TL;DR
This paper introduces a method for automatically discovering "kind of" relationships (hyponymy) between words in large, unrestricted text corpora. By identifying specific linguistic patterns, the approach can build up lexical knowledge without relying on pre-existing dictionaries, offering a way to augment and critique existing resources like WordNet.
What Problem Does This Paper Address?
Building comprehensive lexical resources for natural language processing (NLP) is a significant challenge. While Machine-Readable Dictionaries (MRDs) have been useful, they are inherently limited by their fixed entries. To expand beyond MRDs and capture the vast, dynamic nature of language, researchers need methods to automatically extract lexical information from large, general-purpose text corpora. The core problem is how to reliably identify specific semantic relationships, like hyponymy (e.g., "dog" is a kind of "animal"), from the messy, diverse, and often implicit language found in everyday texts, without requiring extensive pre-encoded knowledge or complex parsers.
Methods
The authors propose a pattern-based approach to identify hyponymy. The key idea is to find recurring lexico-syntactic patterns in text that reliably signal this relationship. They identify several such patterns, exemplified by:
- "NP such as NP1, NP2, ..., NPn": For instance, "The bow lute, such as the Bambara ndang, is plucked..." implies "Bambara ndang" is a kind of "bow lute."
- "NP {, NP}*, {,} or other NP": For example, "Bruises, wounds, broken bones or other injuries" implies "bruise," "wound," and "broken bone" are kinds of "injury."
These patterns are designed to be:
- Frequent and genre-spanning: Occurring often across different types of text.
- Reliable: Almost always indicating the hyponymy relation.
- Recognizable with minimal pre-encoded knowledge: Not requiring complex linguistic tools that the acquisition process itself aims to build.
The acquisition algorithm uses a unification-based constituent analyzer, built on top of a part-of-speech tagger, to recognize these patterns. When a pattern is found, the noun phrases involved are extracted. The paper notes that for simplicity in this experiment, they focused on unmodified nouns, though they acknowledge the need for handling modifiers in more advanced versions.
Datasets
- Grolier's American Academic Encyclopedia: Used for running the acquisition algorithm and evaluating its results. This corpus contains approximately 8.6 million words.
- New York Times text: Mentioned as a corpus used in a slightly relaxed version of the experiment (approximately 20 million words).
- WordNet: Used as a benchmark to verify, critique, and augment the results obtained from the encyclopedia corpus. WordNet is a hand-built lexical thesaurus.
Key Insights and Ideas
- Implicit Knowledge in Text: Even without explicit definitions, natural language constructions can implicitly convey semantic relationships like hyponymy. A reader can infer that a "Bambara ndang" is a type of "bow lute" from sentence structure alone.
- Pattern-Based Acquisition is Efficient: Identifying specific, reliable linguistic patterns is a low-cost way to acquire lexical information, avoiding the need for complex parsers or extensive domain knowledge.
- Hyponymy is Particularly Amenable: The "naming" nature of English might make the hyponymy relation especially well-suited for this pattern-based acquisition approach, more so than other relations like meronymy.
- Critiquing and Augmenting Existing Resources: The discovered relations can be used to validate entries in established resources like WordNet (verify), suggest corrections (critique), or propose new entries (augment).
Key Results
- In an experiment on Grolier's Encyclopedia using a restricted version of pattern (1a) (with unmodified nouns), 152 relations were extracted from 7067 sentences containing "such as."
- When these relations were compared against WordNet:
- 180 out of 226 unique words involved in the relations existed in WordNet.
- 61 out of 106 feasible relations (where both terms were in WordNet) were found to be consistent with WordNet's hierarchy.
- The quality of the found relations was generally high, though some issues like metonymy and underspecification were observed.
- The authors suggest that relaxing restrictions on noun phrases (e.g., allowing simple compounds) significantly increases the number of relations found.
Limitations and Future Work
- Modifier Handling: The current experiment focused on unmodified nouns. A significant limitation is the difficulty in determining which modifiers are significant and how to handle them for accurate hyponymy extraction.
- Sense Disambiguation: When inserting new relations into a hierarchy, deciding which senses of polysemous words to link is a challenge. The authors suggest using lexical disambiguation algorithms.
- Pattern Discovery Automation: While they propose a procedure for automatically discovering new patterns, the current implementation is manual. Step 4 of their proposed procedure (finding commonalities among environments) is described as "underdetermined."
- Future Work: The authors plan to test the pattern discovery algorithm on more relations and languages, analyze the role of modifiers, and apply the method to analyzing environmental impact reports. They also aim to improve the automatic updating of lexical hierarchies.
Summary updated