TL;DR
This paper presents a three-step method to automatically collect a dozen technical terms closely related to a given "seed" term from the web. The system leverages search engines to build a corpus, applies automatic term recognition, and then filters candidates to ensure they are relevant and technical, achieving a precision of 85%. This work offers a practical approach to building domain-specific glossaries and technical dictionaries without requiring manually curated datasets.
What Problem Does This Paper Address?
Building comprehensive glossaries or technical dictionaries for specific domains is a labor-intensive process. Traditional methods for automatic term recognition (ATR) often rely on large, manually prepared corpora, which are expensive and time-consuming to create. This paper addresses the challenge of automatically acquiring related technical terms from the vast, unstructured data available on the web, using only a single seed term as input. The goal is to automate the initial stages of glossary compilation, making it more efficient and accessible.
Methods
The proposed system employs a three-step pipeline:
- Compiling Corpus: For a given seed term
s, the system queries search engines (Goo and Infoseek) using variations ofscombined with Japanese functional words like "toha," "toiu," and "ha" to elicit definitional or explanatory content. It collects up to 100 pages per query and further gathers pages linked via anchor text containings. Sentences containingsare then extracted to form the corpusCs. This step aims to gather relevant web content related to the seed term. - Automatic Term Recognition (ATR): This step extracts candidate technical terms from the compiled corpus. The authors adapt Nakagawa's ATR method, which is suitable for Japanese text. The process involves:
- Generating a term list: Identifying all nouns and compound nouns from the corpus.
- Selection by scoring: Ranking the extracted terms using a scoring function that combines Nakagawa's
Imp₁(a measure of term importance) with a frequency factorF(x, L). The formula isscore(x, L) = Imp₁(x, L) × F(x, L)ᵃ, whereais a parameter controlling the influence of frequency (set to 0.5 in experiments). The top 30 terms are selected as candidates. - Filtering: This crucial step removes noisy or irrelevant terms from the ATR candidates. It comprises two tests:
- Technical-term test: This test filters terms that are unlikely to be technical. It uses the web as a resource to estimate term frequency. A term
xis considered technical if its web hit countH(x)(number of pages containingx) falls within a specific range (100 to 100,000, determined empirically). Terms outside this range are deemed too general or too rare to be technical. - Relation test: This test ensures that the candidate terms are closely related to the seed term. It uses conditional probabilities derived from web hit counts:
P(s|x) = H(s ∧ x) / H(x)andP(x|s) = H(s ∧ x) / H(s). A termxis considered related tosif eitherP(s|x)orP(x|s)is greater than a thresholdZ(set to 0.05). The paper also categorizes terms into types (e.g., containings, subsequence ofs) to aid in relation assessment.
Datasets
The paper does not introduce new benchmark datasets. Instead, it utilizes the World Wide Web as its primary data source. For evaluation, the authors created a set of fifty seed terms across five genres: natural language processing, Japanese language, information technology, current topics, and persons in Japanese history. For each seed term, they manually identified target related terms to estimate the system's recall.
Key Insights and Ideas
- Web as a Dynamic Corpus: The paper demonstrates the effectiveness of using the web as a dynamic and readily available source for building domain-specific corpora, bypassing the need for manual curation.
- Hybrid ATR and Filtering: The combination of a statistical ATR method with web-based filtering is key. The ATR extracts potential terms, while the filtering step leverages the scale of the web to validate their technicality and relevance, significantly improving the quality of the output.
- Empirical Thresholds for Technicality: The use of web hit counts (
H(x)) within a defined range (100-100,000) as a proxy for technical term identification is an interesting heuristic. It suggests that technical terms occupy a specific niche in web content frequency. - Conditional Probabilities for Relation: Employing conditional probabilities derived from web search results (
P(s|x)andP(x|s)) provides a robust statistical measure for assessing the semantic relatedness between candidate terms and the seed term.
Key Results
The system achieved a precision of 85% across fifty seed terms. In the evaluation, 519 out of 610 collected terms were deemed correct related terms. The filtering step was particularly effective, with the technical-term test passing 133 out of 210 target terms that were initially identified by ATR. However, the recall was noted as low, with only 20% of target terms being collected, primarily due to failures in the ATR step.
Limitations and Future Work
The authors acknowledge the low recall of the system, attributing it mainly to the limitations of the automatic term recognition (ATR) step. They suggest that the ATR step might be improved or potentially replaced with a more exhaustive term collector from the corpus. Future work could focus on enhancing the ATR component or exploring alternative methods for term extraction. The paper also hints at future integration with a "term explainer" to automatically generate definitions for the collected terms.
Summary updated