2008 ExploitingHyponymy
- (Mandhani & Soderland, 2008) ⇒ Bhushan Mandhani, Stephen Soderland. (2008). “Exploiting Hyponymy in Extracting Relations and Enhancing Ontologies.” In: IEEE Workshop on Natural Language Processing and Ontology Engineering. doi:10.1109/WIIAT.2008.314
Subject Headings: Hyponymy, Relation Extraction Task, Taxonomic Relational Pattern.
Notes
Quotes
Abstract
- Relation extraction systems typically rely on local lexical and syntactic features as evidence. Recent work suggests that for a given relation, there might exist certain patterns which, if present in the graph of relationships between objects, provide additional evidence for that relation. While these relational patterns can be very useful, obtaining them on a per-relation basis can be difficult. We propose template relational patterns based on the hyponymy relation. These patterns are applicable for all relations. Existing resources like WordNet are a rich and reliable source of hyponymy relationships between entities. We present techniques for making use of these template patterns for extracting relationships and incorporating them into the ontology. Our experiments show performance improvements in both tasks.
1. Introduction
- The problem of extracting relationships between entities
from a corpus has recently received a lot of attention. It is an important information extraction task, and has many uses. It is a key component in building ontologies and structured knowledge bases from the web [6]. It also has uses in more traditional applications like question answering and information retrieval. A variety of techniques have been proposed for relation extraction. They primarily use lexical and syntactic information as evidence. A range of classifiers and probabilistic models have been used for deciding whether or not the relation holds, given the evidence.
- However, there has been little work on employing semantic
evidence in relation extraction. The most common such feature seems to be the class of the entities involved in the relationship [8, 3]. This feature is available only when the entity recognizer also assigns class labels to the entities. [2] propose a new source of semantic evidence, which they call “relational patterns”. The idea here is that for a given relation, there might exist a pattern of relationships involving other relations, such that if this pattern holds between two entities, then the original relation is also likely to hold. We illustrate this with an example. Suppose our set of relations includes “Mother”, “Father” and “Wife”. A relational pattern predictive of entity X having mother Z would be that X has father Y and Y has wife Z. So, if these last two relationships are known, we have additional evidence for the “Mother” relation between X and Z. Note that this evidence is not only semantic, but also global. It is based on relationships extracted from elsewhere in the corpus.
- It is instructive to express the semantics of relational patterns
more formally. They represent Horn clauses in firstorder logic. The above pattern can be written as follows:
- ∀x, y, z Father(x, y) ∧ Wife(y, z) ⇒ Mother(x, z)
- Unlike first-order logic, these have a probabilistic interpretation.
When the clause body fires, the head of the clause can’t be “inferred”. Instead, we have more evidence that it is true. We formalize this later in the paper.
- Relational patterns can help improve precision and recall
in relation extraction. [2] obtained improvements in FMeasure when they incorporated them into their extraction model. However, obtaining them on a per-relation basis is a challenge, especially when the number of relations is very large. Mining the set of extracted relationships seems like a promising approach. However, for a web-scale system with millions of extracted relationships, learning such rules may not be computationally feasible.
- We propose Taxonomic Relational Patterns (TRPs), which are template patterns that can be applied for each
distinct relation. Entities in extracted relationships are often mapped to objects in an existing ontology or semantic taxonomy like WordNet. This provides us with an accurate and largely complete set of hyponymy relationships between these entities. TRPs look to exploit these relationships.
- We present three TRPs in this paper.
,
| Author | volume | Date Value | title | type | journal | titleUrl | doi | note | year | |
|---|---|---|---|---|---|---|---|---|---|---|
| 2008 ExploitingHyponymy | Stephen Soderland Bhushan Mandhani | Exploiting Hyponymy in Extracting Relations and Enhancing Ontologies | http://www.cs.washington.edu/homes/soderlan/mandhani-Hyponymy.pdf | 10.1109/WIIAT.2008.314 |