Unsupervised Training Set Generation for Automatic Acquisition of Technical Terminology in Patents

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 290–300, Dublin, Ireland, August 23-29 2014. Unsupervised Training Set Generation for Automatic Acquisition of Technical Terminology in Patents Alex Judea1 Hinrich Sch¨utze2 S¨oren Br¨ugmann3 1Heidelberg Institute for Theoretical Studies, Heidelberg, Germany 2Center for Information and Language Processing, University of Munich, Germany 3Br¨ugmann Software GmbH, Papenburg, Germany Abstract NLP methods for automatic information access to rich technological knowledge sources like patents are of great value. One important resource for accessing this knowledge is the tech- nical terminology of the patent domain. In this paper, we address the problem of automatic terminology acquisition (ATA), i.e., the problem of automatically identifying all technical terms in a document. We analyze technical terminology in patents and define the concept of technical term based on the analysis. We present a novel method for labeling large amounts of high-quality training data for ATA in an unsupervised fashion. We train two ATA methods on this training data, a term candidate classifier and a conditional random field (CRF), and investigate the utility of different types of features. Finally, we show that our method of automatically generating train- ing data is effective and the two ATA methods successfully generalize, considerably increasing recall while preserving high precision relative to a state-of-the-art baseline. 1 Introduction A large part of our technological knowledge is encoded in patents. Methods for automatically finding information in patents and inferring information from patents are thus of great value. An important step in getting access to patent information is identification of technical terminology, i.e., finding the linguistic expressions that denote the technical concepts of a patent: the methods, processes, substances and objects that are part of the invention or modified by it. In the example “The present invention relates to a charging apparatus of a bicycle dynamo”, the bolded compound nouns are the main content words and refer to specific technological concepts. We call such linguistic expressions (technical) terms or terms and their totality the (technical) terminology of a document or domain. We address the task of automatic terminology acquisition (ATA), the task of finding technical terms in texts without reliance on existing resources that list terms of the domain. In contrast to this stands automatic terminology recognition (ATR), which we define as finding known terms and their variants (Jacquemin and Bourigault, 2003). ATA provides input to downstream components like automatic sum- marization, machine translation, ontology building, information extraction and retrieval. terms ex- tracted by ATA can be semantically classified or mapped to entries in a semantic database (Krauthammer and Nenadic, 2004), but we focus on identifying them without further classification in this paper. Our main contributions are as follows. (i) We present a method for automatically labeling large amounts of training data for ATA. (ii) We show that two types of statistical classifiers trained on this training data beat a state-of-the-art baseline, indicating that the automatic labeling is of high quality. (iii) We study different feature types for ATA and investigate how much …

Content truncated to 500 words