A computational framework for intellectual property litigation: from causal text analysis to LLM-based prediction systems

Abstract

Intellectual property (IP) litigation poses critical challenges for corporate innovation and judicial efficiency in knowledge-based economies. Yet measuring associated IP litigation risks and predicting litigation outcomes remain formidable challenges. This thesis develops a computational framework spanning causal inference, predictive modelling, and an LLM-based framework to address these challenges using China’s judicial text resources.
Effectively measuring latent intellectual property litigation risk is essential for optimizing innovation portfolio decisions for listed firms. The thesis starts with employing the online disclosure of judicial decisions (China Judgments Online) as a quasi-natural experiment and using a novel approach to construct a firm-level measure of intellectual property litigation risk through the Structural Topic Model. We aim to examine how firms adjust their innovation strategies under judicial disclosure reforms. The findings indicate that judicial disclosure incentivizes firms with lower litigation risk to engage in high-quality innovation, with this positive effect being more significant among firms with higher information transparency and no political connections. Mechanism analysis reveals that this effect operates primarily through two mechanisms: enhanced social trust and increased R&D investment. Furthermore, through causal pathway decomposition, we demonstrate that these two mechanisms coexist and reinforce each other in motivating innovation.
Given that the above approach cannot provide direct predictions of IP litigation outcomes for listed firms, the thesis extends the research by proposing a novel approach to effectively predict the IP litigation outcomes for listed firms using a machine learning technique. By integrating textual and financial data, we innovatively apply the Structural Topic Model (STM) in IP litigation outcome prediction models rather than the topic discovery task. Specifically, the thesis evaluates the effectiveness of STM, Latent Dirichlet Allocation (LDA), and Word Frequency (WF) models in analyzing and predicting the outcomes of IP litigation, leveraging data from China Judgement Online, which contains over 140 million lawsuits spanning 2010 to 2021. Our findings indicate that the STM, which integrates both financial and textual data, significantly enhances the accuracy of predictive models compared to those relying on a single input type. The STM exhibits strong predictive performance across a broad range of cases involving both plaintiffs and defendants, and across different IP types, which include copyrights, trademarks, and patents. It is also effective in predicting IP litigation outcomes of cross-border U.S. or European companies. In addition, we identify leading thematic topics associated with the win/loss outcome of IP litigation in both the full sample and the cross-border subsample using the STM framework, which illustrates the STM’s practical feasibility for predicting the IP litigation outcome. Our finding that STM outperforms LDA for IP litigation cases reflects STM’s superior strength in handling structured textual documents with covariates (e.g., financial variables).
The thesis further extends this research by developing predictive frameworks applicable to all IP litigation cases (not only to listed firms), thereby overcoming a limitation of prior prediction frameworks, namely their dependence on accessible financial covariates, and expanding applicability to the broader universe of intellectual property disputes. To be specific, we propose a Regional Judicial Multi-Agent Architecture (RJ-MAA), which leverages Large Language Models’ inherent semantic comprehension capabilities to eliminate dependency on external financial constraints while maintaining predictive accuracy. Simultaneously, by constructing specialized court agents for different prefecture-level jurisdictions, RJ-MAA enables empirical verification of whether adjudication patterns differ across regions under China’s unified legal framework. By inputting identical cases into court agents calibrated to distinct jurisdictions, we observe significant variations in predicted outcomes.

Date of Award	18 Jul 2026
Original language	English
Awarding Institution	University of Nottingham
Supervisor	Anthony Graham Bellotti (Supervisor), Xiuping Hua (Supervisor) & Hang Zhou (Supervisor)