Laypeople who are not medical expert may formulate short query using words from their discharge summaries or long query that explain their health conditions. The di erent query styles should be treated with di erent query expansion mechanisms. This work is an adaptive query expansion based on the coherence among query terms. To provide users with more readability documents, the document complexity is analiyzed on the word lenght and is used as a re-ranking method. The baseline retrieves using Lucene v.4.6 with default con guration. Overall retrieval performance of the baseline is better than the adaptive query expansion (P@10, MAP, and NDCG). Since the complexity of document uses only length of word, the readability performance after re-ranking is poorly performing.
Health information is now available on the web and easy to access by
nonprofessionals who associate to health care system and called laypeople. The
2015 CLEF eHealth [
From the example, we assume that the medical term(s) should be presented in the relevant web page(s) and may be absented from the query. There is a gap between query and relevant web pages. Query expansion (QE) technique is commonly used to handle this situation. Since there are many approaches for doing expansion, we examine the e ectiveness of di erent query expansion approaches for di erent query styles. This work examines three local-based and one globalbased query expansion mechanisms. The local-based QE uses documents from the initial query retrieval and the global-based QE uses all documents from the collection for term selection.
The degree of coherence among query terms is used to estimate performance of the query retrieval, called QPPpair. We assume that the query with performing well should be expanded with terms from the small-size of top-ranked documents or no need to expand any term. Because these documents are assumed to be the relevance. The query with performing poorly should be expanded with terms from the entire collection or from an external source.
In addtion to retrieve more useful web pages, the results should be easy to read for laypeople. According to the readability requirement, factors used to evaluate the readability [4{6] are sentence length, number of sentences, the number of syllables or number of characters per word. Since our work treats the collection as a bag of words therefore we assume that long word is a complex word and then the complexity of document can be estimated by the frequency of the long words.
In this paper, we use the following techniques to achieve the CLEF 2015
eHealth Task 2 [
Query performance prediction (QPP) [
The survey of pre-retireval QPP predictors [
We focus on the term relatedness of QPP predictor because we assume that
query from laypeople contains more terms, as seen from the query example,
therefore relation among query terms may be re ect to the query performance.
The examples of the term relatedness-based QPP predictors [
Query expansion (QE) is a technique which new terms are added to an original
query and assumes that the new term(s) should be enlarging the document range
for second-pass retrieval. There are two major factors that in uence on the
effectiveness of QE technique; term selection method and reweighting method[
There are three steps of concept selection described in[
Readability of health information is a current issue as seen in many current
works examined the readability of web pages. Each work focuses on di erent
health condition such as epilepsy[
The alternative way to measure readability is using language model[
Given an assumption that characteristic of the query of CLEF 2015 task[
Our QPP method is called the QP P pair, which measures how many pairs of query terms existing in the collection (called P airExist) respect to all possible pairs in the query (called AllP airs). The formula for measuring query coherence is de ned as Equation 1.
QP P pair(Q) = p2QP airExist(p)
AllP air(Q) (1 , if pair exist in collection
0 , otherwise P airExist(p) =
AllP air(Q) = p2QP air(p) and pP RF (xi) = TPFRPFRFsi(zxei) > c1
pP RF (xi) pcollection(xi) In the preliminary experimental, we have examined the QP P pair of query set in three collections; CLEF eHealth 2014, TREC Genomics 2004, and OHSUMED. The results show that the Genomics collection is more consistent with the CLEF collection so that we use the Genomics collection as external source of query expansion. We use the average of the QP P pair value of three collections to determine query expansion mechanism. There are ve query expansion mechanisms; no query expansion, expanded with top-small PRF set, expanded with top-medium PRF of both local and external collection, expanded with top-large PRF of both local and external collection, and expanded with co-occurrence terms in global collection.
documents contain at least one query term. The top-ranked documents should
contain more query terms and assumed that indirectly related to a whole query.
In this PRF-QE method, we select concept terms from terms within the PRF
set based on the work [
Given a set, P RF = (x1; x2; :::; xn) where xi is a term in the PRF set. There are two conditions for nding concept terms; Equations 4 and Equation 5, respectively. The rst condition measures how importance of term respect to the PRF set based on the distribution of term. The second condition measures the importance of term respect to entire collection.
Now we have a number of concepts that dervied from Equation 4-7. Then nally select smaller number of concepts for expansion based on collection frequency and document frequency value as the following equation.
ConceptSpecif icity(c) = log(CF (c))
DF (c) where T FP RF (x) is a total frequency of term x in the PRF set and P RF size is a number of retrieved documents in PRF set. The pcollection(x) is distribution of term x in the collection and measured by total collection frequency of term x respect to total documents in the collection.
To lter important concept, the entropy of a concept in the retrieved set is derived from the following equation.
G(xi) =
pP RF (xi)log(pP RF (xi))
To select most related concepts, the conditional probability of them is used in the following conditions.
pP RF (xijxj ) > sa and pP RF (xj jxi) > sb (7) (6) (8) where CF (c) is the collection frequency of concept c and DF (c) is document frequency that concept c appeared.
documents that retrieved from external collection are also indirectly related with
original query. Some queries are improved by this method as reported in work[
The method to select concepts is based on the PRF-QE method that described in the previous section. In addtion, threshold values of external collection are less than the CLEF collection, called target collection. Because the query is generated for the target collection therefore retreived size and related terms of the external are likely to be less than the target one.
for the query with lowest QP P pair value. This method is based on co-occurrence assumption that two terms frequently occur together in the same context, are likely to be related. The pair is two terms (ti; tj ) within a small window length. This work xes window length at 7 words for the co-occurrence condition. The set of pairs is P airSet = (p1; p2; :::; pk). Each pair pi has the collection frequency ci called cotimes and this value is used to select the most related of a query term. The cotimes set is CoT ime = (< p1; c1 >; < p2; c2 >; :::; < pk; ck >).
The steps of nding related terms from the co-occurrence terms are following. 1. Ascending sort the document frequency of query terms,
DF sorted = (q1; q2; :::; qm). 2. Start from qi, lookup a corresponding pair(s) in the P airSet then take them to the CandidateP air set.
{ if qi does not exits in the P airSet set, increment i. 3. For each candidate pair, lookup the corresponding cotimes in the CoT ime set. 4. Descending sort candidate pairs according to the cotimes value, { if the pair pi consists of one query term as (qi; qi), then cotimes ci of the pair pi is used to select other term(s), { otherwise, select top 5 terms from the sorted CandidateP air set. 3.3
To return most readability documents to user, we measure the complexity of document based on word length with the assumption that long word is the complex word. The document contains more complex words may be high complexity and uneasy to understand.
DocComplex(d) =
(9)
After retreival we re-rank each 200 documents in the results set with the following equation.
ComplexV al(d) = 1
(10) 4
We submit four runs where Run2 examines both query analysis method and query expansion method. Run3 and Run4 examine re-ranking method based on complexity of retrieved document.
Run1. The baseline retrieves with original title query using Lucene 4.6 [
StandardAnalyzer con guration and using Lucene's default VSM similarity. Run2. The adaptive query expansion is using the QP P pair to determine the expansion mechanism.
Run3. The re-ranking version of Run1 based on complexity of document. Run4. The re-ranking version of Run2 based on complexity of document.
QP P pair as described in Equation 1. Then each query is assigned with one of ve group and detail as the shown in Table 1. We note that values of QP P pair in each group are from human estimation. One 0.5 <= x No expansion Two 0.2 <= x < 0.5 PRF-QE with top 100 documents Three 0.1 <= x < 0.2 CLEF-GPRF-QE with top 500 documents Four 0 < x < 0.1 CLEF-GPRF-QE with top 1000 documents Five 0 = x Global-QE This task consists of two relevant judgments. The rst one is traditional evaluation meaurement that shows retrieval performance such as P@10 and NDCG. The second measurement is readability-biased evaluation. The comparison of the P@10 with other teams shown in Fig1,2,3, and 4. The map, P@10,NDCG cut 5 and NDCG cut 10 of our four runs are shown in Table 4. The readability performance of all runs are shown in Table 5.
Query Coherence Analysis. The query with more existing pairs in the corpus does not guarantee that the result set will be more relevant as seen that queries Q7,Q15,Q37, and Q51 have the high QppP air value but the P@10 value of these queries is not better than the median retrieval performance. Because Lucene's default similarity function computes weight of individual term not pair of terms.
Using the QPPpair to determine expansion mechanism is still better than
expand all queries with the same mechanism. We believe that there are
alternative ways to take advantage from these existing pairs such as reform the query
to add more weight on the pair or search query with phrase option.
Query Expansion Mechanism. The work [
For expansion based on PRF set, the candidate terms sensitive to the tuning parameter. Even the queries are in the same group, the candidate terms for some query are likely to be related while others are drifting.
For global expansion method, the candidate terms are derived from the cooccurrence terms of the most speci city term in the query. This method is working for some query, for example, Q61. \ ngernail bruises", the co-occurrence terms of " ngernail" are useful for expansion. On the other side, for example, Q59. \heavy and squeaky breath", the co-occurrence terms of the \squeaky" are not useful for expansion. This is the weakness of this method.
Document Complexity Method. The complexity document score is using only word length and then re-ranking retrieved documents. This method gets the worse results. Therefore using more features of the document is necessary to estimate the complexity.