Neural Machine Translation models are extremely data and compute-hungry. However, not all datapoints contribute equally to model training and generalization. Data pruning to remove the low-value data points has the benefit of drastically reducing the compute budget without significantdrop in model performance. In this paper, we propose a new data pruning technique: CheckpointsAcross Time (CAT ), that leverages early model training dynamics to identify the most relevantdata points for model performance. We benchmark CAT against several data pruning techniquesincluding COMET-QE, LASER and LaBSE. We find that CAT outperforms the benchmarks onIndo-European languages on multiple test sets. When applied to English-German, English-Frenchand English-Swahili translation tasks, CAT achieves comparable performance to using the fulldataset, while pruning up to 50% of training data. We inspect the data points that CAT selectsand find that it tends to favour longer sentences and sentences with unique or rare words.
论文检索
输入标题、作者或关键词，从 9,460 篇学术成果中精准定位
Deciphering the Impact of Pretraining Data on Large Language Models through Machine Unlearning
PDF ↗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.
Large language models (LLMs) are notorious for hallucinating, i.e., producing erroneous claims in their output. Such hallucinations can be dangerous, as occasional factual inaccuracies in the generated text might be obscured by the rest of the output being generally factually correct, making it extremely hard for the users to spot them. Current services that leverage LLMs usually do not provide any means for detecting unreliable generations. Here, we aim to bridge this gap. In particular, we propose a novel fact-checking and hallucination detection pipeline based on token-level uncertainty quantification. Uncertainty scores leverage information encapsulated in the output of a neural network or its layers to detect unreliable predictions, and we show that they can be used to fact-check the atomic claims in the LLM output. Moreover, we present a novel token-level uncertainty quantification method that removes the impact of uncertainty about what claim to generate on the current step and what surface form to use. Our method Claim Conditioned Probability (CCP) measures only the uncertainty of a particular claim value expressed by the model. Experiments on the task of biography generation demonstrate strong improvements for CCP compared to the baselines for seven different LLMs and four languages. Human evaluation reveals that the fact-checking pipeline based on uncertainty quantification is competitive with a fact-checking tool that leverages external knowledge.
Open-sourced Large Language Models (LLMs) have achieved great success in various NLP tasks, however, they are still far inferior to API-based models when acting as agents. How to integrate agent ability into general LLMs becomes a crucial and urgent problem.This paper first delivers three key observations: (1) the current agent training corpus is entangled with both formats following and agent reasoning, which significantly shifts from the distribution of its pre-training data; (2) LLMs exhibit different learning speeds on the capabilities required by agent tasks; and (3) current approaches have side-effects when improving agent abilities by introducing hallucinations. Based on the above findings, we propose Agent-FLAN to effectively Fine-tune LANguage models for Agents.Through careful decomposition and redesign of the training corpus, Agent-FLAN enables Llama2-7B to outperform prior best works by 3.5% across various agent evaluation datasets. With comprehensively constructed negative samples, Agent-FLAN greatly alleviates the hallucination issues based on our established evaluation benchmark. Besides, it consistently improves the agent capability of LLMs when scaling model sizes while slightly enhancing the general capability of LLMs. The code and models are available at https://github.com/InternLM/Agent-FLAN.
Fact verification aims to automatically probe the veracity of a claim based on several pieces of evidence. Existing works are always engaging in accuracy improvement, let alone explainability, a critical capability of fact verification systems.Constructing an explainable fact verification system in a complex multi-hop scenario is consistently impeded by the absence of a relevant, high-quality dataset. Previous datasets either suffer from excessive simplification or fail to incorporate essential considerations for explainability. To address this, we present EX-FEVER, a pioneering dataset for multi-hop explainable fact verification. With over 60,000 claims involving 2-hop and 3-hop reasoning, each is created by summarizing and modifying information from hyperlinked Wikipedia documents. Each instance is accompanied by a veracity label and an explanation that outlines the reasoning path supporting the veracity classification. Additionally, we demonstrate a novel baseline system on our EX-FEVER dataset, showcasing document retrieval, explanation generation, and claim verification, and validate the significance of our dataset. Furthermore, we highlight the potential of utilizing Large Language Models in the fact verification task. We hope our dataset could make a significant contribution by providing ample opportunities to explore the integration of natural language explanations in the domain of fact verification.
We present Q-ViD, a simple approach for video question answering (video QA), that unlike prior methods, which are based on complex architectures, computationally expensive pipelines or use closed models like GPTs, Q-ViD relies on a single instruction-aware open vision-language model (InstructBLIP) to tackle videoQA using frame descriptions. Specifically, we create captioning instruction prompts that rely on the target questions about the videos and leverage InstructBLIP to obtain video frame captions that are useful to the task at hand. Subsequently, we form descriptions of the whole video using the question-dependent frame captions, and feed that information, along with a question-answering prompt, to a large language model (LLM). The LLM is our reasoning module, and performs the final step of multiple-choice QA. Our simple Q-ViD framework achieves competitive or even higher performances than current state of the art models on a diverse range of videoQA benchmarks, including NExT-QA, STAR, How2QA, TVQA and IntentQA.
Dialects introduce syntactic and lexical variations in language that occur in regional or social groups. Most NLP methods are not sensitive to such variations. This may lead to unfair behavior of the methods, conveying negative bias towards dialect speakers. While previous work has studied dialect-related fairness for aspects like hate speech, other aspects of biased language, such as lewdness, remain fully unexplored. To fill this gap, we investigate performance disparities between dialects in the detection of five aspects of biased language and how to mitigate them. To alleviate bias, we present a multitask learning approach that models dialect language as an auxiliary task to incorporate syntactic and lexical variations. In our experiments with African-American English dialect, we provide empirical evidence that complementing common learning approaches with dialect modeling improves their fairness. Furthermore, the results suggest that multitask learning achieves state-of-the-art performance and helps to detect properties of biased language more reliably.
Recently, Knowledge Editing has received increasing attention, since it could update the specific knowledge from outdated ones in pretrained models without re-training. However, as pointed out by recent studies, existing related methods tend to merely memorize the superficial word composition of the edited knowledge, rather than truly learning and absorbing it. Consequently, on the reasoning questions, we discover that existing methods struggle to utilize the edited knowledge to reason the new answer, and tend to retain outdated responses, which are generated by the original models utilizing original knowledge. Nevertheless, the outdated responses are unexpected for the correct answers to reasoning questions, which we named as the outdated issue. To alleviate this issue, in this paper, we propose a simple yet effective decoding strategy, i.e., outDated ISsue aware deCOding (DISCO), to enhance the performance of edited models on reasoning questions. Specifically, we capture the difference in the probability distribution between the original and edited models. Further, we amplify the difference of the token prediction in the edited model to alleviate the outdated issue, and thus enhance the model performance w.r.t the edited knowledge. Experimental results suggest that applying DISCO could enhance edited models to reason, e.g., on reasoning questions, DISCO outperforms the prior SOTA method by 12.99 F1 scores, and reduces the ratio of the outdated issue to 5.78% on the zsRE dataset.
Current fact verification methods generally follow the two-stage training paradigm: evidence retrieval and claim verification. While existing works focus on developing sophisticated claim verification modules, the fundamental importance of evidence retrieval is largely ignored. Existing approaches usually adopt the heuristic semantic similarity-based retrieval strategy, resulting in the task-irrelevant evidence and undesirable performance. In this paper, we concentrate on evidence retrieval and propose a Retrieval-Augmented Verification framework RAV, consisting of two major modules: the hybrid evidence retrieval and the joint fact verification. Hybrid evidence retrieval module incorporates an efficient retriever for preliminary pruning of candidate evidence, succeeded by a ranker that generates more precise sorting results. Under this end-to-end training paradigm, gradients from the claim verification can be back-propagated to enhance evidence selection. Experimental results on FEVER dataset demonstrate the superiority of RAV.
Pre-trained language models (PLMs) have shown great dialogue generation capability in different scenarios. However, the huge VRAM consumption when fine-tuning them is one of their drawbacks. PEFT approaches can significantly reduce the number of trainable parameters, which enables us to fine-tune larger dialogue generation models. However, the reduction in parameter quantity can diminish a PLM’s expressive capacity and affect the PLM’s learning from certain specific examples like knowledge-related conversations. Previous works have demonstrated that injecting external knowledge into dialogue generation models can improve the model’s performance in knowledge-related conversations. Nonetheless, these methods are designed for the scenario where most parameters of the entire framework are trainable. In this paper, we propose PEK, a parameter-efficient framework for knowledge-enhanced dialogue generation. It enables PLMs to leverage external knowledge documents and knowledge graphs to enhance its generation capabilities with an acceptable number of trainable parameters. Evaluation results on the Wizard of Wikipedia and CMU_DoG datasets show that our approach outperforms baseline methods on multiple evaluation metrics, which validates the effectiveness of our approach.
Large language models (LLMs) possess immense capabilities but are susceptible to malicious exploitation. To mitigate the risk, safety alignment is employed to align LLMs with ethical standards. However, safety-aligned LLMs may remain vulnerable to carefully crafted jailbreak attacks, but these attacks often face high rejection rates and limited harmfulness. In this paper, we expose the vulnerabilities of safety alignment in open-access LLMs, which can significantly enhance the success rate and harmfulness of jailbreak attacks. Through reverse alignment, achieved by accessing model parameters, we show the feasibility of efficiently fine-tuning LLMs to undermine their inherent safeguards. We investigate two types of reverse alignment techniques: reverse supervised fine-tuning (RSFT) and reverse preference optimization (RPO). RSFT operates by supervising the fine-tuning of LLMs to reverse their inherent values. We also explore how to prepare data needed for RSFT. RPO optimizes LLMs to enhance their preference for harmful content, reversing the models’ safety alignment. Our extensive experiments reveal that open-access high-performance LLMs can be adeptly reverse-aligned to output harmful content, even in the absence of manually curated malicious datasets. Our research acts as a whistleblower for the community, emphasizing the need to pay more attention to safety of open-accessing LLMs. It also underscores the limitations of current safety alignment approaches and calls for research on robust safety alignment methods to counteract malicious fine-tuning attacks.
Are LLMs Capable of Data-based Statistical and Causal Reasoning? Benchmarking Advanced Quantitative Reasoning with Data
PDF ↗Quantitative reasoning is a critical skill to analyze data, yet the assessment of such ability remains limited. To address this gap, we introduce the Quantitative Reasoning with Data (QRData) benchmark, aiming to evaluate Large Language Models’ capability in statistical and causal reasoning with real-world data. The benchmark comprises a carefully constructed dataset of 411 questions accompanied by data sheets from textbooks, online learning materials, and academic papers. To compare models’ quantitative reasoning abilities on data and text, we enrich the benchmark with an auxiliary set of 290 text-only questions, namely QRText. We evaluate natural language reasoning, program-based reasoning, and agent reasoning methods including Chain-of-Thought, Program-of-Thoughts, ReAct, and code interpreter assistants on diverse models. The strongest model GPT-4 achieves an accuracy of 58%, which has much room for improvement. Among open-source models, Deepseek-coder-instruct, a code LLM pretrained on 2T tokens, gets the highest accuracy of 37%. Analysis reveals that models encounter difficulties in data analysis and causal reasoning, and struggle in using causal knowledge and provided data simultaneously. Code and data are in https://github.com/xxxiaol/QRData.
Multilingual neural machine translation models generally distinguish translation directions by the language tag (LT) in front of the source or target sentences. However, current LT strategies cannot indicate the desired target language as expected on zero-shot translation, i.e., the off-target issue. Our analysis reveals that the indication of the target language is sensitive to the placement of the target LT. For example, when placing the target LT on the decoder side, the indication would rapidly degrade along with decoding steps, while placing the target LT on the encoder side would lead to copying or paraphrasing the source input. To address the above issues, we propose a simple yet effective strategy named Language Converter Strategy (LCS). By introducing the target language embedding into the top encoder layers, LCS mitigates confusion in the encoder and ensures stable language indication for the decoder. Experimental results on MultiUN, TED, and OPUS-100 datasets demonstrate that LCS could significantly mitigate the off-target issue, with language accuracy up to 95.28%, 96.21%, and 85.35% meanwhile outperforming the vanilla LT strategy by 3.07, 3,3, and 7.93 BLEU scores on zero-shot translation, respectively.
Cross-lingual chain-of-thought can effectively complete reasoning tasks across languages, which gains increasing attention.Recently, dominant approaches in the literature improve cross-lingual alignment capabilities by integrating reasoning knowledge from different languages. Despite achieving excellent performance, current methods still have two main challenges: (1) Manual language specification: They still highly rely on manually selecting the languages to integrate, severely affecting their generalizability; (2) Static weight allocation: Current methods simply integrate all languages equally. In fact, different language reasoning paths should have different weights to achieve better complementation and integration. Motivated by this, we introduce an Automatic Cross-lingual Alignment Planning (AutoCAP) for zero-shot chain-of-thought to address the above challenges. The core of AutoCAP consists of two components: (1) Automatic Language Selection Prompting to guide LLMs to select appropriate languages and (2) Automatic Weight Allocation Prompting to automatically allocate alignment weight scores to each reasoning path. Extensive experiments on several benchmarks reveal that AutoCAP achieves state-of-the-art performance, surpassing previous methods that required manual effort.
Information extraction plays a critical role in natural language processing. When applying large language models (LLMs) to this domain, we discover an unexpected phenomenon: LLMs’ spurious associations. In tasks such as relation extraction, LLMs can accurately identify entity pairs, even if the given relation (label) is semantically unrelated to the pre-defined original one. To find these labels, we design two strategies in this study, including forward label extension and backward label validation. We also leverage the extended labels to improve model performance. Our comprehensive experiments show that spurious associations occur consistently in both Chinese and English datasets across various LLM sizes. Moreover, the use of extended labels significantly enhances LLM performance in information extraction tasks. Remarkably, there is a performance increase of 9.55%, 11.42%, and 21.27% in F1 scores on the SciERC, ACE05, and DuEE datasets, respectively.
RRNorm: A Novel Framework for Chinese Disease Diagnoses Normalization via LLM-Driven Terminology Component Recognition and Reconstruction
PDF ↗The Clinical Terminology Normalization aims at finding standard terms from a given termbase for mentions extracted from clinical texts. However, we found that extracted mentions suffer from the multi-implication problem, especially disease diagnoses. The reason for this is that physicians often use abbreviations, conjunctions, and juxtapositions when writing diagnoses, and it is difficult to manually decompose. To address this problem, we propose a Terminology Component Recognition and Reconstruction strategy that leverages the reasoning capability of large language models (LLMs) to recognize the components of terms, enabling automated decomposition and transforming original mentions into multiple atomic mentions. Furthermore, we adopt the mainstream “Recall and Rank” framework to apply the benefits of the above strategy to the task flow. By leveraging the LLM incorporating the advanced sampling strategies, we design a sampling algorithm for atomic mentions and train the recall model using contrastive learning. Besides the information about the components is also used as knowledge to guide the final term ranking and selection. The experimental results show that our proposed strategy effectively improves the performance of the terminology normalization task and our proposed approach achieves state-of-the-art on the experimental dataset. We release our code and data on the repository https://github.com/yuugaochyan/RRNorm.
Direct speech-to-speech translation (S2ST) has achieved impressive translation quality, but it often faces the challenge of slow decoding due to the considerable length of speech sequences. Recently, some research has turned to non-autoregressive (NAR) models to expedite decoding, yet the translation quality typically lags behind autoregressive (AR) models significantly. In this paper, we investigate the performance of CTC-based NAR models in S2ST, as these models have shown impressive results in machine translation. Experimental results demonstrate that by combining pretraining, knowledge distillation, and advanced NAR training techniques such as glancing training and non-monotonic latent alignments, CTC-based NAR models achieve translation quality comparable to the AR model, while preserving up to 26.81\times decoding speedup.
One major challenge for Large Language Models (LLMs) is completing complex tasks involving multiple entities, such as tool APIs. To tackle this, one approach is to retrieve relevant entities to enhance LLMs in task completion. A crucial issue here is obtaining accurate natural language representations for each entity to aid in retriever precision. In this paper, we propose the Natural Language Representation Optimization Problem, which aims to refine entity descriptions for improved retrieval and LLM utilization. We introduce the Learning to Represent with Natural Language method, which utilizes LLMs to optimize entity representations consisting of text patterns based on environmental feedback. We iteratively prompt LLMs to enhance or adjust patterns based on entity samples and evaluate their effectiveness through environmental feedback. Our method successfully learns human-understandable representations for classification tasks (e.g., instructions and documents) and API call tasks (e.g., APIbench and Virtual Home), significantly improving GPT-4’s task performance.
Empowering Large Language Models (LLMs) with distinct human-like personality traits has become an innovative task for developing advanced dialog systems.Although LLMs demonstrate impressive capabilities in following instructions, directly prompting them to exhibit certain personalities through manually crafted instructions may result in sub-optimal performance.In this paper, we propose a plug-and-play prompting method to manipulate the LLMs’ personality traits.Specifically, we append discrete personalized suffixes, automatically generated through an aggregated gradient-based search method, to the user query or dialog histories and induce LLMs to respond with target personalities.In addition, due to the high redundancy of the search space, we adopt a reward-based strategy to prune the vocabulary and focus exclusively on influential tokens.Experiment results on four models ranging from 1.1B to 13B show that our method achieves 79.9% accuracy in customizing LLMs’ personalities, significantly outperforming other prompting methods (65.5%) and model editing methods.Our method also excels in generation fluency and quality with the lowest generation perplexity and the highest GPT-4 evaluation scores.