ABSTRACT. Highlighting the prediction of technology evolution as two patterns: technological development (i.e., the expansion and contraction of a technological system) and technological change (i.e., the disruption of a technological system and the recombination of multiple technological systems), this study develops a temporal graph learning-based method, including a multi-source heterogeneous data fusion model to integrate multiple bibliometric data sources and external knowledge sources, and a temporal-aware graph attention mechanism to capture the evolving relationships among technical domains over time dynamically. We examine the models by comparing with certain SOTA baselines in an artificial intelligence-related dataset with records retrieved from DBLP, USPTO, and Wikipedia.

ABSTRACT. This study translates traditional tech mining methods into a neural structure, enabling a unifying review of the respective methodologies. While the neural or tensor paradigm as embodied in BERT appears to be distinct from statistics, it can in fact be used to express a wide variety of traditional statistical and machine learning algorithms. This common structure enables a full comparison and contrast with BERT, an important class of LLM.

ABSTRACT. This study explores the use of Large Language Models (LLMs) to enhance Literature-Based Discovery in tech mining applications. Using robotics research and UN Sustainable Development Goals as a case study, we evaluate LLMs' capability to filter semantically similar but practically irrelevant technology-social issue pairs. Through comparative analysis of AI and human evaluations across 357 research topics and 62 social issues, we find that LLMs can effectively reduce manual validation requirements, particularly for environmental and poverty-related issues, though performance varies across different social domains. Our results suggest a promising path toward semi-automated technology-social issue matching while highlighting the continued importance of human expertise in final assessments.

ABSTRACT. Project selection in science, technology, and innovation (ST&I) is crucial for efficiently allocating resources but faces challenges such as data overload, evaluator bias, and high assessment costs. This study introduces the birddog R package, which employs an unsupervised computational model to detect technological and scientific trajectories empirically. Grounded in Dosi’s definition of technological trajectories as co-evolving processes, the proposed method integrates insights from neo-Schumpeterian economics and the scientific emergence literature to overcome limitations of existing approaches like main path analysis. The methodology follows ten steps: the initial four focus on data selection, collection, and organization, followed by citation network construction through bibliographic coupling and direct citation. Subsequent steps apply natural language processing and text mining to identify group content, capture group attributes such as growth rate and key actors, and use structural topic modeling to determine subgroups based on textual similarities. Additionally, the method develops indicators of prestige, novelty, and technological cycles before visualizing trajectories using the Sugiyama algorithm to map their emergence, convergence, divergence, and dormancy. By enabling empirical measurement of technological evolution, this approach enhances decision-making in ST&I policy and strategy, ensuring a more precise and data-driven allocation of resources.

ABSTRACT. Knowledge innovation drives social and scientific progress, yet its core mechanisms remain unclear. This study explores the concept of "Gain New Insights by Reviewing the Past" (温故知新) from The Analects of Confucius, which posits that revisiting existing knowledge fosters new insights and innovation. Using classic citation data from the Web of Science (1936–1987), we analyze how old knowledge influences new knowledge creation. Our multi-dimensional dataset includes 1,898 classic citation reviews, 1,924 classic citations, and over 1.2 million citing references. We employ natural language processing to classify knowledge novelty (e.g., new methods, theories, results) and calculate the Consolidating/Disruptive (CD) index to assess the impact of classic knowledge on innovation. The study addresses three questions: (1) Does reviewing old knowledge lead to new insights? (2) What types of old knowledge promote new knowledge, and what types emerge? (3) How do knowledge flow pathways influence innovation? By examining intermediary citation patterns and novelty types, we reveal how classic knowledge drives both consolidating and disruptive innovation, offering insights into the mechanisms of knowledge evolution.

ABSTRACT. The integration of technological innovation into finance has propelled banking fintech, offering significant potential for reducing income inequality and enhancing financial inclusion. This study uses semantic-based machine learning to assess fintech innovation quality in Chinese commercial banks from 2010 to 2022, examining its effects on income distribution both regionally and between urban and rural areas. The results show explosive growth in banking fintech patent applications, but steady improvements in innovation quality. The innovation had a trickle-down effect on regional income disparity but a Matthew effect on urban-rural inequality. The trickle-down effect was stronger in areas with credit shortages and higher non-agricultural productivity, while the Matthew effect was more prominent in regions with scarce credit and lower agricultural productivity. Additionally, fintech innovation's impact on income inequality varied by bank type. Mechanism analysis suggests fintech innovation influences income distribution by narrowing regional entrepreneurial disparities but widening urban-rural gaps. Bridging digital divides, especially in digital literacy, can enhance the trickle-down effect, while increasing rural access to digital resources may reduce the Matthew effect. This study provides insights into the banking sector's transformation in the digital economy and the role of digital finance in promoting inclusive growth.

ABSTRACT. So far, innovation statistics and innovation analytics have mostly used a patent-counting approach (one patent = one technology/invention). However, reality might look different. To tackle this problem, the idea is to identify bundles of patents per company (patent clusters). This presentation proposes an AI-based alternative to simply counting patents by clustering patents per company/enterprise according to the underlying technologies they represent. We developed an approach using the LLaMA-Model to identify technologies in patent documents (titles and abstracts). We follow a two-step approach to accomplish our task of clustering patents and to compare them across companies/enterprises. The first step is a "keyword extraction" and a "synonym grouping". In a second step we perform the patent clustering based on a similarity search, combined with a keyword search. The oral presentation will start from the conceptual discussion of patents as an innovation indicator and put the idea of patent clusters in that perspective. The methodology will be explained against the background of the challenges arising from the task as such, from the huge dataset (scalability), and the validation process. Next to the methods and their implementation, we will also provide results in the field of “wind energy”. Additional steps to be taken are comparisons of patent clusters per firm between different firms. As this is work in progress, we might only give an outlook on this step.

ABSTRACT. Can open IP strategies promote innovation among competitors, thereby advancing the development of the technology field? This empirical puzzle has been a focal point of debate in the open innovation literature. This study evaluates the impact of open IP strategies adopted by leading firms on technological advancement. Tesla and Toyota's patent pledges acted as exogenous shocks, allowing for the analysis of how open IP strategies influence technological development in the Low Emission Vehicles industry. We utilized a Difference-in-Differences (DID) model analyzing patent data from 2010 to 2019 to measure the effects on technological performance across firms. Our results indicate that open IP strategies significantly enhance technological output, including quantity, quality, and novelty, especially benefiting start-ups, and to a lesser extent, firms with rich knowledge bases. This study contributes to understanding the role of open innovation in fostering technological competition.

ABSTRACT. This research proposes the main path of a technology system network as core technology from the aspect of technology intrinsic attributes and creates a framework to identify the core technology. In this framework, technology system network is constructed with technology topics which are clustered from patent abstracts and titles. Then three methods are tried to identified the main path of the network,includes SPLC,GNN and RL. Finally explore the core technology of MCU chip as a case study. While most existing research identify core technology from the aspect of technology external attributes with patent data, such as citation number, patent family number,patent claims number and so on. Those research can find the high value technology, however may omit some components of the technology system skeleton. This study explore patent text which reflects technology intrinsic information and will be a useful complement to identify core technology.

ABSTRACT. This research analyzes the patent landscape of Autonomous Mobile Robots (AMR) across some critical industries. AMRs are equipped with advanced navigation, multi-sensor fusion, and AI-driven systems, enabling autonomous operations in diverse environments. The AMR industry involves key stakeholders: core technology providers, manufacturers, system integrators, and end-users. This research first defines AMR ontology schema and patent search strategies to systematically identify innovation landscape. With patent portfolio and sub-domain s-curve analysis, we classify the technological lifecycles of AMRs into phases of growth, maturity, or decline. Current findings suggest that, among the AMR industry chain, the domains of navigation systems, hardware/software integration, and control systems independently are in mature stages. To provide actionable insights for stakeholders aiming to innovate or expand within the AMR industry, the research mapping key patent holders and analyzing their roles within the AMR industry chain. The study further identifies top assignees across each technological sub-domain. This strategic mapping offers critical guidance for selecting optimal supply chain partners, fostering collaborations, and enhancing technological development. A case study on Hyundai Motor Group further illustrates how a leading enterprise leverages its expertise in automotive manufacturing to strategically expand into the AMR market. The case study highlights Hyundai Motor Group's strategic use of patents as claimed innovation IPs to establish a competitive edge in the AMR market, emphasizing cross-domain collaborations and advanced technological applications. This research aims to contribute to the strategic development and advancement of AMRs by aligning patent portfolio with industry chain eco-systems, supporting stakeholders in competitive AMR applications.

ABSTRACT. Meeting the global food demand by 2050 requires sustainable agricultural practices, with biofertilizers emerging as an alternative to chemical fertilizers. Despite their importance, the main research fronts and their interconnections remain unclear. This study identifies scientific trajectories in biofertilizer research using citation network analysis on publications indexed in the Web of Science up to 2023. We applied an unsupervised computational approach in ten steps, integrating natural language processing, structural topic modeling, and Sugiyama’s algorithm to detect and visualize technological trajectories. Results indicate that biofertilizer research has grown annually at 20.1%, significantly outpacing the overall scientific production growth (6.7%). Six research groups were identified from 4,604 documents. The largest groups focus on stress-tolerant microbes, phosphate-solubilizing bacteria, biofertilizer production from waste, and organic manure integration. The most recent group explores soil microbiomes, exhibiting the highest growth rate (31% per year). India leads research in four groups, while Brazil and China lead in biofertilizer production and soil microbiomes, respectively. The rapid expansion of biofertilizer research aligns with global sustainability goals, emphasizing the need for research-informed policies. The findings provide valuable insights for policymakers, researchers, and industry stakeholders, facilitating strategic funding and innovation decisions in sustainable agriculture.

ABSTRACT. In recent years, non-terrestrial network technologies like Unmanned Aerial Vehicle (UAV/drone) and satellite communications have emerged as core drivers of ICT industry innovation. These technologies significantly expand network coverage, enhance data transmission speed, and enable new applications. UAVs offer rapid deployment for communication in remote or disaster areas, while Low Earth Orbit (LEO) satellites ensure stable internet access in maritime and remote regions. Together, they integrate with terrestrial networks to form robust systems crucial for IoT, smart transportation, and smart cities. This study systematically reviews relevant literature to construct a technology ontology framework for UAVs and satellites. Using a patent retrieval strategy, a comprehensive patent database was developed for macro-level analyses, focusing on trends, major assignees, IPC classifications, and co-occurrence heatmaps to uncover technology relationships. Advanced text mining methods, including Natural Language Processing (NLP), topic modeling, and keyword extraction, revealed technological hotspots and innovation trends. A key methodological highlight is the application of Generative Topographic Mapping (GTM), a probabilistic model that transforms high-dimensional patent data into interpretable low-dimensional feature maps. GTM enables reverse analysis, revealing synergies between technologies, identifying gaps in corporate technological frameworks, and pinpointing innovation opportunities through potential technology combinations. By integrating these insights with industry needs and emerging trends, the study identifies promising areas for innovation in UAVs and satellite technologies. These findings provide strategic guidance for academia and industry, highlighting pathways for collaboration, innovation, and advancement in 6G and ICT ecosystems.

ABSTRACT. Generative AI (GenAI) and Large Language Models (LLMs) have significantly advanced in generating textual content, including summaries, technical reports, and articles tailored for specific domains and purposes. These technologies have revolutionized labor-intensive tasks, particularly in patent-related processes such as drafting, filing, rebutting, and litigation. However, patent documents often contain highly technical descriptions and complex semantics, posing challenges for academia and industry in interpreting them, visualizing innovative processes, and analyzing cross-patent technological similarities. This study addresses these challenges by evaluating the analytical capabilities of advanced GenAI and LLM technologies developed by leading enterprises. It also introduces an AI-driven flowchart generation system capable of interpreting patent semantics, extracting innovative processes, and visualizing logical steps and procedures. To validate this approach, we tested 30 patents related to advanced 5G/B5G communication technologies in the Information and Communication Technology (ICT) industry, demonstrating the accuracy and feasibility of the methodology and system. Moreover, this research proposes a multi-level similarity analysis framework to evaluate patent process similarities across three dimensions: semantics, behaviors, and structures. This framework provides valuable insights and decision support for patentability analysis, validity verification, and potential infringement investigations. Therefore, this study enhances the visualization and understanding of patent innovation processes, identifies key technical steps—particularly in cross-disciplinary scenarios—and uncovers the hidden value of technical details. Integrating GenAI, language modeling technologies, and domain expertise establishes a foundation for systematic technological innovation and patent analysis, improving efficiency and strengthening intellectual property protection. This solution offers valuable support for researchers and practitioners.

ABSTRACT. The goal of this paper is to understand the relationship between gender, institutional affiliation and research visibility as measured by the number of citations. It tests two ideas frequently present in the literature on sociology citations: the Matthew Effect, which posits that research from more prestigious departments will receive more citations, and the Matilda Effect, which posits that research by female scientists will be less visible. Using papers as the unit of analysis, I use full records from Web of Science for articles published at the 4 leading generalist sociology journals between 1992 and 2005, coded for the gender of the leading author. Using Generalized Structural Equation Models, a number of existing explanations for the gender based differences in citations are tested. Findings show that gender based differences in citations are partially explained by publication venue, authorship patterns, and publication subfield within sociology. That said, even after accounting for these factors, an emerging pattern appears, that is mediated by the interaction of gender and top 5 institutional affiliation, where papers with male first authors benefit from top 5 departmental affiliation, but not papers with female first authors.

ABSTRACT. The increasing integration of artificial intelligence (AI) in education has led to new methods for assessing student engagement and learning outcomes. Aligned with the Global TechMining Conference's focus on utilizing text-mining techniques for research and analytics, this study uses AI-driven methods to analyze over 7,000 student discussion posts from 203 students in a Research and Business Writing course. By combining Packback's Curiosity Score with natural language processing (NLP) metrics, such as Term Frequency-Inverse Document Frequency (TF-IDF) and proportions of Hapax Legomena, we examine the impact of AI-driven feedback on student participation, language diversity, and overall learning behaviors.

ABSTRACT. How does the green technology innovation system evolve? How do institutional factors stimulate the transformation of green technology? This study investigates the evolution of the green technology innovation system, which adopts the technological innovation system (TIS) framework. Combining Scientometric Mapping, social network analysis, Multi-Criteria Decision-Making (MCDM), and in-depth interviews in the green technology sector, this research attempted to map out the emergence of the green technology innovation system in order to identify the key determinants of the systematic transformation in the green technology sector. The findings show that research institutions play an important role in driving the emergence of technological developments. The firms contribute actively to the acquisition, creation, and diffusion of knowledge, which triggers the development of the institutions. Possible implications for policy are indicated at the end of this talk.

ABSTRACT. Cultural shifts are often reflected in the interpretations used in academic discourse, playing a pivotal role in characterizing scientific innovation. Recent advances in natural language processing now allow researchers to analyze large text corpora, deepening our understanding of how language evolves over time. In this study, we examine semantic changes in journal article abstracts spanning over a century, focusing on publications involving German institutions. These institutions offer a unique natural experiment due to significant historical milestones in Germany—such as World War I, World War II, Division, and Reunification.

To conduct our analysis, we scraped abstracts from OpenAlex, the largest open-source scholarly database available. Our primary objective is to quantify language certainty in these abstracts. We developed a measurement framework that identifies and counts uncertain words (e.g., “may”, “possible”, “could”), where higher frequencies of such terms correspond to lower overall certainty scores. This metric enables us to track shifts in linguistic expression that may indicate broader cultural and intellectual transitions.

Preliminary findings suggest that fluctuations in language certainty are more pronounced in disciplines such as physics and mathematics than in the social sciences. Notably, these computational natural sciences exhibited marked language changes in the aftermath of World War II, implying a sensitivity to the cultural and political upheavals of the period. Our study provides empirical evidence for the interplay between cultural context and scientific discourse evolution, demonstrating the potential of text-analysis methodologies to uncover nuanced patterns in scholarly communication. These results illuminate how cultural forces shape scientific narratives effectively.

ABSTRACT. This study compares LDA-based topic modeling outcomes from two text sources: metaverse-related tech news and patents. Study A uses patent clustering and LDA to create a knowledge ontology from patent data, while Study B employs LDA and bibliometric analysis on tech news to identify emerging trends. The structured nature of patents contrasts with the challenges of clustering tech news articles, leading to differing analytical results. The findings reveal three common topics—AR/VR, AI, and encryption technology—while highlighting differences in focus. Patents emphasize technical methods and system architectures, while tech news centers on user experience, accessibility, and social implications. Keyword analysis of "ar/vr," "virtual," and "digital" shows differing contextual uses across both sources. Overall, while patents offer technological blueprints, news articles reflect market and societal trends, highlighting the need for careful selection of text sources in tech mining. This approach enhances strategic intelligence in Science, Technology, and Innovation (ST&I) research.

ABSTRACT. Large-scale, open-source bibliometric databases are essential to quantify the scientific enterprise and shape science policy. However, despite their widespread use, the metadata quality within these repositories has not been systematically scrutinized. This study examines the journal publication metadata extracted from OpenAlex.

Our investigation centers on the prevalence of missing reference data in journal publications, a deficiency that can undermine the validity of citation analysis and raise questions about the scholarly rigor. We found that the absence of references remains a persistent issue over time, and approximately 60% of total journal publications indexed in OpenAlex lack data on references.

We further examined the reference data from journal publications and uncovered significant disparities across several dimensions. Publications from higher-reputation institutions, regions with higher incomes, and larger research teams generally show fewer issues with missing references. Notably, one striking observation is that articles originating from higher-prestige institutions consistently have fewer incomplete references. The disparity is more conspicuous in recent years between low-prestige and high-prestige institutions.

The implications of these findings are far-reaching. Incomplete reference metadata can lead to biased assessments of scholarly impact, potentially skewing policy decisions and evaluations. As the reliance on open-source bibliometric databases grows, it is crucial for researchers and policymakers to be aware of these underlying disparity issues and to exercise caution when sampling and interpreting such data. Our study advocates for a more nuanced approach to the use of bibliometric databases, suggesting the development of standardized quality control measures.

ABSTRACT. This paper introduces a novel methodology and indicator framework to enhance the measurement of scientific novelty, addressing a critical gap in the assessment of transformation policies. Leveraging advancements in Artificial Intelligence (AI), I develop the "novelty vector," a new concept that learns from expert assessments to predict novelty scores with over 90% accuracy—far surpassing traditional text-based novelty metrics. This approach departs from conventional bibliometric proxies by integrating expert evaluations into a machine learning model, providing a more precise measure of scientific novelty. The novelty vector retains statistical significance in econometric models even after controlling for proposal quality and applicants’ credentials. Applied to a dataset of published papers, it demonstrates high predictive power in identifying Nobel-winning research. This breakthrough has significant implications for science policy, offering a more accurate tool to monitor and evaluate funding strategies aimed at fostering transformative research. By shifting away from biased impact-factor-based indicators, this methodology enables better detection of high-risk, high-gain research and informs peer review processes. Ultimately, the novelty vector provides a pathway to enhance our understanding of scientific breakthroughs, supporting policymakers in their efforts to promote innovative and societally impactful research.

ABSTRACT. Artificial intelligence (AI) is set to reshape the practice of science, stimulating debate about its impacts on research productivity, novelty, reproducibility, intellectual property rights, workforce dynamics, risk, bias, and ethics. These discussions have become even more prominent with the rapid progression of Generative Artificial Intelligence (GenAI), a branch of AI that uses massive amounts of training data in machine learning models to produce text, visualizations, code and complex data analyses. To inform these debates, we suggest that probing not only the extent but also the nature of AI and GenAI use in science is important in understanding its applications and implications. Prior studies construct AI bibliometric search terms to classify AI. However, while such techniques recognize AI in papers, they do not reliably identify what use is made of AI. In this paper, we advance a novel GenAI methodology for classifying AI usage types in scientific papers. In doing so, we also probe whether we can reliably automate scientific practices using DeepSeek-R1 and GPT-4 to enhance human annotation and assist data processing. By leveraging large-scale bibliometric data and GenAI techniques, this paper offers fresh methodological and empirical insights which provide a basis for understanding AI’s evolving role in science and its implications for research policy.

ABSTRACT. Knowledge management is an essential challenge in a high-technology world. Universities in particular face particular institutional challenges in aligning diverse disciplines and bases of knowledge. The traditional structure of a university is based on a bureau or hierarchy structure which supports disciplinary learning, specialised centres of research, and financial accountability. However universities are increasingly introducing networked forms of coordination both within and without the university. This addition to traditional governance structures is in a response to an increasingly transdisciplinary and international landscape for research and development. This extended abstract utilises tech mining approaches to support research network management within universities.

ABSTRACT. We combine scientometrics and generative AI to analyze the intersection of AI and green technologies. Leveraging technological trajectories built using the entire EPO data, the study identifies current common themes and projects future directions of twin (AI and green) technologies. Generative AI summarizes patent information, revealing technological paradigms and their expected future evolution. Initial findings highlight key areas like automation, environmental management, and IoT integration. In ongoing work we are expanding the analysis, enhancing replicability, and comparing results with other NLP methods for a more robust representation of past and future technological developments.

ABSTRACT. How does the emerging innovation ecosystem evolve? How does the high-tech sector in small emerging economies engage with the global innovation ecosystem? To what extent can science and technology policy contribute to strengthening innovation networks in the global AI innovation ecosystem? Combining scientometric mapping, social network analysis (SNA), and system dynamic analytic techniques, this research aims to map out the emergence of the AI innovation ecosystem in the technology followers. The study highlights the impact of key developments of international institutions, such as the EU AI Act, which has made a significant impact on the development of the AI innovation ecosystem of the technology followers. Research findings indicate that Taiwan’s high-tech sector engages with the global ecosystem by leveraging its strengths in semiconductor manufacturing and expanding partnerships with global institutions. Taiwan demonstrates significant strengths in AI-related hardware infrastructure, particularly in patents for semiconductors and sensors, which are essential for AI computing power. The evolving landscape of AI management regulations necessitates adaptive policies that foster innovation while ensuring stability. The government should promote academia-industry research collaboration through mission-oriented policies to enhance the commercialization of scientific research outcomes. Possible implications for future policy implications will be noted.