United States

In this paper, we perform an error analysis on human and LLM annotation data from the recent GUMBridge corpus for varieties of bridging anaphora. We explore the distribution of precision and recall errors made by annotators and how that distribution correlates with bridging subtypes. We find that while LLMs perform substantially worse than human annotators, they are more balanced in their precision and recall scores than humans, whose performance strongly favors precision. With regard to subtypes, we find that comparison and meronomy relations are easier to reliably annotate than the more broadly construed entity relations for both human and LLM annotators, but that LLM errors are more distributed across subtypes than human errors. Analyzing these results, we provide insights for future annotation projects on bridging anaphora.

ACL 2026

Cracks in the Bridge—or A Bridge Too Far? Comparing Human and LLM Errors in the Annotation of Bridging Anaphora

error annalysis

bridging anaphora

llm annotation

llms

human annotation

bridging

annotation

It is my great pleasure and honor to welcome you to the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026), held in San Diego, California, USA, July 2–7, 2026. ACL 2026 continues our fields tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.

It is exciting to see the many developments in the field reflected at our conference, such as the advancement of large language models (born out of our community) into many application directions. This is not without challenges, partly due to unsolved problems, and partly due to misuse of the technology. All this is reflected well at ACL 2026. The conference also reflects that the field is not only about advancing technology, but also the science of fundamentally understanding language and how it is use by many different speaker communities. <br>
[Full Message from General Chair Philipp Koehn](https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing) <br><br><br>
<html><button style="display: inline-flex; align-items: center; justify-content: center; white-space: nowrap; border-radius: 9999px; font-weight: bold; background: #7c3aed; color: white; font-family: 'Space Grotesk', sans-serif; height: 40px; font-size: 16px; padding: 0 20px; border: none; cursor: pointer" onclick="window.open('https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing')">Conference Handbook</button></html>
<br><br><br><br>

To access this event page, you need to log in with the **email address you registered with**. <br>Access credentials will be sent to your email from Underline -  subject line "Welcome to ACL 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://2026.aclweb.org/registration/) first.

Registration Required

The 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026) will take place in San Diego, California, from July 2 - 7, 2026.

workshop paper

This paper introduces a method for the automatic detection of annotation errors and corrections using a novel two-stage dimension reduction of dense sentence embeddings. We first find the top-\textit{n} principal components of an embedding and then use UMAP for second-stage, non-linear dimension reduction and clustering using different distance metrics. We analyze these clusters using silhouette scores to flat outlier mentions for correction. Using the corrections in the CoNLL\# dataset as a benchmark, all of the top five outliers needed correction, as did 7 of the top 10, and 32 of the top 50. This method offers a relatively low-effort way to leverage text embeddings and dimensionality reduction to identify likely annotation errors.

Clustering Analysis for Error Detection in Named Entity Recognition Datasets

Linguistic annotation of high-stakes narrative data is often constrained by data confidentiality, domain expertise, and the lack of large-scale multi-annotator pipelines. We present a human-in-the-loop framework for auditing annotation discrepancies in crash narratives, combining structured labels, narrative-based annotation, and expert adjudication. Using 9,387 crash reports, we conduct a multi-layer analysis of disagreement across annotation sources. Nearly half of the records (49.4\%) exhibit discrepancies between structured and narrative labels, driven mainly by unsupported structured assignments. In contrast, narrative-based annotation achieves near-perfect agreement with adjudication ($\kappa = 0.990$), indicating strong consistency when grounded in textual evidence. We introduce a taxonomy of discrepancies, showing refinement opportunities and missing details are the most common, while linguistic factors such as hedging and underspecification contribute to ambiguity. We further show that annotator-reported uncertainty strongly predicts annotation difficulty, with uncertain records nearly nine times more likely to disagree with structured labels. These findings highlight limitations of administrative coding and support a scalable, uncertainty-guided annotation paradigm for restricted-access domains.

When Ground Truth Disagrees: A Human-in-the-Loop Audit of Annotation Errors in High-Stakes Crash Narratives

The behavior of contemporary generative Large Language Models (LLMs) is directly shaped by prompts, unstructured texts that describe the desired output and model behavior. In this paper we argue that prompts are linguistic objects that merit investigation in their own right. To this end, we collect 57.5K unique samples of prompts from GitHub. Specifically, we focus on transactional prompts: reproducible natural language instructions that are integrated into software. To enable the empirical, quantitative study of prompts, we introduce a structured ontology, capturing the properties of prompts as well as their formal and semantic. components. Based on this ontology, we transform prompts from unstructured raw texts into richly structured linguistic objects. Analysis of these structured data reveals significant diversity of usage patterns across languages, domains, tasks, and modalities, in a typical Zipf-like distribution where some clearly prevail and others, more diverse, appear in the long tail. To validate the reliability of the ontology-based annotation of the prompts, we perform a comprehensive error analysis across all fields, providing a detailed assessment of annotation quality. We release the dataset together with a browsing and exploration interface.

Prompts in the Wild: A Large Analyzed Collection of Transactional Prompts in Code

Fine-grained morphosyntactic error annotation is important in clinical and developmental language research, yet it is labour-intensive, expert-dependent, and difficult to scale. We present TalkTag, an LLM-based lightweight tool fine-tuned to automate CHAT-style error annotation in spoken-language transcripts. Developed under conditions of extreme data scarcity using children’s narrative data, the system shows the feasibility of linguistic analysis in low-resource settings. Our evaluation demonstrates that TalkTag produces encouragingly precise annotation while effectively identifying instances where linguistic ambiguity makes automated tagging genuinely complex. In summary, with TalkTag, we provide a scalable alternative to manual error annotation and practically viable support for morphosyntactic error annotation.

TalkTag: Fine-Grained Morphosyntactic Error Annotation for Transcribed Speech

High-quality datasets are critical for training and evaluating reliable NLP models. In tasks like natural language inference (NLI), human label variation (HLV) arises when multiple labels are valid for the same instance, making it difficult to separate annotation errors from plausible variation. An earlier framework VARIERR (Weber-Genzel et al., 2024) asks multiple annotators to explain their label decisions in the first round and flag errors via validity judgments in the second round. However, conducting two rounds of manual annotation is costly and may limit the coverage of plausible labels or explanations. Our study proposes a new framework, EVADE, for generating and validating explanations to detect errors using large language models (LLMs). We perform a comprehensive analysis comparing human- and LLM-detected errors for NLI across distribution comparison, validation overlap, and impact on model fine-tuning. Our experiments demonstrate that LLM validation refines generated explanation distributions to more closely align with human annotations, and that removing LLM-detected errors from training data yields improvements in fine-tuning performance than removing errors identified by human annotators. This highlights the potential to scale error detection, reducing human effort while improving dataset quality under label variation.

Evade: LLM-Based Explanation Generation and Validation for Error Detection in NLI

We present semantic-pragmatic specification and annotation (ellipsis, coreference, bridging and discourse relations, information structure, scope of negation) in the multi-layer, genre-diversified, 3+ million-token Enriched PDT. While morphology and syntax work almost exclusively on sentence level, the semantic-pragmatic phenomena are often related to two or more neighbouring sentences and possibly to an extra-linguistic context. In the contribution, we describe these phenomena from both the linguistic perspective (form of expression, relation to syntax and morphology) and the cognitive perspective (relation to context, real world knowledge, as well as to the related processes such as thinking or reasoning) -- classifying the possible relations between the semantic-pragmatic units into cognitively plausible, distinguishable, and human-understandable categories. We have applied our results to the corpus, by annotating it in its entirety. The resulting corpus is publicly and freely available, to serve for verification and further investigation of (not only) these phenomena.

Semantic-pragmatic Annotations in the Prague Dependency Treebank

Large-scale ASR systems such as Whisper achieve competitive aggregate Word Error Rate (WER) on multilingual benchmarks, but this aggregate conceals systematic disparities across speaker populations. We evaluate Whisper large-v3 on 276 recordings from the \textit{Corpus Oral y Sonoro del Espa\~{n}ol Rural} (COSER), a dialectological archive of elderly rural speakers across all Spanish provinces. WER is computed separately for Informants and Interviewers within each recording, revealing that mixed-role evaluation underestimates Informant WER in the majority of provinces, with the largest corrections in southern areas. Negative Binomial regression with cluster-robust standar errors shows that Andalusia and Extremadura generate significantly more Informant errors than the Castilian heartland (Andalusia IRR\,=\,1.20, $p < 0.001$; Extremadura IRR\,=\,1.24, $p = 0.020$), while no geographic predictor reaches significance for Interviewers sharing the same recording environment. Male Informants generate 12.5\% more errors than females after geographic adjustment ($p < 0.001$), consistent with differential vernacular retention in traditional rural communities. The geographic pattern aligns with established dialectological classifications of Peninsular Spanish.

Beyond Acoustics: Isolating Dialectal and Sociolinguistic Bias in Spanish ASR

The rise of online platforms has facilitated covert forms of explicit content, which pose significant challenges for detection and regulation. Often using coded language to bypass moderation, this content erodes user trust and may be associated with scam-related risks, posing direct financial and personal risks. In this study, we map the landscape of online explicit content posts, focusing on their categorization, linguistic strategies, and temporal and behavioral patterns as they appear within mainstream platform reddit. We investigated five distinct content categories including Virtual Services (VS), Physical Services (PS), Exhibitionism (Ex), Couples and Group Interactions (CGI), and Content Creation and Sales (CCS) and performedmed large-scale experimentation using state-of-the-art large language models (LLMs) such as GPT-4, LLaMA 3.3-70B-Instruct, Gemini 1.5 Flash, Mistral 8×7B, Qwen 2.5 Turbo, and Claude 3.5 Haiku. Our work demonstrates that a nuanced classification of these services requires moving beyond simple keywords, and we establish that expressive signals such as sentiment, emotion, and tone are critical features for accurate detection. Our analysis reveals the distinct behavioral and psychosocial expression patterns that characterize each service category, providing a robust framework for future moderation.

Mapping the Landscape of Unregulated eXplicit Contents on Reddit

This paper offers a new perspective on the computational complexity of reduplication. Since one-way deterministic transducers cannot model reduplication in a straightforward way, the phenomenon has long been considered the outlier of morphology from a complexity perspective. Drawing on algebraic methods, I show that the vast majority of reduplicative processes belong to a few remarkably simple
classes of subregular functions. A detailed study of the RedTyp database (Dolatian and Heinz, 2019) reveals that 100% of the
surveyed reduplicative processes correspond to string-to-string functions in the class DA, while over 98% are locally testable (LJ1 ) and over
87% are locally trivial (L1). These results indicate a new upper bound on the complexity of reduplication that is comparable to that of morphological processes in general.

Algebraic Classification of Reduplicative Processes

Cue-based interference theory predicts that distractors similar in syntax or semantics can disrupt retrieval. Predictors derived from attention weights in selective heads or aggregated across layers remain opaque to support clear mechanistic interpretations. We propose a transparent, factorized alternative that models retrieval as an asymmetric linear transformation from a cue space to a target space along syntactic and semantic dimensions. On the materials manipulating syntactic similarity and semantic plausibility, the two submodels yield independent interference signals. The results demonstrate that a simple linear combination of these signals reproduces the graded interference pattern in the experimental design while providing a clearer and more mechanistically interpretable account of cue-based retrieval than attention-based predictors from large neural models.

Downloads

Next from ACL 2026

Clustering Analysis for Error Detection in Named Entity Recognition Datasets

Similar lecture

What's in a Bridge?: A Descriptive, Multi-Genre Analysis of the GUMBridge Corpus for Varieties of Bridging Anaphora

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from ACL 2026

Clustering Analysis for Error Detection in Named Entity Recognition Datasets

Similar lecture

What's in a Bridge?: A Descriptive, Multi-Genre Analysis of the GUMBridge Corpus for Varieties of Bridging Anaphora

Downloads