United States

We present H-RAG, our submission to SemEval-2026 Task 8 (MTRAGEval), addressing both Task A (Retrieval) and Task C (Generation with Retrieved Passages). Task A evaluates standalone retrieval quality, while Task C assesses end-to-end retrieval-augmented generation (RAG) in multi-turn conversational settings, requiring both accurate answer generation and faithful grounding in retrieved evidence. Our approach implements a hierarchical parent--child RAG pipeline that separates fine-grained child-level retrieval from parent-level context reconstruction during generation. Documents are segmented into overlapping sentence-based child chunks, while full documents are preserved as parent units to provide coherent context. weighting, and embedding-based similarity rescoring over child chunks. Retrieved evidence is aggregated at the parent level and supplied to an instruction-tuned language model for response generation. H-RAG achieves an nDCG@5 score of 0.4271 on Task A and a harmonic mean score of 0.3241 on Task C (RB_agg: 0.2488, RL_F: 0.2703, RB_llm: 0.6508), underscoring the importance of retrieval configuration and parent-level aggregation in multi-turn RAG performance.

ACL 2026

H-RAG at SemEval-2026 Task 8: Hierarchical Parent–Child Retrieval for Multi-Turn RAG Conversations

mtrageval

parent–child chunking

multi-turn rag

hierarchical retrieval

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 presents a novel agentic LLM pipeline for SemEval-2026 Task 10 that jointly extracts psycholinguistic conspiracy markers and detects conspiracy endorsement. Unlike traditional classifiers that conflate semantic reasoning with structural localization, our decoupled design isolates and addresses these challenges separately. For marker extraction, we propose Dynamic Discriminative Chain-of-Thought (DD-CoT) with deterministic anchoring to resolve semantic ambiguity and character-level brittleness. For conspiracy detection, an ``Anti-Echo Chamber`` architecture, consisting of an adversarial Parallel Council adjudicated by a Calibrated Judge, overcomes the ``Reporter Trap``, where models falsely penalize objective reporting. Our system achieves 0.24 Macro F1 (+100% over baseline) on S1 and 0.79 Macro F1 (+49%) on S2, ranking 3rd on the S1 development leaderboard and 8th on the test set, demonstrating that structured agentic deliberation is an effective alternative to fine-tuning for interpretable psycholinguistic NLP.

AILS-NTUA at SemEval-2026 Task 10: Agentic LLMs for Psycholinguistic Marker Extraction and Conspiracy Endorsement Detection

In this paper, we present AILS-NTUA system for Track-A of SemEval-2026 Task 3 on Dimensional Aspect-Based Sentiment Analysis (DimABSA), which encompasses three complementary problems: Dimensional Aspect Sentiment Regression (DimASR), Dimensional Aspect Sentiment Triplet Extraction (DimASTE), and Dimensional Aspect Sentiment Quadruplet Prediction (DimASQP) within a multilingual and multi-domain framework. Our methodology combines fine-tuning of language-appropriate encoder backbones for continuous aspect-level sentiment prediction with language-specific instruction tuning of large language models using LoRA for structured triplet and quadruplet extraction. This unified yet task-adaptive design emphasizes parameter-efficient specialization across languages and domains, enabling reduced training and inference requirements while maintaining strong effectiveness. Empirical results demonstrate that the proposed models achieve competitive performance and consistently surpass the provided baselines across most evaluation settings.

AILS-NTUA at SemEval-2026 Task 3: Efficient Dimensional Aspect-Based Sentiment Analysis

This paper presents the NYCU-NLP system for SemEval-2026 Task 9 on online polarization analysis. Our approach explores the effectiveness of instruction-tuned small language models (SLMs), including Phi-4 (14B), Mistral-small-3.2 (24B), and Gemma-3 (27B), with task-specific prompting strategies and combined them via a stacking ensemble to leverage complementary modeling capacities. Evaluated across 22 languages and three subtasks, our system achieved macro-averaged F1 scores of 0.8071 for Polarization Detection (Subtask 1), 0.6108 for Polarization Type Classification (Subtask 2), and 0.5111 for Polarization Manifestation Identification (Subtask 3). Notably, our approach ranked first in 15, second in 12, and third in 10 of the 62 language-specific leaderboards, demonstrating the robustness and competitiveness of stacking-based SLM ensembles in multilingual settings.

NYCU-NLP at SemEval-2026 Task 9: Stacking Small Language Models for Multilingual, Multicultural and Multievent Polarization Detection

We describe the AILS-NTUA system for SemEval-2026 Task 8 (MTRAGEval), addressing all three subtasks of multi-turn retrieval-augmented generation: passage retrieval (A), reference-grounded response generation (B), and end-to-end RAG (C).

Our approach is based on two main design principles. First, we adopt a query-diversity-over-retriever-diversity strategy, where multiple complementary LLM-based query reformulations are issued to a single corpus-aligned sparse retriever and combined using a variance-aware nested Reciprocal Rank Fusion scheme. Second, we employ an agentic generation pipeline that decomposes grounded response generation into evidence span extraction, dual-candidate drafting, and calibrated multi-judge selection.

The proposed system achieves strong performance across subtasks, ranking first in Task A and second in Task B in the official evaluation. Our empirical findings indicate that query diversity over a well-aligned retriever is more effective than heterogeneous retriever ensembling, and that answerability calibration—rather than retrieval coverage—emerges as the primary bottleneck in end-to-end performance.

AILS-NTUA at SemEval-2026 Task 8: Evaluating Multi-Turn RAG Conversations

This paper presents our system for SemEval-
2026 Task 3, which identifies four-part opinion
details in product reviews. We used a sequence
of pairs of BERT encoder models connected
by cross-attention layers. The cross-attention
mechanism provided marginally better results
than a self-attention equivalent, failing to show-
case a significant improvement. Error propaga-
tion through the pipeline hurt the correctness of
the outputs, with certain stages collapsing the
scores. The pipeline architecture's performance
was largely independent of model size, sug-
gesting that small modular encoders for down-
stream tasks are an efficient alternative to large
decoder models. Our best model got a cF1
score of 0.53 on restaurant data and 0.26 on
laptop data.

Team BOBW (Best Of Both Worlds) at SemEval-2026 Task 3: Modular Cross-Attention Encoders for Dimensional Aspect-Based Sentiment Analysis

This paper describes our system for SemEval-2026 Task 3, which focuses on predicting continuous valence and arousal scores. The task poses significant challenges due to variations in data scale and pragmatic ambiguities across languages. To address these disparities, we propose an Adaptive Dual-Track Framework that dynamically selects modeling strategies based on task characteristics. For semantically stable tasks, we apply a robust single baseline optimized with layer-wise learning rate decay (LLRD) to ensure stability. For high-ambiguity scenarios such as the Environmental Protection domain, we adopt a heterogeneous ensemble strategy to mitigate prediction variance. Experimental results demonstrate that our system consistently outperforms the initial standard baseline across all subtasks. Furthermore, our lightweight approach exhibits remarkable parameter efficiency, achieving highly competitive performance against newly introduced large language model (LLM) baselines. Additionally, ablation studies reveal that under regression settings, conventional regularization techniques, cross-lingual data transfer, and homogeneous ensemble learning can lead to negative transfer, confirming the necessity of strategically diverging approaches tailored to linguistic characteristics.

SCU_Mesclab at SemEval-2026 Task 3: An Adaptive Dual-Track Framework for Dimensional Aspect-Based Sentiment Analysis

This paper describes our system for the SemEval-2026 Task 3: Dimensional Aspect-Based Sentiment Analysis (DimABSA). We participate in Track A (DimABSA) and Track B (DimStance), both of which involve Subtask 1 – predicting continuous valence–arousal (VA) scores for given text–aspect pairs in English and Chinese.
Our system combines pre-trained multilingual transformers with aspect-marker input encoding and dual regression heads for VA prediction, trained with a 5-fold cross-validation ensemble. We select XLM-RoBERTa-large as the backbone for Track A and mDeBERTa-v3-base for Track B based on systematic model comparison on the development sets. On the official test sets, our system substantially outperforms the organizer-provided baselines across all language domain settings. On the unofficial postevaluation leaderboard, the system achieves strong results on Chinese subsets, ranking 1st on zho-env (Track B) and 2nd on zho-fin (Track A).

YangS_team at SemEval-2026 Task 3: Transformer-Based Aspect-Aware Regression for Dimensional Sentiment and Stance Analysis

We present our system for SemEval-2026 Task 9: Multilingual Polarization Detection, a binary classification task spanning 22 languages. Our approach fine-tunes separate Gemma 3 models (12B and 27B parameters) per language using Low-Rank Adaptation (LoRA), augmented with synthetic data generated by a large language model (LLM). We employ three synthetic data strategies (direct generation, paraphrasing, and contrastive pair creation) using GPT-4o-mini, with a multi-stage quality filtering pipeline including embedding-based deduplication. We find that per-language threshold tuning on the development set yields 2 to 4% F1 improvements without retraining. We also use weighted ensembles of 12B and 27B model predictions with per-language strategy selection. Our final system achieves a mean macro-F1 of 0.811 across all 22 languages, ranking 2nd overall out of 60 participating teams, with 1st place finishes in 2 languages and top-3 in 8 languages. We also find that alternative architectures (XLM-RoBERTa, Qwen3) that showed strong development set performance suffered 30 to 50% F1drops on the test set, highlighting the importance of generalization.

PSK at SemEval-2026 Task 9: Multilingual Polarization Detection Using Ensemble Gemma Models with Synthetic Data Augmentation

This paper describes Team SoloSemantics' submissions to SemEval-2026 Task 4: Narrative Story Similarity and Narrative Representation Learning. We began with lightweight neuro-symbolic knowledge-graph baselines, but a triplet-tuned MPNet bi-encoder produced stronger semantic separation in our experiments. We adopted a shared dense encoder family across both tracks and kept the KG and fusion variants as diagnostic baselines. Team SoloSemantics ranked 22nd on Track A and 9th on Track B. Our reproducibility audit further shows that the KG branch was often too sparse on short summaries to represent abstract narrative relations reliably under the current extraction pipeline.

SoloSemantics at SemEval-2026 Task 4: Triplet-Tuned MPNet for Story Similarity

SemEval-2026 Task 3 (DimABSA) includes Dimensional Aspect Sentiment Quadruplet Extraction (DimASQP), which requires extracting structured tuples—aspect term, aspect category, and opinion term—together with continuous valence–arousal (VA) values from reviews (Yu et al., 2026a). In this work, we participate in Track A, Subtask 3. We describe NYCU Speech Lab's submission for the Chinese Restaurant and Laptop domains. Our system is a post-processing ensemble over heterogeneous architectures: LoRA/QLoRA fine-tuned decoder-only LLMs, a fine-tuned encoder-only model, and (optionally) prompted API-based LLMs. To improve robustness under the continuous F1 (cF1) metric, we use validation-calibrated weighted voting for tuple selection and weighted VA fusion for numerical aggregation, with strict output validation to enforce task constraints. Experiments on a held-out validation split show consistent gains over single models and clarify the precision–recall trade-offs induced by the voting threshold. On the organizers' released (tentative) test leaderboard snapshot, our submission ranks first in both domains.

NYCU Speech Lab at SemEval-2026 Task 3: Heterogeneous Model Ensemble with Adaptive Weighted Voting for Dimensional Aspect Sentiment Quadruplet Extraction

Downloads

Next from ACL 2026

AILS-NTUA at SemEval-2026 Task 10: Agentic LLMs for Psycholinguistic Marker Extraction and Conspiracy Endorsement Detection

Similar lecture

IITKanBDone at SemEval-2026 Task 8: MTRAGEval - Evaluating Multi-Turn RAG Conversations

.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

AILS-NTUA at SemEval-2026 Task 10: Agentic LLMs for Psycholinguistic Marker Extraction and Conspiracy Endorsement Detection

Similar lecture

IITKanBDone at SemEval-2026 Task 8: MTRAGEval - Evaluating Multi-Turn RAG Conversations

Downloads