Javascript must be enabled for the correct page display

Text-based Patent-Quality Prediction Using Multi-Section Attention

Krant, Xabi (2023) Text-based Patent-Quality Prediction Using Multi-Section Attention. Master's Thesis / Essay, Artificial Intelligence.


Download (3MB) | Preview
[img] Text
Restricted to Registered users only

Download (129kB)


The number of patents has increased tremendously in recent years and statistics derived from patents have become the standard measurement for innovation. Patents statistics are widely avail- able and correlate well with patent valuation and quality. They do however suffer from inherent biases and flaws, for instance caused by the characteristics of patent offices and their employees. The major drawback of using patent statistics is substantial time needs to have passed before they can be used and there are no ex-ante indicators available. This thesis aims to solve this problem, using current advances in machine learning and language models. A new text model is introduced that can predict the innovation and market value of patents based solely on the patent text. Current state of the art machine learning text models like BERT are however not a perfect fit for patents, as patents can contain very long text and they have a multi-section structure. This thesis proposes a new model, called MSABERT, that is able to handle longer texts and the multi-section structure. Each section is handled separately, after which they are combined using attentitive pooling. This attentitive pooling also adds a layer of explainability to the model, showing the relative importance of each section. This new model is compared to models that lack some of these capabilities in several experi- ments. The results show that this newly introduced model achieves similar performance as existing models when th

Item Type: Thesis (Master's Thesis / Essay)
Supervisor name: Schomaker, L.R.B. and Steinberg, P.J.
Degree programme: Artificial Intelligence
Thesis type: Master's Thesis / Essay
Language: English
Date Deposited: 08 May 2023 09:35
Last Modified: 08 May 2023 09:35

Actions (login required)

View Item View Item