Proceedings of the 16th International Conference on Parsing Technologies and the IWPT 2020 Shared Task on Parsing into Enhanced Universal Dependencies

Gosse Bouma, Yuji Matsumoto, Stephan Oepen, Kenji Sagae, Djamé Seddah, Weiwei Sun, Anders Søgaard, Reut Tsarfaty, Dan ZemanACL 2020

TL;DR

This document presents the proceedings of IWPT 2020, focusing on the 16th International Conference on Parsing Technologies and a shared task on parsing into Enhanced Universal Dependencies (EUD). The conference highlights advancements in syntactic parsing, particularly the novel EUD task which pushes the boundaries of representing richer syntactic structures across multiple languages.

What Problem Does This Paper Address?

The field of natural language processing (NLP) has long sought to accurately represent the grammatical structure of sentences, a task known as parsing. Traditional dependency parsing, while powerful, often simplifies or omits certain linguistic phenomena. Enhanced Universal Dependencies (EUD) aims to address this by providing a richer, more expressive representation of syntactic structures, capturing nuances that standard dependency trees might miss. This conference and its associated shared task tackle the challenge of developing and evaluating parsers capable of generating these more complex EUD structures across a wide range of languages, pushing the state-of-the-art in syntactic analysis.

Methods

The papers presented cover a diverse range of methods for syntactic parsing, with a particular emphasis on neural network architectures and graph-based approaches. Techniques explored include:

  • Distilling Neural Networks: Methods for creating smaller, faster models from larger, more complex ones while retaining performance.
  • End-to-End Parsing: Architectures that directly map sentences to syntactic structures without intermediate steps.
  • Graph-Based and Transition-Based Parsers: Exploring hybrid approaches that combine the strengths of different parsing paradigms.
  • Variational Autoencoders (VAEs): Utilizing VAEs for semi-supervised parsing, enabling learning from limited labeled data.
  • Memory-bounded Architectures: Developing parsers that can handle long-range dependencies efficiently.
  • Privacy-preserving Parsing: Investigating methods to obfuscate syntactic information for privacy concerns.
  • Latent-Variable Models: Employing models with unobserved variables to capture complex linguistic phenomena.
  • Transformer Encoders: Adapting multilingual transformer models for robust EUD parsing.
  • Second-Order Inference: Incorporating higher-order dependencies for more accurate parsing.

Datasets

The primary focus of the IWPT 2020 Shared Task was Enhanced Universal Dependencies (EUD) over 17 languages. This represents a significant expansion in scale for evaluating graph-based representations of syntactic structures. While specific training and evaluation datasets for each paper are not detailed in this overview, the shared task itself established a new benchmark for EUD parsing.

Key Insights and Ideas

  • The Rise of Enhanced Universal Dependencies (EUD): The conference strongly signals a shift towards richer syntactic representations beyond standard dependency trees, enabling more nuanced linguistic analysis.
  • Scalability of EUD: The shared task's evaluation across 17 languages demonstrates the growing feasibility and importance of EUD for diverse linguistic contexts.
  • Hybrid Parsing Approaches: Combining different parsing strategies (e.g., graph-based and transition-based) shows promise in overcoming individual limitations.
  • Efficiency in Neural Parsing: A significant theme is the pursuit of faster and more resource-efficient neural parsing models, crucial for real-world applications.
  • The Impact of Contextualized Embeddings: The "Deep Contextualized Era" highlights how pre-trained language models are revolutionizing parsing performance.
  • Privacy in Syntactic Analysis: The emergence of research on privacy-preserving parsing indicates a growing awareness of data sensitivity in NLP.

Key Results

The proceedings showcase a variety of results, including:

  • Improved Parsing Accuracy: Many papers report state-of-the-art or near state-of-the-art performance on various parsing tasks, particularly for EUD.
  • Faster Parsing Speeds: Several methods demonstrate significant speedups in parsing, making them more practical for large-scale applications.
  • Successful Shared Task Participation: Ten teams submitted systems to the EUD shared task, providing a valuable benchmark and comparison of current capabilities.
  • Analysis of EUD Coverage: Research explores how much of the richer EUD information is already captured by standard UD, providing insights into the added value of EUD.

Limitations and Future Work

  • Computational Cost: While progress is being made, many advanced neural parsing models still require substantial computational resources for training and inference.
  • Data Availability for EUD: The development of comprehensive EUD datasets for a wider range of languages remains an ongoing challenge.
  • Interpretability of Neural Models: Understanding why certain neural parsers perform well, especially in complex EUD scenarios, is an area for continued research.
  • Robustness Across Languages: Ensuring that parsing models generalize well and perform consistently across diverse linguistic typologies is a key future direction.
  • Integration with Downstream Tasks: Further work is needed to fully leverage the richer information provided by EUD in downstream NLP applications.

Summary updated

BibTeX
@inproceedings{bouma2020proceedings,
  author = {Gosse Bouma and Yuji Matsumoto and Stephan Oepen and Kenji Sagae and Djamé Seddah and Weiwei Sun and Anders Søgaard and Reut Tsarfaty and Dan Zeman},
  title = {{Proceedings of the 16th International Conference on Parsing Technologies and the IWPT 2020 Shared Task on Parsing into Enhanced Universal Dependencies}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2020},
}
Figure