Preserving contextual information in cultural heritage metadata through multidimensional knowledge graphs
为解决文化遗产元数据中上下文信息丢失的问题，本文提出多维知识图谱方法来同时建模和查询跨多个社会、文化和政治维度的数据。
0 citationsRead paper
Selected work
Representative Papers
AI Overviews in Academic Search: Evaluating AI-generated Summaries of Search Results in a Domain-specific Search Engine
This study addresses the high cognitive load associated with relevance assessment in academic search. Focusing on social science domains, it introduces AI-generated search engine results page (SERP)-level summaries within an interface designed according to information foraging theory and analyzes user behavior. The work proposes a six-category error taxonomy tailored to academic contexts and outlines five deployment safeguards, leveraging both commercial and open-source large language models to generate summaries. User experiments reveal that while AI-generated summaries do not significantly improve primary effectiveness metrics, they consistently reduce perceived mental workload and frustration, decrease click-through and query reformulation rates, and support early-stage screening through enhanced informational cues. These findings highlight their potential as context-sensitive assistive tools in scholarly search environments.
0 citationsRead paper
Analysing Calls to Order in German Parliamentary Debates
This study addresses the lack of systematic analysis of Calls to Order—a form of incivility in parliamentary discourse—within the German Bundestag. The authors construct a novel dataset spanning 72 years of parliamentary debates and propose a rule-based natural language processing approach to systematically define, annotate, and classify the triggers of such interventions for the first time. Their analysis reveals that offensive speech constitutes the primary catalyst for Calls to Order, with male legislators and members of opposition parties disproportionately targeted, particularly during debates on government business. The findings underscore the subjective nature of these procedural interventions and demonstrate their significant association with partisanship, gender, and the presiding role of the speaker, thereby offering new empirical insights and a foundational framework for research on parliamentary norms and discursive civility.
0 citationsRead paper
TriTopic: Tri-Modal Graph-Based Topic Modeling with Iterative Refinement and Archetypes
This work addresses key limitations of existing topic models—such as BERTopic—including stochastic instability, ambiguous embeddings, and reliance on a single data modality. To overcome these issues, we propose a trimodal graph-based topic modeling approach that integrates semantic embeddings, TF-IDF features, and metadata. Our method constructs a denoised graph using Mutual kNN and shared nearest neighbors, followed by consensus Leiden clustering and iterative centroid refinement. Additionally, we introduce a boundary-case-driven prototypical representation for topics. Evaluated across multiple benchmark datasets, the proposed method achieves an average Normalized Mutual Information (NMI) score of 0.575, substantially outperforming state-of-the-art alternatives. It ensures 100% corpus coverage with no outliers and is publicly available as an open-source PyPI library.
0 citationsRead paper
Optimizing Sensory Neurons: Nonlinear Attention Mechanisms for Accelerated Convergence in Permutation-Invariant Neural Networks for Reinforcement Learning
To address low training efficiency and slow convergence in reinforcement learning—particularly in sensor neuron systems requiring permutation invariance—this paper proposes an enhanced Sensory Neuron architecture. The core methodological innovation is the first incorporation of a nonlinear key vector mapping (K → K′) into the attention mechanism, enabling richer nonlinear cross-sensor feature interactions while strictly preserving permutation invariance. Critically, this modification introduces no additional parameters or inference overhead. Empirical evaluation demonstrates that the proposed architecture accelerates policy learning substantially: average convergence steps decrease by 37%, and total training time is significantly reduced. Moreover, policy performance matches or exceeds that of the original Sensory Neuron architecture and leading baselines across multiple RL benchmark tasks. This work establishes a new paradigm for efficient, structure-aware joint perception-decision modeling under strict architectural constraints.
0 citationsRead paper
Recent publications
Latest Papers
Preserving contextual information in cultural heritage metadata through multidimensional knowledge graphs
为解决文化遗产元数据中上下文信息丢失的问题，本文提出多维知识图谱方法来同时建模和查询跨多个社会、文化和政治维度的数据。
0 citationsRead paper
AI Overviews in Academic Search: Evaluating AI-generated Summaries of Search Results in a Domain-specific Search Engine
This study addresses the high cognitive load associated with relevance assessment in academic search. Focusing on social science domains, it introduces AI-generated search engine results page (SERP)-level summaries within an interface designed according to information foraging theory and analyzes user behavior. The work proposes a six-category error taxonomy tailored to academic contexts and outlines five deployment safeguards, leveraging both commercial and open-source large language models to generate summaries. User experiments reveal that while AI-generated summaries do not significantly improve primary effectiveness metrics, they consistently reduce perceived mental workload and frustration, decrease click-through and query reformulation rates, and support early-stage screening through enhanced informational cues. These findings highlight their potential as context-sensitive assistive tools in scholarly search environments.
0 citationsRead paper
Analysing Calls to Order in German Parliamentary Debates
This study addresses the lack of systematic analysis of Calls to Order—a form of incivility in parliamentary discourse—within the German Bundestag. The authors construct a novel dataset spanning 72 years of parliamentary debates and propose a rule-based natural language processing approach to systematically define, annotate, and classify the triggers of such interventions for the first time. Their analysis reveals that offensive speech constitutes the primary catalyst for Calls to Order, with male legislators and members of opposition parties disproportionately targeted, particularly during debates on government business. The findings underscore the subjective nature of these procedural interventions and demonstrate their significant association with partisanship, gender, and the presiding role of the speaker, thereby offering new empirical insights and a foundational framework for research on parliamentary norms and discursive civility.
0 citationsRead paper
TriTopic: Tri-Modal Graph-Based Topic Modeling with Iterative Refinement and Archetypes
This work addresses key limitations of existing topic models—such as BERTopic—including stochastic instability, ambiguous embeddings, and reliance on a single data modality. To overcome these issues, we propose a trimodal graph-based topic modeling approach that integrates semantic embeddings, TF-IDF features, and metadata. Our method constructs a denoised graph using Mutual kNN and shared nearest neighbors, followed by consensus Leiden clustering and iterative centroid refinement. Additionally, we introduce a boundary-case-driven prototypical representation for topics. Evaluated across multiple benchmark datasets, the proposed method achieves an average Normalized Mutual Information (NMI) score of 0.575, substantially outperforming state-of-the-art alternatives. It ensures 100% corpus coverage with no outliers and is publicly available as an open-source PyPI library.
0 citationsRead paper
Optimizing Sensory Neurons: Nonlinear Attention Mechanisms for Accelerated Convergence in Permutation-Invariant Neural Networks for Reinforcement Learning
To address low training efficiency and slow convergence in reinforcement learning—particularly in sensor neuron systems requiring permutation invariance—this paper proposes an enhanced Sensory Neuron architecture. The core methodological innovation is the first incorporation of a nonlinear key vector mapping (K → K′) into the attention mechanism, enabling richer nonlinear cross-sensor feature interactions while strictly preserving permutation invariance. Critically, this modification introduces no additional parameters or inference overhead. Empirical evaluation demonstrates that the proposed architecture accelerates policy learning substantially: average convergence steps decrease by 37%, and total training time is significantly reduced. Moreover, policy performance matches or exceeds that of the original Sensory Neuron architecture and leading baselines across multiple RL benchmark tasks. This work establishes a new paradigm for efficient, structure-aware joint perception-decision modeling under strict architectural constraints.
0 citationsRead paper