Abstract When you use a large language model (LLM) in your research, you often need to formulate a prompt to elicit some relevant output from the LLM. This step is challenging since (1) LLMs are known to be brittle and their results may vary drastically between different prompts; and (2) for any given task, there are infinitely many possible prompts. Thus we end up with the following problem: If you cannot try out infinitely many prompts, how do you justify your selected prompt? This article discusses different ways in which authors may motivate their choices, and urges authors to consider how their prompting strategy might be justified.
Abstract Reproductions in NLP/ML indicate worryingly low levels of reproducibility, but their results are hard to interpret, compare, and learn from, because they are reported in qualitative terms with varying, often unstated, criteria. We propose QRA, a quantitative approach to reproducibility assessment that (i) produces continuous-valued degree of reproducibility assessments at three levels of granularity; (ii) utilizes reproducibility measures that are comparable across different studies; and (iii) grounds expectations about degree of reproducibility in similarity between experiments. QRA facilitates more informative reproducibility assessments and conclusions about the causes of better/poorer reproducibility. We demonstrate its benefits by applying QRA to three sets of comparable experiments, obtaining clear evidence that degree of reproducibility depends on similarity of experiment properties, system type, and evaluation method.
Abstract Automatic methods for scoring text readability have been studied for over a century, and are widely used in research and in user-facing applications in many domains. Thus far, the development and evaluation of such methods have primarily relied on two types of offline human behavioral data, performance on reading comprehension tests and ratings of text readability levels. In this work, we instead focus on a fundamental and understudied aspect of readability, real-time reading ease, captured with online reading measures using eye tracking. We introduce a new cognitive evaluation framework for readability scoring methods that quantifies their ability to account for reading ease, while controlling for content variation across texts. Applying this evaluation to prominent traditional readability formulas, NLP-based methods, commercial systems used in education, and frontier LLMs suggests that they are all poor predictors of English reading ease in adults as compared to word properties commonly used in psycholinguistics for the prediction of reading times. This outcome holds across L1 and L2 speakers, different reading regimes, and textual units of different lengths. Our results reveal an important limitation of a wide range of methods for readability scoring, highlight the utility of real-time behavioral benchmarks for readability research, and call for new, cognitively driven readability scoring approaches that can better account for how humans experience texts in real time.
Abstract This study examines how language models access and recover linguistic information associated with the correct antecedents of two types of anaphoric elements: Anaphoric one and Noun Phrase Ellipsis (NPE), both of which require antecedent resolution for proper interpretation. In this study, we created manually annotated datasets (collectively referred to as AO-NPE datasets) from English Wikipedia, containing annotated instances of Anaphoric one and NPE. We assessed model performance by fine-tuning BERT and RoBERTa, and by evaluating GPT-4o mini, LLaMA 3.2, and Gemma-2 in a zero-shot setting. We conducted experiments using the forced-choice task and open-ended antecedent identification task on both humans and models with the same datasets. Notably, GPT-4o mini performed comparably to humans and outperformed other language models. However, results from both types of tasks revealed that language models performed better in identifying the antecedent for NPE than for Anaphoric one, a pattern opposite to that observed in human data. Linear distance and number agreement between antecedents and anaphoric elements also affected the performance, with language models more strongly influenced by these factors than humans. These findings suggest that the underlying mechanisms of anaphoric comprehension in language models may not fully align with those employed by humans: while humans flexibly integrate syntactic, semantic, and discourse information, LLMs rely more heavily on surface-level cues such as linear distance, number agreement, and syntactic parallelism. Additionally, model performance dropped significantly in the open-ended identification task compared to the forced-choice task, while human performance remained relatively stable. This further suggests potential limitations in anaphoric processing.
Abstract Recent reasoning-focused large language models (LLMs) (e.g., GPT-o1 and DeepSeek-R1) combine chain-of-thought prompting with reinforcement learning to generate explicit, multi-step derivations. However, their practical advantages over standard instruction-tuned models remain underexplored. In this article, we present a comprehensive empirical study to address these gaps. We compare reasoning-focused and non-reasoning LLMs across several critical dimensions: (1) their performance on mathematical and general reasoning benchmarks; (2) their generalization capabilities across 30 sub-disciplines spanning six scientific domains; (3) the qualitative and quantitative differences in their underlying solution processes; and (4) the impact of model compression via distillation on core reasoning capabilities. Crucially, to reliably conduct our in-depth process analysis, we address the critical issue of LLM-as-a-Judge reliability by proposing and validating a multi-level evaluation framework against human experts. Our experimental results demonstrate that reasoning-focused models significantly outperform non-reasoning counterparts on logic-intensive tasks such as mathematics and physics, while the latter remain competitive on factual or observational problems (e.g., biology, geography). Specifically, our validated process analysis reveals distinct styles: DeepSeek-R1 generates more elaborate reasoning processes, correlating with greater effectiveness on harder tasks, despite potential redundancy. In contrast, GPT-o1 exhibits greater conciseness and frequently presents solutions with superior clarity and efficiency, particularly on simpler problems, while struggling with complex scenarios. Additionally, reasoning capabilities can be partially preserved through distillation, with distilled reasoning models consistently outperforming their non-reasoning counterparts across all domains. These findings provide new insights into when and how explicit reasoning contributes to LLM performance, offering practical guidance for model development and deployment.
Abstract Even though Transformers are the standard to extract information from text, they have considerable limitations that make them sub-optimal for certain tasks. Graph neural networks (GNNs) are an alternative that mitigate some of these limitations, with recent work focusing on combining information from pre-trained language models (PLMs) with graph structures. However, most existing methods use fixed-size sliding windows to construct graphs, ignoring long-distance relationships between words. We propose a memory-efficient method to construct token graphs that exploits semantic and relational information inside Transformers via their attention coefficients, mitigating the limitations of using sliding windows. Additionally, each graph can fully encode documents longer than the Transformer’s context window via a chunk-and-stride mechanism, while also reducing memory usage. Our method surpasses the performance of sliding windows when both approaches are compared under the same model architecture, particularly in longer documents, and reaches statistically identical performance in shorter ones. It also surpasses GNN-based state-of-the-art techniques. When compared with fine-tuning, our method reaches marginally lower performance in documents that fit inside the pre-trained model’s context window. However, our method surpasses fine-tuning in documents that do not fit inside these windows, which comes as a result of fully encoding each document using our chunk-and-stride mechanism.
Abstract Overcoming the trade-off between quality and latency is a challenge in simultaneous speech translation. Common approaches in previous work have been to segment the source sentence or align target sentences with the source’s syntax as closely as possible, enabling faster translations while maintaining quality. However, a major limitation in these studies is the reliance on existing translation test data, which often includes reordering and are unsuitable for simultaneous settings with low latency. Alternatively, some use interpretation data transcribed from interpreters, which is also problematic due to translation errors and omissions, making both inadequate for fully evaluating simultaneous models. In this work, we introduce a construction, verification, and analysis of a new test set specifically designed for simultaneous settings, with a focus on maintaining word and phrase order consistency with the source. The test set comprises three language pairs representing different levels of word order similarity to the source by leveraging large language models, with quality verified by professional interpreters. This provides an interpreter-grounded perspective, and empirically shows the ideal level of monotonicity and other sentence style characteristics including syntax simplicity and sentence length. It also reveals the capabilities and limitations of LLMs on monotonic translation. Experiments revealed that existing test data tends to underestimate a model’s performance, while the proposed test set, simul-tst-COMMON, offers a more appropriate evaluation of simultaneous models. Moreover, the quality gap between wait-k and Local Agreement suggests that the adaptive policy more closely resembles the monotonic translation behavior of human interpreters. Finally, the analysis highlights the limitations of current metrics, which may not be fully suitable for evaluating simultaneous tasks.
Abstract Mandelkern and Linzen (2024) argue that words generated by language models (LMs) are linked to causal histories of use within human linguistic communities, and ultimately to their referents. Therefore, LMs’ words refer. We qualify Mandelkern and Linzen’s claim as applicable to a narrow class of expressions. Thus qualified, the claim is valid and motivated by the need to evaluate LMs’ outputs for relevance and truth. Next, we discuss the actual scope of their claim, and we conclude that the bounds of sense and reference in LMs are more restricted than in humans. We close with some considerations on the status of LMs as members of human linguistic communities.
After the introduction of Large Language Models (LLMs), there have been substantial improvements in the performance of Natural Language Generation (NLG) tasks, including Text Summarization and Machine Translation. However, LLMs still produce outputs containing hallucinations, that is, content not grounded in factual information. Therefore, developing methods to assess the factuality of LLMs has become urgent. Indeed, resources for factuality evaluation have recently emerged. Although challenging, these resources face one or more of the following limitations: (i) they are tailored to a specific task or domain; (ii) they are limited in size, thereby preventing the training of new factuality evaluators; (iii) they are designed for simpler verification tasks, such as claim verification. To address these issues, we introduce LLM-Oasis, to the best of our knowledge the largest resource for training end-to-end factuality evaluators. LLM-Oasis is constructed by extracting claims from Wikipedia, falsifying a subset of these claims, and generating pairs of factual and unfactual texts. We then rely on human annotators to both validate the quality of our dataset and to create a gold standard test set for benchmarking factuality evaluation systems. Our experiments demonstrate that LLM-Oasis presents a significant challenge for state-of-the-art LLMs, with GPT-4o achieving up to 60% accuracy in our proposed end-to-end factuality evaluation task, highlighting its potential to drive future research in the field.
I was honored to receive the Association for Computational Linguistics Lifetime Achievement Award in 2025. I especially want to thank the people who nominated me for the award as I know nominations require time and effort. This retrospective is a rough transcript of the speech I gave accepting the award at the conference in Vienna, Austria. In the talk, I look back at my research at early stages of my career and then look at the arc that research takes and how it relates to work that I still carry out today. I look at the trajectories of four areas of my research: language generation, text summarization, social media analysis, and multimodal analysis of artwork. In the talk, I featured videos of my current students speaking about their research and where they think the field is heading. I dedicate the talk and this article to the amazing students I have had the honor to work with over the years.
Various encodings have been proposed to cast constituent parsing in terms of a sequence labeling task. However, unlike in the case of dependency parsing, existing comparisons have not been entirely homogeneous and, to the best of our knowledge, there is no systematic evaluation of these encodings under uniform configurations. A homogeneous evaluation needs to account for various aspects that could influence results, either by controlling for these aspects to ensure uniformity (e.g., network architecture, parameter settings, postprocessing of ill-formed output), or by systematically analyzing their impact (e.g., the impact of binary versus arbitrary structures). In this article, we: (1) compare different encodings comprehensively both theoretically and empirically, on a modern neural architecture and across nine languages, and (2) introduce new encodings and variants, including an encoding that our analysis finds particularly accurate and compact.
This article introduces the concept of multimodal oxymorons. Multimodal oxymorons extend the traditional oxymoron theory by constructing and communicating meaning through the interplay of multiple modalities (such as visual and textual) rather than relying solely on language. We argue that multimodal oxymorons are central mechanisms of meaning-making in contemporary communication, as evidenced by the use of memes as an example. While textual oxymorons have long been the subject of analysis in order to ascertain their role in shaping thought and meaning, multimodal oxymorons demonstrate how human cognitive process transcends linguistic boundaries, integrating different modalities (e.g., visual) in order to convey complex ideas. To encourage further study, we present a curated multilingual dataset of Multimodal OXYmoron (MOXY), which can be used as a foundation for further analysis and experimentation. Furthermore, we propose a methodical approach for the identification of multimodal oxymorons along with a pipeline for automated generation. Through illustrative examples and a detailed methodology, this work establishes a comprehensive framework for understanding, identifying, and generating multimodal oxymorons, paving the way for advancements in computational linguistics, artificial intelligence, and figurative language studies.
Recent advances in large language models have demonstrated promising capabilities in following simple instructions through instruction tuning. However, real-world tasks often involve complex, multi-step instructions that remain challenging for current NLP systems. Robust understanding of such instructions is essential for deploying LLMs as general-purpose agents that can be programmed in natural language to perform complex, real-world tasks across domains like robotics, business automation, and interactive systems. Despite growing interest in this area, there is a lack of a comprehensive survey that systematically analyzes the landscape of complex instruction understanding and processing. Through a systematic review of the literature, we analyze available resources, representation schemes, and downstream tasks related to instructional text. Our study examines 181 papers, identifying trends, challenges, and opportunities in this emerging field. We provide AI/NLP researchers with essential background knowledge and a unified view of various approaches to complex instruction understanding, bridging gaps between different research directions and highlighting future research opportunities.
Interpretability provides a toolset for understanding how and why language models behave in certain ways. However, there is little unity in the field: Most studies use ad-hoc evaluations and do not share theoretical foundations, making it difficult to measure progress and compare the pros and cons of different techniques. Furthermore, while mechanistic understanding is frequently discussed, the basic causal units underlying these mechanisms are often not explicitly defined. In this article, we propose a perspective on interpretability research grounded in causal mediation analysis. Specifically, we describe the history and current state of interpretability taxonomized according to the types of causal units (mediators) utilized, as well as methods used to search over mediators. We discuss the pros and cons of each mediator, providing insights as to when particular kinds of mediators and search methods are most appropriate. We argue that this framing yields a more cohesive narrative of the field and helps researchers select appropriate methods based on their research objective. Our analysis yields actionable recommendations for future work, including the discovery of new mediators and the development of standardized evaluations tailored to these goals.
Linguistic steganography (stego) seeks to conceal secret information within natural language text. However, existing methods often struggle to balance stego text quality with embedding efficiency, largely due to limitations in generation strategies and coding mechanisms. We propose SA-ANS, a self-adaptive linguistic steganography framework based on a self-adjusting Asymmetric Numeral System. SA-ANS allows user-specified embedding rates and uses probabilistic coding with adaptive candidate selection, dynamically tailoring the token pool to the language model’s probability distribution. This design produces fluent, semantically coherent stego text while preserving statistical indistinguishability from natural language. Extensive experiments on multiple benchmark datasets, evaluated across embedding efficiency, linguistic quality, statistical similarity, robustness to steganalysis, and human judgment, show that SA-ANS consistently outperforms state-of-the-art methods, demonstrating both effectiveness and practicality.
Removing personally identifiable information (PII) from texts is necessary to comply with various data protection regulations and to enable data sharing without compromising privacy. However, recent works show that documents sanitized by PII removal techniques are vulnerable to reconstruction attacks. Yet, we suspect that the reported success of these attacks is largely overestimated. We critically analyze the evaluation of existing attacks and find that data leakage and data contamination are not properly mitigated, leaving the question whether or not PII removal techniques truly protect privacy in real-world scenarios unaddressed. We investigate possible data sources and attack setups that avoid data leakage and conclude that only truly private data can allow us to objectively evaluate vulnerabilities in PII removal techniques. However, access to private data is heavily restricted - and for good reasons - which also means that the public research community cannot address this problem in a transparent, reproducible, and trustworthy manner.
This article investigates the ability of large language models (LLMs) to evaluate semantic relations between word pairs by examining their alignment with human-generated semantic ratings. Semantic relations represent the degree of connection (e.g., relatedness or similarity) between linguistic elements and are traditionally validated against human-annotated datasets. Due to the challenges of building such datasets and recent progress in LLMs' capacity to model humanlike understanding, we explore whether LLMs can serve as reliable substitutes for traditional human ratings. We conducted experiments using multiple LLMs from OpenAI, Google, Mistral, and Anthropic, evaluating their performance across diverse English and Portuguese semantic relations datasets. We included in the analysis PAP900, a recently published dataset of semantic relations in Portuguese, to examine the influence of prior exposure to the dataset on LLM training. The results show that the LLM predictions correlate strongly with human ratings. The findings reveal the potential of LLMs to supplement or replace traditional semantic measure algorithms and crowd-sourced human annotations in semantic tasks.
Natural Language Processing (NLP) metrics for bilingual code-switching (CS) have, until now, used words as the token level. However, the assumption that any two words constitute an equally likely switch point is erroneous. In spoken language, a major delimiter of CS is a prosodic chunk known as the Intonation Unit (IU). Switch points are far more likely between words at IU boundaries than between words in the same IU. The word as an elementary NLP unit is thus incommensurate with bilingual speech patterns. Here, we put forward an IU-based adaptation of a familiar metric of CS probability. We then compare the token levels on this metric for ten bilingual datasets featuring multi-word CS. Our comparison shows that the currently standard two-significant-figure precision of the word-based metric is insufficient, as the token level compresses the range of values by inflating the universe of CS. More discerning CS probability values can be obtained by normalizing word-based counts using mean IU length.
Recent years have seen an increase in polarized discourse worldwide, on various platforms. We propose a novel method for quantifying polarization, based on the emotional style of the discourse rather than on differences in ideological stands. Using measures of Valence, Arousal and Dominance, we detect signals of emotional discourse and use them to operationalize the concept of affective polarization. Applying this method to a recently released corpus of proceedings of the Knesset, the Israeli parliament (in Hebrew), we find that the emotional style of members of government differs from that of opposition members; and that the level of affective polarization, as reflected by this style, is significantly increasing with time.
