NAACL 2024 Papers

Textual backdoor attacks, characterized by subtle manipulations of input triggers and training dataset labels, pose significant threats to security-sensitive applications. The rise of advanced generative models, such as GPT-4, with their capacity for human-like rewriting, makes these attacks increasingly challenging to detect. In this study, we conduct an in-depth examination of black-box generative models as tools for backdoor attacks, thereby emphasizing the need for effective defense strategies. We propose BGMAttack, a novel framework that harnesses advanced generative models to execute stealthier backdoor attacks on text classifiers. Unlike prior approaches constrained by subpar generation quality, BGMAttack renders backdoor triggers more elusive to human cognition and advanced machine detection. A rigorous evaluation of attack effectiveness over four sentiment classification tasks, complemented by four human cognition stealthiness tests, reveals BGMAttack’s superior performance, achieving a state-of-the-art attack success rate of 97.35% on average while maintaining superior stealth compared to conventional methods. The dataset and code are available: https://github.com/JiazhaoLi/BGMAttack.

Vision-language models (VLMs) are achieving increasingly strong performance on multimodal tasks. However, reasoning capabilities remain limited particularly for smaller VLMs, while those of large-language models (LLMs) have seen numerous improvements. We pro-pose a technique to transfer capabilities from LLMs to VLMs. On the recently introduced ChartQA, our method obtains state-of-the-artperformance when applied on the PaLI3-5B VLM by Chen et al. (2023c), while also enabling much better performance on PlotQA and FigureQA.We first improve the chart representation by continuing the pre-training stage using an improved version of the chart-to-table translation task by Liu et al. (2023a). We then propose constructing a 20x larger dataset than the original training set. To improve general reasoning capabilities and improve numerical operations, we synthesize reasoning traces using the table representation of charts. Lastly, our model is fine-tuned using the multitask loss introduced by Hsieh et al. (2023).Our variant ChartPaLI-5B outperforms even 10x larger models such as PaLIX-55B without using an upstream OCR system, while keeping inference time constant compared to the PaLI3-5B baseline. When rationales are further refined with a simple program-of-thought prompt (Chen et al., 2023a), our model outperforms the recently introduced Gemini Ultra and GPT-4V.

Recent advancements in large language models’ (LLMs) capabilities have yielded few-shot, human-comparable performance on a range of tasks. At the same time, researchers expend significant effort and resources gathering human annotations. At some point, LLMs may be able to perform some simple annotation tasks, but studies of LLM annotation accuracy and behavior are sparse. In this paper, we characterize OpenAI’s GPT-3.5’s judgment on a behavioral task for implicit object categorization. We characterize the embedding spaces of models trained on human vs. GPT responses and give similarities and differences between them, finding many similar dimensions. We also find that despite these similar dimensions, augmenting humans’ responses with GPT ones drives model divergence across the sizes of datasets tested.

In this paper, we examine the collaborative dynamics between humansand language models (LMs), where the interactions typically involveLMs proposing text segments and humans editing or responding to theseproposals. Productive engagement with LMs in such scenarios necessitates that humans discern effective text-based interaction strategies, such as editing and response styles, from historical human-LM interactions. This objective is inherently causal, driven by the counterfactual ‘what-if’ question: how would the outcome of collaboration change if humans employed a different text editing/refinement strategy? A key challenge in answering this causal inference question is formulating an appropriate causal estimand: the conventional average treatment effect (ATE) estimand is inapplicable to text-based treatments due to their high dimensionality. To address this concern, we introduce a new causal estimand– *Incremental Stylistic Effect (ISE)*, which characterizes the average impact of infinitesimally shifting a text towards a specific style, such as increasing formality. We establish the conditions for the non-parametric identification of ISE. Building on this, we develop *CausalCollab*, an algorithm designed to estimate the ISE of various interaction strategies in dynamic human-LM collaborations. Our empirical investigations across three distinct human-LM collaboration scenarios reveal that *CausalCollab* effectively reduces confounding and significantly improves counterfactual estimation over a set of competitive baselines.

The dynamic nature of knowledge in an ever-changing world presents challenges for language models trained on static data; the model in the real world often requires not only acquiring new knowledge but also overwriting outdated information into updated ones. To study the ability of language models for these time-dependent dynamics in human language, we introduce a novel task, EvolvingQA, a temporally evolving question-answering benchmark designed for training and evaluating LMs on an evolving Wikipedia database. The construction of EvolvingQA is automated with our pipeline using large language models. We uncover that existing continual learning baselines suffer from updating and removing outdated knowledge. Our analysis suggests that models fail to rectify knowledge due to small weight gradients. In addition, we elucidate that language models particularly struggle to reflect the change of numerical or temporal information. Our work aims to model the dynamic nature of real-world information, suggesting faithful evaluations of the evolution-adaptability of language models. Our data construction code and dataset files are available at https://github.com/kimyuji/EvolvingQA_benchmark.

Supervised classification heavily depends on datasets annotated by humans. However, in subjective tasks such as toxicity classification, these annotations often exhibit low agreement among raters. Annotations have commonly been aggregated by employing methods like majority voting to determine a single ground truth label. In subjective tasks, aggregating labels will result in biased labeling and, consequently, biased models that can overlook minority opinions. Previous studies have shed light on the pitfalls of label aggregation and have introduced a handful of practical approaches to tackle this issue. Recently proposed multi-annotator models, which predict labels individually per annotator, are vulnerable to under-determination for annotators with few samples. This problem is exacerbated in crowdsourced datasets. In this work, we propose Annotator Aware Representations for Texts (AART) for subjective classification tasks. Our approach involves learning representations of annotators, allowing for exploration of annotation behaviors. We show the improvement of our method on metrics that assess the performance on capturing individual annotators’ perspectives. Additionally, we demonstrate fairness metrics to evaluate our model’s equability of performance for marginalized annotators compared to others.

We study (differentially) private federated learning (FL) of language models. The language models in cross-device FL are relatively small, which can be trained with meaningful formal user-level differential privacy (DP) guarantees when massive parallelism in training is enabled by the participation of a moderate size of users. Recently, public data has been used to improve privacy-utility trade-offs for both large and small language models. In this work, we provide a systematic study of using large-scale public data and LLMs to help differentially private training of on-device FL models, and further improve the privacy-utility tradeoff by techniques of distillation. Moreover, we propose a novel distribution matching algorithm with theoretical grounding to sample public data close to private data distribution, which significantly improves the sample efficiency of (pre-)training on public data. The proposed method is efficient and effective for training private models by taking advantage of public data, especially for customized on-device architectures that do not have ready-touse pre-trained models.

Conversational moderation of online communities is crucial to maintaining civility for a constructive environment, but it is challenging to scale and harmful to moderators. The inclusion of sophisticated natural language generation modules as a force multiplier to aid human moderators is a tantalizing prospect, but adequate evaluation approaches have so far been elusive. In this paper, we establish a systematic definition of conversational moderation effectiveness grounded on moderation literature and establish design criteria for conducting realistic yet safe evaluation. We then propose a comprehensive evaluation framework to assess models’ moderation capabilities independently of human intervention. With our framework, we conduct the first known study of language models as conversational moderators, finding that appropriately prompted models that incorporate insights from social science can provide specific and fair feedback on toxic behavior but struggle to influence users to increase their levels of respect and cooperation.

We present a novel approach to automatically synthesize “wayfinding instructions” for an embodied robot agent. In contrast to prior approaches that are heavily reliant on human-annotated datasets designed exclusively for specific simulation platforms, our algorithm uses in-context learning to condition an LLM to generate instructions using just a few references. Using an LLM-based Visual Question Answering strategy, we gather detailed information about the environment which is used by the LLM for instruction synthesis. We implement our approach on multiple simulation platforms including Matterport3D, AI Habitat and ThreeDWorld, thereby demonstrating its platform-agnostic nature. We subjectively evaluate our approach via a user study and observe that 83.3% of users find the synthesized instructions accurately capture the details of the environment and show characteristics similar to those of human-generated instructions. Further, we conduct zero-shot navigation with multiple approaches on the REVERIE dataset using the generated instructions, and observe very close correlation with the baseline on standard success metrics (< 1% change in SR), quantifying the viability of generated instructions in replacing human-annotated data. We finally discuss the applicability of our approach in enabling a generalizable evaluation of embodied navigation policies. To the best of our knowledge, ours is the first LLM-driven approach capable of generating “human-like” instructions in a platform-agnostic manner, without training.

The contemporary LLMs are prone to producing hallucinations, stemming mainly from the knowledge gaps within the models. To address this critical limitation, researchers employ diverse strategies to augment the LLMs by incorporating external knowledge, aiming to reduce hallucinations and enhance reasoning accuracy. Among these strategies, leveraging knowledge graphs as a source of external information has demonstrated promising results. In this survey, we comprehensively review these knowledge-graph-based augmentation techniques in LLMs, focusing on their efficacy in mitigating hallucinations. We systematically categorize these methods into three overarching groups, offering methodological comparisons and performance evaluations. Lastly, this survey explores the current trends and challenges associated with these techniques and outlines potential avenues for future research in this emerging field.

Many recent studies examining the knowledge capacity of large language models (LLM) have focused on knowledge explicitly learned from the pretraining data or implicitly inferable from similar contexts. However, the extent to which an LLM effectively captures corpus-level statistical trends of concepts for reasoning, especially long-tail ones, is largely underexplored. In this study, we introduce a novel few-shot question-answering task (CPopQA) that examines LLMs’ statistical ranking abilities for long-tail cultural concepts (e.g., holidays), particularly focusing on these concepts’ popularity in the United States and the United Kingdom, respectively. We curate a dataset of 457 holidays across 58 countries, generating a total of 9,000 QA testing pairs. Experiments on four strong LLMs show that open-sourced LLMs still lag way behind close LLM API (e.g., GPT-3.5) in statistical ranking of cultural concepts. Notably, GPT-3.5 exhibited its potential to identify geo-cultural proximity across continents.

Knowledge graph completion (KGC) aims to infer missing facts based on existing facts within a KG. Recently, research on generative models (GMs) has addressed the limitations of embedding methods in terms of generality and scalability. However, GM-based methods are sensitive to contextual facts on KG, so the contextual facts of poor quality can cause GMs to generate erroneous results. To improve the performance of GM-based methods for various KGC tasks, we propose a COntextual FactS GuIded GeneratioN (COSIGN) model. First, to enhance the inference ability of the generative model, we designed a contextual facts collector to achieve human-like retrieval behavior. Second, a contextual facts organizer is proposed to learn the organized capabilities of LLMs through knowledge distillation. Finally, the organized contextual facts as the input of the inference generator to generate missing facts. Experimental results demonstrate that COSIGN outperforms state-of-the-art baseline techniques in terms of performance.

We present COPAL-ID, a novel, public Indonesian language common sense reasoning dataset. Unlike the previous Indonesian COPA dataset (XCOPA-ID), COPAL-ID incorporates Indonesian local and cultural nuances, and therefore, provides a more natural portrayal of day-to-day causal reasoning within the Indonesian cultural sphere. Professionally written by natives from scratch, COPAL-ID is more fluent and free from awkward phrases, unlike the translated XCOPA-ID. In addition, we present COPALID in both standard Indonesian and in Jakartan Indonesian–a dialect commonly used in daily conversation. COPAL-ID poses a greater challenge for existing open-sourced and closedstate-of-the-art multilingual language models, yet is trivially easy for humans. Our findings suggest that general multilingual models struggle to perform well, achieving 66.91% accuracy on COPAL-ID. South-East Asian-specific models achieve slightly better performance of 73.88% accuracy. Yet, this number still falls short of near-perfect human performance. This shows that these language models are still way behind in comprehending the local nuances of Indonesian.

A wave of new task-based virtual assistants has been fueled by increasingly powerful large language models (LLMs), such as GPT-4 (OpenAI, 2023). A major challenge in deploying LLM-based virtual conversational assistants in real world settings is ensuring they operate within what is admissible for the task. To overcome this challenge, the designers of these virtual assistants rely on an independent guardrail system that verifies the virtual assistant’s output aligns with the constraints required for the task. However, relying on commonly used, prompt-based guardrails can be difficult to engineer correctly and comprehensively. To address these challenges, we propose CONSCENDI. We use CONSCENDI to exhaustively generate training data with two key LLM-powered components: scenario-augmented generation and contrastive training examples. When generating conversational data, we generate a set of rule-breaking scenarios, which enumerate a diverse set of high-level ways a rule can be violated. This scenario-guided approach produces a diverse training set and provides chatbot designers greater control. To generate contrastive examples, we prompt the LLM to alter conversations with violations into acceptable conversations to enable fine-grained distinctions. We then use this data, generated by CONSCENDI, to train a smaller model. We find that CONSCENDI results in guardrail models that improve over baselines in multiple dialogue domains.

Recently, instruction-tuned large language models (LLMs) are showing prominent performance on various tasks, such as question answering. However, the majority of instruction-tuned LLMs are English-centric, which hinders their application to low-resource language QA. In this paper, we propose COde-Mixed Multilingual Instruction Tuning (COMMIT) to adapt English-centric LLM to low-resource language QA. We point out two main causes of English-centricness: imbalance of unlabeled data, and English-centric instruction tuning datasets. To deviate from English-centric instruction tuning, we propose to specialize code-mixing for instruction tuning, which blocks code-mixing in English templates, to leverage the potential of its superiority. To overcome data imbalance, we perform cross-lingual alignment. The majority of cross-lingual alignment works focused on making representations similar, which is not desirable to decoder-based LLMs, such as LLaMA. Therefore, we propose code-mixed continual causal language modeling to align the decoder. COMMIT improves the exact match score of low-resourced language QA by up to 32x. Code is publicly available.

Named entities – typically expressed via proper nouns – play a key role in Natural Language Processing, as their identification and comprehension are crucial in tasks such as Relation Extraction, Coreference Resolution and Question Answering, among others. Tasks like these also often entail dealing with concepts – typically represented by common nouns – which, however, have not received as much attention. Indeed, the potential of their identification and understanding remains underexplored, as does the benefit of a synergistic formulation with named entities. To fill this gap, we introduce Concept and Named Entity Recognition (CNER), a new unified task that handles concepts and entities mentioned in unstructured texts seamlessly. We put forward a comprehensive set of categories that can be used to model concepts and named entities jointly, and propose new approaches for the creation of CNER datasets. We evaluate the benefits of performing CNER as a unified task extensively, showing that a CNER model gains up to +5.4 and +8 macro F1 points when compared to specialized named entity and concept recognition systems, respectively. Finally, to encourage the development of CNER systems, we release our datasets and models at https://github.com/Babelscape/cner.

Large Language Models (LLMs) provide a possibility to make a great breakthrough in medicine. The establishment of a standardized medical benchmark becomes a fundamental cornerstone to measure progression. However, medical environments in different regions have their local characteristics, e.g., the ubiquity and significance of traditional Chinese medicine within China. Therefore, merely translating English-based medical evaluation may result in contextual incongruities to a local region. To solve the issue, we propose a localized medical benchmark called CMB, a Comprehensive Medical Benchmark in Chinese, designed and rooted entirely within the native Chinese linguistic and cultural framework. While traditional Chinese medicine is integral to this evaluation, it does not constitute its entirety. Using this benchmark, we have evaluated several prominent large-scale LLMs, including ChatGPT, GPT-4, dedicated Chinese LLMs, and LLMs specialized in the medical domain. We hope this benchmark provide first-hand experience in existing LLMs for medicine and also facilitate the widespread adoption and enhancement of medical LLMs within China. Our data and code are publicly available at https://github.com/FreedomIntelligence/CMB.

Neural Text-to-Speech (TTS) systems find broad applications in voice assistants, e-learning, and audiobook creation. The pursuit of modern models, like Diffusion Models (DMs), holds promise for achieving high-fidelity, real-time speech synthesis. Yet, the efficiency of multi-step sampling in Diffusion Models presents challenges. Efforts have been made to integrate GANs with DMs, speeding up inference by approximating denoising distributions, but this introduces issues with model convergence due to adversarial training. To overcome this, we introduce CM-TTS, a novel architecture grounded in consistency models (CMs). Drawing inspiration from continuous-time diffusion models, CM-TTS achieves top-quality speech synthesis in fewer steps without adversarial training or pre-trained model dependencies. We further design weighted samplers to incorporate different sampling positions into model training with dynamic probabilities, ensuring unbiased learning throughout the entire training process. We present a real-time mel-spectrogram generation consistency model, validated through comprehensive evaluations. Experimental results underscore CM-TTS’s superiority over existing single-step speech synthesis systems, representing a significant advancement in the field.

Recently, contrastive learning has begun to gain popularity in multimodal sentiment analysis (MSA). However, most of existing MSA methods based on contrastive learning lacks more detailed learning of the distribution of sample pairs with different sentiment intensity differences in the contrastive learning representation space. In addition, limited research has been conducted on the fusion of each modality representation obtained by contrastive learning training.In this paper, we propose a novel framework for multimodal sentiment analysis based on Contrastive Learning Guided by Sentiment Intensity (CLGSI). Firstly, the proposed contrastive learning guided by sentiment intensity selects positive and negative sample pairs based on the difference in sentiment intensity and assigns corresponding weights accordingly.Subsequently, we propose a new multimodal representation fusion mechanism, called Global-Local-Fine-Knowledge (GLFK), which extracts common features between different modalities’ representations. At the same time, each unimodal encoder output is separately processed by a Multilayer Perceptron (MLP) to extract specific features of each modality. Finally, joint learning of the common and specific features is used to predict sentiment intensity. The effectiveness of CLGSI is assessed on two English datasets, MOSI and MOSEI, as well as one Chinese dataset, SIMS. We achieve competitive experimental results, which attest to the strong generalization performance of our approach. The code for our approach will be released in https://github.com/AZYoung233/CLGSI