Identifying Research Hotspots and Research Gaps in Specific

Research Area Based on Fine-Grained Information Extraction via

Large Language Models

Yuling Sun

1,2 a

, Xuening Cui

, Aning Qin

1,*

and Jiatang Luo

National Science Library, Chinese Academy of Sciences, Beijing, China

Department of Information Resources Management, School of Economics and Management, University of Chinese

Academy of Sciences, Beijing, China

School of Advanced Interdisciplinary Sciences, University of Chinese Academy of Sciences, China

Keywords: Large Language Model, Information Extraction, Scientific Data, Research Hotspots, Research Gaps.

Abstract: This paper constructs a fine-grained scientific data indicator framework using LLMs to conduct knowledge

mining in a specific field of natural science and technology, with empirical analysis carried out in the domain

of carbon dioxide conversion and utilization technology. Firstly, based on the characteristics of the technical

field, we systematically established four key scientific data dimensions: products, technologies, materials, and

performance. Subsequently, six key scientific data indicators were selected to characterize these dimensions.

Finally, the extracted scientific data were employed to analyse research hotspots and gaps in the field. This

approach effectively addresses the inherent limitations of traditional technology topic analysis, such as overly

coarse metric granularity and the lack of quantitative features. Moreover, since these scientific data

dimensions and indicators are generalizable to natural science and technology fields aimed at product

development, the proposed methodology demonstrates broad applicability.

1 INTRODUCTION

Identification of research hotspots and gaps is essential

for understanding disciplinary dynamics, optimizing

resource allocation, and formulating policies.

Scientific papers hold significant academic value and

function as indicators of a field’s developmental level.

Consequently, the hotspots and cutting-edge directions

of disciplinary research can be achieved through

knowledge mining of scientific papers.

Existing studies typically integrate thematic

dimensions (e.g., methodologies, products, research

mechanisms) to uncover the aggregation and

evolution, they exhibit two critical limitations: a)

Inability to conduct detailed, in-depth analyses of

specific key scientific data indicators; b) Neglect of

fine-grained performance parameters in hotspot/gap

identification. With the rapid growth of scientific

publications, intelligence research now demands

more refined and intelligent methods to process and

https://orcid.org/0000-0001-6836-5530

* Corresponding author

analyze vast bibliographic data. Recent advances in

natural language processing (NLP) techniques,

specifically fine-grained data mining and large

language models (LLMs), offer novel approaches for

intelligence studies. Fine-grained data indicators

provide deeper insights into research specifics.

Leveraging their proficiency in scientific text

comprehension, knowledge reasoning, and

multimodal processing, LLMs are capable of

undertaking sophisticated tasks related to text

generation and information extraction.

This study aims to construct a multi-level

knowledge network for specific scientific research

area and leverage large language models to extract

fine-grained, multi-labeled scientific data, thereby

forming a research dataset of key domain indicators.

Furthermore, we establish an analytical framework

for identifying research hotspots and gaps based on

fine-grained scientific indicator data. Through an

empirical analysis in the research area of carbon

capture, utilization and storage (CCUS), the

Sun, Y., Cui, X., Qin, A. and Luo, J.

Identifying Research Hotspots and Research Gaps in Speciﬁc Research Area Based on Fine-Grained Information Extraction via Large Language Models.

DOI: 10.5220/0013824900004000

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2025) - Volume 1: KDIR, pages 473-480

473

effectiveness of the proposed framework is validated.

This study enriches the methods and perspectives of

knowledge mining in disciplinary fields.

2 LITERATURE REVIEW

2.1 Identification of Research Hotspots

and Discovery of Research Gaps

Traditional approaches primarily include keyword

co-occurrence analysis, citation network analysis and

topic models.

Keyword co-occurrence analysis can reveal

thematic clusters and evolution by extracting

frequently co-occurring terms. Wang et al. (Wang et

al., 2023) conducted bibliometric analysis on 4,922

articles in the field of carbon neutrality based on the

Web of Science (WoS) database, using Citespace and

Bibliometrix functions for descriptive statistics and

co-occurrence analysis of keywords. Xu et al. (Xu et

al., 2016) combined keyword co-occurrence with a

cosine similarity algorithm, integrating academic

papers and patents to identify research frontier

hotspots in the LED field.

Citation network analysis uncover core research

and directional evolution through highly cited

publications and citation chains. Morris et al. Morris

et al.,2003) employed bibliographic coupling analysis

to construct a timeline of anthrax-research hotspots,

visualizing the evolution of active themes. Chang et

al. (Chang et al., 2015) combined keywords,

bibliographic-coupling and co-citation analyses to

explore the evolution of hotspots in library and

information science over two decades.

Topic models automatically identify latent topic

distributions from enormous text by applying various

text analysis techniques. The predominant method is

the Latent Dirichlet Allocation (LDA) model. Liu

(Liu, 2025) applied the LDA model to 559 articles on

biosecurity legislation (1996–2023) and identified

nine key hotspots and significant trends. Tan and

Xiong (Tan and Xiong, 2021) extracted topics via

LDA model from core data-mining journals in CNKI

and Web of Science, combining topic life cycles with

time-slicing to map evolutionary paths.

Based on their activity levels and persistence,

research hotspots can be categorized into three kinds:

sustained, emerging, and potential hotspots. Liu et al.

(Liu et al., 2023) identified sustained hotspots in

computer science by measuring keywords survival

metrics (time/frequency), applying logistic regression

to analyse influencing factors of keyword survival

patterns. Hu et al. (Hu et al., 2024) leveraged the

global blockchain patent literature, integrating LDA,

Word2Vec and BERT to construct a fine-grained

topic-mining framework that surfaced emerging

technological hotspots. Thakuria and Deka (Thakuria

and Deka, 2024) utilized topic modelling to identify

prevalent potential hotspots in Library and

Information Science (LIS) journals between 2013 and

2022, and reveal unknown research themes.

Gap Analysis is a strategic analysis method used to

evaluate the differences between the current state and

the expected or target state. In this paper, research gaps

refer to important issues that have not been adequately

studied, received insufficient attention, or have become

disconnected from policy or industry expectations in

the existing literature. Currently, the main approaches

to identifying research gaps include systematic

literature reviews (Anton et al., 2022) and expert

consultation (Mohtasham et al., 2023). However, these

qualitative methods suffer from strong subjectivity and

low efficiency, making it difficult to rapidly and

accurately pinpoint gaps within the massive body of

literature. Some data-driven quantitative techniques

have also been applied to gap analysis. Westgate et al.

(Westgate et al., 2015) employed LDA model and

statistical methods (cluster analysis, regression, and

network analysis) to investigate trends and identify

potential research gaps within the scientific literature.

The common limitation of existing hotspot

identification methods is that they mainly rely on a

single type of data (such as keywords, citation

relationships or subject terms), and the analytical

perspective is concentrated on the macro aggregation

of disciplinary themes. There is a lack of in-depth

mining of fine-grained performance indicators that

support these macro themes, as well as an overall

correlation framework for integration analysis of

multi-dimensional indicators. Quantitative analytical

methods for identifying research gaps remain scarce.

Existing strategies mostly adopt qualitative approaches

that infer “under-studied” areas from bibliographic

coverage or topic popularity. Quantitative comparison

between the specific performance parameters reported

in the literature and the targets set by policy plans,

industry technical standards, or real-world application

is rarely considered. Consequently, substantial

performance gaps at the technology level cannot be

pinpointed with precision.

2.2 Fine-Grained Information

Extraction Based on Large

Language Models

Information Extraction typically consists of three sub-

tasks: Named Entity Recognition (NER), Relation

KDIR 2025 - 17th International Conference on Knowledge Discovery and Information Retrieval

474

Extraction (RE), and Event Extraction (EE). Fine-

grained information extraction (FGIE) can extract

semantically specific and application-oriented data

(e.g., research methods, experimental data,

performance metrics) from massive academic

literature. These granular data indicators provide deep

technical insights.

Early approaches relied mainly on manually

defined rules and pattern matching, which were

highly interpretable but had poor generalization

ability. With the rise of deep learning, neural-

network-based methods (e.g., LSTM, BiLSTM,

CNN) the automatically learn textual features have

become mainstream. Yu et al. (Yu et al., 2019)

improved the Bootstrapping algorithm and built a

deep learning model to extract four fine-grained

knowledge units from the abstracts of 17,756 ACL

papers. Onishi et al. (Onishi et al.,2018) constructed

a weakly supervised ML framework that uses CNNs

trained on a materials-microstructure corpus to

extract “processing-structure-property” triplets.

Rodríguez et al. (Rodríguez et al.,2022) proposed an

attention mechanism based on the noun-type

syntactic elements, combining BPEmb vectors and

the Flair model to address two sub-tasks of fine-

grained NER: Named Entity Detection and Named

Entity Typing.

Based these advances, pre-trained encoder models

such as BERT and RoBERTa marked a new stage in

FGIE. By pre-training on large general corpora and

fine-tuning on task-specific datasets (e.g., SQuAD).

They achieved stronger transferability and became

the dominant approach for entity-centric extraction.

Domain-adapted variants such as MatSciBERT have

further demonstrated effectiveness in specialized

areas like materials science (Gupta et al., 2022).

Similarly, RoBERTa-based architectures have been

widely applied in fine-grained biomedical IE and

radiology text analysis (Datta & Roberts, 2022; Yin

et al., 2021). This pre-train–fine-tune paradigm

greatly reduced reliance on handcrafted rules and

task-specific feature engineering, and for a time they

became the mainstream solution for FGIE. However,

these encoder-only models still had critical

limitations: they required large-scale labelled data for

downstream adaptation, had limited zero-shot

generalization, and captured domain-specific

knowledge only weakly.

Recently, large language models such as GPT-

3.5/4, PaLM 2, DeepSeek and Claude have developed

powerful language understanding and generation

capabilities, enabling them to efficiently extract key

fine-grained information from scientific literature.

Fine-grained information extraction based on LLMs

mainly includes the following methods: a) Prompt

engineering, designing appropriate prompts to guide

LLMs to directly extract the required information

from the text in zero-shot or few-shot scenarios. For

instance, Wu et al. (Wu et al.,2025) enhanced LLMs

for miRNA information extraction through

diversified prompt strategies and systematically

compared the performance of GPT-4o, Gemini, and

Claude in NER and RE. b) Task decomposition,

breaking down complex outputs into a series of

simpler questions and answers, which enhances the

reasoning ability of LLMs. Wei et al. (Wei et al.,

2023) implemented zero-shot FGIE via ChatGPT

multi-turn question answering (QA), decomposing

the complex IE task into two conversation rounds: a)

identifying entity, relation, and event types in a

sentence; b) converting unlabelled text directly into

fine-grained structured knowledge through chain-of-

question templates. Qiao et al. (2025) introduced a

novel FGER-GPT method, which employs multiple

inference chains and a hierarchical strategy for

recognizing fine-grained entities, significantly

enhancing the performance of LLM in fine-grained

entity recognition and effectively alleviating the

problems of label lack and hallucination. c) Fine-

tuning, further training the general LLMs on specific

scietific research area labeled data to adapt it to

specific IE tasks. Dagdelen(2024) utilized fine-tuned

LLMs such as GPT-3 and Llama-2 for joint named

entity and relation extraction from scientific texts,

and verified its effectiveness on materials science

texts.

Overall, neural network methods advanced the

automation of fine-grained information extraction by

learning textual features directly, while pre-trained

encoder models such as BERT and RoBERTa further

improved performance through large-scale pre-

training and task-specific fine-tuning, reducing the

need for handcrafted features. However, these

approaches typically depend heavily on large labelled

datasets and exhibit limited zero-shot generalization.

In contrast, large language models offer far greater

flexibility and contextual understanding, enabling

effective fine-grained information extraction without

task-specific supervision, though the issue of

hallucination in their outputs remains a critical

challenge.

The LLM-driven FGIE framework introduced in

this study transforms the descriptive macro analysis

of research hotspots into a fine-grained quantitative

assessment, and compares literature data with policy

demands to accurately identify research gaps. By

integrating prompt engineering and progressive

optimization strategies, the method reduces the

Identifying Research Hotspots and Research Gaps in Speciﬁc Research Area Based on Fine-Grained Information Extraction via Large

Language Models

475

Figure 1: The research framework of this paper.

reliance on labelled data and enhancing the complex

text processing capability of LLMs. This approach

offers a transferable pathway for specific scientific

research knowledge mining.

3 RESEARCH FRAMEWORK

We have chosen the technology research and

development field with product as the research object.

Based on the scientific data characteristics of these

specific fields, we developed a four scientific data

dimensions that including products, technology,

materials and key performance. Then, based on the

characteristics of specific technology research and

development fields and expert consultation, we

further selected specific scientific data indicators that

can characterize the above dimensions.

Leveraging semantic understanding techniques,

the DeepSeek-V3 large language model was

employed to automate indicator extraction,

establishing a comprehensive specific scientific

research fine-grained database. This database enables

multidimensional bibliometric analyses of

methodological innovation, material applications,

product development, and their interdisciplinary

intersections. Concurrently, it facilitates comparative

analysis between research metrics and policy targets,

systematically evaluating alignment with current

research trajectories. Figure 1 presents the research

framework.

Despite empirical validation within CO₂

conversion and utilization, the framework exhibits

high modularity and prompt-level portability,

enabling analogous fine-grained indicator extraction

and knowledge mining across product-oriented

technology research area—notably chemical

engineering and materials discovery.

3.1 Data Source

Data were extracted from the Web of Science Core

Collection (WOS). A systematic retrieval was

conducted for Science Citation Index (SCI) and

Social Sciences Citation Index (SSCI) articles (2020-

2024), resulting in a total of 15,695 publications. We

additionally collected 81 strategic planning

documents related to CCUS from the official

government website of major economies (e.g., the

United States, the European Union, the United

Kingdom, Germany, and France).

3.2 Key Indicator Entity

Focusing on CO

conversion and utilization as an

empirical testbed, through expert consultation, we

further selected six key scientific data indicators to

characterize the four dimensions mentioned above

(Table 1). Specifically, The final product of CO

conversion in research papers is used as a product

such as methanol, ethanol, ethylene, etc; the

technological pathway for CO

conversion is

characterized by techniques such as electrocatalysis,

KDIR 2025 - 17th International Conference on Knowledge Discovery and Information Retrieval

476

photocatalysis, etc; catalysts are used to characterize

materials; three indicators including faradic

efficiency, target product selectivity, and conversion

rate are used to characterize performance.

Table 1: Key dimensions and indicators.

No. Dimensions Indicato

1 Products Conversion Products

2 Technolog

Technical Pathways

3 Materials Catalyst

Performance

Faradic Efficienc

Target Product

Selectivit

6 CO₂ Conversion Rate

3.3 LLM-Based Entity Extraction

To compare the performance of a BERT-style pre-

trained model and a large language model on FGIE,

we conducted a preliminary experiment focusing on

a single entity type, namely product. Using a unified

question answering (QA) framework, the document

was provided as context, and the extraction target was

expressed as a natural-language query (e.g., “What

products are formed from CO₂ conversion?”). Within

this setup, RoBERTa (Liu et al., 2019) achieved only

around 42% exact-match accuracy, whereas

DeepSeek-V3 (zero-shot) reached close to 99%,

corresponding to a +57 percentage-point and 2.4×

relative improvement (Table 2). Although this is only

a pilot study restricted to one entity type, the results

already highlight the substantial advantage of large

language models over BERT-style encoders in QA-

based FGIE.

Table 2: Comparison of product-recognition performance

between DeepSeek and RoBERTa.

Model Accuracy(%) Diffrence(%)

RoBERTa 41.89

——

DeepSee

-V3 98.65 +56.76

Leveraging the large language model DeepSeek-

V3, we construct a target entity extraction system

using prompt engineering (Figure 2). Our three-stage

progressive optimization strategy includes: First, the

model learns basic extraction paradigms using a small

set of sample examples, with prompts iteratively

refined through manual verification. Next, building

upon the verified baseline model, incremental

optimization is performed using a medium-sized

sample set. This stage involves dynamically adjusting

the prompt and supplementing it with 5% sampling

for quality verification. Finally, the full dataset is

processed automatically, supported by a 2% sampling

review mechanism to ensure output stability. This

phased optimization approach significantly enhances

the model’s adaptability to complex texts.

Figure 2: LLM-based prompt engineering for named entity

extraction in scientific literature.

3.4 Knowledge Graph Construction

Within diverse technological pathways for CO₂

conversion and utilization, catalyst selection and

design exhibit significant differences. Different

conversion methods (such as electrocatalysis,

photocatalytic, thermocatalytic, and bioconversion)

require specific types and properties of catalysts due

to their distinct reaction mechanisms, operating

conditions, and target products. Therefore, clarifying

the applicable types of efficient, stable catalysts for

each technical pathway is critical to optimizing

reaction performance, lowering energy consumption

and costs, and advancing the practical application of

specific CO₂ conversion routes. By extracting the

entity associations between technical methods and

catalysts and conducting network visualization using

Gephi software, we have successfully constructed a

knowledge graph illustrating the interconnections

between them.

4 RESULTS AND ANALYSIS

4.1 Research Hotspots Analysis

Statistical analysis of the conversion products (Figure

3) reveals that C1 compounds dominate, accounting

for 72% of the publications surveyed, significantly

higher than the share for C2+ products (28%). Among

C1 products, carbon monoxide (CO), methane (CH₄),

methanol (CH₃OH), formic acid (HCOOH), and

formate are the most extensively studied. Research on

Identifying Research Hotspots and Research Gaps in Speciﬁc Research Area Based on Fine-Grained Information Extraction via Large

Language Models

477

CO₂ conversion to C2 products primarily focuses on

ethylene, ethanol, cyclic carbonates, acetate,

dimethyl ether, acetic acid, and ethane.

Figure 3: Research hotspots of CO

conversion products.

From the technical method–catalyst knowledge

map (Figure 4), we can identify the main catalysts

employed by different technologies. It can be

observed that different CO₂ conversion technologies

prioritize distinct types of catalysts, owing to

variations in their reaction mechanisms and

operational conditions. Electrocatalysis CO₂

reduction primarily employs metals (e.g., Cu, Ag) and

their oxides, bimetallics, metal-organic frameworks

(MOFs), molecular catalysts (e.g., metal porphyrins),

and nitrogen-doped carbon materials. Photocatalytic

CO₂ reduction uses metal oxides (e.g., TiO₂, Cu₂O),

nitrides (e.g., g-C₃N₄), sulfides (e.g., CdS), MOFs,

molecular complexes (e.g., phthalocyanines), and

hybrid heterojunctions. CO₂ hydrogenation relies on

metal oxides (e.g., CeO₂, In₂O₃), metals (e.g., Cu, Fe,

Ru), bifunctional zeolites, bimetallic/ternary catalysts

(e.g., Cu/ZnO/Al₂O₃), and carbides. Cycloaddition

catalysts include ionic liquids, MOFs, porous

polymers, metal complexes, and covalent organic

frameworks (COFs). CO₂ bioconversion utilizes

enzymes (e.g, formate dehydrogenase, carbonic

anhydrase) and microbial cells (e.g., Clostridium,

engineered strains).

4.2 Research Gap Analysis

Through bibliometric analysis, this study identifies

key characteristics and persistent challenges in

current CO₂ conversion research. The product

distribution reveals a pronounced focus on C1

chemicals (carbon monoxide, methane, and

methanol), while industrially critical higher-carbon

Figure 4: The knowledge map of CO

conversion technical pathways and catalysts.

KDIR 2025 - 17th International Conference on Knowledge Discovery and Information Retrieval

478

Figure 5: Faraday efficiency, selectivity, conversion rate distribution.

compounds (e.g., olefins, aromatics like

toluene/xylene, and synthetic fuels) remain

understudied. Performance metrics (Figure 5) reveal

significant limitations: only 12% of studies achieve≥

99% product selectivity (predominantly for C1

species), while advanced carbon products exhibit

substantially lower selectivity. CO₂ conversion

efficiencies exceed 90% occur in merely 4% of cases,

and full conversion (100%) in merely 1%. Notably,

only 37% of catalytic systems demonstrate Faradaic

efficiencies above 90%, indicating substantial energy

efficiency deficits.

The current technical status lags significantly

behind the strategic objectives of major economies.

For instance, Japan’s updated Carbon Recycling

Technology Roadmap prioritizes commercializing

polycarbonate and bio-jet fuel by 2030, followed by

industrial-scale olefin, aromatic compound, and

synthetic fuel production by ~2040. The significant

gap between current research progress and these

industrial targets underscores fundamental hurdles in

CO₂ conversion, particularly in precise selectivity

control, energy utilization optimization, and

multicomponent product.

5 CONCLUSION AND PROSPECT

This study integrates LLM-based semantic parsing

with prompt engineering for in-depth knowledge

mining in the specific research area. The

effectiveness of the method has been empirically

verified in the research area of CO₂ conversion and

utilization. By developing a multi-stage prompt

framework, we systematically extracted key scientific

indicators to construct a fine-grained research area

database. Based on this methodological system, we

have identified some new research hotspots and gaps

in the field of carbon dioxide conversion and

utilization, which is of great significance for

optimizing the research layout and direction in this

area. This approach effectively overcomes limitations

of traditional technical theme analysis, such as coarse

indicator granularity and insufficient quantitative

characterization, and enriches the methods and

perspectives of knowledge discovery in the field of

scientific research.

Current limitations include suboptimal accuracy

in extracting specific key performance indicators.

Further research will prioritize optimizing prompt

engineering to enhance extraction precision.

Additionally, while this study focuses on academic

literature, we plan to expand data sources to include

patents and industrial reports, thereby building a more

comprehensive dataset. Leveraging this foundation,

we will advance knowledge graph development for

CO₂ conversion technology, emphasizing core

functionalities such as technology development

roadmap and patent-technology correlation analysis,

ultimately supporting strategic advancements in this

critical field.

ACKNOWLEDGMENT

The authors gratefully acknowledge the support

provided by Strategic Priority Research Program of

the Chinese Academy of Sciences, Grant No.

XDC0230605 and Special Project of Strategic

Research and Decision Support System Construction

of Chinese Academy of Sciences (GHJ-ZLZX-2025-

07).

REFERENCES

Anton, E., Oesterreich, T. D., & Teuteberg, F. (2022). the

Property of Being Causal–the Conduct of Qualitative

Comparative Analysis in Information Systems

Research. Information & Management, 59(3), 103619.

Chang, Y. W., Huang, M. H., & Lin, C. W. (2015).

Evolution of Research Subjects in Library and

Information Science Based on Keyword,

Bibliographical Coupling, and Co-Citation Analyses.

Scientometrics, 105(3), 2071-2087.

Identifying Research Hotspots and Research Gaps in Speciﬁc Research Area Based on Fine-Grained Information Extraction via Large

Language Models

479

Dagdelen, J., Dunn, a., Lee, S., Walker, N., Rosen, a. S.,

Ceder, G., ... & Jain, a. (2024). Structured Information

Extraction from Scientific Text with Large Language

Models. Nature Communications, 15(1), 1418.

Datta, S., & Roberts, K. (2022). Fine-Grained Spatial

Information Extraction in Radiology as Two-Turn

Question Answering with BERT. International Journal

of Medical Informatics, 162, 104754.

Gupta, T., Zaki, M., Krishnan, N. M. a., & Mausam. (2022).

Matscibert: a Materials Domain Language Model for

Text Mining and Information Extraction. Npj

Computational Materials, 8, 141.

Hu, Z., Wang, M., & Han, Y. (2024). Multidimensional

Indicator Identification and Evolution Analysis of

Emerging Technology Topics Based on LDA2Vec-

BERT——a Case Study of Blockchain Technology in

the Field of Disruptive Technology. Journal of Modern

Information, 44(9), 42-58. (in Chinese)

Liu, Y. (2025). Discovering Topics and Trends in

Biosecurity Law Research: A Machine Learning

Approach. One Health, 20, 100964.

Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., … &

Stoyanov, V. (2019). Roberta: a Robustly Optimized

BERT Pretraining Approach. Arxiv Preprint

Arxiv:1907.11692. Https://Arxiv.Org/Abs/1907.11692.

Liu, Z., Yang, J., & Ma, Y. (2023). Research on Keyword

Survival Model from the Perspective of Term Function.

Journal of Intelligence, 42(8), 150-156+176. (in

Chinese)

Mohtasham, F., Yazdizadeh, B., & Mobinizadeh, M.

(2023). Research Gaps Identified in Iran’s Health

Technology Assessment Reports. Health Research

Policy and Systems, 21(1), 132.

Morris, S. a., Yen, G., Wu, Z., & Asnake, B. (2003). Time

Line Visualization of Research Fronts. Journal of the

American Society for Information Science and

Technology, 54(5), 413-422.

Onishi, T., Kadohira, T., & Watanabe, I. (2018). Relation

Extraction with Weakly Supervised Learning Based on

Process-Structure-Property-Performance

Reciprocity. Science and Technology of Advanced

Materials, 19(1), 649-659.

Qiao, X., Gu, S., Cheng, J., Peng, C., Xiong, Z., Shen, H.,

& Jiang, G. (2025). Fine-Grained Entity Recognition

via Large Language Models. IEEE Transactions on

Neural Networks and Learning Systems.

Rodríguez, a. J. C., Castro, D. C., & García, S. H. (2022).

Noun-Based Attention Mechanism for Fine-Grained

Named Entity Recognition. Expert Systems with

Applications, 193, 116406.

Tan, C., & Xiong, M. (2021). Contrastive Analysis at Home

and Abroad on the Evolution of Hot Topics in the Field

of Data Mining Based on LDA Model. Information

Science, 39(4), 174-185. (in Chinese)

Thakuria, a., & Deka, D. (2024). a Decadal Study on

Identifying Latent Topics and Research Trends in Open

Access LIS Journals Using Topic Modeling

Approach. Scientometrics, 129(7), 3841-3869.

Wang, G., Shi, R., Cheng, W., Gao, L., & Huang, X. (2023).

Bibliometric Analysis for Carbon Neutrality with

Hotspots, Frontiers, and Emerging Trends Between

1991 and 2022. International Journal of Environmental

Research and Public Health, 20(2), 926.

Wei, X., Cui, X., Cheng, N., Wang, X., Zhang, X., Huang,

S., ... & Han, W. (2023). Chatie: Zero-Shot Information

Extraction via Chatting with Chatgpt. Arxiv Preprint

Arxiv:2302.10205.

Westgate, M. J., Barton, P. S., Pierson, J. C., &

Lindenmayer, D. B. (2015). Text Analysis Tools for

Identification of Emerging Topics and Research Gaps

in Conservation Science. Conservation Biology, 29(6),

1606-1614.

Wu, R., Zong, H., Wu, E., Li, J., Zhou, Y., Zhang, C., ... &

Shen, B. (2025). Improving Large Language Models

for Mirna Information Extraction via Prompt

Engineering. Computer Methods and Programs in

Biomedicine, 109033.

Xu, X., Zheng, Y., & Liu, Z. (2016). Study on the Method

of Identifying Research Fronts Based on Scientific

Papers and Patents. Library and Information Service, 60

(24), 97-106. (in Chinese)

Yin, X., Zheng, S., & Wang, Q. (2021). Fine-Grained

Chinese Named Entity Recognition Based on Roberta-

WWM-Bilstm-CRF Model. 6th International

Conference on Intelligent Computing and Signal

Processing.

Yu, L., Qian, L., Fu, C., & Zhao, H. (2019). Extracting

Fine-Grained Knowledge Units from Texts with Deep

Learning. Data Analysis and Knowledge Discovery,

3(1), 38-45. (in Chinese)

Zhang, X., Li, P., & Li, H. (2020). AMBERT: a Pre-Trained

Language Model with Multi-Grained Tokenization.

Arxiv Preprint Arxiv:2008.11869.

KDIR 2025 - 17th International Conference on Knowledge Discovery and Information Retrieval

480