Job Market Seminar

Abstract: Patents can account for a substantial share of firm value, making the prediction of patent validity litigation outcomes economically important. Predicting these outcomes requires interpreting interdependent texts and aggregating information across multiple validity challenges. Using a newly constructed dataset of UK patent litigation, I develop a multiple-instance learning architecture tailored to the structure of these challenges. The architecture uses cross-attention, constructing queries from representations of the challenged patent claim and keys and values from representations of the texts cited against it. To account for textual dependencies, I construct expanded text representations that explicitly incorporate relevant content from the texts on which they depend. Together, these components address how individual texts are interpreted and how information across challenges is combined into an overall validity prediction. Empirical results show that the proposed architecture outperforms benchmark architectures from the literature as well as ChatGPT in this setting.