Learning Arguments and Supertypes of Semantic Relations Using Recursive Patterns

Zornitsa Kozareva, Eduard H. HovyACL 2010 · OpenReview

Abstract

A challenging problem in open information extraction and text mining is the learning of the selectional restrictions of semantic relations. We propose a minimally supervised bootstrapping algorithm that uses a single seed and a recursive lexico-syntactic pattern to learn the arguments and the supertypes of a diverse set of semantic relations from the Web. We evaluate the performance of our algorithm on multiple semantic relations expressed using "verb", "noun", and "verb prep" lexico-syntactic patterns. Human-based evaluation shows that the accuracy of the harvested information is about 90%. We also compare our results with existing knowledge base to outline the similarities and differences of the granularity and diversity of the harvested knowledge.

Introduction

In Natural Language Processing (NLP), understanding a relation like "flies to" requires identifying its selectional restrictions. For example, while a person or an airline can "fly to" a city, a "sandwich" generally cannot. Traditionally, acquiring this knowledge has relied on either unsupervised clustering, which often lacks human-interpretable granularity, or pattern-based methods that require extensive manual seeding. This research introduces a bootstrapping algorithm that requires only a single seed word to discover both valid arguments and their general categories (supertypes) with high precision.

Methodology: Recursive Lexico-Syntactic Patterns

The core innovation of this approach is the use of recursive lexico-syntactic patterns to explore the vastness of the Web.

Argument and Supertype Harvesting

Instead of a simple pattern like "X flies to Y," the system utilizes a conjunction: "X and [Seed] fly to Z." This leverages the linguistic principle that items in a list are typically semantically similar. Once the system identifies a new argument (e.g., finding "Mary" via the seed "John"), it uses that new argument as a seed for the next iteration. This creates a breadth-first expansion of the relation's argument space.

To categorize these arguments, the system employs an inverse Doubly-Anchored Pattern (DAP): "* such as X and Y." By plugging in discovered argument pairs (e.g., "such as John and Mary"), the system harvests the leading word to identify the supertype (e.g., "people" or "passengers").

Figure 1: Illustration of recursive argument and supertype harvesting for the 'fly to' relation.

Graph-Based Ranking

To filter out noise inherent in Web data, harvested terms are modeled as a directed graph. A term is considered valid if it frequently discovers other terms and is frequently discovered by them, ensuring the system remains robust against typos or parsing errors.

Experimental Results

The system was tested on 14 different relations involving various parts of speech, including verbs ("cause"), nouns ("person"), and verb-prepositions ("work for", "live in").

Performance and Diversity

Human evaluators found that the harvested arguments and supertypes were correct approximately 90% of the time. A key advantage of this recursive method is its ability to capture the "long tail" of language. While traditional knowledge bases like Yago focus on famous entities, this system discovers valid but "non-famous" arguments, such as living in "peace" or "harmony," or working for a "local boss."

Table 1: Summary of recursive patterns, seeds, and the volume of arguments learned across various relations.

Conclusion

This research demonstrates that a single seed word is sufficient to trigger a chain reaction of semantic discovery across the Web. By using conjunctions as semantic anchors, the algorithm effectively learns the "who" and "what" of diverse relations. Future work involves using this framework to build a global semantic repository to support complex tasks like Question Answering and Textual Entailment by providing the common-sense constraints that machines currently lack.

Summary updated

Abstract

A challenging problem in open information extraction and text mining is the learning of the selectional restrictions of semantic relations. We propose a minimally supervised bootstrapping algorithm that uses a single seed and a recursive lexico-syntactic pattern to learn the arguments and the supertypes of a diverse set of semantic relations from the Web. We evaluate the performance of our algorithm on multiple semantic relations expressed using "verb", "noun", and "verb prep" lexico-syntactic patterns. Human-based evaluation shows that the accuracy of the harvested information is about 90%. We also compare our results with existing knowledge base to outline the similarities and differences of the granularity and diversity of the harvested knowledge.
BibTeX
@inproceedings{kozareva2010learning,
  author = {Zornitsa Kozareva and Eduard H. Hovy},
  title = {{Learning Arguments and Supertypes of Semantic Relations Using Recursive Patterns}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2010},
  url = {https://openreview.net/forum?id=S1-uNTguWB},
}
Figure