Curriculum
Computer Science and Innovation for Societal Challenges, XXXV series
Grant sponsor
Fondazione Bruno Kessler
Supervisor

Alberto Lavelli
Co-supervisor

Sabrina Cipolletta

Project: Natural Language Processing for Technology Foresight Summarization and Simplification: the case of patents
Full text of the dissertation book can be downloaded from: https://www.research.unipd.it/handle/11577/3498347.

Abstract: Technology foresight aims to anticipate possible developments, understand trends, and identify technologies of high impact. To this end, monitoring emerging technologies is crucial. Patents – the legal documents that protect novel inventions – can be a valuable source for technology monitoring. Millions of patent applications are filed yearly, with 3.4 million applications in 2021 only. Patent documents are primarily textual documents and disclose innovative and potentially valuable inventions. However, their processing is currently underresearched. This is due to several reasons, including the high document complexity: patents are very lengthy and are written in an extremely hard-to-read language, which is a mix of technical and legal jargon. This thesis explores how Natural Language Processing – the discipline that enables machines to process human language automatically – can aid patent processing. Specifically, we focus on two tasks: patent summarization (i.e., we try to reduce the document length while preserving its core content) and patent simplification (i.e., we try to reduce the document’s linguistic complexity while preserving its original core meaning). We found that older patent summarization approaches were not compared on shared benchmarks (making thus it hard to draw conclusions), and even the most recent abstractive dataset presents important issues that might make comparisons meaningless. We try to fill both gaps: we first document the issues related to the BigPatent dataset and then benchmark extractive, abstraction, and hybrid approaches in the patent domain. We also explore transferring summarization methods from the scientific paper domain with limited success. For the automatic text simplification task, we noticed a lack of simplified text and parallel corpora. We fill this gap by defining a method to generate a silver standard for patent simplification automatically. Lay human judges evaluated the simplified sentences in the corpus as grammatical, adequate, and simpler, and we show that they can be used to train a state-of-the-art simplification model. This thesis describes the first steps toward Natural Language Processing-aided patent summarization and simplification. We hope it will encourage more research on the topic, opening doors for a productive dialog between NLP researchers and domain experts.