Introduction
Semantic parsing—the translation of natural language into machine-readable queries—has traditionally faced a trade-off between precision and coverage. Traditional parsers are highly accurate within a fixed schema (e.g., Freebase) but fail on out-of-vocabulary terms. Conversely, open-vocabulary parsers use large text corpora for broad coverage but lack the structured "ground truth" of a formal Knowledge Base (KB). This paper introduces a hybrid approach that bridges this gap by treating formal KB queries as features for a distributional model.
Methodology
The authors propose a system that combines Combinatory Categorial Grammar (CCG) parsing with structured knowledge features. The parser first generates "surface" logical forms based on the input text.
Subgraph Feature Extraction (SFE)
To integrate the KB, the model uses Subgraph Feature Extraction (SFE). This technique identifies paths in the Freebase graph (e.g., DESIGNED → LOCATED_IN) and converts them into binary features. By treating the KB as a source of features rather than a rigid destination, the model remains robust; it leverages structured data when available and falls back on distributional statistics otherwise.
The Combined Model
The probability that an entity fits a predicate is calculated using two distinct components: distributional vectors learned from text co-occurrence and weighted KB features extracted via SFE. To manage dimensionality, Pointwise Mutual Information (PMI) is used to select the top 100 most relevant features per predicate.
Results
The model was evaluated on a fill-in-the-blank task using the ClueWeb09 corpus and Freebase. The results demonstrate that structured knowledge provides a massive boost to open-vocabulary systems. The combined model achieved a Mean Average Precision (MAP) of 0.370, representing a 127% relative improvement over the distributional-only baseline of 0.163.
Key insights include:
- Connectivity: The KB provides critical data for rare entities that lack sufficient text mentions.
- Compositionality: Unlike many neural approaches, this model maintains a compositional structure, allowing it to understand how sentence components combine into complex queries.
Discussion and Limitations
Despite its success, the model has limitations. The current edge-sequence features cannot always capture fine-grained nuances, such as specific player positions in sports. Additionally, the system still struggles with "zero-mapping" predicates that have no correlate in the KB. Future work may focus on more expressive path features and improved candidate entity generation.
Summary updated