TL;DR
This paper presents a hybrid system for predicting "enhanced" Universal Dependencies (UD) that go beyond basic syntactic relationships. By combining a graph-based parser with a rule-based system and a learned classifier, the approach achieves high accuracy and language independence, ranking second in the EUD Parsing Shared Task.
What Problem Does This Paper Address?
Traditional dependency parsing focuses on surface syntactic structure, representing sentences as trees. However, many linguistic phenomena, like coordinations, control structures, and elided words, require richer annotations to be fully captured. Enhanced Universal Dependencies (EUD) aim to address this by adding implicit syntactic links and enriching existing ones with more detail (e.g., prepositions, cases). The challenge lies in creating these EUD annotations, which are complex, time-consuming to create manually, and can vary significantly across languages, making it difficult to build robust, language-independent parsing systems.
Methods
The authors propose a hybrid, two-step approach:
- Graph-Based Parsing: A standard graph-based parser is used to generate a basic syntactic dependency tree. This step leverages contextual word embeddings (like BERT and XLM-R) for improved performance.
- Rule-Based Enhanced Dependency Prediction: A set of linguistic rules, derived from manual analysis of EUD annotations, is applied to the basic dependency tree. These rules aim to predict enhanced dependencies.
- Language Independence: The rules primarily rely on UPOS tags and the structure of the dependency tree, which are shared across languages, promoting language independence.
- Classifier for Rule Application: To optimize rule application and handle context-specific nuances, a classifier (ICSIBoost) is trained. This classifier predicts the suitability of a rule in a given context, using features like rule name, language, dependency label, and distances between words.
- Pipeline Approach: The system operates as a pipeline, where errors in earlier stages (tokenization, parsing) can propagate to later stages.
Datasets
The system was trained and evaluated on Universal Dependencies (UD) treebanks. Specifically, the authors used data from the IWPT 2020 Shared Task on Enhanced Dependency Parsing, which is based on UD v2.5. They also refer to results on the CoNLL 2018 Shared Task data (UD v2.2) for comparison.
Key Insights and Ideas
- Hybrid Approach Effectiveness: Combining rule-based systems with machine learning (via the classifier) proves effective for EUD, balancing the precision of rules with the adaptability of learned models.
- Language Independence is Achievable: By focusing on universal features like UPOS tags and dependency structures, the rule-based component can be largely language-independent, reducing the need for extensive language-specific training data for the rule generation itself.
- Classifier Improves Robustness: The ICSIBoost classifier helps to filter rule applications, leading to better performance by avoiding incorrect rule applications in specific contexts.
- Elided Word Prediction is Challenging: The authors acknowledge that predicting enhanced dependencies for elided words is particularly complex and was abandoned for the shared task.
Key Results
- The system ranked 2nd in the Shared Task on Enhanced Dependency Parsing, achieving an average ELAS of 82.6% and EULAS of 84.6%.
- Evaluations on gold UPOS and basic dependencies showed very high precision for enhanced dependency extraction.
- The use of contextual embeddings (XLM-R, BERT) significantly improved basic parsing performance.
- The ICSIBoost classifier provided a small but noticeable gain in ELAS for most languages, with a more significant increase for Estonian.
Limitations and Future Work
- Elided Word Prediction: The authors explicitly state they abandoned predicting enhanced dependencies for elided nodes due to difficulties. This remains an area for future work.
- Pipeline Errors: The pipeline architecture means errors in earlier stages (tokenization, parsing) can negatively impact the prediction of enhanced dependencies.
- Data Heterogeneity: The training data was described as heterogeneous and partially incomplete, which is a general challenge in EUD annotation.
- Future Work: The authors suggest that a possible extension would be to process elided nodes. They also propose that their language-independent system can be used to pre-annotate existing UD treebanks, which could then serve as a basis for purely ML-based approaches to EUD.
Summary updated