PaTRIZ: A framework for mining TRIZ contradictions in patents

Abstract

Patents are a significant source of information about inventions. However, understanding the content of a patent with the aim of using it for an automatic solution search is still an unsolved challenge. To achieve this purpose, a model based on the TRIZ theory (Altshuller, 1984) has been developed. This theory introduces the notion of contradiction, which is a reliable and domain-independent technique to formulate the problem solved by each patent through an opposition between parameters of a system. Each patent is considered a solution concept to a contradiction. Mining contradictions, therefore, means characterizing solution concepts.In this paper, we propose a new approach called PaTRIZ, a complete framework for patent analysis based on a combination of sentences and word-level deep neural networks. The word-level network, called ParaBERT, comprises a novel Conditional Random Field structure, developed to integrate syntactic information. The idea is to mine the patent’s motivating problem (aka contradiction), which is fundamental to understanding the invention and identifying for which purpose it could be used. The models are evaluated on built-in real-world datasets.

BibTeX

@article{guarino:hal-03752549,
  TITLE = {{PaTRIZ: A framework for mining TRIZ contradictions in patents}},
  AUTHOR = {Guarino, Guillaume and Samet, Ahmed and Cavallucci, Denis},
  URL = {https://hal.science/hal-03752549},
  JOURNAL = {{Expert Systems with Applications}},
  PUBLISHER = {{Elsevier}},
  VOLUME = {207},
  PAGES = {117942},
  YEAR = {2022},
  DOI = {10.1016/j.eswa.2022.117942},
  KEYWORDS = {Patent ; Deep learning ; NLP ; Contradiction ; TRIZ},
  HAL_ID = {hal-03752549},
  HAL_VERSION = {v1},
}