United States

RAG systems retrieve documents optimized for answering *one query at a time*. Yet enterprise users arrive with *sessions*, that is, coherent episodes of related questions that span semantically distant parts of the knowledge base. We show that a single retrieval call over a standard knowledge base covers only 41% of a user&#39;s session-level information need. To close this gap, we reorganize the KB offline using co-occurrence-aware clustering and expand retrieval candidates through cluster neighborhoods at query time. On WixQA (6,221 enterprise support articles), our method raises single-query session coverage to 58% (+17% absolute; 95% CI: [14.1, 20.4]), reduces retrieval calls to 70% coverage by 34%, and compresses the KB to 20% of its original size, all consistently across four embedding models and six functional domains. We argue that session-level coverage, not single-query recall, should be the primary metric for enterprise RAG evaluation.

ACL 2026

One Retrieval to Cover Them All: Co-occurrence-Aware Knowledge Base Reorganization for Session-Level RAG

knowledge base reorganization

session-level evaluation

retrieval-augmented generation

## Welcome to EMNLP 2024! 
We are excited to welcome you to one of the most prominent conferences in the field of Natural Language Processing. This year, EMNLP 2024 is being held in a hybrid format,
offering both virtual and in-person participation in beautiful Miami. Due to a record-breaking number of submissions, we've expanded the total number of accepted papers to accommodate more cutting-edge research from around the globe.
### [Conference Handbook](https://drive.google.com/file/d/1WPROgxjLAC96AJL7Ugy0tEnYm7dkrbHt/view?usp=sharing)

You are required to register for this event. **Please register [here](https://2024.emnlp.org/registration/).** The EMNLP 2024 event page on Underline will be open to public one week prior to the event.

Please register!

EMNLP 2024

EMNLP 2024 will take place in Miami, Florida from Nov 12th to Nov 16th, 2024, at the Hyatt Regency Miami Hote and on Underline for remote participants.

RAG systems retrieve documents optimized for answering *one query at a time*. Yet enterprise users arrive with *sessions*, that is, coherent episodes of related questions that span semantically distant parts of the knowledge base. We show that a single retrieval call over a standard knowledge base covers only 41% of a user's session-level information need. To close this gap, we reorganize the KB offline using co-occurrence-aware clustering and expand retrieval candidates through cluster neighborhoods at query time. On WixQA (6,221 enterprise support articles), our method raises single-query session coverage to 58% (+17% absolute; 95% CI: [14.1, 20.4]), reduces retrieval calls to 70% coverage by 34%, and compresses the KB to 20% of its original size, all consistently across four embedding models and six functional domains. We argue that session-level coverage, not single-query recall, should be the primary metric for enterprise RAG evaluation.

workshop paper

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. 
[Full Message from General Chair Philipp Koehn](https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing) 
<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>

To access this event page, you need to log in with the **email address you registered with**. 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.

Object recognition models are well studied on benchmark datasets, typically focusing on performance in retrieving objects that exist in images. However, in real-life scenarios there is no prior knowledge of an object's existence, and current research fails to assess model performance in these situations. This research aims to shed light on this problem by testing three Open-World models, YOLO-World, Grounding Dino and GPT-4o, on the LVIS, Open Images, and JUS datasets. We design an experiment where models are confronted with impossible prompts by instructing them to retrieve non-existing objects. This allows us to observe the models' uncertainty performance. Overall, GPT-4o performed poorest with regard to object recognition and uncertainty estimation. GPT-4o showed to be highly overconfident. In contrast, YOLO-World and Grounding Dino are slightly underconfident, but they are superior in their uncertainty calibration in comparison to GPT-4o. However, all three models occasionally assign high confident predictions to non-existing objects. Showing that improvement can still be made to the uncertainty estimation of these models when confronted with impossible prompts.

Tricking Open-World Object Recognition Models: Uncertainty in Out-of-Distribution Detection

In factual question answering, many errors are not failures of access but failures of commitment: the system retrieves relevant evidence, yet still settles on the wrong answer. We present CounterRefine, a lightweight inference-time repair layer for retrieval-grounded question answering. CounterRefine first produces a short answer from retrieved evidence, then gathers additional support and conflicting evidence with follow-up queries conditioned on that draft answer, and finally applies a restricted refinement step that outputs either KEEP or REVISE, with proposed revisions accepted only if they pass deterministic validation. In effect, CounterRefine turns retrieval into a mechanism for testing a provisional answer rather than merely collecting more context. On the full SimpleQA benchmark, CounterRefine improves a matched GPT-5 Baseline-RAG by 5.8 points and reaches a 73.1 percent correct rate, while exceeding the reported one-shot GPT-5.4 score by roughly 40 points. These findings suggest a simple but important direction for knowledgeable foundation models: beyond accessing evidence, they should also be able to use that evidence to reconsider and, when necessary, repair their own answers.

CounterRefine: Answer-Conditioned Counterevidence Retrieval for Inference-Time Knowledge Repair in Factual Question Answering

Large language models (LLMs) are commonly described as acquiring knowledge through large scale pretraining on textual corpora. This view underestimates the epistemic consequences of post training safety mechanisms. Modern LLMs undergo extensive safety alignment via curated datasets, human annotations, and reinforcement learning from human feedback (RLHF), processes that do not merely constrain outputs but actively restructure how propositional and procedural knowledge is represented, accessed, and expressed. We argue that safety alignment should be understood as a systematic form of knowledge editing at scale. Annotation frameworks used to construct safety datasets function as normative ontologies that partition language into categories of acceptable and unacceptable content, and alignment training embeds these categorical distinctions into model parameters. We introduce the Safety Knowledge Pipeline (SKP), a four stage conceptual framework describing how pretraining knowledge is progressively filtered, reframed, and constrained through annotation and alignment mechanisms. We identify three mechanisms of knowledge modification: suppression, reframing, and substitution, each with distinct diagnostic signals, and we operationalize them in a cross lingual evaluation protocol. Case studies spanning harmful instruction queries, hate speech annotation in Arabic dialects, and culturally variable discourse illustrate the framework. We further discuss how treating annotator disagreement as a training signal rather than noise can mitigate the culturally hegemonic effects of current alignment pipelines. Our analysis has direct implications for LLM factuality, cross cultural equity, and the governance of AI safety systems.

Annotation Frameworks Shape Model Knowledge: Safety Alignment in Large Language Models

We present an annotation schema for Haitian Creole medical translation that makes clinical risk and sociolinguistic variation explicit while remaining lightweight enough for small expert teams. The schema includes binary fields for overall acceptability, severity of potential misunderstanding, and foreign-influence cues, along with conditional error tags aligned with Multidimensional Quality Metrics (MQM), commonly used in the medical domain, for interoperability. Through three rounds of annotation and adjudication we achieve stable inter-annotator agreement and release a gold dataset of 152 EN$\rightarrow$HC medical sentence pairs. A simple classifier–labeller baseline demonstrates that acceptability and severity are reliably learnable under data scarcity, while foreign-influence judgments remain limited by prevalence. These results show that clinically oriented, variety-sensitive annotation can both support immediate screening of patient-facing translations and provide reward-ready signals for future preference-based MT and LLM fine-tuning.

Annotating Clinical Risk and Variation in Haitian Creole Medical Translation

We propose a simplified human-in-the-loop workflow for second language (L2) Korean morphosyntactic annotation by leveraging agreement between two domain-adapted parsers. We first evaluate whether parser agreement can serve as a proxy for annotation correctness by comparing it with independent human judgments. The results show strong correspondence between parser and human judgments, supporting the feasibility of semi-automatic L2-Korean UD annotation. Further analysis demonstrates that parser disagreements cluster in linguistically predictable domains such as grammatical-relation distinctions and clause-boundary ambiguity. While many disagreement cases are tractable for iterative model refinement, others reflect deeper representational challenges inherent in parsing and tagging L2-Korean corpora.

Parser agreement and disagreement in L2 Korean UD: Implications for human-in-the-loop annotation

In this paper, we present a set of linguistic rules, employed to enhance the readability of legal texts. The rules were compiled and implemented as a rule-based module of PONK, an advisory tool that contributes to simplification and higher clarity of Czech legal texts, especially those destined to non-expert audience. Based on recurring phenomena in authentic texts and relevant scientific sources, the rules cover mainly the domains of syntax and lexicon. In addition, we present the results of application of the rules to a corpus of authentic legal texts, evaluated by a human annotator, and examine their impact.

Rules-based system for Czech legal text readability

Large annotated datasets inevitably contain errors that are costly to identify via manual review. We study a human-AI annotation error auditing workflow using frontier Large Language Models (LLMs), focusing on Hebrew diacritization. We take the the EACL 2023 Hebrew Homograph Challenge Set as our test case. In a focused evaluation on 12 of the homograph sets with 271 confirmed errors (verified through exhaustive manual review of all 7,241 sentences), Gemini 3 Pro achieves 83.6% recall (95% Confidence Interval: [79.3%, 88.2%]) and 99.1% precision - substantially higher than other frontier LLMs. Furthermore, two independent human experts achieved 62.4% and 42.8% recall respectively, a 20-percentage-point spread that reflects the difficulty of sparse-target error search. Even the union of both experts' findings (73.4% recall) falls short of a single LLM run (83.6%), while LLM-aided auditing reduces review effort by over 95%. We analyze the trade-offs between batch size and recall, and release both a human-verified Gold Standard with per-error difficulty annotations and a globally corrected version of the Challenge Set.

Human-AI Annotation Error Auditing for Hebrew Diacritization with Frontier LLMs

Annotation errors in hate speech corpora are often attributed to annotator disagreement or bias. This paper argues that a substantial and underexamined class of errors originates upstream, from structural weaknesses in annotation guidelines themselves. When guidelines fail to encode the linguistic and cultural properties of the target discourse, they make certain errors structurally inevitable regardless of annotator quality. Focusing on Arabic social media discourse, a challenging setting due to its dialect continuum, culturally embedded insult conventions, sarcasm-heavy pragmatics, and complex religious rhetoric, we identify three mechanisms through which guideline design produces systematic annotation errors: cultural misclassification, when culturally specific hostile expressions fall outside annotation categories; dialectal ambiguity, when lexical meanings shift across regional varieties; and annotation projection, when frameworks developed for English moderation are applied to Arabic without adequate adaptation. Using six illustrative case studies with attested Arabic examples, we show how these mechanisms produce recurrent misannotations in existing datasets. We propose a taxonomy of five guideline-induced error types, an explicit mapping from mechanisms to error types, and a practical four-stage diagnostic framework for dataset builders.

Beyond Annotator Disagreement: Guideline-Induced Errors in Arabic Hate Speech Annotation

Accurate linguistic annotation is crucial for creating high-quality datasets in specialized domains, yet manual labeling is often slow, expensive, and inconsistent. We present a reproducible workflow for evaluating the effectiveness of large language models (LLMs) as annotators of domain-specific health misinformation on social media. Using a data set of 169 Instagram posts on seed oils, expert nutritionists provided gold-standard labels (71% positives), which we compared against the outputs of five open-source LLMs. We introduce a hierarchical error taxonomy that categorizes LLM misclassifications according to the direction, mechanism, and contributing factors of the error, providing interpretable insights into model failures. Our analysis reveals systematic error patterns, including misinterpretation of nuanced claims and overconfidence in predictions, highlighting conditions under which LLM annotations do not align with expert judgment. Although the data set is modest in size and exhibits class imbalance, it reflects real‑world distributions of nutrition‑related Instagram content and motivates the need for a careful evaluation of the robustness of the LLM annotation. This study has implications for the development of frameworks for automated LLM-based annotators in the health and nutrition domains, as well as LLM developers in general.

When LLMs Disagree with Human Experts: Understanding LLM Annotation Failures in Nutrition Misinformation through Hierarchical Error Analysis using Seed Oil Narratives

Large Language Models have demonstrated significant progress in solving mathematical word problems through techniques like Chain-of-Thought (CoT) prompting. However, recent research indicates that these models often rely on statistical regularities and surface-level patterns rather than true logical reasoning, leading to performance drops when faced with minor problem perturbations or irrelevant information. In this study, we introduce Math Discourse Bank (Math-DB), a novel discourse framework and annotated dataset designed to enhance LLM reasoning. Inspired by the Penn Discourse TreeBank (PDTB) and mathematics education research, Math-DB defines a hierarchy of discourse senses designed for quantitative reasoning, including categories such as Change, Combine, Compare, and Equalize. We applied this framework to the GSM-Symbolic dataset of 12,500 problems, yielding 47,815 sense-labeled discourse relations over 11,414 successfully-aligned instances (91.3\% pipeline yield). Our experiments demonstrate that incorporating Math-DB annotations into CoT prompts consistently improves LLM performance across various difficulty levels.

Downloads

Next from ACL 2026

Tricking Open-World Object Recognition Models: Uncertainty in Out-of-Distribution Detection

Similar lecture

Question-Based Retrieval using Atomic Units for Enterprise RAG

.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

Tricking Open-World Object Recognition Models: Uncertainty in Out-of-Distribution Detection

Similar lecture

Question-Based Retrieval using Atomic Units for Enterprise RAG

Downloads