EMNLP 2024 论文列表

Language models are widely used in computational psycholinguistics to test theories that relate the negative log probability (the surprisal) of a region of interest (a substring of characters) under a language model to its cognitive cost experienced by readers, as operationalized, for example, by gaze duration on the region. However, the application of modern language models to psycholinguistic studies is complicated by the practice of using tokenization as an intermediate step in training a model. Doing so results in a language model over *token* strings rather than one over character strings. Vexingly, regions of interest are generally misaligned with these token strings. The paper argues that token-level language models should be (approximately) marginalized into character-level language models before they are used in psycholinguistic studies to compute the surprisal of a region of interest; then, the marginalized character-level language model can be used to compute the surprisal of an arbitrary character substring, which we term a focal area, that the experimenter may wish to use as a predictor. Our proposal of marginalizing a token-level model into a character-level one solves this misalignment issue independently of the tokenization scheme. Empirically, we discover various focal areas whose surprisal is a better psychometric predictor than the surprisal of the region of interest itself.

Reinforcement Learning from Human Feedback (RLHF) is an effective approach for aligning language models to human preferences. Central to RLHF is learning a reward function for scoring human preferences. Two main approaches for learning a reward model are 1) training an EXplicit Reward Model (EXRM) as in RLHF, and 2) using an implicit reward learned from preference data through methods such as Direct Preference Optimization (DPO). Prior work has shown that the implicit reward model of DPO (denoted as DPORM) can approximate an EXRM on the limit infinite samples. However, it is unclear how effective is DPORM in practice. DPORM’s effectiveness directly implies the optimality of learned policy of DPO and also has practical implication for more advanced alignment methods, such as iterative DPO. We compare the accuracy at distinguishing preferred and rejected answers using both DPORM and EXRM. Our findings indicate that even though DPORM can fit the training dataset, it generalizes less effective than EXRM, especially when the validation datasets contain distributional shifts. Across five out-of-distribution settings, DPORM has a mean drop in accuracy of 3% and a maximum drop of 7%. These findings highlight that DPORM has limited generalization ability and substantiates the integration of an explicit reward model in iterative DPO approaches.

We study the presence of heteronormative biases and prejudice against interracial romantic relationships in large language models by performing controlled name-replacement experiments for the task of relationship prediction. We show that models are less likely to predict romantic relationships for (a) same-gender character pairs than different-gender pairs; and (b) intra/inter-racial character pairs involving Asian names as compared to Black, Hispanic, or White names. We examine the contextualized embeddings of first names and find that gender for Asian names is less discernible than non-Asian names. We discuss the social implications of our findings, underlining the need to prioritize the development of inclusive and equitable technology.

Large language models (LLMs) are found to have the ability of in-context generation (ICG): when they are fed with an in-context prompt concatenating a few somehow similar examples, they can implicitly recognize the pattern of them and then complete the prompt in the same pattern. ICG is curious, since language models are usually not explicitly trained in the same way as the in-context prompt, and the distribution of examples in the prompt differs from that of sequences in the pretrained corpora. This paper provides a systematic study of the ICG ability of language models, covering discussions about its source and influential factors, in the view of both theory and empirical experiments. Concretely, we first propose a plausible latent variable model to model the distribution of the pretrained corpora, and then formalize ICG as a problem of next topic prediction. With this framework, we can prove that the repetition nature of a few topics ensures the ICG ability on them theoretically. Then, we use this controllable pretrained distribution to generate several medium-scale synthetic datasets (token scale: 2.1B-3.9B) and experiment with different settings of Transformer architectures (parameter scale: 4M-234M). Our experimental results further offer insights into how the data and model architectures influence ICG.

Large language models, such as ChatGPT, achieve amazing performance on various language processing tasks. However, they can also be exploited for improper purposes such as plagiarism or misinformation dissemination. Thus, there is an urgent need to detect the texts generated by LLMs. One type of most studied methods trains classification models to distinguish LLM texts from human texts. However, existing studies demonstrate the trained models may suffer from distribution shifts (during test), i.e., they are ineffective to predict the generated texts from unseen language tasks or topics which are not collected during training. In this work, we focus on ChatGPT as a representative model, and we conduct a comprehensive investigation on these methods’ generalization behaviors under distribution shift caused by a wide range of factors, including prompts, text lengths, topics, and language tasks. To achieve this goal, we first collect a new dataset with human and ChatGPT texts, and then we conduct extensive studies on the collected dataset. Our studies unveil insightful findings that provide guidance for future methodologies and data collection strategies for LLM detection.

Active learning (AL) techniques optimally utilize a labeling budget by iteratively selecting instances that are most valuable for learning. However, they lack “prerequisite checks”, i.e., there are no prescribed criteria to pick an AL algorithm best suited for a dataset. A practitioner must pick a technique they trust would beat random sampling, based on prior reported results, and hope that it is resilient to the many variables in their environment: dataset, labeling budget and prediction pipelines. The important questions then are: how often on average, do we expect any AL technique to reliably beat the computationally cheap and easy-to-implement strategy of random sampling? Does it at least make sense to use AL in an “Always ON” mode in a prediction pipeline, so that while it might not always help, it never under-performs random sampling? How much of a role does the prediction pipeline play in AL’s success?We examine these questions in detail for the task of text classification using pre-trained representations, which are ubiquitous today.Our primary contribution here is a rigorous evaluation of AL techniques, old and new, across setups that vary wrt datasets, text representations and classifiers. This unlocks multiple insights around warm-up times, i.e., number of labels before gains from AL are seen, viability of an “Always ON” mode and the relative significance of different factors.Additionally, we release a framework for rigorous benchmarking of AL techniques for text classification.

Chain-of-thought (CoT), tree-of-thought (ToT), and related techniques work surprisingly well in practice for some complex reasoning tasks with Large Language Models (LLMs), but why? This work seeks the underlying reasons by conducting experimental case studies and linking the performance benefits to well-established sample and computational complexity principles in machine learning. We experimented with six reasoning tasks, ranging from grade school math, air travel planning, ..., to Blocksworld. The results suggest that (i) both CoT and ToT benefit significantly from task decomposition, which breaks a complex reasoning task into a sequence of steps with low sample complexity and explicitly outlines the reasoning structure; (ii) for computationally hard reasoning tasks, the more sophisticated tree structure of ToT outperforms the linear structure of CoT; (iii) explicitly annotating important variables is important for good performance. These findings provide useful guidelines for using LLM in solving reasoning tasks in practice.

Amidst the rapid advancements in generative language models, the investigation of how training data shapes the performance of GPT models is still emerging. This paper presents GPTfluence, a novel approach that leverages a featurized simulation to assess the impact of training examples on the training dynamics of GPT models. Our approach not only traces the influence of individual training instances on performance trajectories, such as loss and other key metrics, on targeted test points but also enables a comprehensive comparison with existing methods across various training scenarios in GPT models, ranging from 14 million to 2.8 billion parameters, across a range of downstream tasks. Contrary to earlier methods that struggle with generalization to new data, GPTfluence introduces a parameterized simulation of training dynamics, demonstrating robust generalization capabilities to unseen training data. This adaptability is evident across both fine-tuning and instruction-tuning scenarios, spanning tasks in natural language understanding and generation. We make our code and data publicly available at https://github.com/ernie-research/gptfluence.

While learning with limited labelled data can effectively deal with a lack of labels, it is also sensitive to the effects of uncontrolled randomness introduced by so-called randomness factors (i.e., non-deterministic decisions such as choice or order of samples). We propose and formalise a method to systematically investigate the effects of individual randomness factors while taking the interactions (dependence) between them into consideration. To this end, our method mitigates the effects of other factors while observing how the performance varies across multiple runs. Applying our method to multiple randomness factors across in-context learning and fine-tuning approaches on 7 representative text classification tasks and meta-learning on 3 tasks, we show that: 1) disregarding interactions between randomness factors in existing works led to inconsistent findings due to incorrect attribution of the effects of randomness factors, such as disproving the consistent sensitivity of in-context learning to sample order even with random sample selection; and 2) besides mutual interactions, the effects of randomness factors, especially sample order, are also dependent on more systematic choices unexplored in existing works, such as number of classes, samples per class or choice of prompt format.

How far have we come in mitigating performance disparities across genders in multilingual speech recognition? We compare the impact on gender disparity of different fine-tuning algorithms for automated speech recognition across model sizes, languages and gender. We look at both performance-focused and fairness-promoting algorithms. Across languages, we see slightly better performance for female speakers for larger models regardless of the fine-tuning algorithm. The best trade-off between performance and parity is found using adapter fusion. Fairness-promoting fine-tuning algorithms (Group-DRO and Spectral Decoupling) hurt performance compared to adapter fusion with only slightly better performance parity. LoRA increases disparities slightly. Fairness-mitigating fine-tuning techniques led to slightly higher variance in performance across languages, with the exception of adapter fusion.

In this paper, we consider contamination by code generation test sets, in particular in their use in modern large language models.We discuss three possible sources of such contamination and show findings supporting each of them: (i) direct data leakage, (ii) indirect data leakage through the use of synthetic data and (iii) overfitting to evaluation sets during model selection.To address this, we release Less Basic Python Problems (LBPP): an uncontaminated new benchmark of 161 prompts with their associated Python solutions. LBPP is released at https://huggingface.co/datasets/CohereForAI/lbpp

Large language models (LLMs) have emerged as valuable tools for enhancing textual features in various text-related tasks. Despite their superiority in capturing the lexical semantics between tokens for text analysis, our preliminary study on two popular LLMs, i.e., ChatGPT and Llama2, showcases that simply applying the news embeddings from LLMs is ineffective for fake news detection. Such embeddings only encapsulate the language styles between tokens. Meanwhile, the high-level semantics among named entities and topics, which reveal the deviating patterns of fake news, have been ignored. Therefore, we propose a topic model together with a set of specially designed prompts to extract topics and real entities from LLMs and model the relations among news, entities, and topics as a heterogeneous graph to facilitate investigating news semantics. We then propose a Generalized Page-Rank model and a consistent learning criteria for mining the local and global semantics centered on each news piece through the adaptive propagation of features across the graph. Our model shows superior performance on five benchmark datasets over seven baseline methods and the efficacy of the key ingredients has been thoroughly validated.

Is explainability a false promise? This debate has emerged from the insufficient evidence that explanations help people in situations they are introduced for. More human-centered, application-grounded evaluations of explanations are needed to settle this. Yet, with no established guidelines for such studies in NLP, researchers accustomed to standardized proxy evaluations must discover appropriate measurements, tasks, datasets, and sensible models for human-AI teams in their studies. To aid with this, we first review existing metrics suitable for application-grounded evaluation. We then establish criteria to select appropriate datasets, and using them, we find that only 4 out of over 50 datasets available for explainability research in NLP meet them. We then demonstrate the importance of reassessing the state of the art to form and study human-AI teams: teaming people with models for certain tasks might only now start to make sense, and for others, it remains unsound. Finally, we present the exemplar studies of human-AI decision-making for one of the identified tasks — verifying the correctness of a legal claim given a contract. Our results show that providing AI predictions, with or without explanations, does not cause decision makers to speed up their work without compromising performance. We argue for revisiting the setup of human-AI teams and improving automatic deferral of instances to AI, where explanations could play a useful role.

Unsupervised parsing, also known as grammar induction, aims to infer syntactic structure from raw text. Recently, binary representation has exhibited remarkable information-preserving capabilities at both lexicon and syntax levels. In this paper, we explore the possibility of leveraging this capability to deduce parsing trees from raw text, relying solely on the implicitly induced grammars within models. To achieve this, we upgrade the bit-level CKY from zero-order to first-order to encode the lexicon and syntax in a unified binary representation space, switch training from supervised to unsupervised under the contrastive hashing framework, and introduce a novel loss function to impose stronger yet balanced alignment signals. Our model shows competitive performance on various datasets, therefore, we claim that our method is effective and efficient enough to acquire high-quality parsing trees from pre-trained language models at a low cost.

Recent advancements in language and vision assistants have showcased impressive capabilities but suffer from a lack of transparency, limiting broader research and reproducibility. While open-source models handle general image tasks effectively, they face challenges with the high computational demands of complex visually-situated text understanding. Such tasks often require increased token inputs and large vision modules to harness high-resolution information. Striking a balance between model size and data importance remains an open question. This study aims to redefine the design of vision-language models by identifying key components and creating efficient models with constrained inference costs. By strategically formulating datasets, optimizing vision modules, and enhancing supervision techniques, we achieve significant improvements in inference throughput while maintaining high performance. Extensive experiments across models ranging from 160M to 13B parameters offer insights into model optimization.We will fully open-source our codebase, models, and datasets at https://github.com/naver-ai/elva.

Aligning large language models (LLMs) with human preferences has been recognized as the key to improving LLMs’ interaction quality. However, in this pluralistic world, human preferences can be diversified due to annotators’ different tastes, which hinders the effectiveness of LLM alignment methods. This paper presents the first quantitative analysis of the experimental scaling law for reward models with varying sizes, from 1.3 billion to 7 billion parameters, trained with human feedback exhibiting diverse preferences. Our analysis reveals that the impact of diversified human preferences depends on both model size and data size. Larger models with sufficient capacity mitigate the negative effects of diverse preferences, while smaller models struggle to accommodate them. To mitigate the impact of diverse preferences, we introduce a new metric, Expected Calibration Error (ECE), to evaluate RMs and show their obvious positive correlation with the alignment performance of LLMs. Furthermore, we propose a Multi-Objective Reward learning method (MORE) to enhance the calibration performance of RMs on shared preferences. Through experiments on four models and five human preference datasets, we find the calibration error can be adopted as a key metric for evaluating RMs and MORE can obtain superior alignment performance.

Machine translation (MT) in the medical domain plays a pivotal role in enhancing healthcare quality and disseminating medical knowledge. Despite advancements in English-Thai MT technology, common MT approaches often underperform in the medical field due to their inability to precisely translate medical terminologies. Our research prioritizes not merely improving translation accuracy but also maintaining medical terminology in English within the translated text through code-switched (CS) translation. We developed a method to produce CS medical translation data, fine-tuned a CS translation model with this data, and evaluated its performance against strong baselines, such as Google Neural Machine Translation (NMT) and GPT-3.5/GPT-4. Our model demonstrated competitive performance in automatic metrics and was highly favored in human preference evaluations. Our evaluation result also shows that medical professionals significantly prefer CS translations that maintain critical English terms accurately, even if it slightly compromises fluency. Our code and test set are publicly available https://github.com/preceptorai-org/NLLB_CS_EM_NLP2024.

Recent advancements in Large Language Models (LLMs) have expanded their capabilities to multimodal contexts, including comprehensive video understanding. However, processing extensive videos such as 24-hour CCTV footage or full-length films presents significant challenges due to the vast data and processing demands. Traditional methods, like extracting key frames or converting frames to text, often result in substantial information loss. To address these shortcomings, we develop OmAgent, efficiently stores and retrieves relevant video frames for specific queries, preserving the detailed content of videos. Additionally, it features an Divide-and-Conquer Loop capable of autonomous reasoning, dynamically invoking APIs and tools to enhance query processing and accuracy. This approach ensures robust video understanding, significantly reducing information loss. Experimental results affirm OmAgent’s efficacy in handling various types of videos and complex tasks. Moreover, we have endowed it with greater autonomy and a robust tool-calling system, enabling it to accomplish even more intricate tasks.

Employing Large Language Models (LLMs) to assess the quality of generated responses has become a widely adopted evaluation method. Specifically, instruct-tuned models and fine-tuned judge models based on open-source LLMs have been reported. While it is known that judge models are vulnerable to certain biases, such as favoring longer answers regardless of content, the specifics of these biases remain under-explored. In this work, we qualitatively identify six types of biases inherent in various judge models. We propose EvalBiasBench as a meta-evaluation collection of hand-crafted test cases for each bias type. Additionally, we present de-biasing dataset construction methods and the associated preference dataset OffsetBias. Experimental results demonstrate that fine-tuning on our dataset significantly enhances the robustness of judge models against biases and improves performance across most evaluation scenarios. We release our datasets and the fine-tuned judge model to public.