TL;DR
This paper introduces a novel "Semantic Scattering" model for automatically identifying the semantic relationships expressed by English genitives (e.g., "child's mother" vs. "car door"). By leveraging WordNet's hierarchical structure and a corpus-driven approach, their model achieves significantly higher accuracy than traditional machine learning methods, demonstrating the power of tailored semantic analysis.
What Problem Does This Paper Address?
Natural Language Processing (NLP) systems often struggle to understand the nuanced meanings embedded in language. A prime example is the English genitive construction (like "X's Y" or "Y of X"), which can express a wide array of semantic relationships (e.g., possession, kinship, authorship, part-whole). Accurately interpreting these relationships is crucial for tasks like question answering, text summarization, and building knowledge bases. While previous work has explored identifying specific relations like PART-WHOLE, this paper tackles the broader challenge of automatically classifying the diverse semantic relations conveyed by genitives in general text.
Methods
The core of this paper is the Semantic Scattering Model. The authors propose a supervised learning approach where the goal is to map genitive constructions to one of 35 predefined semantic relations.
Here's a breakdown of their key methods:
- Feature Engineering: They extract features from the genitive's noun phrases, focusing on:
- Word Sense Disambiguation (WSD): Using a state-of-the-art WSD system, they determine the correct WordNet sense for both the modifier and head nouns. This is crucial because the meaning of "Mary's book" can vary greatly depending on whether Mary is the author, owner, or reader.
- Semantic Class of Nouns: They leverage the WordNet synsets (sets of synonyms representing a concept) of the modifier and head nouns. The semantic class of a noun (e.g., "entity," "possession") provides a higher-level abstraction that influences the genitive's meaning.
- Semantic Scattering Model: This model is built on the idea that the semantic relationships encoded by genitives are related to the position of their constituent noun senses within the WordNet hierarchy.
- Boundary Detection: They iteratively define a "boundary" (G*) within the WordNet hierarchy. This boundary separates noun concept pairs that uniquely map to a single semantic relation from those that are ambiguous. The process starts with a general boundary and specializes it based on the training data.
- Specialization: Ambiguous features (pairs of noun senses) are specialized by moving down the WordNet hierarchy to more specific senses, effectively "scattering" the potential semantic relations until each feature pair uniquely maps to a relation.
- Classification Algorithm: Once the boundary is established, a genitive is classified by:
- Processing the sentence (WSD, parsing).
- Identifying the head and modifier noun concepts.
- Mapping these concepts to the closest feature pair on the learned boundary.
- Assigning the semantic relation associated with that feature pair on the boundary.
Datasets
- Training and Testing Corpus: The authors used a corpus of approximately 20,000 sentences randomly selected from the LA Times articles (TREC-9). This corpus contained 3,255 genitive instances (2,249 of-genitives and 1,006 s-genitives).
- Annotation: Genitive instances were manually annotated with one of 35 predefined semantic relations (or "OTHER") by two graduate students. Noun components were disambiguated using WordNet 2.0 senses or named entities.
- Data Split: 80% of the corpus was used for training (including boundary detection) and 20% for testing.
Key Insights and Ideas
- Genitive Ambiguity is Hierarchical: The semantic relations expressed by genitives are not arbitrary but are structured by the hierarchical nature of WordNet. More general noun concepts lead to more ambiguous relations, while more specific concepts tend to resolve to a single relation.
- Tailored Models Outperform General Ones: A model specifically designed for genitive interpretation (Semantic Scattering) significantly outperforms general-purpose machine learning algorithms (like Decision Trees, Naive Bayes, and SVMs) when applied to this task. This highlights the importance of domain-specific feature engineering and modeling.
- WSD is Critical: The performance gains from incorporating Word Sense Disambiguation are substantial, underscoring its role in unlocking the semantic potential of genitives.
- S-genitives vs. Of-genitives Differ: The analysis reveals distinct semantic preferences for s-genitives (e.g., "man's brother") and of-genitives (e.g., "dress of silk"), suggesting they should be treated separately.
Key Results
The Semantic Scattering model achieved impressive results:
- Overall F-measure: 79% on the test set.
- Of-genitives: 81.82% precision and 77.88% recall.
- S-genitives: 83.53% precision and 73.96% recall.
These results are significantly better than:
- C5.0 Decision Trees: 40.60% accuracy (of-genitives), 47.0% (s-genitives).
- Naive Bayes: 42.31% accuracy (of-genitives), 43.7% (s-genitives).
- Support Vector Machines (SVM): 31.45% accuracy (of-genitives), 23.51% (s-genitives).
The paper also shows that the specialization process (moving from G¹ to G²) significantly improved accuracy, with the G² boundary being the best approximation.
Limitations and Future Work
- Error Sources: The authors identified several error sources, including missing features, general semantic classes, issues with the WSD system, reversed constituent order, and missing WordNet senses. Addressing these could further improve performance.
- "OTHER" Category: While the 35 relations were found to be representative, a significant number of examples fell into the "OTHER" category, suggesting that the predefined list might not be exhaustive or that some novel relations are being expressed.
- Future Work: The authors suggest that further improvements could come from addressing the identified error sources and potentially expanding the set of semantic relations. The Semantic Scattering model's generality makes it suitable for extending to other types of noun phrase constructions.
Summary updated