Authored by Brooke Cowan and Michael Collins, this research explores the adaptation of statistical parsing techniques—originally optimized for English—to the morphologically rich landscape of the Spanish language. By incorporating language-specific morphological features and a global reranking step, the authors significantly improve the accuracy of Spanish constituency parsing.
The Challenge of Morphological Richness
Most early work in statistical parsing focused on English, a language often described as "morphologically impoverished." In English, word order is the primary driver of syntax. In contrast, Spanish morphology provides critical structural cues. Nouns, adjectives, and determiners must agree in gender and number, and verbs must agree with their subjects.
Standard parsers often ignore these details, treating variations like gato (cat, masc. sing.) and gatas (cats, fem. plur.) as unrelated tokens. This paper addresses how to effectively incorporate these "local" morphological constraints and "global" structural features into a unified parsing framework.
Methodology
The researchers employed a two-stage strategy to move beyond a standard lexicalized Probabilistic Context-Free Grammar (PCFG) baseline.
1. Morphological Integration
The authors modified the Part-of-Speech (POS) tagset of their baseline parser to encode attributes like number, gender, and verb mode. By enriching tags, the parser can use agreement as a filter to rule out ungrammatical dependencies. For example, if the model encounters a plural noun and a singular verb, the enriched tags allow it to assign that structure a very low probability based on general patterns, even if it hasn't seen those specific words together before.
Experimentation with 22 different morphological features revealed a "Goldilocks" effect: adding every possible feature led to data sparsity issues. The best results came from a subset including verb mode/number and the number of adjectives, nouns, and pronouns.
2. Global Reranking
Because a PCFG makes local decisions, it can miss broader structural patterns. The authors took the top $n$ parse candidates from their morphological model and applied a separate reranking algorithm (an exponentiated gradient algorithm). This stage utilized "global" features—complex properties of the entire tree—that are too computationally expensive to include in the initial search. Interestingly, many of these features were originally developed for English, suggesting that high-level structural features are effective across different languages.
Datasets and Preprocessing
The study utilized the Spanish 3LB treebank, containing approximately 3,500 hand-annotated trees. Because the treebank is relatively small, the authors performed systematic modifications to the trees to make linguistic patterns easier for the statistical model to learn. This included restructuring coordinate phrases and subordinate clauses.
Key Results
The improvements were consistent across all metrics, including Labeled Constituents and Dependency scores:
- Baseline (English-style PCFG): 81.2% F1 score.
- Morphological Model: 83.6% F1 score. The biggest gains occurred in identifying subjects and sentential heads, where agreement is a primary signal.
- Final Reranking Model: 85.1% F1 score.
The jump from 81.2% to 85.1% represents a significant reduction in error, demonstrating that morphology is a core component of Spanish syntax that statistical models must account for to achieve high accuracy.
Summary updated