Abstract
In the biomedical domain, researchers often rely on automated tools to navigate vast amounts of literature. This paper explores a syntactic approach to sentence compression designed to optimize the presentation of related article titles in search engines like PubMed. By automatically trimming inessential fragments from titles, the proposed system allows for more results to be displayed within limited screen space. Through a prototype applying seven syntactic rules to parse trees, the study evaluates the effectiveness of this method using both BLEU scores and human assessments. Results indicate that titles can be compressed by approximately 30% while maintaining the "interest judgment" accuracy of users—the ability to quickly decide if an article is worth further investigation.
The Challenge of Information Density
Information retrieval interfaces, particularly in the medical field, face a constant struggle between providing comprehensive information and maintaining a clean, usable interface. In PubMed, when a user views a specific abstract, the system generates a list of "Related Links" in a sidebar. However, the physical constraints of the screen limit the number of titles that can be shown.
Traditionally, interfaces handle this by either truncating titles (which can lead to a loss of critical context) or showing fewer results. This research proposes sentence compression as a middle ground: removing linguistically inessential fragments to preserve the "indicative" and "informative" nature of the title while reducing its length. The goal is to facilitate "interest judgments"—the quick decision a researcher makes about whether to click on a related article.
Methodology: A Rule-Based Syntactic Approach
Unlike statistical compression models that require massive training datasets, this system utilizes a syntactic trimming approach. It processes titles through the Stanford Parser to generate parse trees and then applies a sequence of seven linguistically-motivated rules in order of increasing complexity:
- Subtitles: Deleting text following colons or dashes.
- Determiners: Removing articles such as "the," "a," and "an."
- Participial/Gerund Phrases: Trimming phrases beginning with "-ing" verbs.
- Serial Prepositional Phrases (PPs): In a sequence of PPs, only the first is retained.
- Nested PPs: Removing prepositional phrases embedded three or more levels deep in the parse tree.
- Conjoined Noun Phrases (NPs): Removing the second part of a "Noun and Noun" structure.
- Simple NPs: Removing adjectives unless the "head" noun is considered "lightweight" (determined by a low inverse document frequency (IDF) score).
The system was tested in two configurations: Variant A (Rules 1–6) and Variant B (Rules 1–7).
Evaluation and Dataset
To evaluate the system, the researchers created a gold-standard dataset of 400 article titles sampled from the TREC 2005 Genomics track. Two subject-matter experts (PhDs in bioinformatics and genetics) manually compressed these titles to provide a reference.
The evaluation focused on two primary metrics:
- BLEU Scores: To measure the fluency and n-gram overlap between machine-generated and human-generated compressions.
- Interest Judgments: A summative evaluation where human assessors determined if a paper was worth reading based on the compressed title, comparing their decisions against those made using the original full titles.
Key Results and Findings
The experiments yielded several significant insights into the viability of syntactic compression in specialized domains:
- Compression Efficiency: Variant A successfully reduced title length by approximately 30%. Variant B achieved even higher compression, matching the levels of the human experts.
- Task Performance: Crucially, there was no significant difference in the accuracy of "interest judgments" between the original titles and the Variant A compressed titles. This suggests that the trimmed fragments were indeed inessential for the user's decision-making process.
- Fluency: Human assessors rated the machine-compressed titles in Variant A as being as fluent as, and occasionally more fluent than, the manually compressed versions.
- The Limit of Trimming: While Variant B achieved higher compression, it began to negatively impact the precision and recall of user interest judgments. This indicates that Rule 7 (adjective removal) sometimes strips away vital descriptive information.
Conclusion
This research demonstrates that a syntactic approach to sentence compression is highly effective for the biomedical domain. By leveraging the structure of language rather than just statistical frequency, the system can significantly reduce title length without compromising the utility of search results. Future work aims to refine these rules and potentially apply them to other constrained formats, such as GeneRIFs (concise phrases describing gene functions), to further facilitate rapid access to critical biological data.
Summary updated