TL;DR
This paper explores how to automatically extract noun-phrase (NP) chunks from Universal Dependencies (UD) trees, a task previously reliant on constituency parsing. They demonstrate that these automatically derived NP-chunks can improve performance on related NLP tasks like Part-of-Speech (PoS) tagging and dependency parsing, particularly when trained in a multi-task learning setting.
What Problem Does This Paper Address?
Syntactic chunking, the task of identifying phrases like noun-phrases (NPs) and verb-phrases (VPs), has historically been a valuable preprocessing step for improving syntactic parsing. However, traditional chunking methods often relied on constituency parse trees. With the rise of Universal Dependencies (UD) for cross-linguistic syntactic parsing, which provides dependency trees rather than constituency trees, a gap emerged: gold-standard chunk annotations are not readily available for UD. This paper addresses this by investigating how to automatically derive NP-chunks from UD dependency trees and evaluating their utility for other NLP tasks.
Methods
The core of the paper's methodology involves two main parts:
- Automatic NP-Chunk Deduction: The authors propose a rule-based system to extract NP-chunks directly from UD dependency trees. This system identifies core noun tokens and then recursively attaches their direct and indirect children based on specific dependency relation labels (e.g.,
compound,det,amod). They explicitly aim for "minimal" NP-chunks, meaning embedded prepositional phrases are not included within the NP. They also validate their automatic deduction method by manually annotating a portion of the English Web Treebank (EWT) and achieving a high accuracy of 98.7%. - Multi-Task Learning and Feature Integration:
- Sequence Labeling: For tasks like PoS-tagging, feature-tagging, and NP-chunking, they employ a deep recurrent neural network architecture based on bidirectional Long Short-Term Memory (bi-LSTM). They experiment with two multi-task learning strategies: "shared" (where tasks share parameters across the network) and "stacked" (where tasks are trained sequentially at different levels).
- Dependency Parsing: They adapt an existing arc-hybrid transition-based dependency parser. The key innovation here is incorporating the predicted NP-chunk tags (along with PoS tags) as features for the parser. These tags are converted into embeddings and concatenated with word and character embeddings to represent tokens.
Datasets
The experiments were conducted on three English treebanks from the Universal Dependencies project (version 2.1):
- EWT (English Web Treebank): The largest dataset (254k tokens).
- LinES: A smaller dataset (82k tokens). This treebank notably lacks feature annotations (morpho-syntactic tags), so feature-tagging was excluded from evaluations on this dataset.
- ParTUT: The smallest dataset (49k tokens).
Key Insights and Ideas
- UD as a Source for Chunking: The paper successfully demonstrates that UD dependency trees, despite not being constituency trees, contain sufficient information to automatically derive meaningful NP-chunks. This opens up chunking capabilities for UD-based pipelines.
- Synergy in Multi-Task Learning: Training PoS-tagging and NP-chunking together in a multi-task learning setting (especially with the "shared" strategy) consistently benefits at least one of the tasks, and often both. This highlights the inherent linguistic connections between these tasks.
- NP-Chunks as Valuable Features for Dependency Parsing: While the impact on dependency parsing was mixed across datasets, the results suggest that NP-chunk features can be particularly beneficial for smaller treebanks. This indicates that even imperfectly predicted chunks can provide useful structural information.
- Faster Training with Features: Incorporating features like PoS and NP-chunks into the dependency parser led to faster convergence during training, suggesting that these features help the model learn more efficiently.
Key Results
- Tagging Results (Table 1): Multi-task learning generally improved PoS-tagging and NP-chunking performance across the treebanks. The "shared" multi-task learning strategy often yielded the best results. For instance, on the ParTUT treebank, multi-task learning provided improvements of up to +0.6 for PoS-tagging, +0.7 for feature-tagging, and +1.73 for NP-chunking.
- Dependency Parsing Results (Table 2):
- Adding PoS and NP-chunk tags as features significantly improved dependency parsing performance on the smallest treebank, ParTUT (e.g., +0.96 LAS).
- For the LinES dataset, NP-chunks alone provided a small benefit (+0.22 LAS), but PoS tags were more impactful.
- The largest treebank, EWT, showed the baseline outperforming enriched models, but the "gold" NP-chunk experiment indicated that the potential for improvement exists if chunk prediction quality were higher.
- Training convergence was faster when using features (17.6 epochs on average) compared to the baseline (25.8 epochs).
Limitations and Future Work
- Mixed Impact on Dependency Parsing: The paper acknowledges that the benefit of NP-chunk features for dependency parsing was mixed across datasets, suggesting that the quality of the automatically deduced chunks (or PoS tags) might not always be sufficient to significantly improve parsing on larger, more complex datasets.
- Minimal NP-Chunks: The current system extracts "minimal" NP-chunks, excluding embedded PPs.
- Future Work: The authors plan to extend their research to other languages, investigate the possibility of extracting embedded chunks, and further explore the utility of NP-chunks as features for training on smaller datasets. They also note that further experiments are needed to fully understand the potential of NP-chunks as features.
Summary updated