NAACL 2024 Papers

Existing watermarked generation algorithms employ token-level designs and therefore, are vulnerable to paraphrase attacks. To address this issue, we introduce watermarking on the semantic representation of sentences. We propose SemStamp, a robust sentence-level semantic watermarking algorithm that uses locality-sensitive hashing (LSH) to partition the semantic space of sentences. The algorithm encodes and LSH-hashes a candidate sentence generated by a language model, and conducts rejection sampling until the sampled sentence falls in watermarked partitions in the semantic embedding space. To test the paraphrastic robustness of watermarking algorithms, we propose a “bigram paraphrase” attack that produces paraphrases with small bigram overlap with the original sentence. This attack is shown to be effective against existing token-level watermark algorithms, while posing only minor degradations to SemStamp. Experimental results show that our novel semantic watermark algorithm is not only more robust than the previous state-of-the-art method on various paraphrasers and domains, but also better at preserving the quality of generation.

Language models (LMs) are indispensable tools for natural language processing tasks, but their vulnerability to adversarial attacks remains a concern. While current research has explored adversarial training techniques, their improvements to defend against word-level attacks have been limited. In this work, we propose a novel approach called Semantic Robust Defence (SemRoDe), a Macro Adversarial Training strategy to enhance the robustness of LMs. Drawing inspiration from recent studies in the image domain, we investigate and later confirm that in a discrete data setting such as language, adversarial samples generated via word substitutions do indeed belong to an adversarial domain exhibiting a high Wasserstein distance from the base domain. Our method learns a robust representation that bridges these two domains. We hypothesize that if samples were not projected into an adversarial domain, but instead to a domain with minimal shift, it would improve attack robustness. We align the domains by incorporating a new distance-based objective. With this, our model is able to learn more generalized representations by aligning the model’s high-level output features and therefore better handling unseen adversarial samples. This method can be generalized across word embeddings, even when they share minimal overlap at both vocabulary and word-substitution levels. To evaluate the effectiveness of our approach, we conduct experiments on BERT and RoBERTa models on three datasets. The results demonstrate promising state-of-the-art robustness.

Modern Neural Machine Translation systems exhibit strong performance in several different languages and are constantly improving. Their ability to learn continuously is, however, still severely limited by the catastrophic forgetting issue. In this work, we leverage a key property of encoder-decoder Transformers, i.e. their generative ability, to propose a novel approach to continually learning Neural Machine Translation systems. We show how this can effectively learn on a stream of experiences comprising different languages, by leveraging a replay memory populated by using the model itself as a generator of parallel sentences. We empirically demonstrate that our approach can counteract catastrophic forgetting without requiring explicit memorization of training data. Code will be publicly available upon publication.

Sample diversity depends on the task; within mathematics, precision and determinism are paramount, while storytelling thrives on creativity and surprise. This paper presents a simple self-regulating approach where we adjust sample diversity inference parameters dynamically based on the input prompt—in contrast to existing methods that require expensive and inflexible setups, or maintain static values during inference. Capturing a broad spectrum of sample diversities can be formulated as a straightforward self-supervised inference task, which we find significantly improves the quality of responses generically without model retraining or fine-tuning. In particular, our method demonstrates significant improvement in all supercategories of the MMLU multitask benchmark (GPT-3.5: +4.4%, GPT-4: +1.5%), which captures a large variety of difficult tasks covering STEM, the humanities and social sciences.

Recently, there has been a growing availability of pre-trained text models on various model repositories. These models greatly reduce the cost of training new models from scratch as they can be fine-tuned for specific tasks or trained on large datasets. However, these datasets may not be publicly accessible due to the privacy, security, or intellectual property issues. In this paper, we aim to develop a lightweight student network that can learn from multiple teacher models without accessing their original training data. Hence, we investigate Data-Free Knowledge Amalgamation (DFKA), a knowledge-transfer task that combines insights from multiple pre-trained teacher models and transfers them effectively to a compact student network. To accomplish this, we propose STRATANET, a modeling framework comprising: (a) a steerable data generator that produces text data tailored to each teacher and (b) an amalgamation module that implements a self-regulative strategy using confidence estimates from the teachers’ different layers to selectively integrate their knowledge and train a versatile student. We evaluate our method on three benchmark text classification datasets with varying labels or domains. Empirically, we demonstrate that the student model learned using our STRATANET outperforms several baselines significantly under data-driven and data-free constraints.

Open-Domain Question Answering (ODQA) aims to answer questions without explicitly providing specific background documents. This task becomes notably challenging in a zero-shot setting where no data is available to train tailored retrieval-reader models.While recent Large Language Models (LLMs) like GPT-3 have demonstrated their effectiveness in zero-shot ODQA using direct prompting methods, these methods still fall short of fully harnessing the potential of LLMs when implicitly invoked.In this paper, we propose a Self-Prompting framework to explicitly utilize the massive knowledge encoded in the parameters of LLMs and their strong instruction understanding abilities. Concretely, we prompt LLMs step by step to generate multiple pseudo QA pairs with background passages and explanations entirely from scratch.These generated elements are then utilized for in-context learning. Experimental results show that our method significantly surpasses previous state-of-the-art zero-shot methods on three widely-used ODQA datasets and even achieves comparable performance with various customized fine-tuned models on full training data. Our code is available at https://github.com/lockon-n/self-prompting.

Exploring the application of powerful large language models (LLMs) on the named entity recognition (NER) task has drawn much attention recently. This work pushes the performance boundary of zero-shot NER with LLMs by proposing a training-free self-improving framework, which utilizes an unlabeled corpus to stimulate the self-learning ability of LLMs. First, we use the LLM to make predictions on the unlabeled corpus using self-consistency and obtain a self-annotated dataset. Second, we explore various strategies to select reliable annotations to form a reliable self-annotated dataset. Finally, for each test input, we retrieve demonstrations from the reliable self-annotated dataset and perform inference via in-context learning. Experiments on four benchmarks show substantial performance improvements achieved by our framework. Through comprehensive experimental analysis, we find that increasing the size of unlabeled corpus or iterations of self-improving does not guarantee further improvement, but the performance might be boosted via more advanced strategies for reliable annotation selection.

Large language models (LLMs) have shown promising abilities of in-context learning (ICL), adapting swiftly to new tasks with only few-shot demonstrations. However, current few-shot methods heavily depend on high-quality, query-specific demos, which are often lacking. When faced with out-of-demonstration (OOD) queries, methods that rely on hand-crafted demos or external retrievers might fail. To bridge the gap between limited demos and OOD queries, we propose Self-Demos, a novel prompting method that elicits the inherent generalizability in LLMs by query-aware demo generation. The generated demos strategically interpolate between existing demos and the given query, transforming the query from OOD to ID. To evaluate the effectiveness of our approach, we manually constructed OOD-Toolset, a dataset in the tool-using scenario with over 300 real-world APIs and 1000 instances, each consisting of three tool-use cases as demos and an OOD query. Thorough experiments on our dataset and two public math benchmarks have shown that our method can outperform state-of-the-art baselines in the OOD setting. Moreover, we conduct a range of analyses to validate Self-Demos’s generalization and provide more insights.

To achieve state-of-the-art performance, one still needs to train NER models on large-scale, high-quality annotated data, an asset that is both costly and time-intensive to accumulate. In contrast, real-world applications often resort to massive low-quality labeled data through non-expert annotators via crowdsourcing and external knowledge bases via distant supervision as a cost-effective alternative. However, these annotation methods result in noisy labels, which in turn lead to a notable decline in performance. Hence, we propose to denoise the noisy NER data with guidance from a small set of clean instances. Along with the main NER model we train a discriminator model and use its outputs to recalibrate the sample weights. The discriminator is capable of detecting both span and category errors with different discriminative prompts. Results on public crowdsourcing and distant supervision datasets show that the proposed method can consistently improve performance with a small guidance set.

Fact-checking is an essential task in NLP that is commonly utilized to validate the factual accuracy of a piece of text. Previous approaches mainly involve the resource-intensive process of fine-tuning pre-trained language models on specific datasets. In addition, there is a notable gap in datasets that focus on fact-checking texts generated by large language models (LLMs). In this paper, we introduce Self-Checker, a plug-and-play framework that harnesses LLMs for efficient and rapid fact-checking in a few-shot manner. We also present the BingCheck dataset, specifically designed for fact-checking texts generated by LLMs. Empirical results demonstrate the potential of Self-Checker in the use of LLMs for fact-checking. Compared to state-of-the-art fine-tuned models, there is still significant room for improvement, indicating that adopting LLMs could be a promising direction for future fact-checking research.

Image–text models (ITMs) is the prevalent architecture to solve video question–answering tasks, which requires only a few input frames to save huge computational cost compared to video–language models.However, we find existent ITM video question–answering solutions either 1) adopt simplistic and unintentional sampling strategies, which may miss key frames to offer the answer clues; or 2) sample a large number of frames into divided groups, which the computational sources can not accommodate. In this work, we aim at an efficient sampling method towards the few-frame situations.We first summarize a family of prior sampling methods based on question–frame correlation into a unified one, dubbed *Most Implied Frames* (MIF). Through some primary results and analysis, Through analysis, we form a hypothesis that question-aware sampling is not necessary, from which we further propose the other method *Most Dominant Frames* (MDF).Experimental results on four public datasets and three advanced ITMs demonstrate that our proposed strategies can boost the performance for image–text pretrained models, and have a wide application scenario in terms of model architectures and dataset types. Our code is available at https://github.com/declare-lab/Sealinghttps://github.com/declare-lab/Sealing.

The latest large language models (LMs) support increasingly longer contexts. While this trend permits using substantial amounts of text with SOTA LMs, requiring these large LMs to process potentially redundant or irrelevant data needlessly increases inference time and cost. To remedy this problem, we propose BLINDER, a method that leverages a small finetuned LM to sample the minimal set of input features that maximizes the performance of a downstream LM. BLINDER trains an LM with a value head to estimate the likelihood of optimal outputs from a downstream LM given an input. We evaluate BLINDER on embodied decision making tasks with notoriously verbose state descriptions: NetHack and robot planning. BLINDER reduces the length of LM actor input by 87% and 99% while improving task success rates by 158% and 54% on NetHack and robot planning respectively which represents substantial inference cost savings while actually increasing performance.

Abstractive summarization for long-form narrative texts such as movie scripts is challenging due to the computational and memory constraints of current language models. A movie script typically comprises a large number of scenes; however, only a fraction of these scenes are salient, i.e., important for understanding the overall narrative. The salience of a scene can be operationalized by considering it as salient if it is mentioned in the summary. Automatically identifying salient scenes is difficult due to the lack of suitable datasets. In this work, we introduce a scene saliency dataset that consists of human-annotated salient scenes for 100 movies. We propose a two-stage abstractive summarization approach which first identifies the salient scenes in script and then generates a summary using only those scenes. Using QA-based evaluation, we show that our model outperforms previous state-of-the-art summarization methods and reflects the information content of a movie more accurately than a model that takes the whole movie script as input.

Large language models (LLMs) have notably revolutionized many domains within natural language processing due to their exceptional performance. Their security has become increasingly vital. This study is centered on protecting LLMs against unauthorized access and potential theft. We propose a simple yet effective protective measure wherein a unique key prompt is embedded within the LLM. This mechanism enables the model to respond only when presented with the correct key prompt; otherwise, LLMs will refuse to react to any input instructions. This key prompt protection offers a robust solution to prevent the unauthorized use of LLMs, as the model becomes unusable without the correct key. We evaluated the proposed protection on multiple LLMs and NLP tasks. Results demonstrate that our method can successfully protect the LLM without significantly impacting the model’s original function. Moreover, we demonstrate potential attacks that attempt to bypass the protection mechanism will adversely affect the model’s performance, further emphasizing the effectiveness of the proposed protection method.

In the E-Commerce domain, search query refinement reformulates malformed queries into canonicalized forms by preprocessing operations such as “term splitting” and “term merging”. Unfortunately, most relevant research is rather limited to English. In particular, there is a severe lack of study on search query refinement for the Japanese language. Furthermore, no attempt has ever been made to apply refinement methods to data improvement for downstream NLP tasks in real-world scenarios.This paper presents a novel query refinement approach for the Japanese language. Experimental results show that our method achieves significant improvement by 3.5 points through comparison with BERT-CRF as a baseline. Further experiments are also conducted to measure beneficial impact of query refinement on named entity recognition (NER) as the downstream task. Evaluations indicate that the proposed query refinement method contributes to better data quality, leading to performance boost on E-Commerce specific NER tasks by 11.7 points, compared to search query data preprocessed by MeCab, a very popularly adopted Japanese tokenizer.

We present SeaEval, a benchmark for multilingual foundation models. In addition to characterizing how these models understand and reason with natural language, we also investigate how well they comprehend cultural practices, nuances, and values. Alongside standard accuracy metrics, we investigate the brittleness of foundation models in the dimensions of semantics and multilinguality. Our analyses span both open-sourced and closed models, leading to empirical results across classic NLP tasks, reasoning, and cultural comprehension. Key findings indicate (1) Many models exhibit varied behavior when given paraphrased instructions. (2) Many models still suffer from exposure bias (e.g., positional bias, majority label bias). (3) For questions rooted in factual, scientific, and commonsense knowledge, consistent responses are expected across multilingual queries that are semantically equivalent. Yet, most models surprisingly demonstrate inconsistent performance on these queries. (4) Multilingually-trained models have not attained “balanced multilingual” capabilities. Our endeavors underscore the need for more generalizable semantic representations and enhanced multilingual contextualization. SeaEval can serve as a launchpad for more thorough investigations and evaluations for multilingual and multicultural scenarios.

Despite the success of multilingual pretrained language models (mPLMs) for tasks such as dependency parsing (DEP) or part-of-speech (POS) tagging, their coverage of 100s of languages is still limited, as most of the 6500+ languages remains “unseen”. To adapt mPLMs for including such unseen langs, existing work has considered transliteration and vocabulary augmentation. Meanwhile, the consideration of combining the two has been surprisingly lacking. To understand why, we identify both complementary strengths of the two, and the hurdles to realizing it. Based on this observation, we propose ScriptMix, combining two strengths, and overcoming the hurdle.Specifically, ScriptMix a) is trained with dual-script corpus to combine strengths, but b) with separate modules to avoid gradient conflict. In combining modules properly, we also point out the limitation of the conventional method AdapterFusion, and propose AdapterFusion+ to overcome it. We empirically show ScriptMix is effective– ScriptMix improves the POS accuracy by up to 14%, and improves the DEP LAS score by up to 5.6%. Our code is publicly available.

We describe our system for authorship attribution in the IARPA HIATUS program. We describe the model and compute infrastructure developed to satisfy the set of technical constraints imposed by IARPA, including runtime limits as well as other constraints related to the ultimate use case. One use-case constraint concerns the explainability of the features used in the system. For this reason, we integrate features from frame semantic parsing, as they are both interpretable and difficult for adversaries to evade. One trade-off with using such features, however, is that more sophisticated feature representations require more complicated architectures, which limit usefulness in time-sensitive and constrained compute environments. We propose an approach to increase the efficiency of frame semantic parsing through an analysis of parallelization and beam search sizes. Our approach results in a system that is approximately 8.37x faster than the base system with a minimal effect on accuracy.

Reinforcement learning from human feedback (RLHF) is a vital strategy for enhancing model capability in language models. However, annotating preference data for RLHF is a resource-intensive and creativity-demanding process, while existing automatic generation methods face limitations in data diversity and quality. In response, we present Safer-Instruct, a novel pipeline for automatically constructing large-scale preference data. Our approach leverages reversed instruction tuning, instruction induction, and expert model evaluation to efficiently generate high-quality preference data without human annotators. To verify the effectiveness of Safer-Instruct, we apply the pipeline to construct a safety preference dataset as a case study. Finetuning an Alpaca model on this synthetic dataset not only demonstrates improved harmlessness but also outperforms models fine-tuned on human-annotated safety preference data, all the while maintaining a competitive edge in downstream tasks. Importantly, our Safer-Instruct framework is versatile and can be applied to generate preference data across various domains, extending its utility beyond safety preferences. It addresses the challenges in preference data acquisition and advances the development of more capable and responsible AI systems. For dataset and code implementation, see https://github.com/uscnlp-lime/safer-instruct/.