TL;DR
This paper introduces a novel three-phase approach to automatically construct domain-specific web corpora for term recognition, a crucial task in ontology learning. By intelligently partitioning the web based on topical relevance, the authors demonstrate that these dynamically generated corpora can achieve term recognition performance comparable to manually curated local corpora, with significantly less effort.
What Problem Does This Paper Address?
Ontology learning, a field focused on building knowledge bases, heavily relies on large text corpora to identify terms and their relationships. However, obtaining such corpora is a significant bottleneck. Existing approaches often use small, static, manually created local corpora, which are limited in scope and adaptability. While the World Wide Web offers an abundance of data, using it directly as a corpus is challenging due to its noisy, dynamic, and unorganized nature. This paper addresses the need for an automated, on-demand method to construct large, dynamic web corpora that are suitable for term recognition tasks, bridging the gap between the vastness of the web and the specific requirements of ontology learning.
Methods
The core of this paper is a novel three-phase web partitioning approach:
- Phase 1: Probabilistic URI Filtering: This phase starts with seed terms (e.g., "molecular biology") to query search engines like Google. It then filters the returned web page URIs based on three "odds":
- Odds of Popularity (Op): Measures how popular a website is in relation to the seed term, using PageRank as an indicator.
- Odds of Vocabulary Coverage (Oc): Estimates a site's ability to cover the domain's vocabulary by looking at the proportion of pages containing the seed term.
-
Odds of Vocabulary Specificity (Os): Assesses how specific a site's vocabulary is to the domain, by examining the variation in page content. These odds are combined into an Odds of Domain Characterisation (OD) score, and sites exceeding a threshold (ODT) are selected.
-
Phase 2: URI Clustering: The selected URIs are then clustered to group sites that represent different aspects or sub-topics within the domain. This phase uses a two-pass Tree-Traversing Ants (TTA) algorithm with a featureless similarity measure (NGD) to group similar URIs and refine the clusters, aiming to identify internal aspectual polarization.
- Phase 3: URI Cluster Selection: The final phase refines the clusters by removing those with poor domain characterization quality. This involves calculating:
- β-score: Measures the relatedness of a cluster to the seed terms.
- α-score: Measures the inter-cluster relatedness. These are combined into N-score (neighborhood score) and I-score (individual score), which are then used to compute a final Θ-score. Clusters with a Θ-score above a dynamically determined threshold are retained. The output is a set of URIs representing a focused domain corpus.
Datasets
The paper uses the following for evaluation:
- GENIA Corpus: A semantically annotated corpus for the "molecular biology" domain, used as a gold standard for term recognition. It contains approximately 400,000 words.
- Local Contrastive Corpus: A collection of about 6,700,000 words from online news articles.
- Constructed Web Corpora: The paper describes constructing two web corpora for the "molecular biology" domain using their proposed approach:
- A web domain corpus (55 URIs, over 22 million web pages).
- A web contrastive corpus.
Key Insights and Ideas
- Web as a Dynamic Corpus: The paper strongly advocates for the web as a source for dynamic, on-demand corpora, moving beyond static, manually curated datasets.
- Topical Partitioning is Key: The core innovation lies in the multi-phase partitioning strategy that goes beyond simple keyword matching to assess popularity, vocabulary coverage, and specificity, ensuring the constructed corpus is truly representative of the domain.
- Aspectual Clustering: The URI clustering phase is crucial for identifying different facets of a domain, allowing for a more nuanced understanding and potentially enabling more targeted term recognition.
- Performance vs. Effort Trade-off: The results suggest that automatically constructed web corpora can achieve performance comparable to manually curated ones, with a significantly lower investment of human effort. This implies that the sheer size and evolutionary nature of web corpora can compensate for the lack of manual vetting.
Key Results
The paper demonstrates the effectiveness of their approach through term recognition experiments using the GENIA corpus as a gold standard.
- Comparable Performance: The web corpora constructed using their approach achieved term recognition performance (precision, recall, F0.1, F1, accuracy) comparable to, and in some cases slightly lower (1-4%) than, manually curated local corpora.
- Efficiency: The approach significantly reduces the effort required to build domain-specific corpora, making it feasible to construct large, dynamic collections on demand.
- Empirical Evidence: Figure 3 shows that even with a minimal effort in corpus construction, the performance is promising, especially considering the web corpora are not manually vetted for text quality.
Limitations and Future Work
- Limitations:
- The paper acknowledges that the performance of web corpora is slightly lower than local corpora, which are manually vetted for text quality.
- The derivation of the overall threshold ODT is noted as being beyond the scope of the paper.
- The paper focuses on term recognition; its applicability to other ontology learning tasks like relation discovery is mentioned as future work.
- Future Work:
- Investigating the effect of different seed terms on web partitioning.
- Examining how the content of web corpora evolves over time and its impact on term recognition.
- Extending the use of web corpora to other applications requiring contrastive analysis and other ontology learning tasks, such as relation discovery.
Summary updated