TL;DR
This paper introduces UD-Japanese-BCCWJ, the largest UD-annotated Japanese corpus to date, created through an automatic conversion of the existing Balanced Corpus of Contemporary Written Japanese (BCCWJ). This work addresses the need for a large-scale, consistently annotated Japanese dataset to advance cross-linguistic NLP research, particularly within the Universal Dependencies framework.
What Problem Does This Paper Address?
The field of Natural Language Processing (NLP) is increasingly focused on multilingual and cross-linguistic research. The Universal Dependencies (UD) project aims to standardize linguistic annotation across languages by defining common part-of-speech (PoS) tags and dependency relations. However, creating high-quality UD-annotated corpora for all languages is a significant undertaking. Japanese, in particular, presents challenges due to its unique linguistic structure, such as the absence of explicit word boundaries (spaces) and the use of "bunsetsu" (syntactic phrases) as a primary unit for dependency annotation in existing Japanese corpora. While previous UD Japanese resources exist, they are often smaller or require manual correction. This paper tackles the challenge of creating a large-scale, automatically annotated UD Japanese corpus by converting the comprehensive BCCWJ, which already contains bunsetsu-level dependency information.
Methods
The core of this paper is the automatic conversion of the BCCWJ corpus to the UD annotation schema. This process involved several key steps:
- Word Unit Detection: Japanese text doesn't use spaces, so identifying "syntactic words" is crucial. The authors chose to use "Short Unit Words" (SUWs) as the fundamental linguistic unit for UD-Japanese-BCCWJ, aligning with BCCWJ's basic PoS system. SUWs are the smallest atomic units, which can be combined into "Long Unit Words" (LUWs) and then "bunsetsu" (phrases).
- PoS Tag Conversion: The paper describes mapping BCCWJ's UniDic PoS tags to the Universal Dependencies PoS tagset. While the SUW PoS system was generally used, for certain ambiguous cases (like nominal verbs or adjectives), they leveraged LUW PoS tags based on usage principles to achieve better consistency with UD standards and other languages.
- Dependency Structure Conversion: The BCCWJ corpus provides bunsetsu-level dependency structures. The conversion to UD's word-level dependency involved identifying the head word within each bunsetsu and attaching all other words in the bunsetsu to that head word. A key difference noted is that UD dependency arrows point from head to dependent, while BCCWJ's point from dependent to head.
- UD Relation Label Assignment: Since BCCWJ's dependency roles are less detailed than UD's, the authors derived UD relation labels by combining information from bunsetsu structure, case-marking particles, and predicate-argument structure information from BCCWJ-PAS.
Datasets
The primary dataset used is the Balanced Corpus of Contemporary Written Japanese (BCCWJ). This is a large-scale corpus (104.3 million words) covering diverse genres like books, magazines, newspapers, blogs, and legal statutes. The UD-Japanese-BCCWJ corpus is derived from the "core data" of BCCWJ, which consists of 1,980 documents, 57,109 sentences, and approximately 1.27 million words across six distinct domains.
Key Insights and Ideas
- Automatic Conversion is Feasible for Large-Scale Corpora: The paper demonstrates that a fully automatic conversion process can yield a substantial UD-annotated corpus, significantly reducing the manual effort typically required.
- SUW vs. LUW for Japanese UD: While UD guidelines favor "syntactic words," the authors chose SUWs for their UD-Japanese-BCCWJ. They acknowledge that LUWs might be more aligned with UD's "syntactic words" concept but cite practical difficulties in producing LUW segmentations and issues with certain multi-word expressions as reasons for their choice. This highlights an ongoing debate in UD Japanese annotation.
- Bridging Bunsetsu and Word-Level Dependencies: The conversion method effectively bridges the gap between Japanese bunsetsu-level dependencies and UD's word-level structure, a crucial step for cross-linguistic comparability.
- Leveraging Existing Semantic Information: The successful mapping of predicate-argument structures from BCCWJ-PAS to UD relations shows the potential of integrating semantic information for richer syntactic annotation.
Key Results
- UD-Japanese-BCCWJ: The resulting corpus is the largest UD Japanese corpus and the second-largest UD corpus overall, containing 1,980 documents, 57,109 sentences, and 1,273k words across six domains.
- Parsing Performance: The paper presents parsing results (UAS and LAS) by genre, showing that for genres with over 200K tokens, using in-domain data for training yielded better results than using the entire corpus. This suggests genre-specific models can be beneficial.
Limitations and Future Work
- Clause Handling: The current rules are unable to effectively identify clauses and handle clause-related UD labels (csubj, advcl, acl). This is attributed to the vagueness of clauses in Japanese compared to English, and the authors plan to address this by establishing criteria for identifying clauses in the future.
- Coordinate Structures: While BCCWJ-DepPara contains information on coordinate structures, the current conversion rules do not yet define specific rules for UD labels like
ccandconj. This is an area for future work. - LUW-based Corpus: The authors plan to implement and release a UD Japanese corpus based on LUWs, acknowledging their potential alignment with UD's "syntactic words" definition.
- Ongoing Discussion on Word Units: The paper notes that the choice of apposite word units for Japanese UD is still a subject of ongoing discussion.
Summary updated