In this paper, we describe our participation in the CLEF eHealth 2016 task 3: Patient-Centred Information Retrieval focusing on the clinical web documents based on user queries in the health forum. In our participation, we submitted three runs in ad-hoc search and two runs in query variation search subtasks. In ad-hoc search, the main challenge is to retrieve high quality clinical documents based on user query. For ad-hoc search, we employ multiple features based unsupervised reranking method to the documents retrieved by a baseline system. During the query variation search, the main challenge is to generate a ranked list of documents covering the di erent variations of the query. To tackle the query variation problem, rst we formulate a query and a set of information needs from the query variation. Then, we re-rank the documents retrieved for the formulated query by focusing on the set of information needs.
Users often seek health related information and issue their information needs in the Web search engine. Unfortunately, the documents on the Web are mostly lowquality and contain a lot of spam information. Therefore, laypeople are usually unsuccessful in getting their health related answers. CLEF eHealth Evaluation Lab [1{5] has been organizing a task for the last couple of years to help the laypeople to obtain health related information. In 2016, task 3 includes three subtasks such as ad-hoc search, query variation search, and multilingual search. We participated in the ad-hoc search and query variation search. In particular, we desire to evaluate the following research question in both of the subtasks: 1. Is the spam identi cation method su ciently reliable for the ad-hoc retrieval task? 2. Is multiple features based ranking method enough to estimate the topical relevance of the query and documents? 3. Is the diversity based ranking successfully handling query variation?
We submitted a total ve runs where three runs in an ad-hoc retrieval and two
runs in query variation. We make use of Clueweb12-B13 [
To prepare three runs in ad-hoc search, we apply some common procedures. Given a query, rst, we tokenize the query and format it for ad-hoc retrieval with the Indri Search engine. Second, we retrieve at most 1000 documents from the Clueweb12-B13 corpus using a query likelihood model with the Dirichlet smoothing model as baseline retrieval. Third, document with a spam score less than 70 is ltered out of the retrieved documents. The three runs are described as follows:
Run 1 (KDEIM EN Run1): following the common procedures described above, we re-rank the documents by fusing the page rank and documents' baseline (language model) scores and take the top 200 documents.
Run 2 (KDEIM EN Run2): following the common procedures stated above,
we extract multiple query-independent and query-dependent features including
reciprocal rank, topic cohesiveness [
Run 3 (KDEIM EN Run3): following the common procedures stated above, in comparison to the previous two runs, we tokenize the query and format it for expert retrieval, where we consider document title, header, body, and anchor text. Then, we re-rank the documents by combining page rank and documents' baseline (language model) score, and take the top 200 documents. 2.2
To prepare two runs in query variation search, we apply some common procedures. First, we tokenize all the six query variations and formulate a vector space model representation of the query from the query variations. We also consider all the query variations as information needs (sub-queries) of the users. Second, we retrieve at most 1000 documents from the Clueweb12-B13 corpus based on the formulated query as baseline retrieval. Third, document with a spam score less than 70 is ltered out of the retrieved documents.
Run 1 (KDEIM EN Run1): following the common procedures described above, we re-rank the documents using page rank and documents' baseline (language model) scores as a relevance based ranking. By considering the query variations as sub-queries (aka, information needs), we employ an explicit diversi cation algorithm [14] and take the top 100 documents.
Run 2 (KDEIM EN Run2): following the common procedures stated above,
we re-rank the documents using multiple query-independent and query-dependent
features including page rank, reciprocal rank, topic cohesiveness [
In this paper, we described the participation of KDEIR at CLEF eHealth 2016 Patient-Centred Information Retrieval task, where we proposed our approaches to ad-hoc search and query variations of clinical documents.
This research was partially supported by the HORI FOUNDATION of JAPAN, Grant-in-Aid C114. 10. Salton, G., Buckley, C.: Term-weighting approaches in automatic text retrieval.
Information processing & management 24(5) (1988) 513{523 11. Robertson, S., Zaragoza, H.: The probabilistic relevance framework: BM25 and beyond. Now Publishers Inc (2009) 12. Amati, G.: Probability models for information retrieval based on divergence from randomness. PhD thesis, University of Glasgow (2003) 13. La erty, J., Zhai, C.: Document language models, query models, and risk minimization for information retrieval. In: Proceedings of the 24th annual international ACM SIGIR conference on Research and development in information retrieval, ACM (2001) 111{119 14. Santos, R.L., Macdonald, C., Ounis, I.: Exploiting query reformulations for web search result diversi cation. In: Proceedings of the 19th international conference on World wide web, ACM (2010) 881{890