Statistical Deep Parsing for Spanish Using Neural Networks

Luis Chiruzzo, Dina WonseverACL 2020

TL;DR

This paper introduces a novel deep parsing system for Spanish that leverages HPSG grammar and LSTM neural networks to capture both syntactic and semantic information in sentence analysis. The system achieves state-of-the-art results on Spanish parsing tasks, outperforming existing baselines in both syntactic constituency and dependency metrics, as well as semantic role labeling.

What Problem Does This Paper Address?

Syntactic analysis is a crucial step in Natural Language Processing (NLP) pipelines, often preceding deeper semantic understanding. While traditional approaches have focused on either constituency or dependency parsing, there's a growing interest in "deep parsing" that integrates both syntactic and semantic information. This paper tackles the challenge of applying these advanced deep parsing techniques, specifically using the Head-Driven Phrase Structure Grammar (HPSG) formalism, to the Spanish language. Previous work has primarily focused on English and other formalisms like CCG, leaving a gap in deep parsing capabilities for Spanish.

Methods

The authors developed a two-step deep parsing architecture using LSTM neural networks:

  1. Top-down Syntactic Parsing (Split and Rule Calculation): This step takes a sentence as input and recursively splits it into binary constituents, assigning a corresponding HPSG grammar rule to each split. This process is guided by a neural network that predicts the probability of splitting at each point and the most likely rule to apply. The method greedily selects the split point with the highest probability.
  2. Argument Calculation: Once the basic syntactic tree is formed, this step identifies and labels semantic arguments for predicates (verbs, nouns, adjectives). A separate neural network takes the head word, its potential argument, and the surrounding constituent as input to predict the probability of it being a semantic argument and its corresponding role label (e.g., ARG0, ARG1).

The system uses a simplified HPSG grammar adapted for Spanish, focusing on feature structures that encode morphological, syntactic valence, and semantic role information. For semantics, they adopt the PropBank notation.

Datasets

The authors utilized the AnCora Spanish corpus, which contains approximately 17,000 sentences with CFG-style syntactic annotations and semantic argument structures. This corpus was semi-automatically transformed into their HPSG grammar. The training, development, and test partitions were based on the standard AnCora partitions used for the CoNLL-2009 shared task, comprising around 418K words for training, 50K for development, and 50K for testing.

Key Insights and Ideas

  • Neural Networks for HPSG Deep Parsing in Spanish: The paper demonstrates the effectiveness of applying LSTM-based neural networks for deep parsing with a complex grammar formalism like HPSG in Spanish, a language less explored in this domain.
  • Two-Stage Approach for Syntactic and Semantic Integration: The separation into syntactic tree construction and subsequent argument identification allows for a modular and effective integration of both levels of linguistic analysis.
  • Robustness to Spanish Linguistic Phenomena: The grammar and parser are designed to handle specific Spanish characteristics, such as clitic pronouns and relative clauses, leading to improved performance on these phenomena.
  • Trade-off between Speed and Accuracy: While neural network approaches offer significant speed advantages over traditional CKY parsers, the paper highlights the ongoing challenge of optimizing execution time, especially for complex grammars.

Key Results

  • Syntactic Parsing: The LSTM top-down parser achieved the best performance in both constituency and dependency metrics, with an unlabeled constituency F1 of 87.57% and labeled dependency accuracy of 88.96%. It outperformed a CKY baseline and other established Spanish parsers.
  • Semantic Role Labeling (SRL): The parser achieved strong SRL results, with an unlabeled F1 of 87.68% and labeled F1 of 80.66%, outperforming most baselines.
  • Performance on Specific Phenomena: The parser showed good performance on phenomena like clitics, relative clauses, and coordinations, often outperforming baselines.
  • Execution Time: The LSTM top-down parser was significantly faster than the CKY baselines, with an average parsing time of 86.1 ms per sentence, compared to over 288 ms for CKY with gold tags.

Limitations and Future Work

  • Execution Time: While faster than CKY, the execution time is still an area for improvement. The authors plan to develop a unified architecture that performs both syntactic and argument steps simultaneously to potentially reduce execution times and improve generalization.
  • Post-processing for Evaluation: For certain linguistic phenomena, the results of other parsers had to be post-processed to enable comparison, which could introduce noise into the evaluation.
  • Future Directions: The authors suggest extending this approach to other languages like English and Italian, and further exploring its application to languages with similar characteristics to Spanish. They also aim to improve the network's generalization for SRL label prediction.

Summary updated

BibTeX
@inproceedings{chiruzzo2020statistical,
  author = {Luis Chiruzzo and Dina Wonsever},
  title = {{Statistical Deep Parsing for Spanish Using Neural Networks}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2020},
}
Figure