Send message to Author

Paper short abstract

This talk outlines the design of BCCWJ2, a 200-million-word expansion of BCCWJ. We describe NDC-stratified annual book sampling (5M words/year; ~1,000 books), streamlined metadata/annotation and copyright-aware policies, and demo the release data in Chunagon.

Paper long abstract

This presentation reports on the design of BCCWJ2, the ongoing expansion of the Balanced Corpus of Contemporary Written Japanese (BCCWJ) from approximately 100 million to 200 million words. The core addition is a large-scale collection of book data published between 2006 and 2025, sampled at a fixed rate of 5 million words per year. To preserve the principle of balance while ensuring transparency and replicability, sampling is stratified by the Nippon Decimal Classification (NDC), resulting in approximately 1,000 sampled books per year.

We focus on design choices that balance continuity with BCCWJ1 and practical feasibility under current conditions. First, we explain the revised sampling workflow: BCCWJ2 defines the population and sampling unit at the book level, simplifying procedures compared to earlier page- or character-based randomization while keeping NDC-based stratification. Because complete bibliographic information for the entire 2006–2025 period was not available at the outset of the project, the sampling frame is operationalized year by year, allowing steady annual growth and timely releases.

Second, we outline metadata and annotation policies. To improve construction efficiency while maintaining usability, we prioritize the metadata items exposed in the Chunagon interface and streamline less-used structural tag sets. The talk also introduces new metadata for safer secondary use (e.g., adult-content filtering in an rated R scheme) and explains how these decisions support research and educational contexts.

Third, we address copyright-aware design. Building on Japan’s post-2018 legal environment for text and data mining (TDM), BCCWJ2 adopts operational constraints such as excluding short poetic works and carefully controlling the length of contexts displayed in web services.

The presentation concludes with a live demonstration of the currently constructed subset (focusing on the 2006–2010 portion) in Chunagon, showing continuity with BCCWJ1 workflows and illustrating how users in and outside Japan can access and make use of the expanded corpus.