event detection

This paper addresses key challenges in temporal information processing—including temporal intent identification, temporal expression normalization, event chronology modeling, and dynamic fact reasoning—particularly for time-sensitive domains such as news, historical archives, and scientific literature. It provides a systematic survey of recent advances in Temporal Information Retrieval (TIR) and Temporal Question Answering (TQA), unifying traditional approaches with modern large language model (LLM)-driven techniques for chronological modeling, multi-hop temporal reasoning, and retrieval-augmented generation (RAG)-enhanced timely QA. The work introduces a novel evaluation paradigm centered on temporal robustness and constructs a spatiotemporal IR/TQA knowledge graph encompassing benchmark datasets, evaluation metrics, and methodological taxonomies. These contributions offer both theoretical foundations and practical guidelines for developing temporally aware AI systems.

This work addresses the limited time-series modeling capability of large language models (LLMs) stemming from their text-only pretraining. We propose a systematic cross-modal alignment framework specifically designed for time-series analysis. Our approach innovatively introduces a four-category taxonomy of time-series-oriented textual data, systematically synthesizes cross-modal alignment, feature fusion, and joint encoding strategies, and validates effectiveness through prompt engineering and multi-domain benchmarking on forecasting and anomaly detection tasks. Results demonstrate substantial improvements in LLMs’ understanding and generalization over time-series data; uncover principled mappings between textual data types and optimal modeling strategies; and identify key research directions—including scalable architectures and domain-adaptive learning. This study provides both theoretical foundations and practical methodologies for empowering LLMs with robust time-series analytics capabilities.

Contemporary NLP methods often treat text as temporally homogeneous, neglecting semantic evolution over time and thereby introducing temporal semantic bias; moreover, existing tools for temporal text analysis are fragmented and lack reproducibility. To address these limitations, we propose the first unified framework that systematically integrates multi-granular temporal text analysis, enabling both topic evolution modeling and lexical semantic shift detection. Implemented in Python, the framework incorporates dynamic topic modeling, temporal word embedding alignment, sliding-window LDA, and interactive visualization modules. It ensures end-to-end reproducibility and achieves significant improvements in topic trend identification accuracy on cross-year news and social media datasets. Our core contribution is the first standardized, open-source, and extensible toolkit for temporal NLP—bridging the gap between theoretical models of language evolution and empirical, large-scale diachronic analysis.

This study addresses the scarcity of annotated historical Italian news corpora by proposing an unsupervised method to automatically identify major socio-political turning points. Building on a diachronic corpus of approximately 600,000 articles from *La Repubblica* (1985–2000), the approach integrates natural language processing, word embeddings, semantic change modeling, complex network analysis, and tools from statistical physics. For the first time, complex systems theory is applied to diachronic media analysis in the Italian context. Without relying on any prior labels, the method successfully detects abrupt shifts in media discourse corresponding to pivotal events such as the transition from Italy’s First to Second Republic, the Gulf War, and the Kosovo War. This work offers a novel paradigm for digital humanities and computational social science by demonstrating how unlabeled textual data can reveal historically significant societal transformations.

This work addresses the challenge of modeling the dynamic and highly abstract evolution of information narratives during crisis events, a task where existing approaches are largely confined to static snapshots. We propose the first framework that integrates situated cognition theory with unsupervised temporal modeling, enabling adaptive representation of narrative entity trajectories within a shared semantic space. By combining semantic embeddings, density-based clustering, and rolling time-window linkage, our method requires no predefined labels and captures fine-grained narrative lifecycles, revealing heterogeneous evolution patterns characterized by coexisting transient fragments and stable anchors. Experiments on real-world crisis data demonstrate high clustering consistency and the ability to effectively identify diverse narrative evolution pathways, offering interpretable temporal representations for dynamic information monitoring and decision-making.

Existing data discovery methods largely overlook the temporal dynamics of data relationships, particularly hindering timeliness-aware retrieval when explicit temporal metadata is absent. To address this, we propose “temporal-effective data discovery”—a novel paradigm introducing the first time-integrated data discovery system. Its core comprises four techniques: version discovery, temporal lineage inference, change-log synthesis, and time-aware querying. Crucially, the system models semantic evolution of data over time without relying on original timestamps, enabling dynamic lineage tracing and version-level precise retrieval. Experimental evaluation on enterprise and public data lakes demonstrates significant improvements in accuracy and practicality for timeliness-sensitive queries. Our work establishes a new research direction in data discovery centered on temporal evolution, advancing both theoretical foundations and real-world applicability.

Traditional pre-trained language models (e.g., BERT) lack explicit temporal modeling capabilities, limiting their effectiveness on complex time reasoning tasks. To address this, we propose BiTimeBERT 2.0, a temporally aware language model pretrained on a large-scale chronological news corpus. Our method introduces three novel time-aware pretraining objectives: Event Temporal Alignment Masked Language Modeling (ETAMLM), Document Date Prediction (DD), and Time-Sensitive Entity Replacement (TSER), complemented by an efficient corpus preprocessing strategy that accelerates training by 53%. Evaluated on diverse temporal reasoning benchmarks—including datasets with long temporal spans—BiTimeBERT 2.0 consistently outperforms BERT and other baselines. These results empirically validate that explicit temporal modeling yields substantial gains in linguistic understanding and temporal reasoning capability.

Existing time-series analysis methods that integrate large language models (LLMs) with textual label supervision often over-rely on textual cues while neglecting intrinsic temporal dynamics, leading to outputs inconsistent with temporal context. To address this, we propose TimeSense—a novel multimodal framework that restores the original sequential structure via a time-aware reconstruction module and introduces coordinate-based positional embeddings to explicitly model spatial relationships among time points. This design enables deep coupling between linguistic reasoning and temporal dynamics while preserving LLMs’ language understanding capabilities and enhancing temporal feature learning. Evaluated on the EvalTS benchmark across ten diverse tasks, TimeSense achieves state-of-the-art performance, particularly excelling in complex, multivariate time-series reasoning where it significantly outperforms existing approaches.

Traditional multi-document summarization struggles with integrating multi-perspective narrative texts (e.g., legal testimonies, historical accounts) due to its overemphasis on concision, compromising temporal coherence and factual completeness. To address this, we formally introduce *Narrative Consolidation*—a novel task requiring precise chronological ordering, comprehensive content coverage, and seamless integration of complementary details. We propose the Temporal Alignment Event Graph (TAEG), a graph-based model that explicitly unifies event alignment with temporal structure modeling. TAEG incorporates graph centrality measures (e.g., PageRank) to automatically select authoritative narrative versions. Evaluated on the Four Gospels dataset, our method achieves perfect temporal consistency (Kendall’s Tau = 1.000) and improves ROUGE-L F1 by 357.2% over baselines, demonstrating that explicit temporal backbone modeling is essential for effective narrative consolidation.

This work addresses the challenges of fragile entity and event extraction from unstructured data, heavy reliance on costly ontology engineering in knowledge graph construction, and limited cross-domain generalization. To overcome these limitations, the authors propose an end-to-end multidimensional information extraction framework that leverages spatiotemporal context as a universal anchor. The approach employs large language models (e.g., GPT-4o-mini, Qwen3-8B) for context-aware entity and event extraction, enhanced by document-level memory, geocoding correction, and quality validation mechanisms. It further supports user-defined analytical dimensions and interactive exploration, including clustering, burst detection, and entity network analysis. Evaluated on a public health benchmark, the method achieves F1 score improvements of 4.37% and 3.60% for spatial and temporal entity extraction, respectively. The code and an online demo platform are publicly released.

This work addresses the limitation of existing time series forecasting methods, which predominantly focus on numerical data and struggle to effectively incorporate accompanying textual information, thereby constraining their ability to model complex real-world scenarios. To overcome this, the authors propose a multimodal fusion framework that leverages event-driven reasoning and historical context learning to guide large language models in semantic inference. The framework introduces an endogenous text alignment mechanism and an adaptive frequency-domain fusion strategy to achieve deep integration of textual information at both representation and prediction levels. Extensive experiments across ten real-world datasets spanning diverse domains demonstrate that the proposed method significantly outperforms current state-of-the-art approaches, confirming that effective utilization of textual cues can substantially enhance time series forecasting performance.

This work addresses the challenge of users struggling to pinpoint specific moments in meeting discussions based solely on content. To overcome this, the paper proposes a novel approach that reframes timestamp prediction as a constrained candidate selection task. Instead of directly generating timestamps, large language models such as Mistral-7B-Instruct are guided to select the most relevant segment from a set of retrieved, timestamped meeting excerpts, thereby avoiding unsupported or invalid predictions. Integrating retrieval-augmented generation (RAG) with a constrained selection mechanism, the method demonstrates significant improvements on a dataset of 200 municipal meetings and 420 queries: Recall@5 increases from 31.9% to 50.0%, mean absolute error decreases to 761 seconds, and the number of valid outputs rises from 373 to 419, substantially enhancing both accuracy and reliability in temporal localization.