Multi-turn retrieval-augmented generation (RAG) is challenging due to evolving user intent, conversational noise, and strict context limits. We propose a training-free hybrid retrieval pipeline for SemEval-2026 Task 8 that combines dense and sparse retrieval with controlled query rewriting and cross-encoder reranking. On the official test set of Task A, our system achieves 0.5453 nDCG@5, ranking third among 38 teams and outperforming the strongest baseline score of 0.4795. For Task C, we reuse the documents retrieved for Task A and apply a lightweight generation pipeline guided by the official prompt, achieving 0.5312 as the harmonic mean of relevance and faithfulness and ranking 15th among 29 teams. All retrieval components are open-source, while query rewriting and answer generation rely on LLM APIs.

Large language models (LLMs) offer significant potential for constructing commonsense knowledge graphs from text, demonstrating adaptability across diverse domains. However, their effectiveness varies significantly with domain-specific language, highlighting a critical need for specialized benchmarks to assess and optimize knowledge graph construction sub-tasks like named entity recognition, relation extraction, and entity linking. Currently, domain-specific benchmarks are scarce. To address this gap, we introduce SynEL, a novel benchmark developed for evaluating text-based knowledge extraction methods, validated using customer support dialogues. We present a comprehensive methodology for benchmark construction, propose two distinct approaches for generating synthetic datasets, and evaluate accumulated hallucinations. Our experiments reveal that existing LLMs experience a significant performance drop, with micro-F1 scores decreasing by up to 25 absolute points when extracting low-resource entities compared to high-resource entities from sources like Wikipedia. Furthermore, by incorporating synthetic datasets into the training process, we achieved an improvement in micro-F1 scores of up to 10 absolute points. We publicly release our benchmark and generation code to demonstrate its utility for fine-tuning and evaluating LLMs.

We present RuSemCor, an openWord Sense Disambiguation (WSD) corpus for Russian. The corpus was constructed by manually linking tokens from the OpenCorpora corpus to senses in the Russian wordnet RuWordNet. It consists of 869 documents with 121,710 tokens of which 51,588 are wordnet annotated. The resource is represented using the NIF, OLiA, OntoLex, and Global WordNet ontologies and integrated into the Linguistic Linked Open Data cloud. We used RuSemCor as a diagnostic benchmark to evaluate a range of WSD methods. Our experiments yielded three main findings. 1) Generative LLMs substantially outperform traditional knowledge-based methods such as Personalized PageRank. 2) Despite their strengths, generative LLMs do not surpass encoder-based models specifically trained for WSD. 3) Incorporating lexical-semantic relations from RuWordNet produces mixed results: it enhances the performance of encoder-based models and leading LLMs like GPT-4, DeepSeek, and Mistral 24B, but tends to degrade accuracy for smaller generative models such as GPT-3 and Mistral 7B. The resource is distributed under the CC BY-SA open license and is available at: http://github.com/llod-ru/rusemcor.

This paper investigates the internal mechanisms of large language models (LLMs) through the lens of Mechanistic Interpretability (MI). We present novel findings on how information is processed and propagated within these models. Our key contributions include: (1) providing evidence for the localized nature of fact storage and information propagation from subject tokens; (2) introducing a new observation that hidden states remain largely unchanged across multiple middle layers, which we call the "plateau" phenomenon; and (3) developing a manually crafted diagnostic dataset of factual prompts. Our work complements and extends prior research on transformer information flow by demonstrating that, contrary to the prevailing assumption of sequential representation enrichment across layers, subject token states stabilize early and remain functionally static throughout multiple middle layers while containing all necessary information for the final prediction. These insights advance our understanding of how transformers process factual information and suggest a more complex pattern of layer specialization than previously identified.

Fact-checking is a crucial yet challenging task that continues to gain importance. In an effort to address this issue, the FEVER large-scale dataset was developed to facilitate evidence-based fact-checking using Wikipedia as a reference. Despite numerous proposed approaches and evaluations on this dataset, a comprehensive understanding of the errors made by these approaches is still lacking. Here, we aim to bridge this gap. We introduce a diagnostic taxonomy and a generative framework to enhance FEVER-style fact-checking. We establish a taxonomy of errors and we construct a diagnostic dataset that enables the analysis of the errors made by state-of-the-art models as well as their distribution within the FEVER dataset. Additionally, we provide a set of prompts to generate examples within this taxonomy. Our experiments demonstrate promising results through the utilization of these generated examples for fine-tuning.

The prevalence of information manipulation online has created a need for propaganda detection systems. Such systems have typically focused on the surface words, ignoring the linguistic structure. Here we aim to bridge this gap. In particular, we present the first attempt at using discourse analysis for the task. We consider both paragraph-level and token-level classification and we propose a discourse-aware Transformer architecture. Our experiments on English and Russian demonstrate sizeable performance gains compared to a number of baselines. Moreover, our ablation study emphasizes the importance of specific types of discourse features, and our in-depth analysis reveals a strong correlation between propaganda instances and discourse spans.

Large Language Models (LLM) often produce text with incorrect facts and hallucinations. To address this issue, we developed a fact-checking system Truth-O-Meter12 which verifies LLM results on the Internet and other sources of information to detect wrong claims/facts and proposes corrections for them. NLP and reasoning techniques such as Abstract Meaning Representation and syntactic alignment are applied to match hallucinating sentences with truthful ones. To handle inconsistent sources while fact-checking, we rely on argumentation analysis in the form of defeasible logic programming, selecting the most authoritative source. Our evaluation shows that LLM content can be substantially improved for factual correctness and meaningfulness on an industrial scale.

In spite of great progress in recent years, deep learning (DNN) and transformers have strong limitations for supporting human–machine teams due to a lack of explainability, information on what exactly was generalized, and machinery to be integrated with various reasoning techniques, and weak defense against possible adversarial attacks of opponent team members. Due to these shortcomings, stand-alone DNNs have limited support for human–machine teams. We propose a Meta-learning/DNN → kNN architecture that overcomes these limitations by integrating deep learning with explainable nearest neighbor learning (kNN) to form the object level, having a deductive reasoning-based meta-level control learning process, and performing validation and correction of predictions in a way that is more interpretable by peer team members. We address our proposal from structural and maximum entropy production perspectives.

In this research, we manually create high-quality datasets in the digital humanities domain for the evaluation of language models, specifically word embedding models. The first step comprises the creation of unigram and n-gram datasets for two fantasy novel book series for two task types each, analogy and doesn't-match. This is followed by the training of models on the two book series with various popular word embedding model types such as word2vec, GloVe, fastText, or LexVec. Finally, we evaluate the suitability of word embedding models for such specific relation extraction tasks in a situation of comparably small corpus sizes. In the evaluations, we also investigate and analyze particular aspects such as the impact of corpus term frequencies and task difficulty on accuracy. The datasets, and the underlying system and word embedding models are available on github and can be easily extended with new datasets and tasks, be used to reproduce the presented results, or be transferred to other domains.

We extend the concept of a discourse tree (DT) in the discourse representation of text towards data of various forms and natures. The communicative DT to include speech act theory, extended DT to ascend to the level of multiple documents, entity DT to track how discourse covers various entities were defined previously in computational linguistics, we now proceed to the next level of abstraction and formalize discourse of not only text and textual documents but also various kinds of accompanying data. We call such discourse representation Multimodal Discourse Trees (MMDTs). The rational for that is that the same rhetorical relations that hold between text fragments also hold between data values, sets and records, such as Reason, Cause, Enablement, Contrast, Temporal sequence. MMDTs are evaluated with respect to the accuracy of recognition of criminal cases when both text and data records are available. MMDTs are shown to contribute significantly to the recognition accuracy in cases where just keywords and syntactic signals are insufficient for classification and discourse-level information needs to be involved.

We address the problem of argument detection by investigating discourse and communicative text structure. A formal graph-based structure called communicative discourse tree (CDT) is used. It consists of a discourse tree (DT) with additional labels on edges, which stand for verbs. These verbs represent communicative actions. Discourse trees are based on rhetoric relations, extracted from a text according to Rhetoric Structure Theory. The problem is tackled as a binary classification task, where the positive class corresponds to texts with arguments and the negative class corresponds to texts with no argumentation. The feature engineering for the classification task is conducted, deciding which discourse and communicative features are better associated with argumentation. New Intense Argumentation dataset is built and described. Mixed dataset including different types of argumentation and different text genres is collected. Evaluation on this mixed dataset is provided.

Recent transformer-based approaches to multi-party conversation generation may produce syntactically coherent but discursively inconsistent dialogues in some cases. To address this issue, we propose an approach to integrate a dialogue act planning stage into the end-to-end transformer-based generation pipeline. This approach consists of a transformer fine-tuning procedure based on linearized dialogue representations that include special discourse tokens. The obtained results demonstrate that incorporating discourse tokens into training sequences is sufficient to significantly improve dialogue consistency and overall generation quality. The suggested approach performs well, including for automatically annotated data. Apart from that, it is observed that increasing the weight of the discourse planning task in the loss function accelerates learning convergence.

Abstract We propose a novel way of the conversational recommendation where instead of asking user questions to acquire her preferences, the recommender tracks her conversations with other people, including customer support agents (CSA) and joins the conversation only when there is something important to recommend and the time is correct to do so. Building a recommender that joins a human conversation (CRS), we propose information extraction, discourse and argumentation analyses, as well as dialogue management techniques to compute a recommendation for a product and service that is badly needed by the customer, as inferred from the conversation. A special case of such conversations is considered where the customer raises his problem with a CSA in an attempt to resolve at, along with receiving a recommendation for a product with features addressing this problem. Performance of CRS is evaluated in a number of human-human and human-chatbot dialogues and demonstrates that CRS is efficient and less intrusive way to provide high relevance and persuasive recommendations.

We present a sense-annotated corpus for Russian. The resource was obtained my manually annotating texts from the OpenCorpora corpus, an open corpus for the Russian language, by senses of Russian wordnet RuWordNet. The annotation was used as a test collection for comparing unsupervised (Personalized Pagerank) and pseudo-labeling methods for Russian word sense disambiguation.

In this chapter, we explore the role of discourse analysis in ontology construction. Extracting candidate phrases to form ontology entries from text, it is important to pay attention to which discourse units these phrases occur in. It turns out that not all discourse units are equal in terms of their contribution to forming ontology entries; satellites are usually good, and nuclei are not. We survey text mining (TM) and ontology information extraction (IE) techniques in the medical domain and select the ones where advanced linguistic analysis including discourse processing is leveraged the most to produce a robust and efficient ontology. We evaluate the consistency of the resultant ontology and its role in assuring high search relevance.

The use of contrastive loss for representation learning has become prominent in computer vision, and it is now getting attention in Natural Language Processing (NLP). Here, we explore the idea of using a batch-softmax contrastive loss when fine-tuning large-scale pre-trained transformer models to learn better task-specific sentence embeddings for pairwise sentence scoring tasks. We introduce and study a number of variations in the calculation of the loss as well as in the overall training procedure; in particular, we find that data shuffling can be quite important. Our experimental results show sizable improvements on a number of datasets and pairwise sentence scoring tasks including classification, ranking, and regression. Finally, we offer detailed analysis and discussion, which should be useful for researchers aiming to explore the utility of contrastive loss in NLP.

Supported decision trees that have been first proposed to boost the performance and the explainability of the expert systems built upon the texts can become a great basis for the machine reading comprehension (MRC) systems. The supported decision tree is based on building and combining the corresponding discourse trees for the text passage. In this work, we build an environment of supported decision trees for the MRC task. Each answer is represented by a path of a supported decision tree and the whole corpus of answers is then form a lattice of supported decision trees. This environment gives a boost to MRC performance, handling cases where it is nontrivial to determine which document/passage MRC needs to be applied to.