A comparative study of abstractive and extractive summarization techniques to label subgroups on patent dataset

Patents are an important source of information for measuring the technological advancement of a specific knowledge domain. To facilitate the search for information in patent datasets, classification systems separate documents into groups according to the area of knowledge, and designate names to define their content. The increase in the number of patented inventions leads to the need to subdivide these groups. Since these groups belong to a restricted knowledge domain, naming the generated subcategories can be extremely laborious. This work aims to compare the performance of abstractive and extractive summarization techniques in the task of generating sentences directly associated with the content of patents. The abstractive summarization model was composed by a Seq2Seq architecture and a LSTM network. The training was conducted with a dataset of patent titles and abstracts. The validation process was performed using the ROUGE set of metrics. The results obtained by the generated model were compared with the sentence resulting from an extractive summarization algorithm applied to the task of naming patent groups. The main idea was to help the specialist to name new patent groups created by the clustering systems. The naming experiments were performed on the dataset of abstracts of patent documents. Comparative experiments were conducted using four subgroups of the United States Patent and Trademark Office, which uses the Cooperative Patent Classification system.