Development of text analytics methods for supporting intellectual property assessment

Abstract

In modern economies where knowledge and technology are central sources of growth, intellectual property (IP) serves as a core legal framework for defining and protecting intangible assets. Among various forms of IP, patents and trademarks are two major instruments that govern the treatment of technological inventions and commercial identifiers. As the number and complexity of patent and trademark filings have grown, the need for systematic approaches to assessing such IP assets has become a critical concern for firms, public institutions, and policymakers. This dissertation frames these issues under the notion of IP assessment, defined as the systematic examination and evaluation of IP assets for legal, strategic, and policy purposes, and develops text-based analytical methods to support three key stakeholder groups: evaluators, examiners, and applicants. The first study addresses the needs of patent evaluators who seek to measure the technological novelty of inventions in a more fine-grained manner than is possible with conventional indicators based on classifications or citations in prior studies. It adopts a patent text analysis in which patent claim texts are converted into numerical vectors in a continuous space using a doc2vec model and applies the local outlier factor (LOF) technique to compute novelty scores that reflect how atypical each patent is relative to its local neighbors. An empirical analysis of 1,877 medical imaging patents filed at the United States Patent and Trademark Office between 1984 and 2019 shows that the resulting novelty scores are significantly associated with established patent indicators and that patents identified as highly novel exhibit higher technological impact on average. The second study focuses on the needs of trademark examiners, who must assess the similarity between the goods and services designated in new applications and those in existing trademarks. Using refined product descriptions provided by the Korean Intellectual Property Office (KIPO), it develops TM-BERT, a domain-adaptive Sentence-BERT model trained on 41,587 goods and 15,344 services. Goods and services are embedded into a low-dimensional semantic space using the trained model, and similarity scores are computed based on cosine similarity of the resulting vectors. Empirical analyses using product descriptions from KIPO confirm that TM-BERT effectively differentiates goods and services within and across similar group codes, demonstrating strong alignment with the European Union Intellectual Property Office CF Similarity database. The proposed model enhances accuracy and consistency in similarity assessments, serving as a complementary tool for trademark examiners in evaluating the similarity of goods and services. The third study turns to the context of trademark applicants, who must decide whether to file trademark applications under refusal uncertainty and resource constraints. It proposes a prescreening approach that learns from historical decisions on refused and registered trademarks to evaluate the distinctiveness of candidate trademarks under the Korean Trademark Act. Trademark examination records from the Korean Intellectual Property Right Information Service are organized into a balanced question–answer dataset aligned with eight distinctiveness-related refusal categories derived from Articles 33 and 34, and the generative language model, Gemma-3, is adapted through parameter- efficient fine-tuning using Low-Rank Adaptation (LoRA). The fine-tuned model produces, for each category, a binary judgment label and legally grounded explanation based on the given trademark and its designated goods or services. Validation on a test set shows that the model achieves moderate accuracy with relatively high precision in identifying refusal cases, and an LLM-as-judge assessment indicates that its refusal explanations attain a nontrivial level of legal and linguistic quality, while remaining below examiner-written explanations. Taken together, the three studies seek to advance text-based methods for IP assessment by addressing stakeholder-specific problems for patent evaluators, trademark examiners, and trademark applicants. From a theoretical perspective, the dissertation extends research on patent novelty, goods and services similarity, and trademark distinctiveness by formulating these issues as text-based assessment tasks that are explicitly grounded in how patents and trademarks are examined and used in practice. From a methodological perspective, it demonstrates how doc2vec combined with LOF, a domain-adaptive Sentence-BERT model, and a fine-tuned generative language model with parameter- efficient adaptation can be applied to patent and trademark texts to construct more fine-grained approaches to support IP assessment. From a practical perspective, the proposed approaches are intended to serve as complementary tools that help patent evaluators scan large patent sets for novel technologies, support trademark examiners in evaluating the similarity of goods and services and enable trademark applicants to anticipate refusal risks and prepare more informed filing strategies.