Hitachi at MRP 2019: Unified Encoder-to-Biaffine Network for Cross-Framework Meaning Representation Parsing

Yuta Koreeda, Gaku Morio, Terufumi Morishita, Hiroaki Ozaki, Kohsuke Yanai

Reference: Yuta Koreeda, Gaku Morio, Terufumi Morishita, Hiroaki Ozaki, Kohsuke Yanai. Proceedings of the Shared Task on Cross-Framework Meaning Representation Parsing at the 2019 Conference on Natural Language Learning.

2019.

Abstract

Hitachi researchers developed a unified neural architecture designed to parse five distinct semantic graph frameworks simultaneously. By combining a shared encoder with specialized biaffine and pointer networks, they demonstrated that a single model can handle diverse linguistic abstractions and benefit from multi-task learning across heterogeneous meaning representations.

Introduction

Meaning Representation Parsing (MRP) is the task of transforming natural language sentences into formal semantic graphs. The challenge lies in the diversity of "frameworks"—different linguistic theories (such as AMR, UCCA, or PSD) that represent meaning with varying levels of abstraction and different graph structures. Historically, researchers have built specialized, "bespoke" parsers for each framework, which limits the ability to share architectural improvements or leverage commonalities between different semantic theories. The Hitachi team addressed this by creating a unified system capable of handling five different frameworks at once, bridging the gap between frameworks that map directly to tokens and those that introduce abstract or anchorless nodes.

Methodology

The core of the system is a Unified Encoder-to-Biaffine Network. The architecture is modular, allowing it to adapt to the specific requirements of each framework while sharing a common backbone.

Shared Encoder

The model starts with a rich feature extraction layer combining surface forms, lemmas, POS tags, and named entity labels with pre-trained embeddings (GloVe, ELMo, and BERT). These are processed by multi-layered bidirectional LSTMs (biLSTMs) to create contextualized token representations.

Biaffine Networks

Regardless of the framework, the system uses biaffine classifiers to predict the probability of edges between nodes and to assign labels to those edges. Biaffine operations are excellent at filtering node candidates and predicting labels across all five graph types.

Framework-Specific Node Generation

Since some frameworks have nodes that do not correspond 1-to-1 with input tokens, specific strategies were implemented:

  • UCCA: Uses a Pointer Network to identify spans of text and generate non-terminal nodes.
  • AMR: Employs an Extended Pointer-Generator Network to handle nodes that are either copied from the source, copied from previously generated nodes, or generated from a fixed vocabulary.
  • EDS: Uses a heuristic conversion from DM graphs followed by a neural anchor predictor for abstract nodes.
Overview of the proposed unified encoder-to-biaffine network architecture

Evaluation

The system was evaluated on the CoNLL 2019 MRP shared task dataset, which includes five frameworks: DM, PSD, EDS, UCCA, and AMR. The team also utilized external linguistic resources for post-processing, such as the ERG SEM-I for DM and CzEngVallex for PSD.

Key Results

The Hitachi system was highly competitive, ranking 5th overall in the formal evaluation with a macro-averaged MRP F1 score of 0.7604.

MRP F1 scores for the formal run across different frameworks and participating teams

The system performed particularly well on PSD, where it ranked 2nd among all participants. It significantly outperformed the TUPA baseline (a unified transition-based parser) across almost all frameworks. Post-evaluation experiments showed that incorporating multi-task learning and fine-tuning with BERT further boosted performance, reaching an average F1 of 0.7507 even without the ensembling used in the formal run.

Discussion and Insights

One of the most important findings was that training on all frameworks simultaneously (multi-task learning) improved performance. This suggests that the shared encoder learns a more robust representation of language by trying to satisfy multiple semantic theories at once.

However, a significant performance gap was noted in AMR compared to top-ranked systems. This was attributed to missing specific features like "Wikification" (linking entities to Wikipedia) and potential implementation nuances in the pointer-generator network that were present in state-of-the-art specialized AMR parsers.

Conclusion

The Hitachi unified parser demonstrates that a single neural architecture can effectively handle diverse semantic frameworks through a shared encoder and standardized edge prediction. While node generation still relies on some framework-specific logic, the success of the biaffine network suggests a path toward more universal parsing. Future work will focus on designing a truly unified node generation method to further streamline cross-framework meaning representation parsing.

Summary updated

BibTeX
@inproceedings{koreeda2019hitachi,
  author = {Yuta Koreeda and Gaku Morio and Terufumi Morishita and Hiroaki Ozaki and Kohsuke Yanai},
  title = {{Hitachi at MRP 2019: Unified Encoder-to-Biaffine Network for Cross-Framework Meaning Representation Parsing}},
  booktitle = {CoNLL Shared Task},
  year = {2019},
  url = {https://openreview.net/forum?id=GbD2dQhucl},
}
Figure