Constructing a Public Meeting Corpus

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 1934–1940 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 1934 Constructing a Public Meeting Corpus Koji Tanaka†, Chenhui Chu‡, Haolin Ren∗, Benjamin Renoust‡, Yuta Nakashima‡, Noriko Takemura‡, Hajime Nagahara‡, Takao Fujikawa∗∗ † Graduate School of Information Science, Osaka University, ‡ Institute for Datability Science, Osaka University, ∗ National Institute of Informatics, ∗∗ Graduate School of Letters, Osaka University tanaka.koji@ist.osaka-u.ac.jp, {chu, renoust, n-yuta, takemura, nagahara}@ids.osaka-u.ac.jp, haolinren@gmail.com fuji@let.osaka-u.ac.jp Abstract In this paper, we propose a method for constructing a large corpus about a century of public meetings in historical Australian newspapers, and analyze the constructed corpus. The corpus construction method is based on image processing and Optical Character Recognition (OCR). We digitize and transcribe texts of the specific topic of public meeting. Experiments show that our proposed method achieves a F-score of 71.5% with a high recall of 97.5% for corpus construction. This allows us to feed a content search tool for temporal and semantic content analysis. Keywords: corpus construction, public meeting, content delivery 1. Introduction Large-scale text corpora are essential for natural language processing (NLP). Most existing corpora are created from text that has already been digitized. For instance, the bench- mark syntactic parsing dataset Penn Treebank (Marcus et al., 1993) is created by labelling part-of-speech tags and syntactic information on the digitized text from the Wall Street Journal newspapers. The parallel corpus Europarl (Koehn, 2005) that has been used for the machine trans- lation shared task workshop, Workshop on Statistical Ma- chine Translation (WMT), is created by aligning parallel sentences from the digitized multilingual European Parlia- ment data. In various fields including literature and humanities, many materials to be studied are not digitized, which are stored in a physical medium such as paper or just scanned but not transcribed into text. By digitizing and transcribing such materials into text, and structuring them via extracting spe- cific topics, we can apply many NLP techniques to analyz- ing them automatically. In research fields such as literature, digitization, text transcript and structure can significantly increase the value of the original materials. In this paper, we work on the historical newspaper database Trove (Cassidy, 2016)1 (Trove covers major Australian daily newspapers and local newspapers). We propose a cor- pus construction method based on image processing and Optical Character Recognition (OCR), which achieves a F-score of 71.5% with a high recall of 97.5%. We first identify the rule lines in newspaper images and trim the images into newspaper articles. Next, we apply OCR to the trimmed articles, and extract the articles with specific topic words. Evaluation conducted on manually annotated golden data indicates that the proposed method can extract 14.9% more articles without excess and deficiency, com- pared to a baseline that is based on linguistic features ex- 1https://trove.nla.gov.au tracted from beginning and ending sentences of articles. We extract articles about the specific topic of public meet- ing (Fujikawa, 1990),2 which covers 120 years spanning from 19th …

Content truncated to 500 words