topic modeling

Weak interpretability of topic models and the bias-prone, fragmented nature of traditional word-list-based interpretation hinder reliable thematic analysis. To address this, we propose a model-agnostic, interactive visualization framework that holistically encodes document–word–topic ternary semantic relationships. Methodologically, we introduce a novel decoupled architecture enabling plug-and-play integration of over ten mainstream models—including LDA, BERTopic, and Top2Vec—while unifying multi-view interaction: topic manifold projection, word embedding alignment, and document distribution heatmaps. Semantic consistency is enhanced via UMAP dimensionality reduction and joint normalization of TF-IDF and topic distributions. Experiments across five benchmark corpora demonstrate a 37% improvement in topic discrimination accuracy and a 2.8× increase in user comprehension depth (p < 0.01), significantly advancing cross-model and context-aware topic interpretation.

This paper addresses the low accuracy and poor interpretability of automated identification of latent concepts—such as frames, narratives, and themes—in textual data. To this end, we propose a human-in-the-loop concept discovery framework. Methodologically, we introduce the first open-source large language model (LLM)-driven concept navigation paradigm, integrating iterative prompt-based sampling, cross-domain text embedding and clustering, and an expert validation feedback loop—thereby tightly coupling automated summarization with human-in-the-loop verification. Experiments on AI policy debates, cryptocurrency news, and the 20 Newsgroups dataset demonstrate substantial improvements in political discourse analysis, media frame detection, and fine-grained topic classification. Our approach achieves a superior trade-off between accuracy and interpretability, offering a robust, transparent, and reproducible pathway for concept modeling in computational social science.

Existing topic model evaluation metrics—such as perplexity—suffer from fundamental limitations: poor cross-model comparability, unidimensionality, and misalignment with human judgments, hindering comprehensive quality assessment. To address these issues, we propose WALM—the first LLM-based *joint* evaluation framework for topic models. WALM simultaneously quantifies both *topic semantic quality* and *document representation quality*, overcoming the dimensional fragmentation inherent in conventional metrics. It achieves semantic alignment through word-level consistency modeling and carefully engineered, task-specific prompting. Extensive experiments demonstrate that WALM achieves strong agreement with human annotations (Spearman ρ > 0.85) across diverse topic models, significantly outperforming existing metrics while enabling reliable, cross-model comparisons. The implementation, including code and a ready-to-use toolkit, is publicly released.

This work addresses the challenge of assessing holistic coherence in long texts. We propose BBScore, a reference-free, training-free metric that requires no human-written references or model fine-tuning. Its core innovation lies in the first application of Brownian bridge stochastic processes to text modeling: sentence embeddings are treated as points along a continuous path, and inter-sentence sequential coordination is quantified via path deviation—thereby unifying local semantic cohesion and global thematic consistency. BBScore relies solely on off-the-shelf sentence embeddings and a lightweight classification head, enabling zero-shot, end-to-end evaluation without parameter optimization. Experiments demonstrate that BBScore matches state-of-the-art supervised methods in human coherence judgment; robustly discriminates human-authored texts from outputs of diverse large language models (LLMs); and generalizes effectively across domains in identifying LLM-specific writing styles.

Existing seed-guided topic discovery methods rely on the closed-vocabulary assumption, rendering them incapable of handling out-of-vocabulary (OOV) user-provided seeds and failing to effectively incorporate semantic knowledge from pretrained language models (PLMs). To address this, we propose SeeTopic—the first framework extending seed-guided topic discovery to OOV scenarios. SeeTopic jointly models global semantic representations from PLMs and local contextual cues from the corpus to achieve semantic alignment and dynamic expansion of OOV seeds, and introduces a jointly optimized topic generation mechanism. Evaluated on three cross-domain real-world datasets, SeeTopic significantly improves topic coherence (+12.3%), accuracy (+9.7%), and diversity (+8.1%), while maintaining strong robustness under OOV seeds. This work establishes a more open and flexible paradigm for user-directed topic discovery.

This work addresses the challenge of generating coherent and interpretable topics in emerging or resource-constrained domains, where traditional topic models often falter due to the absence of external knowledge. The authors propose a novel method that leverages only intrinsic word co-occurrence statistics from the corpus to directly construct the Dirichlet prior parameters for Latent Dirichlet Allocation (LDA), without altering its generative process or relying on external resources. This approach represents the first effective translation of endogenous corpus statistics into a meaningful prior distribution, significantly enhancing topic coherence and interpretability. Empirical results demonstrate that the method achieves performance on par with state-of-the-art models that depend on external knowledge, particularly excelling in data-scarce scenarios, as validated on both textual corpora and single-cell RNA-seq datasets.

This study investigates how to effectively leverage Highlights from academic papers to enhance unsupervised keyword extraction performance. Addressing the limitation of existing approaches that predominantly rely on abstracts, this work presents the first systematic evaluation of the complementary value of Highlights for keyword extraction and proposes an input strategy that integrates both Highlights and abstracts. Experiments conducted on datasets from computer science and library and information science, using four classic unsupervised models, demonstrate that Highlights and abstracts exhibit complementary characteristics in terms of keyword coverage and semantic content. The fusion of these two sections consistently yields significantly better performance than using either section alone, thereby substantially improving keyword extraction effectiveness.

Addressing the challenge of generating sentiment-sensitive summaries from unstructured, short social media texts, this paper proposes the first sentiment-aware dual-path summarization framework. The framework integrates a TextRank-based extraction path enhanced with sentiment lexicon augmentation and fine-grained sentiment embeddings, alongside a UniLM-based generation path that jointly models emotional polarity and topical semantics. Unlike conventional summarization models—designed primarily for formal, structured documents—this approach explicitly incorporates sentiment signals to strengthen decision-support capabilities in brand monitoring and market analysis. Experimental results on user-generated content demonstrate substantial improvements: +18.7% in sentiment accuracy and +12.3% in ROUGE-L score (measuring information fidelity), while maintaining real-time processing capability.

This study addresses the limited interactivity and domain adaptability of existing clustering methods for digital humanities scholars working with large-scale unstructured documents. To bridge this gap, the authors propose an analysis-perspective-driven interactive document clustering framework. This framework enables users to define initial semantic lenses through prompt rewriting and instruction embedding, and integrates interactive visualization, on-the-fly cluster adjustment, and online fine-tuning of embedding models into a closed-loop human-in-the-loop feedback process. The approach supports an interpretable, intervenable, and iterative clustering experience, empowering researchers to efficiently uncover latent semantic structures—such as thematic patterns or sentiment signals—and thereby generate high-quality structured data to support in-depth humanities inquiry.

Traditional topic models assign a single topic to an entire document, which struggles to accurately represent multi-topic texts and often leads to topic mixing and reduced interpretability. This work proposes a Segment-Based Topic Assignment (SBTA) framework that, for the first time, refines the granularity of topic modeling from the document level to semantically coherent text segments. To support this approach, we construct the SemEval-STM dataset by combining large language model–based automatic segmentation with human refinement to produce high-quality segments, and introduce a segment-level word intrusion task to enable fine-grained evaluation. Experiments demonstrate that SBTA significantly improves topic clustering quality and interpretability across multiple topic models and evaluation metrics, confirming its effectiveness and scalability.