An AI-Driven Methodology for Patent Evaluation in the IoT Sector:
Assessing Relevance and Future Impact
Lelio Campanile
2 a
, Renato Zona
1 b
, Antonio Perfetti
3
and Franco Rosatelli
3
1
Department of Engineering, Universit
`
a degli Studi della Campania, Via Roma 29, Aversa (CE), Italy
2
Department of Mathematics and Physics, Universit
`
a degli Studi della Campania, Caserta, Italy
3
Fondazione Ricerca e Imprenditorialit
`
a, Napoli, Italy
Keywords:
Patent Evaluation, Patent Classification, Machine Learning, LLM.
Abstract:
The rapid expansion of the Internet of Things has led to a surge in patent filings, creating challenges in
evaluating their relevance and potential impact. Traditional patent assessment methods, relying on manual
review and keyword-based searches, are increasingly inadequate for analyzing the complexity of emerging IoT
technologies. In this paper, we propose an AI-driven methodology for patent evaluation that leverages Large
Language Models and machine learning techniques to assess patent relevance and estimate future impact.
Our framework integrates advanced Natural Language Processing techniques with structured patent metadata
to establish a systematic approach to patent analysis. The methodology consists of three key components:
(1) feature extraction from patent text using LLM embeddings and conventional NLP methods, (2) relevance
classification and clustering to identify emerging technological trends, and (3) an initial formulation of impact
estimation based on semantic similarity and citation patterns. While this study focuses primarily on defining
the methodology, we include a minimal validation on a sample dataset to illustrate its feasibility and potential.
The proposed approach lays the groundwork for a scalable, automated patent evaluation system, with future
research directions aimed at refining impact prediction models and expanding empirical validation.
1 INTRODUCTION
The proliferation of the Internet of Things (IoT) has
led to a marked increase in patent filings, underscor-
ing the rapid growth of innovation win this domain.
As IoT technologies continue to expand across vari-
ous sectors, such as smart cities, healthcare, industrial
automation and transportation systems, the global
number of patents granted worldwide has seen a sig-
nificant increase, that poses significant challenges for
researchers, policymaker, and industry stakeholders
who seek to assess their relevance, technological im-
pact, and potential future influence. The ability to
efficiently assess and evaluate patents is becoming
crucial, as it supports innovation, assures fair intel-
lectually property practices, and informs investment
strategies. However, traditional workflows for patents
evaluation and management shows theirs limitations
in handle large volumes of documents in special way
while capturing the semantic inherent within patents.
a
https://orcid.org/0000-0003-4021-4137
b
https://orcid.org/0000-0001-6718-9387
The evaluation of patents has traditionally relied
on manual review by domain experts and keyword-
based retrieval systems. This approach limits the scal-
ability, rendering it ineffective for large-scale analy-
sis. Manual evaluations and classifications, are prone
to inefficiencies and inconsistencies due to their re-
liance on human judgment. Furthermore, keyword-
based retrieval systems frequently fail to capture the
subtleties and complexities of patent text, particu-
larly within interdisciplinary domains such as IoT,
where the interconnections between concepts can be
highly context-dependent. Traditional impact assess-
ments rely heavily on citation-based metrics, which
may not provide a comprehensive understanding of a
patent’s potential influence or innovation value. This
limitation renders it challenging to accurately evalu-
ate the pertinence, applicability, and overall impact of
patents within rapidly evolving fields.
The integration of Artificial Intelligence (AI) and
Natural Language Processing (NLP) techniques into
patent analysis has activated advancements in auto-
mated and scalable evaluation methodologies. Recent
breakthroughs in Large Language Model (LLM), in-
Campanile, L., Zona, R., Perfetti, A. and Rosatelli, F.
An AI-Driven Methodology for Patent Evaluation in the IoT Sector: Assessing Relevance and Future Impact.
DOI: 10.5220/0013519700003944
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 10th International Conference on Internet of Things, Big Data and Security (IoTBDS 2025), pages 501-508
ISBN: 978-989-758-750-4; ISSN: 2184-4976
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
501

cluding GPT, and domain-specific embeddings, have
demonstrated superior performance in text compre-
hension, semantic similarity analysis, and automated
classification tasks. These LLMs facilitate the ex-
traction of high-dimensional, context-aware represen-
tations of patent texts. The application of LLMs
to patent analysis is a novel research are with unre-
solved challenges. Some of these challenges include
data sparsity, which hinders the development of ro-
bust models; domain adaptation, which necessitates
the customization of models for specific domains; and
explainability of AI-driven decisions, which raises
concerns regarding the interpretability and trustwor-
thiness of results.
In this paper, we present a novel AI-driven
methodology for patent evaluation that harnesses the
power of state-of-the-art LLMs and machine learning
techniques to assess the relevance and impact of IoT
patents. Our approach is structured around four key
components.
We employ LLM-based embeddings and tradi-
tional NLP methods for feature extraction, encoding
the semantic meaning of patents. This stage produces
unstructured data, which is then transformed into a
structured one using a structured data output frame-
work for LLM. The next step involves applying rele-
vance classification and clustering to identify the un-
derlying technological patterns within the IoT patent
landscape. Finally, our framework estimates the pre-
liminary impact of each patent through a combination
of citation analysis, semantic similarity, and time-
series forecasting techniques. Unlike previous ap-
proaches that rely solely on keyword retrieval or cita-
tion counts, our methodology aims to provide a com-
prehensive, data-driven framework for understanding
the significance of patents and saving in a structured
way.
The primary contribution of this research is in
its methodological framework, which joins conven-
tional patent evaluation workflows with AI-driven
techniques. By incorporating LLMs and deep learn-
ing models, we created a novel pipeline that not only
enhances the accuracy of patent analysis but also ad-
dresses the limitations of existing approaches. Fur-
thermore, our work presents an initial validation on a
sample dataset, demonstrating the feasibility and po-
tential of the proposed approach. Although this val-
idation is limited in scope, it serves as a foundation
for future research aimed at refining impact prediction
models and expanding empirical validation to larger
datasets.
The remainder of this paper is organized as fol-
low: Section 2 provides an overview of context and
background on previous work on patent analysis. Sec-
tion 3 describes the proposed methodology. Section
4 presents a minimal validation study, demonstrat-
ing the applicability of our method to a selected IoT
patent set. Section 5 discusses the implications of our
findings, comparing our approach with existing tech-
niques and discussing limitations. Finally, Section 6
concludes the paper and highlights future research di-
rections.
2 RELATED WORKS
Historically, the patent system was established to en-
courage and regulate technical progress and innova-
tion (Frumkin, 1947). However, increasing global
patent applications and rapid technological develop-
ment pose significant challenges to patent offices and
practitioners (Krestel et al., 2021).
Automating these processes not only improves the
efficiency of IP management, but also allows the NLP
community to explore a promising and still under-
researched field. Previous research has examined the
early stages of the use of intelligent, automated meth-
ods for patent analysis (Abbas et al., 2014), the in-
troduction of Deep Learning (DL) techniques, which
have simplified some patent-related tasks (Krestel
et al., 2021), and specific applications, such as patent
retrieval (Shalaby and Zadrozny, 2019).
Here, we focus on two main areas: patent analy-
sis and automatic patent text generation. While anal-
ysis is aimed at understanding and extracting rele-
vant information from this vast knowledge base (Ab-
bas et al., 2014), generative activities aim at the auto-
mated creation of patent content. Several studies have
explored the use of Machine Learning (ML) and NLP
in the patent domain, which is characterized by highly
technical and legal texts (Krestel et al., 2021). In ad-
dition, recent LLMs have demonstrated extraordinary
capabilities across a wide range of tasks in general
domains (Min et al., 2023), suggesting their potential
use in the management and editing of patent literature,
an essential resource for documenting technological
evolution. However, their use in patent-related activi-
ties still remains under-explored, due to the complex-
ity of the texts and the field itself. The work of (Ab-
bas et al., 2014) represents some of the first research
devoted to patent analysis, adopting text mining and
visualization methods, which opened up new perspec-
tives for study. DL has recently gained attention for
knowledge management and patents. (Krestel et al.,
2021) identified eight main task categories for DL-
based methods, including decision support, patent
classification and retrieval, patent evaluation, predic-
tion of emerging technologies, automatic patent text
AI4EIoT 2025 - Special Session on Artificial Intelligence for Emerging IoT Systems: Open Challenges and Novel Perspectives
502

generation, litigation analysis and the use of computer
vision techniques. Some studies have focused on spe-
cific aspects of patent analysis. For example,(Shalaby
and Zadrozny, 2019) examined the problem of patent
retrieval, i.e. the search for relevant documents, an ac-
tivity that, although it may seem similar to traditional
searching, has well-defined constraints and purposes.
Patent literature is an important source of knowledge
on technological advances and is known for its poten-
tial to inspire innovations, especially when technolo-
gies from different fields are combined.
Data science techniques must address these diffi-
culties in order to effectively extract useful informa-
tion for design and innovation (Jiang et al., 2022).
The application of data science to patents not only
supports theoretical and methodological design anal-
ysis, but also provides tools for innovation strategies.
In particular, in the early stages of technology de-
velopment, NLP can support idea generation, predic-
tion of industry trends and the matching of problems
and solutions (Just, 2024). While much research fo-
cuses on analyzing existing patent literature, there is
a growing interest in the generative use of language
models. NLP techniques can be exploited to make
patent language more accessible, generate abstracts
or even create new patent texts from structured in-
put (Casola and Lavelli, 2022). The recent devel-
opment in the fields of LLM are reported in (Chang
et al., 2024). The field of patents also connects to
other areas of intellectual property, such as copyright,
trademarks and industrial designs, as well as interdis-
ciplinary areas such as knowledge management and
the assessment of the economic value of innovations
(Aristodemou and Tietze, 2018). Moreover, the legal
language of patents shares many features with other
legal texts, so much so that most practitioners in the
field have a strong legal background. For this reason,
NLP techniques developed for law in general could
be applied to patent analysis in the future (Katz et al.,
2023).
Although LLMs offer significant potential for
knowledge extraction and textual analysis, their ap-
plication in patenting is still limited. Previous studies
have employed word embeddings such as Word2Vec
(Mikolov et al., 2013) and DL models such as LSTM,
but transformer-based models have opened up new
possibilities (Vaswani et al., 2017). Early attempts
to apply LLMs to patents were based on relatively
small models, such as GPT-2 (Radford, 2018), while
more advanced models have not yet been explored in
depth in this area. A significant obstacle is the lack
of reference datasets and established metrics to as-
sess the performance of LLMs in patent analysis. Al-
though patent offices release raw documents, publicly
available datasets for specific patent analysis tasks are
still limited. Current searches focus mainly on sum-
mary sections of patents, but detailed descriptions and
claims are much more relevant for automated patent
analysis and generation.
3 PROPOSED METHODOLOGY
Patent evaluation and classification is an activity for
both businesses and regulators, as it allows them
to determine the technological, economic and legal
value of an invention. Patent analysis can be con-
ducted through traditional methodologies entrusted to
experts in the field, as well as through the use of auto-
mated tools based on machine learning and artificial
intelligence techniques. The analysis conducted by
human experts is based on a set of established crite-
ria that include novelty, inventive step and industrial
applicability of the patent. These parameters are ex-
amined within the existing patent and scientific liter-
ature to assess whether the invention meets the legal
requirements for patentability. In addition, the analy-
sis considers the clarity of the claims, scope of pro-
tection and potential market coverage. Patent ana-
lysts, often with a technical and legal background,
conduct extensive searches of specialized databases
and compare new applications with prior patents to
identify any overlaps or contrasts with the state of the
art. In parallel, recent years have seen an increas-
ing development of IT tools for patent classification
and evaluation. These tools are based on NLP and
ML techniques, which make it possible to extract rel-
evant information from patent documents. For exam-
ple, the use of text mining algorithms makes it possi-
ble to identify semantic relationships between patents
and automatically classify inventions into technology
categories defined by international standards, such as
the International Patent Classification (IPC) and the
Cooperative Patent Classification (CPC). Another in-
novative approach involves DL models and text em-
bedding techniques to improve the ability to identify
similarities between existing patents and new appli-
cations, facilitating the evaluation process. Further-
more, LLMs are showing significant potential in gen-
erating patent abstracts and translating technical lan-
guage into a more accessible form. Despite these ad-
vances, the application of AI techniques to the patent
field still presents some challenges. One of the main
difficulties concerns the length and complexity of
patent texts, which often exceed the processing ca-
pacity of traditional language models. Furthermore,
patent language is characterized by highly specific
terminology and complex syntactic structures, mak-
An AI-Driven Methodology for Patent Evaluation in the IoT Sector: Assessing Relevance and Future Impact
503

ing automated interpretation difficult without a train-
ing phase on specific corpora. To address these chal-
lenges, scholars are exploring several strategies, in-
cluding increasing the context window in language
models and adopting Retrieval Augmented Genera-
tion (RAG) techniques, which combine information
retrieval from structured sources with text generation.
These developments could improve the ability to clas-
sify and evaluate patents more accurately and effi-
ciently. In conclusion, patent evaluation and classi-
fication today relies on both the expertise of human
analysts and the potential offered by automated tools
based on artificial intelligence. While the former pro-
vide in-depth and contextualized interpretation, the
latter offer a large-scale analysis capability, reducing
the time and cost of the process. The combination
of these two methodologies appears to be the most
promising way to further improve the management
and analysis of patent literature.
3.1 The Data Model
The proposal is an advanced patent analysis tool
based on information extraction technology and se-
mantic, functional and graphical analysis of patent
texts and related search reports. It leverages AI al-
gorithms, trained on a large patent domain, to pro-
vide an in-depth analysis of the individual patent, its
relative positioning and reference context. The refer-
ence patent domain consists of the worldwide dataset
of existing patents. The advantage is the free avail-
ability of the worldwide database of existing patents
as it is in the public domain. The reference patent
domain databases used as a first approximation are
WIPO, ESPO, ESPACENET, GOOGLE PATENT.
Thanks to the advanced analysis of the proposal,
it is possible to:
• Make strategic decisions based on concrete data;
• Identify emerging technology trends;
• Define patent policies and technology improve-
ment drivers;
• Discover new customers, suppliers, partners or
competitors;
• Identify new markets.
The proposal differs from traditional tools in that it
operates on unstructured information databases, mak-
ing consultation and analysis more efficient, reducing
time and associated costs. The proposal for patent
documents is a comparative test prior to patent docu-
ment submission or parallel to the patent process for
patents already filed but still reserved. This tool acts
as a complement to the prior art analysis, supporting
the investment decision-making process and the start
of the formal patenting process. The output of the
proposal is a radar diagram, which assigns a score on
a scale of 1 to 5 for each of the seven patent indexes,
already trained on the patent reference database:
• Feasibility, literally represents the feasibility of
the patent from a realization point of view, taking
into account complexity, availability of materials,
the need for ad hoc structures for relitigation, and
the maturity of the technological environment ;
• Multipurpose intended as the ability of the in-
vention to have applications outside the originally
chosen field ;
• Obsolescence intended as how far it is from the
end of its useful life. This parameter does not co-
incide with the time span of the legal concession;
• Geo Context indicates how attractive a market is
in the world market;
• Protection represents how protected the invention
is;
• Tech Interest, represents how the subject of inven-
tion is of global interest;
• Tech Resonance, represents how much redun-
dancy there is in the world panorama of inventions
with respect to the subject matter.
These indexes assess both the intrinsic value of the
patent and how it compares with the external context:
• Technological state of the art ;
• Territorial extension level ;
• Geographical distribution of comparable patent
families ;
• Population of similar inventions.
A Global Score summarizes the overall patent value
based on the above indexes. The proposal allows:
• Attribution of ‘weighted’ index scales (arithmetic
or logarithmic);
• Categorization and ranking of patents contained
in a portfolio;
• Comparison with predefined minimum thresholds
or derived from repetitive analysis;
The proposal allows to compare patents within
your portfolio or with previously analyzed patents.
In addition, it allows you to generate a customized
Global Score, based on specific weightings, to assess
the overall patent value.
The proposal represents an innovative and repeat-
able system for patent evaluation, in pre-grant stage
and in the whole patent life cycle, providing strategic
AI4EIoT 2025 - Special Session on Artificial Intelligence for Emerging IoT Systems: Open Challenges and Novel Perspectives
504

support for investment decisions, technological im-
provement and competitive positioning. Thanks to its
advanced technology, it is able to offer an in-depth
and predictive view of the value and impact of patents
in an increasingly dynamic and competitive market
environment.
3.2 Feature Extraction and NLP
Techniques
Our methodology is founded on conventional NLP
techniques but extends them by utilizing LLMs, par-
ticularly LLaMA 70B, to obtain and examine patent
data. Although traditional methods are effective for
preliminary text analysis, LLMs provide the capacity
to understand deeper semantic connections and con-
textual subtleties frequently ignored by simpler mod-
els.
A significant advancement in our methodology is
the generation of structured data from LLMs gener-
ated content. Patent documents are intrinsically un-
structured, rendering methodical processing challeng-
ing. We employ advanced structured output genera-
tion algorithms to transform free-text patent descrip-
tions into a standardized format that corresponds with
the main criteria outlined in previous section. Tech-
niques such as prompt engineering, instruction tweak-
ing, and RAG are utilized producing structured repre-
sentations of patents. Recent scholarly research has
proven the efficacy of these strategies in retrieving
domain-specific knowledge.
To guarantee high-quality structured data, we uti-
lize schema-based extraction, directing LLM to pro-
duce data in specified types, such as JSON or XML.
This facilitates easy integration with subsequent ma-
chine learning models. Furthermore, few-shot learn-
ing enhances the dependability of structured extrac-
tion by supplying instances of accurately formatted
data during inference, improving consistency and di-
minishing ambiguity.
Another significant feature is constraint-based ex-
traction, wherein the LLM-generated content is ver-
ified against established rules or knowledge bases.
Post-processing approaches, including regex valida-
tion, ontology-based filtering, and entity linking, are
utilized to enhance structured data and rectify errors.
These procedures guarantee that extracted fields, in-
cluding assignees, citations, patent classification, and
important claims, comply with domain-specific spec-
ifications.
Following the extraction process, the structured
data is consolidated into a centralized dataset de-
signed for ML applications. This dataset supports
patent comparison and review, facilitating more rigor-
ous study. The organized dataset has fields like tech-
nological categorization, novelty indicators, existing
relationships, impact scores, and key entities such as
assignees, inventors, citations, and jurisdictional cov-
erage.
We primarily utilize zero-shot and few-shot learn-
ing paradigms, while also investigating fine-tuning as
a supplementary upgrade. Fine-tuning LLaMA 70B
on specialized corpora, such historical patent records
and annotated datasets, may enhance the model’s pre-
cision and recall. Nevertheless, considering computa-
tional limitations and ethical concerns related to pro-
prietary data, fine-tuning is not the central emphasis
of this study. We utilize adaptive prompting tactics
and context-aware embeddings to enhance model ef-
ficiency without necessitating significant retraining.
We assess various approaches for the practical
implementation of structured data extraction from
LLMs. Among the most promising solutions are
LangChain, Hugging Face’s Transformers, and LLa-
maIndex, each providing comprehensive APIs for
dealing with various LLMs and producing structured
outputs.
Our methodology provides a scalable and auto-
mated framework for obtaining significant insights
from patents by integrating traditional NLP pipelines
with sophisticated NLP capabilities.
3.3 Relevance Classification and Impact
Estimation
Patent relevance classification is essential for assess-
ing the importance of a patent within the IoT do-
main. We plan to utilize a combination of supervised
machine learning models and unsupervised clustering
techniques to evaluate and classify patents according
to their technological and economic worth. We will
apply random forest classifiers, Support Vector Ma-
chines (SVM) , and DL models for relevance classi-
fication to assess the significance of a patent based
on extracted data. These models are trained on his-
torical patent datasets with annotated relevance rat-
ings, ensuring strong performance in patent catego-
rization. Furthermore, we employ hierarchical clus-
tering to discern clusters of patents that display anal-
ogous technological attributes, thereby offering in-
sights into nascent trends in IoT technologies.
In addition to classification, estimating the im-
pact of a patent is crucial for forecasting its long-term
effects. We combine several predictive indicators,
such as citation trajectory analysis, semantic similar-
ity to high-impact patents, and time-series forecasting
of technological uptake. Citation-based methodolo-
gies analyze the historical citation trends of patents to
An AI-Driven Methodology for Patent Evaluation in the IoT Sector: Assessing Relevance and Future Impact
505

Figure 1: Flowchart of the proposal.
evaluate their anticipated future impact. Furthermore,
embedding-based semantic similarity metrics enable
the assessment of the alignment of a specific patent
with historically significant patents.
Our methodology integrates classification and ef-
fect estimation, offering a thorough evaluation frame-
work that connects traditional and AI-driven patent
examination. Utilizing ML and insights from LLM,
we develop a system that efficiently categorizes
patents and forecasts their significance in influencing
future technical progress.
4 PRELIMINARY VALIDATION
We evaluate the feasibility of our proposed method-
ology through a preliminary validation utiliz-
ing a representative case study based on patent
WO2020193804A1, widely described in (Di Gennaro
et al., 2024; Minutolo et al., 2020; Gennaro et al.,
2022), which relates to an IoT-related invention. This
validation demonstrates the efficacy of our structured
extraction pipeline and classification models in as-
sessing patent relevance and estimating potential im-
pact.
WO2020193804A1, Transducer, is selected as a
case study due to its considerable relevance in the
IoT domain. The patent outlines a new method for
distributed data management in IoT networks, posi-
tioning it as a suitable candidate to assess the efficacy
Figure 2: Index scoring of the evaluated patent.
of our approach in extracting structured insights and
forecasting its technological implications.
The results in terms of patent indexes are reported,
see Figure 2.
The structured data extraction module is utilized
on Transducer, producing a machine-readable rep-
resentation that includes essential attributes such as
technological classification, key claims, prior art ref-
erences, and citation trajectories. Our classification
model predicts the relevance score of a patent by an-
alyzing its textual content and assessing its semantic
similarity to high-impact patents within our dataset.
The evaluation of classification performance is
AI4EIoT 2025 - Special Session on Artificial Intelligence for Emerging IoT Systems: Open Challenges and Novel Perspectives
506

conducted through three primary metrics: best fit
similarity to established high-impact patents within
an embedding space, semantic overlap with glsiot-
specific patent clusters identified via hierarchical
clustering, and citation-based influence indicators ob-
tained from historical patent citation trends.
We compare the structured outputs produced by
LLaMA 70B, with manually annotated patent data to
validate our methodology. Factors considered include
the accuracy of extracted entities, such as assignees,
inventors, and technological components; the com-
pleteness of structured data fields in comparison to
human-extracted insights; and the consistency of rele-
vance classification against expert-labeled patent cat-
egories.
This case study illustrates our approach, yet it has
certain limitations. The validation relies on a singular
patent, necessitating expansion to a broader dataset
for enhanced generalization. Estimating impact based
on citations necessitates longitudinal data, which may
not be completely accessible for recent patents. Fur-
thermore, although LLMs offer structured insights,
their dependability in patent-specific contexts neces-
sitates additional benchmarking against evaluations
by domain experts.
This preliminary validation indicates that our
methodology effectively extracts structured informa-
tion from IoT-related patents, establishing a founda-
tion for evaluating relevance and impact. Further re-
search is necessary to enhance the classification mod-
els, broaden the dataset, and strengthen the reliability
of impact predictions. Future research will concen-
trate on expanding the validation scale and incorpo-
rating supplementary benchmark datasets to enhance
the evaluation comprehensiveness.
5 DISCUSSION
The proposed AI-driven methodology for patent eval-
uation presents notable improvements in the anal-
ysis of intellectual property. The integration of
LLM with ML techniques signifies a significant shift
from traditional methods, providing enhanced accu-
racy, scalability, and automation. This method im-
proves the assessment of patents through more ac-
curate relevance classification and impact forecast-
ing. LLM embeddings enhance the comprehension
of patent texts by effectively capturing semantic re-
lationships that keyword-based methods frequently
miss. This advancement facilitates more efficient
clustering of patents, thereby assisting in the iden-
tification of emerging technological trends. The in-
tegration of citation analysis and semantic similarity
measures enhances impact prediction, offering a com-
prehensive perspective on a patent’s significance over
time.
This AI-driven evaluation framework provides ad-
vantages to stakeholders, including patent offices, re-
search institutions, and corporations engaged in IoT
innovation. Patent examiners can utilize the proposed
method to enhance the efficiency of the patent ap-
proval process by more effectively identifying rel-
evant prior art. Research institutions may employ
the methodology to monitor technological advance-
ments and evaluate the potential impact of specific
patents on future research trajectories. Businesses
can employ the system to inform decisions regarding
patent investments and strategic collaborations, uti-
lizing predicted impact metrics. AI-based evaluation
serves as a transformative instrument in the manage-
ment of intellectual property.
While this approach offers certain advantages, it
is essential to recognize several limitations. The effi-
cacy of LLM is contingent upon the quality and di-
versity of the training data utilized. LLMs trained
on general corpora exhibit robust performance across
various NLP tasks; however, fine-tuning on domain-
specific patent texts may be necessary for optimal out-
comes. Future research should investigate the cre-
ation of specialized LLMs that are specifically trained
on patent data to improve accuracy and relevance.
The preliminary validation study is constrained in
scope due to its reliance on a limited dataset of IoT
patents. A comprehensive evaluation utilizing larger
and more diverse patent datasets is essential to vali-
date the generalization of the methodology. The se-
lection of impact prediction metrics must be refined
to ensure robustness across various technological do-
mains.
A significant challenge in AI-driven patent eval-
uation is the interpretability of the results. LLMs
offer robust semantic representations; however, their
decision-making processes frequently lack trans-
parency. Enhancing the usability and trustworthi-
ness of relevance classification and impact prediction
can be achieved through the development of explain-
able AI techniques that elucidate the reasoning behind
these processes. The comprehensibility and valida-
tion of AI-generated evaluations by human experts is
a critical focus for future development.
6 CONCLUSIONS AND FUTURE
WORKS
This study introduces an AI-based approach for as-
sessing IoT-related patents, including sophisticated
An AI-Driven Methodology for Patent Evaluation in the IoT Sector: Assessing Relevance and Future Impact
507

NLP techniques and ML models. Our methodology
offers a scalable and automated solution for patent
analysis through the integration of LLM-based struc-
tured data extraction, relevance categorization, and
impact calculation. The findings from our initial vali-
dation demonstrate that the methodology successfully
extracts significant insights from patents, presenting a
viable alternative to conventional review techniques.
The suggested methodology improves patent anal-
ysis by utilizing advanced LLMs to extract structured
information, categorizing patents according to their
technological relevance, and assessing their potential
impact through predictive modeling. Our research il-
lustrates that the integration of NLP approaches with
machine learning allows for a more thorough evalua-
tion of patents, aiding in the detection of significant
advances in the IoT sector.
Notwithstanding the encouraging outcomes, nu-
merous obstacles persist. The initial validation was
performed on a restricted dataset, and subsequent re-
search should aim to broaden the methodology to
encompass a wider and more varied collection of
patents. Furthermore, enhancing the effect estima-
tion model through advanced time-series forecasting
and the integration of external elements like market
adoption patterns could improve predictive accuracy.
A vital topic for enhancement is the augmentation
of the interpretability of AI-driven patent evaluation,
guaranteeing that the methodology yields practical in-
sights for academics, enterprises, and policymakers.
The next steps will focus on augmenting the
dataset, refining domain-specific LLM models for
patent analysis, and assessing the methodology across
several technological areas outside IoT. Additionally,
incorporating expert feedback into the review process
could enhance the trustworthiness of automated as-
sessments. This research addresses these problems,
enhancing AI-assisted intellectual property analysis
and facilitating more efficient, data-driven decision-
making in patent review.
REFERENCES
Abbas, A., Zhang, L., and Khan, S. U. (2014). A literature
review on the state-of-the-art in patent analysis.
Aristodemou, L. and Tietze, F. (2018). The state-of-the-art
on intellectual property analytics (ipa): A literature
review on artificial intelligence, machine learning and
deep learning methods for analysing intellectual prop-
erty (ip) data.
Casola, S. and Lavelli, A. (2022). Summarization, simplifi-
cation, and generation: The case of patents.
Chang, Y., Wang, X., Wang, J., Wu, Y., Yang, L., Zhu,
K., Chen, H., Yi, X., Wang, C., Wang, Y., Ye, W.,
Zhang, Y., Chang, Y., Yu, P. S., Yang, Q., and Xie,
X. (2024). A survey on evaluation of large language
models. ACM Transactions on Intelligent Systems and
Technology, 15.
Di Gennaro, L., de Cristofaro, M., Loreto, G., Minutolo, V.,
Olivares, L., Zona, R., and Frunzio, G. (2024). In-
situ load testing of an ancient masonry structure using
fibre optics. Structures, 70:107567.
Frumkin, M. (1947). Early history of patents for innovation.
Transactions of the Newcomen Society, 26.
Gennaro, L. D., Damiano, E., Cristofaro, M. D., Netti,
N., Olivares, L., Zona, R., Iavazzo, L., Coscetta, A.,
Mirabile, M., Giarrusso, G., D’Ettore, A., and Min-
utolo, V. (2022). An innovative geotechnical and
structural monitoring system based on the use of nsht.
Smart Materials and Structures, 31.
Jiang, S., Sarica, S., Song, B., Hu, J., and Luo, J. (2022).
Patent data for engineering design: A critical review
and future directions. Journal of Computing and In-
formation Science in Engineering, 22.
Just, J. (2024). Natural language processing for innovation
search – reviewing an emerging non-human innova-
tion intermediary. Technovation, 129.
Katz, D. M., Hartung, D., Gerlach, L., Jana, A., and Bom-
marito, M. J. (2023). Natural language processing in
the legal domain. SSRN Electronic Journal.
Krestel, R., Chikkamath, R., Hewel, C., and Risch, J.
(2021). A survey on deep learning for patent analy-
sis. World Patent Information, 65.
Mikolov, T., Chen, K., Corrado, G., and Dean, J. (2013).
Efficient estimation of word representations in vector
space. In 1st International Conference on Learning
Representations, ICLR 2013 - Workshop Track Pro-
ceedings.
Min, B., Ross, H., Sulem, E., Veyseh, A. P. B., Nguyen,
T. H., Sainz, O., Agirre, E., Heintz, I., and Roth, D.
(2023). Recent advances in natural language process-
ing via large pre-trained language models: A survey.
ACM Computing Surveys, 56.
Minutolo, V., Cerri, E., Coscetta, A., Damiano, E., Cristo-
faro, M. D., Gennaro, L. D., Esposito, L., Ferla,
P., Mirabile, M., Olivares, L., and Zona, R. (2020).
Nsht: New smart hybrid transducer for structural and
geotechnical applications. Applied Sciences (Switzer-
land), 10.
Radford, A. (2018). Improving language understanding
by generative pre-training. Homology, Homotopy and
Applications, 9.
Shalaby, W. and Zadrozny, W. (2019). Patent retrieval: a
literature review.
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones,
L., Gomez, A. N., Łukasz Kaiser, and Polosukhin,
I. (2017). Attention is all you need. In Advances
in Neural Information Processing Systems, volume
2017-December.
AI4EIoT 2025 - Special Session on Artificial Intelligence for Emerging IoT Systems: Open Challenges and Novel Perspectives
508