Introduction
This 1986 paper from the AAAI proceedings addresses one of the most persistent challenges in Natural Language Processing (NLP): prepositional phrase (PP) attachment ambiguity. By leveraging a structured "Kind Types" system of commonsense knowledge, the authors demonstrate a method to filter out semantically nonsensical parses, achieving an impressive 93% to 98% accuracy rate.
The core innovation lies in using a "commonsense" knowledge base of ontological categories and prototypical features to resolve syntactic ambiguities. Instead of relying solely on syntax, the system uses the semantic properties of nouns and verbs to determine whether a phrase like "with the key" modifies an object (e.g., "the lock") or an action (e.g., "buying").
The Problem of PP Attachment Ambiguity
In English, the sentence structure NP-V-Det-N-PP is notoriously ambiguous for syntactic parsers. Consider the sentence: “John bought the lock with the key.”
Without world knowledge, a standard parser might generate three valid syntactic trees:
- NP Constituency: The key is a component or attribute of the lock (The lock-with-a-key).
- VP Constituency: The key was the instrument used to perform the action of buying (Buying-with-a-key).
- S-Modification: The entire event of buying took place while a key was present.
While humans instantly recognize that keys belong to locks (NP attachment), a computer lacks the "world knowledge" to realize that a physical key is not a standard currency for a purchase. This paper aims to bridge that gap by providing the parser with a "commonsense" filter to prune semantically implausible trees.
Methodology: The Kind Types (KT) System
The authors utilize the Kind Types (KT) System, which moves beyond simple dictionary definitions to include four layers of knowledge:
- Ontological Database: A hierarchy of basic distinctions (e.g., is an object REAL or ABSTRACT? Is a person SENTIENT?).
- Generic Lexicon: A collection of prototypical and inherent features. For example, the entry for "chicken" includes that it "lays eggs" and is "handleable."
- Feature Typing Database: Categorizes features into types like color, size, behavior, and function.
- Preference Strategy Rules: A set of seven global rules and thirteen preposition-specific rules.
When a parser generates multiple trees, the KT system applies these rules. For example, a "with-rule" might check the Generic Lexicon to see if the object of the preposition (the key) is a prototypical part of the direct object (the lock). If it is, it prefers NP-attachment. Conversely, if the object is a prototypical instrument for the verb (e.g., "cut with a knife"), it prefers VP-attachment.
Experimental Evaluation
The system was developed and tested on two primary datasets:
- Development Set: 4,500 words of schoolbook geography texts used to build the initial lexicon and ontology.
- Test Set: Three diverse excerpts totaling roughly 1,300 words, including a novel, a history of science text, and a technical article.
Key Results
- High Accuracy: The system achieved a 100% success rate on the initial geography corpus.
- Generalization: On the three unseen test texts, the success rate was 93%.
- Error Analysis: When excluding "idiomatic phrases" (like "at all" or "in effect"), which the system was not designed to handle, the accuracy jumped to 98%.
- The "Of" Factor: The authors noted that "of-phrases" make up about 32% of all PPs and almost universally attach to the noun to their left, significantly simplifying the task.
Discussion and Key Insights
- Prototypes over Definitions: Disambiguation requires knowing how people typically view the world (prototypes), not just formal logic. Knowing that a "wrench" is a tool is more useful for parsing than a complex dictionary definition.
- The "Size" Heuristic: To prevent the system from thinking "John mixed the salad in the salad bowl" means the entire event of John happened inside the bowl, the system uses "reality-related size ranges." It assumes that for a PP to modify a whole sentence (S-modification) as a location, the location must be at least two size-orders larger than the subject.
- Lexical "Elsewhere" Conditions: The system uses a tiered approach—trying specific, high-confidence rules first and falling back to "elsewhere" defaults (like assuming locative PPs modify NPs if no other rule triggers).
Limitations and Future Work
The primary failure point was idiomatic expressions. The authors acknowledge that a separate system for handling non-literal language would be necessary for a production-level parser. Additionally, some sentences remain truly ambiguous even with commonsense (e.g., "John saw the horse in the city"). Finally, the knowledge engineering bottleneck—the need for hand-coded Ontological Databases—remains a significant challenge for scaling such systems to the entire English language.
Summary updated