective tool for eliciting the power of Large Language Models (LLMs), prompting has recently demonstrated unprecedented abilities across a variety of complex tasks. To further improve the performance, prompt ensemble has attracted substantial interest for tackling the hallucination and instability of LLMs. However, existing methods usually adopt a two-stage paradigm, which requires a pre-prepared set of prompts with substantial manual effort, and is unable to perform directed optimization for different weak learners. In this paper, we propose a simple, universal, and automatic method named PREFER (Prompt Ensemble learning via Feedback-Reflect-Refine) to address the stated limitations. Specifically, given the fact that weak learners are supposed to focus on hard examples during boosting, PREFER builds a feedback mechanism for reflecting on the inadequacies of existing weak learners. Based on this, the LLM is required to automatically synthesize new prompts for iterative refinement. Moreover, to enhance stability of the prompt effect evaluation, we propose a novel prompt bagging method involving forward and backward thinking, which is superior to majority voting and is beneficial for both feedback and weight calculation in boosting. Extensive experiments demonstrate that our PREFER achieves state-of-the-art performance in multiple types of tasks by a significant margin. We have made our code publicly available.
A Comprehensive Analysis of the Effectiveness of Large Language Models as Automatic Dialogue Evaluators
PDF ↗Automatic evaluation is an integral aspect of dialogue system research. The traditional reference-based NLG metrics are generally found to be unsuitable for dialogue assessment. Consequently, recent studies have suggested various unique, reference-free neural metrics that better align with human evaluations. Notably among them, large language models (LLMs), particularly the instruction-tuned variants like ChatGPT, are shown to be promising substitutes for human judges. Yet, existing works on utilizing LLMs for automatic dialogue evaluation are limited in their scope in terms of the number of meta-evaluation datasets, mode of evaluation, coverage of LLMs, etc. Hence, it remains inconclusive how effective these LLMs are. To this end, we conduct a comprehensive study on the application of LLMs for automatic dialogue evaluation. Specifically, we analyze the multi-dimensional evaluation capability of 30 recently emerged LLMs at both turn and dialogue levels, using a comprehensive set of 12 meta-evaluation datasets. Additionally, we probe the robustness of the LLMs in handling various adversarial perturbations at both turn and dialogue levels. Finally, we explore how model-level and dimension-level ensembles impact the evaluation performance. All resources are available at https://github.com/e0397123/comp-analysis.
ConsistentEE: A Consistent and Hardness-Guided Early Exiting Method for Accelerating Language Models Inference
PDF ↗Early Exiting is one of the most popular methods to achieve efficient inference. Current early exiting methods adopt the (weighted) sum of the cross entropy loss of all internal classifiers as the objective function during training, imposing all these classifiers to predict all instances correctly. However, during inference, as long as one internal classifier predicts an instance correctly, it can accelerate without losing accuracy. Thus, there is a notable gap between training and inference. We propose ConsistentEE, an early exiting method that is consistent in training and inference. ConsistentEE formulates the early exiting process as a reinforcement learning problem. A policy network is added to decide whether an instance should exit or continue. The training objective of ConsistentEE only requires each instance to be predicted correctly by one internal classifier. Additionally, we introduce the concept "Memorized Layer" to measure the hardness of an instance. We incorporate the memorized layer into reward function design, which allows "easy'' instances to focus more on acceleration while ``hard'' instances to focus more on accuracy. Experimental results show that our method outperforms other baselines on various natural language understanding and generation tasks using PLMs and LLMs as backbones respectively.
InterpretARA: Enhancing Hybrid Automatic Readability Assessment with Linguistic Feature Interpreter and Contrastive Learning
PDF ↗The hybrid automatic readability assessment (ARA) models that combine deep and linguistic features have recently received rising attention due to their impressive performance. However, the utilization of linguistic features is not fully realized, as ARA models frequently concentrate excessively on numerical values of these features, neglecting valuable structural information embedded within them. This leads to limited contribution of linguistic features in these hybrid ARA models, and in some cases, it may even result in counterproductive outcomes. In this paper, we propose a novel hybrid ARA model named InterpretARA through introducing a linguistic interpreter to better comprehend the structural information contained in linguistic features, and leveraging the contrastive learning that enables the model to understand relative difficulty relationships among texts and thus enhances deep representations. Both document-level and segment-level deep representations are extracted and used for the readability assessment. A series of experiments are conducted over four English corpora and one Chinese corpus to demonstrate the effectiveness of the proposed model. Experimental results show that InterpretARA outperforms state-of-the-art models in most corpora, and the introduced linguistic interpreter can provide more useful information than existing ways for ARA.
Open-sourced large language models (LLMs) have demonstrated remarkable efficacy in various tasks with instruction tuning. However, these models can sometimes struggle with tasks that require more specialized knowledge such as translation. One possible reason for such deficiency is that instruction tuning aims to generate fluent and coherent text that continues from a given instruction without being constrained by any task-specific requirements. Moreover, it can be more challenging to tune smaller LLMs with lower-quality training data. To address this issue, we propose a novel framework using examples in comparison to teach LLMs to learn translation. Our approach involves output comparison and preference comparison, presenting the model with carefully designed examples of correct and incorrect translations and an additional preference loss for better regularization. Empirical evaluation on four language directions of WMT2022 and FLORES-200 benchmarks shows the superiority of our proposed method over existing methods. Our findings offer a new perspective on fine-tuning LLMs for translation tasks and provide a promising solution for generating high-quality translations. Please refer to Github for more details: https://github.com/lemon0830/TIM.
In this paper, we propose a novel method for joint entity and relation extraction from unstructured text by framing it as a conditional sequence generation problem. In contrast to conventional generative information extraction models that are left-to-right token-level generators, our approach is \textit{span-based}. It generates a linearized graph where nodes represent text spans and edges represent relation triplets. Our method employs a transformer encoder-decoder architecture with pointing mechanism on a dynamic vocabulary of spans and relation types. Our model can capture the structural characteristics and boundaries of entities and relations through span representations while simultaneously grounding the generated output in the original text thanks to the pointing mechanism. Evaluation on benchmark datasets validates the effectiveness of our approach, demonstrating competitive results. Code is available at https://github.com/urchade/ATG.
Structured Complex Task Decomposition (SCTD) is the problem of breaking down a complex real-world task (such as planning a wedding) into a directed acyclic graph over individual steps that contribute to achieving the task, with edges specifying temporal dependencies between steps. SCTD is an important component of assistive planning tools, and a challenge for commonsense reasoning systems. We probe how accurately SCTD can be done with the knowledge extracted from pre-trained Large Language Models (LLMs). We introduce a new high-quality human-annotated dataset for this problem and novel metrics to fairly assess performance of LLMs against several baselines. Our experiments reveal that LLMs are able to decompose complex tasks into individual steps effectively, with a relative improvement of 15% to 280% over the best baseline. We also propose a number of approaches to further improve their performance, with a relative improvement of 7% to 37%. However, we find that LLMs still struggle to predict pairwise temporal dependencies, which reveals a gap in their understanding of complex tasks.
Large language models (LLMs) have shown impressive abilities for open-domain NLP tasks. However, LLMs are sometimes too footloose for natural language understanding (NLU) tasks which always have restricted output and input format. Their performances on NLU tasks are highly related to prompts or demonstrations and are shown to be poor at performing several representative NLU tasks, such as event extraction and entity typing. To this end, we present SeqGPT, a bilingual (i.e., English and Chinese) open-source autoregressive model specially enhanced for open-domain natural language understanding. We express all NLU tasks with two atomic tasks, which define fixed instructions to restrict the input and output format but still ``open'' for arbitrarily varied label sets. The model is first instruction-tuned with extremely fine-grained labeled data synthesized by ChatGPT and then further fine-tuned by 233 different atomic tasks from 152 datasets across various domains. The experimental results show that SeqGPT has decent classification and extraction ability, and is capable of performing language understanding tasks on unseen domains. We also conduct empirical studies on the scaling of data and model size as well as on the transfer across tasks. Our models are accessible at https://github.com/Alibaba-NLP/SeqGPT.
Large language models (LLMs) have shown great success in various Natural Language Processing (NLP) tasks, whist they still need updates after deployment to fix errors or keep pace with the changing knowledge in the world. Researchers formulate such problem as Model Editing and have developed various editors focusing on different axes of editing properties. However, current editors can hardly support all properties and rely on heavy computational resources. In this paper, we propose a plug-in Model Editing method based on neuron-indexed dynamic LoRA (MELO), which alters the behavior of language models by dynamically activating certain LoRA blocks according to the index built in an inner vector database. Our method satisfies various editing properties with high efficiency and can be easily integrated into multiple LLM backbones. Experimental results show that our proposed MELO achieves state-of-the-art editing performance on three sequential editing tasks (document classification, question answering and hallucination correction), while requires the least trainable parameters and computational cost.
Automated construction of annotated data holds significant importance in Relation Extraction (RE) tasks due to the hardness and cost of human annotation. In this work, we propose Self-RDGS, a method for Self-supervised Reliable Data Generation and Selection in low-resource RE tasks. At first, we fully utilize the knowledge of triplets as prompts to generate sentences by employing the Large Language Models (LLMs). Since the auto-generated data contains noise, we then propose a ranking-based data selection method to select reliable sentences. Finally, we integrate the data selection and RE model training within a self-supervised iterative framework. Through experimentation on three datasets with low-resource settings, we demonstrate the effectiveness of our proposed approach in constructing annotated data and achieving noteworthy improvements in comparison to multiple baselines. Code, data and models are available at https://github.com/jjyunlp/GenerationRE.