ACL 2024 Papers

Non-compositional expressions are an integral part of natural language and their meanings cannot be directly derived from the meanings of their component words. Recent work has shown how their processing remains a challenge for pre-trained language models. Here we consider the fact that prior knowledge of their component words is inadequate to infer their meaning as a whole and that these expressions constitute a long-tailed process in language (based on their occurrence in corpora and their coming into use as an idiomatic expression in a continual manner). Against this backdrop, this paper studies the ability of recent pre-trained language models to generate non-compositional expressions in English and their continual learning. Formulating this as a mask infilling task termed as CLoNE, the study uncovers the combined challenges of non-compositionality and their continual learning. Using a set of three diverse idiomatic expression datasets repurposed for this task, we benchmark different large pre-trained language models and different continual learning methods on the task of non-compositional expression generation. Our experiments on the CLoNE task show that large pre-trained language models are limited in their ability to generate non-compositional expressions and available continual learning methods are inadequate for our proposed CLoNE task which calls for more effective methods for continual learning of non-compositionality. Our datasets and code will be released publicly upon acceptance.

In non-autoregressive translation (NAT), directed acyclic Transformers (DAT) have demonstrated their ability to achieve comparable performance to the autoregressive Transformers.In this paper, we first show that DAT is essentially a fully connected left-to-right Hidden Markov Model (HMM), with the source and target sequences being observations and the token positions being latent states.Even though generative models like HMM do not suffer from label bias in traditional task settings (e.g., sequence labeling), we argue here that the left-to-right HMM in NAT may still encounter this issue due to the missing observations at the inference stage.To combat label bias, we propose two constrained HMMs: 1) Adaptive Window HMM, which explicitly balances the number of outgoing transitions at different states; 2) Bi-directional HMM, i.e., a combination of left-to-right and right-to-left HMMs, whose uni-directional components can implicitly regularize each other’s biases via shared parameters.Experimental results on WMT’14 EnDe and WMT’17 ZhEn demonstrate that our methods can achieve better or comparable performance to the original DAT using various decoding methods.We also demonstrate that our methods effectively reduce the impact of label bias.

Incorporating every annotator’s perspective is crucial for unbiased data modeling. Annotator fatigue and changing opinions over time can distort dataset annotations. To combat this, we propose to learn a more accurate representation of diverse opinions by utilizing multitask learning in conjunction with loss-based label correction. We show that using our novel formulation, we can cleanly separate agreeing and disagreeing annotations. Furthermore, this method provides a controllable way to encourage or discourage disagreement. We demonstrate that this modification can improve prediction performance in a single or multi-annotator setting. Lastly, we show that this method remains robust to additional label noise that is applied to subjective data.

Healthcare Community Question Answering (CQA) forums offer an accessible platform for individuals seeking information on various healthcare-related topics. People find such platforms suitable for self-disclosure, seeking medical opinions, finding simplified explanations for their medical conditions, and answering others’ questions. However, answers on these forums are typically diverse and prone to off-topic discussions. It can be challenging for readers to sift through numerous answers and extract meaningful insights, making answer summarization a crucial task for CQA forums. While several efforts have been made to summarize the community answers, most of them are limited to the open domain and overlook the different perspectives offered by these answers. To address this problem, this paper proposes a novel task of perspective-specific answer summarization. We identify various perspectives, within healthcare-related responses and frame a perspective-driven abstractive summary covering all responses. To achieve this, we annotate 3167 CQA threads with 6193 perspective-aware summaries in our PUMA dataset. Further, we propose PLASMA, a prompt-driven controllable summarization model. To encapsulate the perspective-specific conditions, we design an energy-controlled loss function for the optimization. We also leverage the prefix tuner to learn the intricacies of the healthcare perspective summarization. Our evaluation against five baselines suggests the superior performance of PLASMA by a margin of ~1.5 - 21% improvement. We supplement our experiments with ablation and qualitative analysis.

While (large) language models have significantly improved over the last years, they still struggle to sensibly process long sequences found, e.g., in books, due to the quadratic scaling of the underlying attention mechanism. To address this, we propose NextLevelBERT, a Masked Language Model operating not on tokens, but on higher-level semantic representations in the form of text embeddings. We pretrain NextLevelBERT to predict the vector representation of entire masked text chunks and evaluate the effectiveness of the resulting document vectors on three types of tasks: 1) Semantic Textual Similarity via zero-shot document embeddings, 2) Long document classification, 3) Multiple-choice question answering. We find that next-level Masked Language Modeling is an effective technique to tackle long-document use cases and can outperform much larger embedding models as long as the required level of detail of semantic information is not too fine. Our models and code are publicly available online.

We present NewsBench, a novel evaluation framework to systematically assess the capabilities of Large Language Models (LLMs) for editorial capabilities in Chinese journalism. Our constructed benchmark dataset is focused on four facets of writing proficiency and six facets of safety adherence, and it comprises manually and carefully designed 1,267 test samples in the types of multiple choice questions and short answer questions for five editorial tasks in 24 news domains. To measure performances, we propose different GPT-4 based automatic evaluation protocols to assess LLM generations for short answer questions in terms of writing proficiency and safety adherence, and both are validated by the high correlations with human evaluations. Based on the systematic evaluation framework, we conduct a comprehensive analysis of eleven popular LLMs which can handle Chinese. The experimental results highlight GPT-4 and ERNIE Bot as top performers, yet reveal a relative deficiency in journalistic safety adherence in creative writing tasks. Our findings also underscore the need for enhanced ethical guidance in machine-generated journalistic content, marking a step forward in aligning LLMs with journalistic standards and safety considerations. The evaluation framework and experimental results are expected to provide an in-depth understanding of the editorial capabilities of LLMs and speed up the development of LLMs in journalism.

While large language models (LLMs) are equipped with longer text input capabilities than before, they are struggling to seek correct information in long contexts. The “lost in the middle” problem challenges most LLMs, referring to the dramatic decline in accuracy when correct information is located in the middle. To overcome this crucial issue, this paper proposes to enhance the information searching and reflection ability of LLMs in long contexts via specially designed tasks called Position-Agnostic Multi-step QA (PAM QA). Trained in this task, our model excels in focusing more precisely on the desired information. Experimental results show substantial improvement in Multi-doc QA and other benchmarks, superior to state-of-the-art models by 13.7% absolute gain in shuffled settings, by 21.5% in passage retrieval task. We release our model and code to promote related research in the community.

We analyze a family of large language models in such a lightweight manner that can be done on a single GPU. Specifically, we focus on the OPT family of models ranging from 125m to 66b parameters and rely only on whether an FFN neuron is activated or not. First, we find that the early part of the network is sparse and represents many discrete features. Here, many neurons (more than in some layers of the 66b model) are “dead”, i.e. they never activate on a large collection of diverse data. At the same time, many of the alive neurons are reserved for discrete features and act as token and n-gram detectors. Interestingly, their corresponding FFN updates not only promote next token candidates as could be expected, but also explicitly focus on removing the information about triggering them tokens, i.e., current input. To the best of our knowledge, this is the first example of mechanisms specialized at removing (rather than adding) information from the residual stream. With scale, models become more sparse in a sense that they have more dead neurons and token detectors. Finally, some neurons are positional: them being activated or not depends largely (or solely) on position and less so (or not at all) on textual data. We find that smaller models have sets of neurons acting as position range indicators while larger models operate in a less explicit manner.

Transformer-based Language Models have become ubiquitous in Natural Language Processing (NLP) due to their impressive performance on various tasks. However, expensive training as well as inference remains a significant impediment to their widespread applicability. While enforcing sparsity at various levels of the model architecture has found promise in addressing scaling and efficiency issues, there remains a disconnect between how sparsity affects network topology. Inspired by brain neuronal networks, we explore sparsity approaches through the lens of network topology. Specifically, we exploit mechanisms seen in biological networks, such as preferential attachment and redundant synapse pruning, and show that principled, model-agnostic sparsity approaches are performant and efficient across diverse NLP tasks, spanning both classification (such as natural language inference) and generation (summarization, machine translation), despite our sole objective not being optimizing performance. NeuroPrune is competitive with (or sometimes superior to) baselines on performance and can be up to 10x faster in terms of training time for a given level of sparsity, simultaneously exhibiting measurable improvements in inference time in many cases.

With the launch of ChatGPT, large language models (LLMs) have attracted global attention. In the realm of article writing, LLMs have witnessed extensive utilization, giving rise to concerns related to intellectual property protection, personal privacy, and academic integrity. In response, AI-text detection has emerged to distinguish between human and machine-generated content. However, recent research indicates that these detection systems often lack robustness and struggle to effectively differentiate perturbed texts. Currently, there is a lack of systematic evaluations regarding detection performance in real-world applications, and a comprehensive examination of perturbation techniques and detector robustness is also absent. To bridge this gap, our work simulates real-world scenarios in both informal and professional writing, exploring the out-of-the-box performance of current detectors. Additionally, we have constructed 12 black-box text perturbation methods to assess the robustness of current detection models across various perturbation granularities. Furthermore, through adversarial learning experiments, we investigate the impact of perturbation data augmentation on the robustness of AI-text detectors. We have released our code and data at https://github.com/zhouying20/ai-text-detector-evaluation.

Large language models are meticulously aligned to be both helpful and harmless. However, recent research points to a potential overkill which means models may refuse to answer benign queries. In this paper, we investigate the factors for overkill by exploring how models handle and determine the safety of queries. Our findings reveal the presence of shortcuts within models, leading to excessive attention to harmful words like ‘kill’ and prompts emphasizing safety will exacerbate overkill. Based on these insights, we introduce Self-Contrastive Decoding (Self-CD), a training-free and model-agnostic strategy, to alleviate this phenomenon. We first extract such excessive attention by amplifying the difference in the model’s output distributions when responding to system prompts that either include or omit an emphasis on safety. Then we determine the final next-token predictions by downplaying the excessive attention via contrastive decoding. Empirical results have indicated that our method has achieved an average reduction of the refusal rate by 20 % while having almost no impact on safety.

Ten years ago a single metric, BLEU, governed progress in machine translation research. For better or worse, there is no such consensus today, and consequently it is difficult for researchers to develop and retain intuitions about metric deltas that drove earlier research and deployment decisions. This paper investigates the “dynamic range” of a number of modern metrics in an effort to provide a collective understanding of the meaning of differences in scores both within and among metrics; in other words, we ask “what point difference x in metric y is required between two systems for humans to notice?”. We conduct our evaluation on a new large dataset, ToShip23, using it to discover deltas at which metrics achieve system-level differences that are meaningful to humans, which we measure by pairwise system accuracy. We additionally show that this method of establishing delta-accuracy is more stable than the standard use of statistical p-values in regards to testset size. Where data size permits, we also explore the effect of metric deltas and accuracy across finer-grained features such as translation direction, domain, and system closeness.

Memory Editing (ME) has emerged as an efficient method to modify erroneous facts or inject new facts into Large Language Models (LLMs). Two mainstream ME methods exist: parameter-modifying ME and parameter-preserving ME (integrating extra modules while preserving original parameters). Regrettably, previous studies on ME evaluation have two critical limitations: (i) evaluating LLMs with single edit only, neglecting the need for continuous editing, and (ii) evaluations focusing solely on basic factual triples, overlooking broader LLM capabilities like logical reasoning and reading understanding. This study addresses these limitations with contributions threefold: (i) We explore how ME affects a wide range of fundamental capabilities of LLMs under sequential editing. Experimental results reveal an intriguing phenomenon: Most parameter-modifying ME consistently degrade performance across all tasks after a few sequential edits. In contrast, parameter-preserving ME effectively maintains LLMs’ fundamental capabilities but struggles to accurately recall edited knowledge presented in a different format. (ii) We extend our evaluation to different editing settings, such as layers to edit, model size, instruction tuning, etc. Experimental findings indicate several strategies that can potentially mitigate the adverse effects of ME. (iii) We further explain why parameter-modifying damages LLMs from three dimensions: parameter changes after editing, language modeling capability, and the in-context learning capability. Our in-depth study advocates more careful use of ME in real-world scenarios.

Reasoning, a fundamental cognitive process integral to human intelligence, has garnered substantial interest within artificial intelligence.Notably, recent studies have revealed that chain-of-thought prompting significantly enhances LLM’s reasoning capabilities, which attracts widespread attention from both academics and industry.In this paper, we systematically investigate relevant research, summarizing advanced methods through a meticulous taxonomy that offers novel perspectives.Moreover, we delve into the current frontiers and delineate the challenges and future directions, thereby shedding light on future research.Furthermore, we engage in a discussion about open questions.We hope this paper serves as an introduction for beginners and fosters future research.Resources have been made publicly available at https://github.com/zchuz/CoT-Reasoning-Survey

Large language models (LLMs) have manifested strong ability to generate codes for productive activities. However, current benchmarks for code synthesis, such as HumanEval, MBPP, and DS-1000, are predominantly oriented towards introductory tasks on algorithm and data science, insufficiently satisfying challenging requirements prevalent in real-world coding. To fill this gap, we propose NaturalCodeBench (NCB), a challenging code benchmark designed to mirror the complexity and variety of scenarios in real coding tasks. NCB comprises 402 high-quality problems in Python and Java, meticulously selected from natural user queries from online coding services, covering 6 different domains. Noting the extraordinary difficulty in creating testing cases for real-world queries, we also introduce a semi-automated pipeline to enhance the efficiency of test case construction. Comparing with manual solutions, it achieves an efficiency increase of more than 4 times. Our systematic experiments on 39 LLMs find that performance gaps on NCB between models with close HumanEval scores could still be significant, indicating a lack of focus on practical code synthesis scenarios or over-specified optimization on HumanEval. On the other hand, even the best-performing GPT-4 is still far from satisfying on NCB. The evaluation toolkit and development set are available at https://github.com/THUDM/NaturalCodeBench.

Efforts to apply transformer-based language models (TLMs) to the problem of reasoning in natural language have enjoyed ever-increasing success in recent years. The most fundamental task in this area to which nearly all others can be reduced is that of determining satisfiability. However, from a logical point of view, satisfiability problems vary along various dimensions, which may affect TLMs’ ability to learn how to solve them. The problem instances of satisfiability in natural language can belong to different computational complexity classes depending on the language fragment in which they are expressed. Although prior research has explored the problem of natural language satisfiability, the above-mentioned point has not been discussed adequately. Hence, we investigate how problem instances from varying computational complexity classes and having different grammatical constructs impact TLMs’ ability to learn rules of inference. Furthermore, to faithfully evaluate TLMs, we conduct an empirical study to explore the distribution of satisfiability problems.

Factual questions typically can be answered correctly at different levels of granularity. For example, both “August 4, 1961” and “1961” are correct answers to the question “When was Barack Obama born?”. Standard question answering (QA) evaluation protocols, however, do not explicitly take this into account and compare a predicted answer against answers of a single granularity level. In this work, we propose GRANOLA QA, a novel evaluation setting where a predicted answer is evaluated in terms of accuracy and informativeness against a set of multi-granularity answers. We present a simple methodology for enriching existing datasets with multi-granularity answers, and create GRANOLA-EQ, a multi-granularity version of the EntityQuestions dataset. We evaluate a range of decoding methods on GRANOLA-EQ, including a new algorithm, called Decoding with Response Aggregation (DRAG), that is geared towards aligning the response granularity with the model’s uncertainty. Our experiments show that large language models with standard decoding tend to generate specific answers, which are often incorrect. In contrast, when evaluated on multi-granularity answers, DRAG yields a nearly 20 point increase in accuracy on average, which further increases for rare entities. Overall, this reveals that standard evaluation and decoding schemes may significantly underestimate the knowledge encapsulated in LMs.

While human speakers use a variety of different expressions when describing the same object in an image, giving rise to a distribution of plausible labels driven by pragmatic constraints, the extent to which current Vision & Language Large Language Models (VLLMs) can mimic this crucial feature of language use is an open question. This applies to common, everyday objects, but it is particularly interesting for uncommon or novel objects for which a category label may be lacking or fuzzy. Furthermore, similar patterns of variation are observed among human speakers for highly context-sensitive expressions, such as the quantifiers ‘few’ or ‘most’. In our work, we evaluate VLLMs (FROMAGe, BLIP-2, LLaVA) on three categories (nouns, attributes, and quantifiers) where humans show great subjective variability concerning the distribution over plausible labels, using datasets and resources mostly under-explored in previous work. Our results reveal mixed evidence on the ability of VLLMs to capture human naming preferences at generation time: while some models are good at mimicking human distributions for nouns and attributes, all of them fail to assign quantifiers, a task that requires more accurate, high-level reasoning.

To address the global issue of online hate, hate speech detection (HSD) systems are typically developed on datasets from the United States, thereby failing to generalize to English dialects from the Majority World. Furthermore, HSD models are often evaluated on non-representative samples, raising concerns about overestimating model performance in real-world settings. In this work, we introduce NaijaHate, the first dataset annotated for HSD which contains a representative sample of Nigerian tweets. We demonstrate that HSD evaluated on biased datasets traditionally used in the literature consistently overestimates real-world performance by at least two-fold. We then propose NaijaXLM-T, a pretrained model tailored to the Nigerian Twitter context, and establish the key role played by domain-adaptive pretraining and finetuning in maximizing HSD performance. Finally, owing to the modest performance of HSD systems in real-world conditions, we find that content moderators would need to review about ten thousand Nigerian tweets flagged as hateful daily to moderate 60% of all hateful content, highlighting the challenges of moderating hate speech at scale as social media usage continues to grow globally. Taken together, these results pave the way towards robust HSD systems and a better protection of social media users from hateful content in low-resource settings.