A Study on Query Expansion with MeSH Terms and Elasticsearch. IMS Unipd at CLEF eHealth Task 3 Giorgio Maria Di Nunzio and Alexandru Moldovan Dept. of Information Engineering – University of Padua giorgiomaria.dinunzio@unipd.it,alexandru.moldovan@studenti.unipd.it Abstract. In this paper, we describe the first participation of the Infor- mation Management Systems (IMS) group at CLEF eHealth 2018 Task 3, Consumer Health Search Task. In particular, we participated in the subtask IRTask 1: Ad-hoc Search which is a standard ad-hoc search task, aiming at retrieving information relevant to people seeking health advice on the web. The goal of our work is to evaluate 1) different query expansion strategies based on the recognition of Medical Subject Headings (MeSH) terms present in the original query; 2) different approaches to combine multiple ranking lists given the query expansions. We used Elasticsearch as search engine and the indexes provided by the organizers of this task. 1 Introduction In this paper, we report the experimental results of our first participation to the CLEF eHealth Lab Task 3 [4, 3]: “Consumer Health Search”. This task in- vestigates the problem of retrieving documents to support information needs of health consumers that are confronted with a health problem. This work is part of a Master Degree thesis in Computer Engineering where the main goal is to test the effectiveness of some variants of query expansion approaches based on the recognition of MeSH terms present in the original query. The contribution of our experiments to this task can be summarized as fol- lows: – A study of several query expansion approaches that takes into account the relationships between MeSH terms [6]; – An evaluation of different document scoring strategies given the multiple ranking list produced by the query expansions [1, 6]. The remainder of the paper will introduce the methodology and a brief sum- mary of the experimental settings that we used in order to create the runs that we submitted for the task. 2 Methodology In this section, we describe the query expansion approaches [6] as well as the document scoring strategies [1] that we used to create the expanded queries and the ranked lists. 2.1 Query expansion approaches For the experiments of the query expansion approach, we used the English ver- sion of the information need. Before running the automatic expansion algorithms, we performed a manual check in order to correct spelling errors. For example, for topic 189001, the term gonorrhea is misspelt as gonhrrea. Identification of MeSH terms After the manual cleaning process, we use the MeshOnDemand1 API to identify the MeSH terms present in the query. For example, for topic 188001 “caffeine high blood pressure” we obtain the following MeSH terms – Caffeine2 – Hypertension3 plus one additional term – Blood pressure4 Finding related MeSH terms For each MeSH term found in the previous step, we use the MeSHRDF5 database to look for semantically related (MeSH) terms. See for example Figure 1 that shows a part of the structure of the tree of relations among concepts related to the MeSH term Caffeine. We choose a subset of all the possible relations (predicates) between terms in the MeSHRDF database6 , and we use this subset of predicates for query expansion in the following way: – Baseline: the original query is used without any additional term. – Simple Expansion (SE): given a MeSH term identified in the first step, all the MesH entries related to that term are kept, except for the predicates ‘meshv:Qualifier’, ‘meshv:seeAlso’, ‘meshv:broader’ e ‘meshv:broaderDescriptor’. Then we re-apply just once (not recursively) a SE for each ‘child’ node. – SE broader: like SE, in addition, only for the original MeSH terms (those that appear in the query) we expand the MesH terms by adding the predicates ‘meshv:broader’ e ‘meshv:broaderDescriptor’. 1 https://meshb.nlm.nih.gov/MeSHonDemand 2 https://meshb.nlm.nih.gov/record/ui?name=Caffeine 3 https://meshb.nlm.nih.gov/record/ui?name=Hypertension 4 https://meshb.nlm.nih.gov/record/ui?name=Blood%Pressure 5 https://id.nlm.nih.gov/mesh/ 6 https://hhs.github.io/meshrdf/predicates Fig. 1. Partial view of the tree structure of the MeSH term Caffeine. – SE also: like SE broader, instead of adding the predicates ‘broader’, we add the MesH terms related with the predicate ‘meshv:seeAlso’. – SE broader also: a combination of SE broader and SE also. – SE child broader: first, apply SE, then for each ‘child’ node search for ‘par- ents’ different from the original MeSH terms. – SE recursive down: like SE, for each ‘child’ we recursively apply SE until leaf nodes are found (no more recursions). – All in one: all the approaches at once. At the end of a query expansion process, we have three main data objects: – the original query q; – a vector m = (m1 , m2 , . . . , mn ) of MeSH terms associated with the original query; – a list t of expanded terms of n elements where each element ti is another vector of terms resulting from the iteration of the expansion approach, ti = (ti1 , ti2 , . . . , tij ). Given these three objects, we create a set of expanded queries that will be used to create a number of ranked lists, as explained in the following subsections. Building expanded query Give the vector of MesH terms m and the list of expanded terms t, we create a set of expanded queries by means of the following procedure: – For each term term mi of the vector of MeSH terms associated with the original query, – we substitute mi with one of the terms in ti , for example ti1 , – then, we build the expanded query by joiningSthe original query with the new vector m∗i = (m1 , . . . , ti1 , . . . , mn ), q ∗ = q m∗i . Therefore, at the end of the process we generate a set V of vectors of expanded queries where the cardinality |V | is the sum of all the elements in the vectors of the list t. For each vector of vk ∈ V we obtain a list lk of ranked documents. Merging ranked lists We use different approaches to merge the |V | ranked list into a single list based on the combination of scores of the documents. – Average: given a document present in one or more lists, the scores associated to the document are averaged. Then the documents are ordered in decreasing order on the basis of this new score. – Sum: given a document present in one or more lists, we sum the scores associated to the documents. – Normalized sum: like Sum but the sum of the scores are normalized by the highest score in the ranked lists. – Round robin: for each rank r, we take the document of each list lk at r and add it to the new ranked list if it has not already been seen. 3 Experiments For the experiments, we used the Elasticsearch search engine7 and the indexes provided by the organizers of the task. We used the BM25 ranking function with default parameters. Given the constraints on the number of runs, four in total, that could be submitted to the task, we submitted one baseline run and three query expansion variant with the same scoring approach. We will evaluate all the other combina- tions as soon as the qrels will be made available. In particular, we submitted the following runs that use the Sum (of the scores) as the document scoring approach: – baseline.exp, a baseline run (plain BM25 with no expansion), – sum score simple.exp, simple expansion, – sum score broader also, simple expansion plus the two predicates broader and also, – sum score recursive.exp, a recursive down approach. In Table ??, we show the number of query variants that are generated per topic by three approaches. 7 https://www.elastic.co/products/elasticsearch 4 Final remarks and Future Work The aim of our first participation to the CLEF eHealth Task 3 was to test the effectiveness of different query expansion approaches that use the MeSH term RDF graph as well as different merging approaches of the ranking lists produced by the query expansion approach. As future work, we will study the combination of the MesH term expansion with the help of medical terminological records [5] for technologically assisted systematic reviews [2]. References 1. Nick J. Belkin, Paul Kantor, Edward A. Fox, and Joseph A. Shaw. Combining the evidence of multiple query representations for information retrieval. Informa- tion Processing & Management, 31(3):431 – 448, 1995. The Second Text Retrieval Conference (TREC-2). 2. Giorgio Maria Di Nunzio. A study of an automatic stopping strategy for tech- nologically assisted medical reviews. In Advances in Information Retrieval - 40th European Conference on IR Research, ECIR 2018, Grenoble, France, March 26-29, 2018, Proceedings, pages 672–677, 2018. 3. Jimmy, Guido Zuccon, Joao Palotti, Lorraine Goeuriot, and Liadh. Kelly, editors. Overview of the CLEF 2018 Consumer Health Search Task. CLEF 2018 Evaluation Labs and Workshop: Online Working Notes. CEUR-WS, September 2018. 4. Hanna Suominen, Liadh Kelly, Lorraine Goeuriot, Evangelos Kanoulas, Leif Az- zopardi, Rene Spijker, Dan Li, Aurélie Névéol, Lionel Ramadier, Aude Robert, Joao Palotti, Jimmy, and Guido Zuccon, editors. Overview of the CLEF eHealth Eval- uation Lab 2018. CLEF 2018 - 8th Conference and Labs of the Evaluation Forum, volume Lecture Notes in Computer Science (LNCS). Springer, September 2018. 5. Federica Vezzani, Giorgio Maria Di Nunzio, and Geneviève Henrot. Trimed: A multilingual terminological database. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation, LREC 2018, Miyazaki, Japan, May 7-12, 2018., 2018. 6. Theodore B Wright, David Ball, and William Hersh. Query expansion using mesh terms for dataset retrieval: Ohsu at the biocaddie 2016 dataset retrieval challenge. Database, 2017:bax065, 2017. Table 1. Number of query variants per approach topic simple broader also recursive 151001 290 412 524 152001 310 330 1402 153001 71 117 127 154001 90 128 139 155001 74 123 184 156001 281 406 547 157001 222 226 671 158001 85 152 85 159001 298 404 1260 160001 84 191 119 161001 104 134 129 162001 42 118 42 163001 168 286 216 164001 38 60 231 165001 181 296 439 166001 253 323 436 167001 120 126 312 168001 241 339 241 169001 16 80 16 170001 144 169 169 171001 22 61 22 172001 96 135 96 173001 15 20 15 174001 112 186 166 175001 82 124 102 176001 42 101 42 177001 76 169 92 178001 236 336 388 179001 100 179 146 180001 83 105 89 181001 28 75 28 182001 257 279 423 183001 85 157 116 184001 117 166 117 185001 96 175 96 186001 245 308 562 187001 85 204 169 188001 125 164 239 189001 18 38 18 190001 377 428 678 191001 6 19 6 192001 169 245 270 193001 53 88 104 194001 95 147 95 195001 47 62 47 196001 13 104 13 197001 109 201 185 198001 51 78 51 199001 131 209 131 200001 129 185 202 average 124.24 183.36 239.94