Skip to main content
Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
Inf Retr Boston. Author manuscript; available in PMC 2009 Sep 21.
Published in final edited form as:
Inf Retr Boston. 2009; 12(1): 69–80.
doi: 10.1007/s10791-008-9074-8
PMCID: PMC2747526
NIHMSID: NIHMS73870
PMID: 19774223

Evaluation of Query Expansion Using MeSH in PubMed

Zhiyong Lu, Won Kim, and W. John Wilbur
Author information Copyright and License information PMC Disclaimer

Abstract

This paper investigates the effectiveness of using MeSH® in PubMed through its automatic query expansion process: Automatic Term Mapping (ATM). We run Boolean searches based on a collection of 64 topics and about 160,000 MEDLINE® citations used in the 2006 and 2007 TREC Genomics Tracks. For each topic, we first automatically construct a query by selecting keywords from the question. Next, each query is expanded by ATM, which assigns different search tags to terms in the query. Three search tags: [MeSH Terms], [Text Words], and [All Fields] are chosen to be studied after expansion because they all make use of the MeSH field of indexed MEDLINE citations. Furthermore, we characterize the two different mechanisms by which the MeSH field is used. Retrieval results using MeSH after expansion are compared to those solely based on the words in MEDLINE title and abstracts. The aggregate retrieval performance is assessed using both F-measure and mean rank precision. Experimental results suggest that query expansion using MeSH in PubMed can generally improve retrieval performance, but the improvement may not affect end PubMed users in realistic situations.

Keywords: Query Expansion, TREC Genomics, PubMed, MeSH

1 Introduction

Query expansion refers to the process of reformulating a query to improve retrieval performance, typically by including additional terminology that is synonymous. PubMed, the search engine for over 17 million biomedical and health citations in MEDLINE, allows its users to manually attach search field tags (also called qualifiers) to search terms in square brackets (i.e., search term [tag]) in order to improve on users’ original searches. However, such advanced MEDLINE search features were seldom used as seen in the PubMed log analysis (Herskovic et al., 2007). Alternatively, PubMed employs a process called Automatic Term Mapping (ATM) that compares and maps untagged terms from the user query to lists of pre-indexed terms in PubMed’s translation tables (in this order): the Medical Subject Headings (MeSH) table (mapping search terms to MeSH concepts), the journals translation table (mapping search terms to journal names), and the author index (mapping search terms to author names). In the context of query expansion, we consider only the translation via MeSH here, which is designed by the National Library of Medicine (NLM) for indexing and searching of the MEDLINE database of journal citations. That is, if a user query includes a term that can be mapped to a MeSH concept, the search will then add the MeSH term to the original query. As a result, the ATM process enables the original query to have an access to the MeSH field of indexed MEDLINE documents. For instance, if a user submits a query tumor with no search tags attached, the query will be automatically mapped to the MeSH term Neoplasms by ATM. Thus, not only documents having the word tumor in the title and abstract will be retrieved, any other documents indexed with the MeSH term Neoplasms (as well as more specific terms beneath it in the MeSH hierarchy) will be returned as well. Furthermore, individual words in MeSH concepts are also indexed and searchable. Take the previous query for example, documents indexed with MeSH terms that include the word tumor (e.g. Tumor Virus Infections) will also be returned.

By default, PubMed applies the Automatic Term Mapping feature to every unqualified user query, thus affecting the search results of millions of user queries each day in spite of the fact that many users may not actually realize its existence. It is designed and implemented by the NLM as a means to improve retrieval performance. Unlike the studies in the past that focused on the accuracy of the mapping process (Gault et al., 2002; Carlin, 2004; Smith, 2004; Shultz, 2006), the main objective of this work is to examine the effectiveness of using the MeSH field of indexed documents for improving retrieval performance.

1.1 Related Work

There is a significant body of literature on query expansion as an effective technique in the general information retrieval (IR) community (Salton and Buckley, 1997; Mitra et al., 1998). However, in the domain of biomedical text retrieval, results have been mixed. Earlier research has suggested the use of this technique could result in elevated retrieval performance (Srinivasan, 1996; Aronson and Rindflesch, 1997), while no performance gain was seen in others (Hersh et al., 2000).

More recently, this technique has been widely implemented in various retrieval systems built by a number of participating teams in TREC Genomics tracks since 2003 (Hersh and Bhupatiraju, 2003). Like previous investigations, there have been some contradictory reports on the benefits of using query expansion. Results from some TREC Genomics ad hoc retrieval task participants (Hersh and Bhupatiraju, 2003; Abdou et al., 2005) pointed to the use of query expansion through multiple concepts: MeSH headings, substance name fields in MEDLINE, and species as accounting for improved performance. On the contrary, other participating teams (Hersh et al., 2003, 2004; Guo et al., 2004; Bacchin and Melucci, 2005) reported no or detrimental impact when expanding original queries with multiple sources including MeSH concepts.

The main differences between the work presented in this paper and the previous studies are (1) the retrieval system being used during evaluation, and (2) the method used for query expansion. All previous analyses based their results on either the vector space model (Salton and Buckley, 1988) or probabilistic models like the Okapi model (Robertson and Walker, 1994), both of which rank retrieved documents on the basis of their relevance to the queries. In contrast, PubMed is a Boolean search system and displays retrievals in reverse chronological order by default. Thus, we implemented our own search system that supports both Boolean and probabilistic text retrievals. As a result, not only can we evaluate the query expansion approach in the same setting as used by PubMed, but we can compare retrieval results with term weighting ranking strategies demonstrated in the previous studies.

The second distinction lies in the methods used for query expansion. The expansion terms in general come from two different sources. First, queries can be expanded using thesauri like the Unified Medical Language System (UMLS) Metathesaurus (Lindberg et al., 1993) via either computational approaches (Srinivasan, 1996; Aronson, 1996; Guo et al., 2004; Abdou et al., 2005) or manual mappings (Hersh et al., 2000). Second, expansion terms can be extracted using retrieval feedback (Srinivasan, 1996; Bacchin and Melucci, 2005). In our investigation, query expansion was executed through ATM, a unique mapping process of PubMed.

Finally, to the best of our knowledge, this is the first formal evaluation on the benefits of applying the query expansion technique to a daily operational search system as opposed to retrieval systems mainly designed and tested in research laboratories. Therefore, our analysis plays a critical role in the understanding of future technology development needs for PubMed, along with its mission to better fulfill the information needs of millions of PubMed users.

2 Methods

2.1 Text Collection

In this work, we used the TREC Genomics data in both 2006 (Hersh et al., 2006) and 2007 (Hersh et al., 2007), which consist of 28 topics and 36 topics (in the form of biological questions), respectively. In both years, a total of 162,048 full-text documents from Highwire Press1 were used. Except for 1,800 instances, most of the documents were successfully mapped to their corresponding PubMed Identifiers (PMIDs). Hereafter, we refer to the remaining set of 160,248 PMIDs as the TREC set in our study. The 36 topics (Topic IDs 200 to 235) in 2007 and 28 (Topic IDs 160 to 187) in 2006 were collected from bench biologists and represent actual information needs in biomedical research. For demonstration purposes, we show three topics 207, 229 and 231 from TREC 2007 in the list below:

  • <207> What toxicities are associated with etidronate?
  • <229> What signs or symptoms are caused by human parvovirus infection?
  • <231> What tumor types are found in zebrafish?

For each topic, a set of relevant documents from the TREC set were produced by the relevance judgments based on pooled results from team submissions. They are assumed as the ground truth or gold standard data in our investigation and referred to as the relevant set in the remainder of this paper.

2.2 Query Construction

For each TREC topic, a user query is required for retrieving relevant documents in PubMed. In order to produce unbiased (i.e. without human interference) user queries in a consistent manner, we chose to automatically select words from questions as queries on the assumption that real users would also intuitively create their queries based on the questions. Specifically, for each question, we first removed stop words (Wilbur and Sirotkin, 1991) from the question and enumerated all possible word combinations, each of which was then expanded by ATM and subsequently searched in PubMed. Next, for each word combination that retrieved a non-empty set of documents, we compared those retrieved documents to the ones in the relevant set and computed the standard IR measures (Hersh, 2003): recall, precision and F-measure (see also: Section 3.1). In the end, the query with the highest F-measure was selected.

Take the topic 229 for example, we first removed stop words what, or, are, by from the question. A total of 63 different user queries were then generated based on the remaining six words (signs, symptoms, caused, human, parvovirus, infection) and subsequently searched in PubMed to obtain a set of relevant documents. Three example queries are shown in Table 1, together with their corresponding IR measures after comparing the retrieved set with the relevant set. The query human parvovirus was finally chosen for further study because it yielded the highest F-measure. We processed all 64 topics in this way and were able to identify query terms for most of the topics except 9 instances (topics 164, 173, 177, 179, 180, 184, 185 in 2006 and topics 207 and 225 in 2007) where either no relevant documents were found in the gold standard or no query terms could be generated to represent meaningful user queries (i.e. using query expansion their F-measures are almost zero). Therefore, these nine topics were excluded from further analysis. The automatically generated queries for the remaining 55 TREC topics varied in length from a single word to a maximum of four words, with a mean of 2.5 words per query.

Table 1

Three potential user queries built automatically by selecting words from topic 229.

Word CombinationRecallPrecisionF-measure
human parvovirus0.470.630.54
human parvovirus infection0.390.850.53
parvovirus infection0.390.690.49

2.3 Preprocessing Search Field Tags After Translation

Once all of the 55 user queries were determined, their translated counterparts through PubMed’s Automatic Term Mapping were obtained. For instance, human parvovirus was translated to: ((“humans”[TIAB] NOT Medline[SB]) OR “humans”[MeSH Terms] OR human[Text Word]) AND (“parvovirus”[MeSH Terms] OR parvovirus[Text Word]). Translated queries may contain a variety of different search field tags. We filtered all but three tags: [MeSH Terms], [Text Words] and [All Fields] because they are (a) the most frequent tags seen after expansion (see also: Section 4.1); and (b) the major ones involving the use of MeSH. Therefore, they are the only important ones in our analysis for comparing results with respect to query expansion. For example, the function of the Subset search field ([SB]) is to restrict retrieval by topic, citation status and journal/citation subsets. Thus, the “human”[TIAB] NOT Medline[SB]” part of the translation for human parvovirus refers to searches for articles in PubMed but not yet indexed for MEDLINE (e.g. in-process citations which have been provided a PMID before it is indexed with MeSH). Since all of the citations in TREC 2006 and 2007 have already been indexed at the time of our experiments, this specific search tag is not applicable to our data. Thus, it was discarded. As a result, the remaining part: “humans”[MeSH Terms] OR human [Text Word]) AND (“parvovirus”[MeSH Terms] OR parvovirus[Text Word]) served as the input to retrieve citations in our system.

All three search tags make use of the MeSH fields. However, the way they use the MeSH field varies. Based on the online help documents2 and discussion with an expert medical librarian at the National Library of Medicine (Kathi Canese, personal communication), we treated the three search tags as follows in this work:

  1. [Text Words]: Match against all words in MEDLINE title, abstract, and MeSH terms (no automatic MeSH explosion).
  2. [MeSH Terms]: Match against the MeSH term and the more specific ones beneath it in the MeSH hierarchy (i.e. automated MeSH explosion).
  3. [All Fields]: Same as [Text Words].

For the search tags [Text Words] and [MeSH Terms], they were handled identically in our analysis as they would be in PubMed. For [All Fields], we limited search fields to text words only because other bibliographic information of a MEDLINE citation (e.g. author name) was not used by our search system. A word or phrase will only be tagged with [All Fields] if no match can be found in existing translation tables. For instance, if a user query is lysosomal, it will be translated to lysosomal[All Fields] since no match can be found for it in any of the translation tables in PubMed.

As can be seen, PubMed would only retrieve documents matching the search keywords in the title or abstracts if the ATM was not used. After applying ATM, PubMed expands its search ability by examining the MeSH field of indexed documents, which could result in returning many relevant documents that would otherwise be missed. For each retrieved document in the following analyses, we were able to characterize whether it was retrieved because of the use of MeSH or because of word match in the title or abstract. In addition, for the ones that were retrieved because of MeSH, we further classified them into three different groups: matching a MeSH term, matching a more specific MeSH term due to automatic MeSH explosion, or matching words of a MeSH term (See also: Section 4.2). In the following experiments, query expansion refers to the inclusion of all three types of MeSH related retrievals.

3 Results

3.1 Evaluation by Recall, Precision and F-measure

Many different measures for evaluating the performance of IR systems have been proposed and used in evaluating the effectiveness of query expansion (Hersh, 2003), two of which are selected in this study: F-measure and mean rank precision.

Retrieval performance was first assessed by recall (the proportion of relevant documents in the collection retrieved by the query) and precision (the proportion of relevant documents retrieved by the query). Then, F-measure was computed as the harmonic mean of recall and precision as follows:

F−measure=2×(precision×recall)precision+recall
(1)

As can be seen in Table 2, results using query expansion are better than those without the query expansion in terms of average F-measures in both TREC 2006 and 2007. Table 3 servers as a query-by-query F-measure analysis, describing the number of queries for which the query expansion technique achieved better, worse, or equal performance levels as done without the technique. The last column in Table 3 lists the results of the non-parametric statistical tests (Noreen, 1989; Wilbur, 1994) we performed through the pair-wise comparisons of all of the individual F-measures. Discrepancies between the results of 2006 and 2007 may be mostly attributed to the differences in topics and possibly the quality of relevance judgments. Our analysis shows that the relevant sets of TREC 2006 and 2007 consist of 1,448 and 2,478 documents, respectively. In contrast, about the same number of total relevant documents were retrieved (e.g., 661 in 2006 and 662 in 2007 when query expansion was used). As a result, performance (as shown in F-measures in Table 2) is considerably higher in 2006 than in 2007.

Table 2

Results of average F-measures on the 2006 and 2007 TREC topics.

TRECTopicsQuery ExpansionAverage F-measures
200621Yes0.406
No0.334
200734Yes0.264
No0.214

Table 3

Query-by-query F-measure analysis: with vs. without query expansion (confidence level = 0.05).

TRECbetterworseequalStatistical Significance
2006759No
20072086Yes

As the statistical test does not account for individual differences in magnitude, Figures 1 and ​and22 show that for topics in TREC 2006 and 2007, the enhancements in F-measure are generally far greater than are the degradations. The largest performance improvement was observed to be over 7 fold. Without using query expansion, the F-measure for Topic 200 (see Table 4) in TREC 2007 was 0.047. With query expansion, the F-measure increased substantially to 0.398.

An external file that holds a picture, illustration, etc. Object name is nihms73870f1.jpg

TREC 2006: Query-by-query differences in F-measures.

An external file that holds a picture, illustration, etc. Object name is nihms73870f2.jpg

TREC 2007: Query-by-query differences in F-measures.

Table 4

The two topics (176 and 200) that display the largest performance improvement and degradation after applying the query expansion technique.

Topic IDQuestionConstructed Query
176How does Sec61-mediated CFTR degradation contribute to cystic fibrosis?cftr degradation cystic fibrosis
200What serum proteins change expression in association with high disease activity in lupusproteins lupus

While we can see performance enhancement for most of the topics, query expansion resulted in degraded performance for some topics. The largest performance decrease happened to Topic 176 (see Table 4) where its F-measure dropped from 0.046 to 0.017 (a 72.6% decrease) after applying the query expansion technique. Further analysis showed that the total number of retrieved documents increased over 10 fold (from 21 to 221) due to an expansion of the term: degradation. PubMed translated the text word degradation to “metabolism”[Subheading] OR (“metabolism”[TIAB] NOT Medline[SB]) OR “metabolism”[MeSH Terms] OR degradation[Text Word]. For this topic, this expansion was harmful to the performance.

3.2 Evaluation by Mean Rank Precision

Our main concern in practice is to enhance retrieval performances among the top few returned documents which users are most likely to examine. To analyze this aspect, we introduce the second measurement for our experiments: Mean Rank Precision (MRP), which is the mean value of the rank precisions (precision at a given rank) computed for all queries. In this study, we chose the cut-of ranks to be 5, 10, and 20 as most of the retrievals occurred before 20 based on our own experience with PubMed users. Thus, the mean rank precision reveals more directly how a ranking strategy affects user retrieval effectiveness in realistic situations.

Note that PubMed returns citations in reverse chronological order by default, displaying more recent citations earlier. In accordance with this display order, we first sorted PMIDs numerically and then reversed the order. As a result, larger PMIDs would appear earlier in a ranked list.

Results in Table 5 show the mean average precisions at given ranks (5, 10, and 20) sorted by the reverse time order, suggesting the existing query expansion feature in PubMed does not always yield improved performance. In fact, the difference between results with and without query expansion is as small as a single relevant document for the TREC 2006 data. Additionally, results from TREC 2007 demonstrate that using such a technique could sometimes be harmful to system performance. This can happen because query expansion is generally considered as a recall-favoring technique (as opposed to precision-favoring) that often results in a large number of additional retrievals. For instance, the number of retrieved documents increased from 1,174 to 3,163 when the query expansion technique was employed in TREC 2007. A much smaller increase in total number of retrievals (from 1,463 to 1,899) was seen in TREC 2006. This may account for the differences between 2006 and 2007, as shown in Table 5.

Table 5

Results of the 2006 and 2007 TREC topics with and without query expansion as measured by mean rank precision. Retrieved documents are sorted in reverse chronological order.

TRECQuery ExpansionP@5P@10P@20
2006Yes0.5560.5400.540
No0.5230.5170.524
2007Yes0.3850.4110.413
No0.4340.4540.452

3.3 Results Under Different Ranking Strategy

We computed a new set of mean rank precisions when sorting retrieved documents under a different ranking strategy. Both previous research (Salton, 1991; Salton and Buckley, 1988) and in particular our own experience with PubMed (Lu et al., 2008) suggest that ranking by relevance can result in better retrieval performance. Thus, we computed TF-IDF scores for retrieved documents (Kim and Wilbur, 2005; Lu et al., 2008) and then ranked them based on these scores. A document with a higher TF-IDF score is returned earlier in a list.

The same conclusions can be drawn with respect to the effect of using the query expansion technique based on the similar results shown in Tables 5 and ​and6.6. Superior results in mean rank precisions are seen in Table 6 compared with their counterparts in Table 5 due to a better ranking strategy. Together with experiments presented in Section 3.2, we confirmed an earlier expectation in (Hersh et al., 2000) where the authors believed that the difference in retrieval models (i.e. Boolean versus statistical-based) should not yield different results when evaluating query expansion techniques.

Table 6

Results of the 2006 and 2007 TREC topics with and without query expansion as measured by mean rank precision. Retrieved documents are sorted under the TF-IDF weighting scheme.

TRECQuery ExpansionP@5P@10P@20
2006Yes0.6220.6160.621
No0.6130.5920.584
2007Yes0.5380.5250.503
No0.5660.5480.519

4 Discussion

4.1 PubMed’s Automatic Term Mapping Process

For all the queries that we used in our work, no search tags were initially assigned. After automatic translation by PubMed’s ATM, the majority of the queries were then associated with either [MeSH Terms], [Text Words] or [All Fields] as shown in Table 7. Note that there could be multiple search tags associated with a single query after translation. Take the previous parvovirus example, its translation included both [MeSH Terms] and [Text Word] (i.e. “parvovirus”[MeSH Terms] OR parvovirus[Text Word]).

Table 7

Results of PubMed’s Automatic Term Mapping. The last column shows the number of queries assigned with at least one of the three search tags listed in the columns 3–5. Note that a single query can be assigned with multiple search tags (e.g. parvovirus[MeSH Terms] OR parvovirus[Text Words]). Thus for each row, the sum of the numbers in columns 3 to 5 is greater than the number in the second column (i.e. Queries). Abbreviations: [MeSH], [MeSH Terms]; [Text], [Text Words]; [ALL], [All Fields]

CollectionQueries[MeSH][Text][ALL][MeSH] OR [Text] OR [ALL]
TREC 20062117171521
TREC 20073428282334
Oct 17, 20052.4M1.5M1.4M1.4M2.0M

In order to show that the automatically generated queries are representative, we processed one day’s worth of PubMed query data3. We discarded users that issued over 50 queries/24 hours as they are likely to represent programmatic searches. For the remaining 2,657,315 queries, we obtained their corresponding PubMed translations through the Entrez Programming Utilities (eutilities) (Geer and Sayers, 2003). Results in the last row of Table 7 show that over 90% (2.4M/2.65M) of user queries are initially search tag free, over 80% (2.0M/2.4M) of which are expanded with at least one of the three tags studied in this work, suggesting that query expansion involving MeSH is a dominant phenomenon under real world circumstances. Other major search fields shown after translation include the ones in the following list. However, none of these search tags are related to MeSH terms. Therefore, they are beyond the scope of discussion in this paper.

  1. TIAB NOT Medline[SB] (776,942): searching for terms in the title and abstract of articles in the databases that are not indexed for MEDLINE.
  2. Author (380,628): searching for author names in the article.
  3. Substance Name (89,510): searching for the name of a chemical in the article.

4.2 Different MeSH Matching Mechanisms

As we pointed out earlier in Section 2.3, the use of MeSH terms can be classified into two distinctive strategies:

  1. If a query term is tagged with [MeSH Terms], the system will retrieve all documents indexed with this particular MeSH term as well as indexed with more specific MeSH terms beneath it in the MeSH hierarchy.
  2. If a query term is tagged with either [Text Words] or [All Fields], our system will retrieve all documents indexed with one or more MeSH terms that include the query term. For instance, if a query term is nervous system and tagged with [Text Words] as ‘nervous system’[Text Words], any documents indexed with the MeSH Term Alcohol Abuse, Nervous System will be retrieved in that ‘nervous system’ is part of this MeSH term.

To characterize the individual contribution of each of the two classes to the overall improved performance, we performed two additional experiments where only one of the two matching mechanisms was exploited. Results are compared to each other in Table 8, and are also compared to the cases when none or both matching mechanisms were used. Results in Table 8 show that directly matching the MeSH term and its more specific terms is the major contribution for performance improvement in TREC 2007 while matching words in MeSH terms contributed significantly in TREC 2006. Therefore, we are unable to draw a conclusion on which one is more important for improving performance based on these experiments.

Table 8

Performance (in F-measure) comparisons using different MeSH matching strategies. Experiments were performed on both TREC 2006 and 2007 data.

TRECNone[MeSH Terms][Text Words] or [All Fields]Both
20060.3340.3460.4070.406
20070.2140.2600.2360.264

As a separate effort, we performed an additional experiment where we turned of the automatic MeSH explosion feature (i.e. more specific MeSH terms are no longer searchable) in the context of evaluating this specific feature. When compared to the results in the third column of Table 8, we see a performance drop from 0.346 to 0.334 in TREC 2006, and from 0.260 to 0.212 in TREC 2007, respectively. This leads us to conclude that the MeSH automatic explosion feature plays a critical role in the first type of MeSH usage.

4.3 Comparison with PubMed Output

The main objective of this work is to discover useful features to improve PubMed search performance. In spite of the fact that our analysis was not based on output directly from PubMed, the retrieved document set by our custom search system was highly consistent with the actual PubMed output as shown in Figure 3. The F-measures are almost identical for the systems. The slight discrepancies in performance come from two sources in our analysis: (a) we limited our searches with only three search field tags; and (b) for the search tag [all fields], our search was limited to text words in title, abstract, and MeSH terms. Any other fields (e.g. author names) were not considered in our computation. The minor performance increase when using the custom-built system suggests (a) the three search tags that we selected to evaluate are critical for PubMed; and (b) that perhaps other search field tags could not further help to retrieve relevant documents for our query set.

An external file that holds a picture, illustration, etc. Object name is nihms73870f3.jpg

The comparisons of F-measures based on outputs from PubMed vs. from our custom search system.

5 Conclusions and Future Work

Based on the results of our large-scale analysis comprised of the 55 real biological questions and independently judged relevant documents, we conclude that query expansion using MeSH through PubMed’s Automatic Term Mapping process can generally result in retrieving more relevant documents. However, this type of improvement may not prove useful for those users looking only at the top ranked returned documents (e.g. the first 20 returned documents).

The comparison results are useful in suggesting changes in PubMed (e.g. the display order). However, as we pointed out earlier, this work makes two assumptions. One assumption is that the automatically generated queries represent real user queries. The other assumption is that the documents in the relevant set are the ground truth and there are no more relevant documents in the TREC set. Our future research goal is to address both issues by involving human experts in evaluations.

Acknowledgments

This research was funded by the Intramural Research Program of NIH, National Library of Medicine. The authors are grateful to the TREC organizers for their efforts in producing and making the text collection and relevance judgments publicly available.

References

  • Abdou S, Ruck P, Savoy J. Evaluation of stemming, query expansion and manual indexing approches for the genomics track. Proceedings of the Fourteenth Text REtrieval Conference 2005 [Google Scholar]
  • Aronson AR. The effect of textual variation on concept based information retrieval. Proc AMIA Annu Fall Symp. 1996:373–377. [PMC free article] [PubMed] [Google Scholar]
  • Aronson AR, Rindflesch TC. Query expansion using the umls metathesaurus. Proc AMIA Annu Fall Symp. 1997:485–489. [PMC free article] [PubMed] [Google Scholar]
  • Bacchin M, Melucci M. Symbol-based query expansion experiments at TREC 2005 genomics track. Proceedings of the Fourteenth Text REtrieval Conference 2005 [Google Scholar]
  • Carlin BG. Pubmed automatic term mapping. J Med Libr Assoc. 2004;92(2):168. [PMC free article] [PubMed] [Google Scholar]
  • Gault LV, Shultz M, Davies KJ. Variations in medical subject headings (mesh) mapping: from the natural language of patron terms to the controlled vocabulary of mapped lists. J Med Libr Assoc. 2002;90(2):173–180. [PMC free article] [PubMed] [Google Scholar]
  • Geer RC, Sayers EW. Entrez: making use of its power. Brief Bioinform. 2003;4(2):179–184. [PubMed] [Google Scholar]
  • Guo Y, Harkema H, Gaizauskas R. Sheffield university and the TREC 2004 gemonics track: query expansion using symomymous terms. Proceedings of the Thirteenth Text REtrieval Conference 2004 [Google Scholar]
  • Hersh W. Information Retrieval: A Health and Biomedical Perspective. 2 Springer-Verlag; 2003. [Google Scholar]
  • Hersh W, Bhupatiraju R, Ross L, Johnson P, Cohen A, Kraemer D. TREC 2004 Genomics Track Overview. Proceedings of the Thirteenth Text REtrieval Conference 2004 [Google Scholar]
  • Hersh W, Bhupatiraju RT. The twelfth Text Retrieval Conference, TREC 2003. Vol. 14. National Institute of Standards and Technology; 2003. TREC Genomics track overview; 23 pp. [Google Scholar]
  • Hersh W, Bhupatiraju RT, Price S. Phrases, boosting, and query expansion using external knowledge resources for genomic information retrieval. Proceedings of The Twelfth Text REtrieval Conference 2003 [Google Scholar]
  • Hersh W, Cohen AM, Roberts P, Rekapalli HK. TREC 2006 Genomics Track Overview. Proceedings of the Fifteenth Text REtrieval Conference 2006 [Google Scholar]
  • Hersh W, Cohen AM, Roberts P, Rekapalli HK. TREC 2007 Genomics Track Overview. Proceedings of the Sixteenth Text REtrieval Conference 2007 [Google Scholar]
  • Hersh W, Price S, Donohoe L. Assessing thesaurus-based query expansion using the umls metathesaurus. Proc AMIA Symp. 2000:344–348. [PMC free article] [PubMed] [Google Scholar]
  • Herskovic JR, Tanaka LY, Hersh W, Bernstam EV. A day in the life of PubMed: analysis of a typical day’s query log. J Am Med Inform Assoc. 2007;14(2):212–220. [PMC free article] [PubMed] [Google Scholar]
  • Kim W, Wilbur WJ. A strategy for assigning new concepts in the MEDLINE database. AMIA Annu Symp Proc. 2005:395–399. [PMC free article] [PubMed] [Google Scholar]
  • Lindberg D, Humphreys B, Mccray A. The unified medical language system. Methods Inf Med. 1993;32(4):281–291. [PMC free article] [PubMed] [Google Scholar]
  • Lu Z, Kim W, Wilbur WJ. Evaluating relevance ranking strategies for medline retrieval. in press. [PMC free article] [PubMed] [Google Scholar]
  • Mitra M, Singhal A, Buckley C. SIGIR ‘98: Proceedings of the 21st annual international ACM SIGIR conference on Research and development in information retrieval. New York, NY, USA: ACM; 1998. Improving automatic query expansion; pp. 206–214. [Google Scholar]
  • Noreen E. Computer Intensive Methods for Testing Hypotheses. John Wiley and Sons; 1989. [Google Scholar]
  • Robertson SE, Walker S. SIGIR ‘94: Proceedings of the 17th annual international ACM SIGIR conference on Research and development in information retrieval. New York, NY, USA: Springer-Verlag New York, Inc.; 1994. Some simple effective approximations to the 2-poisson model for probabilistic weighted retrieval; pp. 232–241. [Google Scholar]
  • Salton G. Developments in automatic text retrieval. Science. 1991;253(5023):974. [PubMed] [Google Scholar]
  • Salton G, Buckley C. Term weighting approaches in automatic text retrieval. Information Processing & Management. 1988;24:513–523. [Google Scholar]
  • Salton G, Buckley C. Readings in Information Retrieval. Morgan Kaufmann Publishers Inc.; San Francisco, CA, USA: 1997. Improving retrieval performance by relevance feedback; pp. 355–364. [Google Scholar]
  • Shultz M. Mapping of medical acronyms and initialisms to medical subject headings (mesh) across selected systems. J Med Libr Assoc. 2006;94(4):410–414. [PMC free article] [PubMed] [Google Scholar]
  • Smith AM. An examination of pubmed’s ability to disambiguate subject queries and journal title queries. J Med Libr Assoc. 2004;92(1):97–100. [PMC free article] [PubMed] [Google Scholar]
  • Srinivasan P. Query expansion and MEDLINE. Inf Process Manage. 1996;32(4):431–443. [Google Scholar]
  • Wilbur W. Non-parametric significance tests of retrieval performance comparisons. Journal of Information Science. 1994;20(4):270–284. [Google Scholar]
  • Wilbur WJ, Sirotkin K. The automatic identification of stop words. Journal of Information Science. 1992;18(1992):45–55. [Google Scholar]
-