NAACL 2024 论文列表

The improvement of LLMs’ instruction-following capabilities relies heavily on the availability of high-quality instruction-response pairs. Unfortunately, the current methods used to collect the pairs suffer from either unaffordable labor costs or severe hallucinations in the self-generation of LLM.To tackle these challenges, this paper proposes a scalable solution.It involves training LLMs to generate instruction-response pairs based on human-written documents, rather than relying solely on self-generation without context.Our proposed method not only exploits the advantages of human-written documents in reducing hallucinations but also utilizes an LLM to wrap the expression of documents, which enables us to bridge the gap between various document styles and the standard AI response.Experiments demonstrate that our method outperforms existing typical methods on multiple benchmarks.In particular, compared to the best-performing baseline, the LLM trained using our generated dataset exhibits a 10% relative improvement in performance on AlpacaEval, despite utilizing only 1/5 of its training data.Furthermore, a comprehensive manual evaluation validates the quality of the data we generated.

Open-vocabulary vision-language models (VLMs) like CLIP, trained using contrastive loss, have emerged as a promising new paradigm for text-to-image retrieval. However, do VLMs understand compound nouns (CNs) (e.g., *lab coat*) as well as they understand nouns (e.g., *lab*)? We curate Compun, a novel benchmark with 400 unique and commonly used CNs, to evaluate the effectiveness of VLMs in interpreting CNs. The Compun benchmark challenges a VLM for text-to-image retrieval where, given a text prompt with a CN, the task is to select the correct image that shows the CN among a pair of distractor images that show the constituent nouns that make up the CN. Next, we perform an in-depth analysis to highlight CLIPs’ limited understanding of certain types of CNs. Finally, we present an alternative framework that moves beyond hand-written templates for text prompts widely used by CLIP-like models. We employ a Large Language Model to generate multiple diverse captions that include the CN as an object in the scene described by the caption. Our proposed method improves CN understanding of CLIP by 8.25% on Compun. Code and benchmark are available.

Prompt-based models have gathered a lot of attention from researchers due to their remarkable advancements in the fields of zero-shot and few-shot learning. Developing an effective prompt template plays a critical role. However, prior studies have mainly focused on prompt vocabulary searching or embedding initialization within a predefined template with the prompt position fixed. In this empirical study, we conduct the most comprehensive analysis to date of prompt position for diverse Natural Language Processing (NLP) tasks. Our findings quantify the substantial impact prompt position has on model performance. We observe that the prompt positions used in prior studies are often sub-optimal, and this observation is consistent even in widely used instruction-tuned models. These findings suggest prompt position optimisation as a valuable research direction to augment prompt engineering methodologies and prompt position-aware instruction tuning as a potential way to build more robust models in the future.

Translate-test is a popular technique to improve the performance of multilingual language models. This approach works by translating the input into English using an external machine translation system before running inference. However, these improvements can be attributed to the use of a separate translation system, which is typically trained on large amounts of parallel data not seen by the language model. In this work, we introduce a new approach called self-translate that leverages the few-shot translation capabilities of multilingual language models. This allows us to analyze the effect of translation in isolation. Experiments over 5 tasks show that self-translate consistently outperforms direct inference, demonstrating that language models are unable to leverage their full multilingual potential when prompted in non-English languages. Our code is available at https://github.com/juletx/self-translate.

The concept of localization in LLMs is often mentioned in prior work; however, methods for localization have never been systematically and directly evaluated. We propose two complementary benchmarks that evaluate the ability of localization methods to pinpoint LLM components responsible for memorized data. In our INJ benchmark, we actively inject a piece of new information into a small subset of LLM weights, enabling us to directly evaluate whether localization methods can identify these “ground truth” weights. In our DEL benchmark, we evaluate localization by measuring how much dropping out identified neurons deletes a memorized pretrained sequence. Despite their different perspectives, our two benchmarks yield consistent rankings of five localization methods. Methods adapted from network pruning perform well on both benchmarks, and all evaluated methods show promising localization ability. On the other hand, even successful methods identify neurons that are not specific to a single memorized sequence.

The integration of Large Language Models (LLMs) in information retrieval has raised a critical reevaluation of fairness in the text-ranking models. LLMs, such as GPT models and Llama2, have shown effectiveness in natural language understanding tasks, and prior works such as RankGPT have demonstrated that the LLMs have better performance than the traditional ranking models in the ranking task. However, their fairness remains largely unexplored. This paper presents an empirical study evaluating these LLMs using the TREC Fair Ranking dataset, focusing on the representation of binary protected attributes such as gender and geographic location, which are historically underrepresented in search outcomes. Our analysis delves into how these LLMs handle queries and documents related to these attributes, aiming to uncover biases in their ranking algorithms. We assess fairness from both user and content perspectives, contributing an empirical benchmark for evaluating LLMs as the fair ranker.

Social media data has been used for detecting users with mental disorders, such as depression. Despite the global significance of cross-cultural representation and its potential impact on model performance, publicly available datasets often lack crucial metadata relatedto this aspect. In this work, we evaluate the generalization of benchmark datasets to build AI models on cross-cultural Twitter data. We gather a custom geo-located Twitter dataset of depressed users from seven countries as a test dataset. Our results show that depressiondetection models do not generalize globally. The models perform worse on Global South users compared to Global North. Pre-trainedlanguage models achieve the best generalization compared to Logistic Regression, though still show significant gaps in performance on depressed and non-Western users. We quantify our findings and provide several actionable suggestions to mitigate this issue

Large Language Models (LLMs) have reshaped natural language processing with their impressive capabilities. However, their ever-increasing size has raised concerns about their effective deployment and the need for LLM compression. This study introduces the Divergent Token Metrics (DTMs), a novel approach to assessing compressed LLMs, addressing the limitations of traditional perplexity or accuracy measures that fail to accurately reflect text generation quality. DTMs measure token divergences that allow deeper insights into the subtleties of model compression, in particular, when evaluating components’ impacts individually. Utilizing the First Divergent Token Metric (FDTM) in model sparsification reveals that 25% of all attention components can be pruned beyond 90% on the Llama-2 model family, still keeping SOTA performance. For quantization, FDTM suggests that more than 80% of parameters can be naively transformed to int8 without special outlier management. These evaluations indicate the necessity of choosing appropriate compressions for parameters individually—and that FDTM can identify those—while standard metrics result in deteriorated outcomes.

Language models pre-trained on general text have achieved impressive results in diverse fields. Yet, the distinct linguistic characteristics of task-oriented dialogues (TOD) compared to general text limit the practical utility of existing language models. Current task-oriented dialogue pre-training methods overlook the one-to-many property of conversations, where multiple responses can be appropriate given the same conversation context.In this paper, we propose a novel dialogue pre-training model called DivTOD, which collaborates with LLMs to learn diverse task-oriented dialogue representations. DivTOD guides LLMs in transferring diverse knowledge to smaller models while removing domain knowledge that contradicts task-oriented dialogues. Experiments show that our model outperforms strong TOD baselines on various downstream dialogue tasks and learns the intrinsic diversity of task-oriented dialogues.

Pre-trained Language Models (PLMs) are known to contain various kinds of knowledge.One method to infer relational knowledge is through the use of cloze-style prompts, where a model is tasked to predict missing subjects orobjects. Typically, designing these prompts is a tedious task because small differences in syntax or semantics can have a substantial impact on knowledge retrieval performance. Simultaneously, evaluating the impact of either prompt syntax or information is challenging due to their interdependence. We designed CONPARE-LAMA – a dedicated probe, consisting of 34 million distinct prompts that facilitate comparison across minimal paraphrases. These paraphrases follow a unified meta-template enabling the controlled variation of syntax and semantics across arbitrary relations.CONPARE-LAMA enables insights into the independent impact of either syntactical form or semantic information of paraphrases on the knowledge retrieval performance of PLMs. Extensive knowledge retrieval experiments using our probe reveal that prompts following clausal syntax have several desirable properties in comparison to appositive syntax: i) they are more useful when querying PLMs with a combination of supplementary information, ii) knowledge is more consistently recalled across different combinations of supplementary information, and iii) they decrease response uncertainty when retrieving known facts. In addition, range information can boost knowledge retrieval performance more than domain information, even though domain information is more reliably helpful across syntactic forms.

As the field of Natural Language Processing (NLP) increasingly adopts transformer-based models, the issue of bias becomes more pronounced. Such bias, manifesting through stereotypes and discriminatory practices, can disadvantage certain groups. Our study focuses on direct and indirect bias in the model explanations, where the model makes predictions relying heavily on identity tokens or associated contexts. We present a novel analysis of bias in model explanation, especially the subtle indirect bias, underlining the limitations of traditional fairness metrics. We first define direct and indirect bias in model explanations, which is complementary to fairness in predictions. We then develop an indirect bias discovery algorithm for quantitatively evaluating indirect bias in transformer models using their in-built self-attention matrix. We also propose an indirect bias mitigation algorithm to ensure fairness in transformer models by leveraging attention explanations. Our evaluation shows the significance of indirect bias and the effectiveness of our indirect bias discovery and mitigation.

We discover alignments of views between interest groups (lobbies) and members of the European Parliament (MEPs) by automatically analyzing their texts. Specifically, we do so by collecting novel datasets of lobbies’ position papers and MEPs’ speeches, and comparing these texts on the basis of semantic similarity and entailment. In the absence of ground-truth, we perform an indirect validation by comparing the discovered alignments with a dataset, which we curate, of retweet links between MEPs and lobbies, and with the publicly disclosed meetings of MEPs. Our best method performs significantly better than several baselines. Moreover, an aggregate analysis of the discovered alignments, between groups of related lobbies and political groups of MEPs, correspond to the expectations from the ideology of the groups (e.g., groups on the political left are more aligned with humanitarian and environmental organisations). We believe that this work is a step towards enhancing the transparency of the intricate decision-making processes within democratic institutions.

Temporal expression (TE) normalization is a well-studied problem. However, the predominately used rule-based systems are highly restricted to specific settings, and upcoming machine learning approaches suffer from a lack of labeled data. In this work, we explore the feasibility of proprietary and open-source large language models (LLMs) for TE normalization using in-context learning to inject task, document, and example information into the model. We explore various sample selection strategies to retrieve the most relevant set of examples. By using a window-based prompt design approach, we can perform TE normalization across sentences, while leveraging the LLM knowledge without training the model.Our experiments show competitive results to models designed for this task. In particular, our method achieves large performance improvements for non-standard settings by dynamically including relevant examples during inference.

Minimum Bayes Risk (MBR) decoding can significantly improve translation performance of Multilingual Large Language Models (MLLMs). However, MBR decoding is computationally expensive. We show how the recently developed Reinforcement Learning technique, Direct Preference Optimization (DPO), can fine-tune MLLMs to get the gains of MBR without any additional computation in inference. Our method uses only a small monolingual fine-tuning set and yields significantly improved performance on multiple NMT test sets compared to MLLMs without DPO.

Previously, non-autoregressive models were widely recognized as being superior in generation efficiency but inferior in generation quality due to the challenges of modeling multiple target modalities.To enhance the multi-modality modeling ability, we propose the diffusion glancing transformer, which employs a modality diffusion process and residual glancing sampling.The modality diffusion process is a discrete process that interpolates the multi-modal distribution along the decoding steps, and the residual glancing sampling approach guides the model to continuously learn the remaining modalities across the layers. Experimental results on various machine translation and text generation benchmarks demonstrate that DIFFGLAT achieves better generation accuracy while maintaining fast decoding speed compared with both autoregressive and non-autoregressive models.

Ensuring the privacy of Large Language Models (LLMs) is becoming increasingly important. The most widely adopted technique to accomplish this is DP-SGD, which trains a model to guarantee Differential Privacy (DP). However, DP-SGD overestimates an adversary’s capabilities in having white box access to the model and, as a result, causes longer training times and larger memory usage than SGD. On the other hand, commercial LLM deployments are predominantly cloud-based; hence, adversarial access to LLMs is black-box. Motivated by these observations, we present Private Mixing of Ensemble Distributions (PMixED): a private prediction protocol for next-token prediction that utilizes the inherent stochasticity of next-token sampling and a public model to achieve Differential Privacy. We formalize this by introducing RD-mollifers which project each of the model’s output distribution from an ensemble of fine-tuned LLMs onto a set around a public LLM’s output distribution, then average the projected distributions and sample from it. Unlike DP-SGD which needs to consider the model architecture during training, PMixED is model agnostic, which makes PMixED a very appealing solution for current deployments. Our results show that PMixED achieves a stronger privacy guarantee than sample-level privacy and outperforms DP-SGD for privacy 𝜖 = 8 on large-scale datasets. Thus, PMixED offers a practical alternative to DP training methods for achieving strong generative utility without compromising privacy.

In the constant updates of the product dialogue systems, we need to retrain the natural language understanding (NLU) model as new data from the real users would be merged into the existing data accumulated in the last updates. Within the newly added data, new intents would emerge and might have semantic entanglement with the existing intents, e.g. new intents that are semantically too specific or generic are actually a subset or superset of some existing intents in the semantic space, thus impairing the robustness of the NLU model.As the first attempt to solve this problem, we setup a new benchmark consisting of 4 Dialogue Version Control dataSets (DialogVCS). We formulate the intent detection with imperfect data in the system update as a multi-label classification task with positive but unlabeled intents, which asks the models to recognize all the proper intents, including the ones with semantic entanglement, in the inference.We also propose comprehensive baseline models and conduct in-depth analyses for the benchmark, showing that the semantically entangled intents can be effectively recognized with an automatic workflow. Our code and dataset are available at https://github.com/Zefan-Cai/DialogVCS.

As sharing images in an instant message is a crucial factor, there has been active research on learning an image-text multi-modal dialogue models.However, training a well-generalized multi-modal dialogue model remains challenging due to the low quality and limited diversity of images per dialogue in existing multi-modal dialogue datasets.In this paper, we propose an automated pipeline to construct a multi-modal dialogue dataset, ensuring both dialogue quality and image diversity without requiring minimum human effort. In our pipeline, to guarantee the coherence between images and dialogue, we prompt GPT-4 to infer potential image-sharing moments - specifically, the utterance, speaker, rationale, and image description. Furthermore, we leverage CLIP similarity to maintain consistency between aligned multiple images to the utterance.Through this pipeline, we introduce DialogCC, a high-quality and diverse multi-modal dialogue dataset that surpasses existing datasets in terms of quality and diversity in human evaluation.Our comprehensive experiments highlight that when multi-modal dialogue models are trained using our dataset, their generalization performance on unseen dialogue datasets is significantly enhanced. We make our source code and dataset publicly available (https://dialogcc.github.io/).

Large language models (LLMs) have achieved remarkable breakthroughs in new dialogue capabilities by leveraging instruction tuning,which refreshes human impressions of dialogue systems. The long-standing goal of dialogue systems is to be human-like enough to establish long-term connections with users. Therefore, there has been an urgent need to evaluate LLMs as human-like dialogue systems. In this paper, we propose DialogBench, a dialogue evaluation benchmark that contains 12 dialogue tasks to probe the capabilities of LLMs as human-like dialogue systems should have. Specifically, we prompt GPT-4 to generate evaluation instances for each task. We first design the basic prompt based on widely used design principles and further mitigate the existing biases to generate higher-quality evaluation instances. Our extensive tests on English and Chinese DialogBench of 26 LLMs show that instruction tuning improves the human likeness of LLMs to a certain extent, but most LLMs still have much room for improvement as human-like dialogue systems. Interestingly, results also show that the positioning of assistant AI can make instruction tuning weaken the human emotional perception of LLMs and their mastery of information about human daily life.