TL;DR
This paper introduces a novel rule-based system, GREW, for transforming standard Universal Dependencies (UD) into Enhanced Universal Dependencies (EUD). By leveraging graph rewriting, the system offers a transparent and adaptable approach to enriching linguistic annotations, achieving competitive results in the IWPT 2021 Shared Task.
What Problem Does This Paper Address?
Universal Dependencies (UD) provides a standardized framework for annotating syntactic dependencies across many languages. However, Enhanced Universal Dependencies (EUD) aims to capture richer linguistic phenomena, such as ellipsis, coordination, and argument structure, which are not fully represented in standard UD. The challenge lies in systematically and accurately converting UD annotations to EUD, especially given the complexity and language-specific nature of these enhancements. Existing approaches often rely on machine learning or complex rule sets, and there's a need for a flexible, rule-based method that can be adapted to different languages and annotation conventions.
Methods
The core of this work is the application of GREW (Graph Rewriting), a generic graph rewriting tool, to perform the transformation from UD to EUD. The system works by defining a set of rules that describe patterns in the UD graph and specify how to modify it to produce EUD.
Key aspects of their method include:
- Rule-based Transformation: Instead of end-to-end machine learning, they use a rule-based system to convert UD to EUD. This offers transparency and allows for direct control over the enhancement process.
- GREW Framework: GREW allows for defining rules with patterns and commands, and a strategy mechanism controls the order and application of these rules. This creates a "Graph Rewriting System" (GRS).
- Six Types of Enhancement: The system is designed to handle six specific types of EUD enhancements, each addressed by a dedicated subset of GREW rules:
- Empty nodes for elided predicates (ellipsis).
- Propagation of incoming dependencies to conjuncts in coordinations.
- Propagation of outgoing dependencies from conjuncts.
- Additional subject relations for raising and control verbs.
- Coreference in relative clause constructions.
- Modifier labels incorporating prepositional or case-marking information.
- Language Adaptation: The GRS is designed to be modular, allowing for language-specific rule packages to be added and integrated into the generic strategy. This enables adaptation to language-specific features and annotation choices.
Datasets
The system was evaluated on data from the IWPT 2021 Shared Task on Parsing into Enhanced Universal Dependencies (EUD). The UD annotations were generated using UDPIPE2, trained on various corpora for different languages (Arabic-PADT, Bulgarian-BTB, Czech-PDT, English-EWT, Estonian-EDT, Finnish-TDT, French-GSD, Italian-ISDT, Lithuanian-ALKSNIS, Dutch-Alpino, Polish-PDB, Slovak-SNK, Swedish-Talbanken, Tamil-TTB, Ukrainian-IU2). The evaluation was performed against gold EUD annotations.
Key Insights and Ideas
- Rule-based transparency: Using graph rewriting provides a clear, interpretable way to understand how UD is transformed into EUD, which is valuable for linguistic analysis and debugging.
- Modularity for language adaptation: The GRS architecture, with its rule packages, is highly effective for adapting to the diverse linguistic phenomena and annotation conventions across different languages.
- Interaction between enhancements: The paper highlights that the order of applying different enhancement rules matters. For instance, ellipsis processing must precede coordination propagation in certain cases to avoid errors.
- Guidelines vs. Gold Data: A significant finding is that gold annotations in the task sometimes deviate from the official EUD guidelines, presenting challenges for rule-based systems that strictly adhere to guidelines. This suggests that learning-based approaches might be more robust to such inconsistencies.
Key Results
The system achieved an ELAS score of 81.58 on the IWPT 2021 Shared Task, starting from UDPIPE annotations which had an LAS score of 85.77. When evaluated on gold UD annotations as input, the system's performance was competitive with previous rule-based systems, often falling between the scores reported by Dehouck et al. (2020) and Heinecke (2020). Manual inspection of discrepancies in French and English revealed that a significant portion of errors stemmed from inconsistencies between the gold annotations and the EUD guidelines, or from the inherent complexity of certain linguistic phenomena.
Limitations and Future Work
- Dependence on UD quality: The system's performance is directly tied to the quality of the input UD annotations from UDPIPE. Errors in the initial UD parsing can propagate to the EUD output.
- Language-specific lexicons: For certain enhancements, like identifying subjects of raised or controlled verbs, language-specific lexicons are needed, which are not always available.
- Deviations from guidelines: The system's adherence to guidelines can be a limitation when gold data deviates from them. The paper suggests that learning-based approaches might be better suited to adapt to such data.
- Future Work: The authors believe their system can be used to identify annotation errors in existing EUD data and serve as a starting point for creating EUD annotations for languages lacking such data, especially where learning-based methods are not feasible. They also acknowledge the need for more precise guidelines, particularly concerning multiple case/mark dependencies on the same head.
Summary updated