TL;DR
This paper proposes a novel query expansion strategy for domain-specific information retrieval, focusing on medical texts. The core idea is to leverage domain-specific knowledge to intelligently decompose medical compounds and use their relevant components to enrich search queries, aiming to improve retrieval effectiveness.
What Problem Does This Paper Address?
Information retrieval (IR) systems aim to find relevant documents for a user's information need. A key challenge is that users may not always express their need using the exact terms found in documents. Query expansion is a technique to address this by modifying or adding terms to a user's query. This paper specifically tackles the problem of expanding queries that contain compounds – words formed by combining two or more meaningful parts (morphemes). In Swedish, compounds are very common, especially in specialized domains like medicine. The challenge lies in deciding which parts of a compound to use for expansion. Simply decomposing all compounds can lead to irrelevant results if the components are either too general or belong to a different domain. The authors hypothesize that using only compound components that are also specific to the target domain (in this case, medicine) will increase retrieval effectiveness without significantly sacrificing precision.
Methods
The proposed strategy involves using a domain-specific thesaurus to identify which components of a compound belong to the target domain. The authors utilize the Swedish MeSH (Medical Subject Headings) thesaurus, which is based on the American MeSH. A "MeSH tagger" is employed to identify medical terms within text. This tagger works by identifying maximal length strings from specific subdomains of MeSH.
The core of the strategy is:
- Identify Medical Compounds: Use the MeSH tagger to identify compounds that are themselves medical terms.
- Decompose and Tag Components: For compounds that are not fully tagged as medical, or for those where only some components are medical, decompose them into their constituent parts.
- Domain-Specific Selection: Use the MeSH tagger to determine if individual components belong to the medical domain.
- Intelligent Expansion: Expand the original query by adding the identified, domain-specific compound components. Crucially, components that are not tagged as medical or are common in standard language are excluded from the expansion.
The paper also explores the impact of pre-decomposing compounds before tagging, finding that it can increase the number of tagged components.
Datasets
- Medlex: A corpus of 5205 compounds extracted from an on-line medical lexicon. This corpus was created by adding medical vocabulary to a learner's dictionary and is stated to be largely from the medical domain.
- Swedish MeSH: Used as the controlled vocabulary for identifying medical terms.
Key Insights and Ideas
- Domain-Specific Decomposition is Key: Not all compound parts are equally useful for query expansion. The effectiveness hinges on selecting components that are semantically relevant to the target domain.
- Thesauri as Domain Guides: Medical thesauri like MeSH are valuable tools for identifying domain-specific terms and can guide intelligent query expansion.
- Compositional vs. Non-Compositional Compounds: The paper implicitly suggests that compositional compounds (where the meaning is derived from the parts) are more amenable to expansion than non-compositional ones (where the meaning has shifted).
- Challenges with Ambiguity: Words that are homonymic, polysemic, or have both medical and non-medical facets pose significant challenges for accurate domain identification and thus for effective query expansion.
- Pre-decomposition Boosts Tagging: Decomposing compounds before applying the MeSH tagger significantly increases the number of tagged components, potentially leading to more effective expansion.
Key Results
- In the initial run with the Medlex list, 895 out of 5205 compounds were tagged as medical.
- Among compounds not tagged as a whole, 233 had one or two components correctly tagged.
- After pre-decomposing the compounds, the number of tagged compounds increased to 1095, with 819 of these being used in a medical sense. This is a substantial improvement over the 233 from the previous run.
- The paper highlights examples where decomposition is beneficial (e.g., narkotikapolitik -> narkotika + politik) and where it is not (e.g., fotboll -> fot + boll).
- The authors identify that a significant portion of components tagged as medical but not used in a medical sense are common words like "hand" and "heart," which also have standard language meanings.
Limitations and Future Work
- Lack of a Test Collection: The paper acknowledges that the effectiveness of the proposed strategy has not been empirically tested due to the absence of a suitable Swedish medical test collection.
- Ambiguity and Facets: Dealing with homonymic, polysemic, and multi-faceted words remains a significant challenge.
- Future Work: The author is currently working on creating a Swedish medical test collection named "MedEval." Future research will involve testing query expansion strategies on this collection and exploring how to apply these strategies not only to queries but also to document indexes. The challenge of handling ambiguous words will continue to be a focus.
Summary updated