TL;DR
This paper introduces COMBO, a novel hybrid approach for Enhanced Universal Dependencies (EUD) parsing that combines machine learning with rule-based methods. COMBO achieved a respectable 4th place in the IWPT 2021 EUD shared task, demonstrating its effectiveness in producing rich, semantically-informed dependency structures for multiple languages.
What Problem Does This Paper Address?
Modern natural language processing relies heavily on understanding the grammatical structure of sentences. While Universal Dependencies (UD) provide a standardized way to represent this structure, Enhanced Universal Dependencies (EUD) go further by incorporating richer semantic information, such as argument structures and coreference relations. However, building EUD parsers is challenging due to the complexity of these enhanced structures and the limited availability of robust, publicly accessible EUD parsing systems. This lack of readily available tools hinders their adoption in downstream NLP tasks like question answering or relation extraction. This paper aims to address this by presenting COMBO, a system designed to produce EUD graphs that is not only effective but also publicly available.
Methods
COMBO employs a hybrid approach, integrating machine learning with rule-based post-processing. The core of the system involves:
- Data Encoder: This component uses a combination of character-based embeddings and pre-trained BERT language models (either language-specific or multilingual XLM-R) to represent input tokens. A scalar mix technique is used to combine embeddings from different BERT layers.
- Morphosyntactic Predictors: These modules predict standard linguistic features like part-of-speech tags, morphological features, and lemmatization. These are crucial for later steps, particularly for enriching EUD labels.
- EUD Predictor: This is the heart of the EUD parsing. It uses two classifiers: an arc classifier to predict the existence of enhanced dependency arcs and a label classifier to assign the correct EUD relation labels.
- Merge Algorithm: A key innovation is the merge algorithm that combines the predicted UD dependency trees with the predicted EUD graphs. This is done by iteratively adding edges from both structures, carefully avoiding cycles to ensure a valid graph.
- Post-processing Rules: Two language-independent rules are applied to refine the EUD graphs:
- Rule 1: Extends EUD edge labels with case information for modifiers and conjuncts, derived from morphological features and syntactic context.
- Rule 2: Corrects erroneous enhanced edges that might be assigned to function words, ensuring structural consistency.
Datasets
The research utilizes data from the IWPT 2021 EUD shared task. This dataset comprises treebanks for 17 languages across four language families, including Indo-European (Bulgarian, Czech, Polish, Russian, Slovak, Ukrainian, Dutch, English, Swedish, French, Italian, Latvian, Lithuanian), Uralic (Estonian, Finnish), Afro-Asiatic (Arabic), and Southern Dravidian (Tamil). The systems were trained and evaluated on this data.
Key Insights and Ideas
- Hybrid Approach is Key: The paper demonstrates that combining data-driven prediction of UD trees and EUD graphs with rule-based merging and refinement significantly improves the quality of the final EUD output.
- Leveraging UD for EUD: By first predicting standard UD trees and then merging them with EUD-specific edges, the system benefits from the robustness of UD parsing while incorporating richer semantic information.
- Language-Independent Rules for Refinement: The development of language-independent rules for adding case information and correcting function word attachments is a significant step towards creating more generalizable EUD parsers.
- Public Availability Matters: The authors emphasize the importance of releasing their code and pre-trained models to facilitate the adoption of EUD parsing in practical NLP applications.
Key Results
COMBO achieved a 4th place ranking in the IWPT 2021 EUD shared task, with an average ELAS (Enhanced Labeled Attachment Score) of 83.79% across languages. In terms of EULAS (Enhanced UD Labeled Attachment Score, ignoring sublabels), it ranked 5th with 85.20%. The system also performed well on predicting standard UD trees, achieving a LAS of 88.91%, placing it second overall. The post-processing step, particularly Rule 1 for case information, showed a notable impact on ELAS scores.
Limitations and Future Work
The authors acknowledge that while their rules are language-independent, some extensions (like conjunct extensions) are only applied to specific languages due to treebank limitations. They also note that the current approach of appending empty nodes at the end of a sentence is a makeshift solution and that identifying proper positions for elided elements remains an open research problem. Future work could involve defining more language-specific correction rules to further improve performance and exploring more sophisticated methods for handling elided elements.
Summary updated