=Paper=
{{Paper
|id=Vol-1391/inv-pap9-CR
|storemode=property
|title=CLEF eHealth Evaluation Lab 2015, Task 2: Retrieving Information About Medical Symptoms
|pdfUrl=https://ceur-ws.org/Vol-1391/inv-pap9-CR.pdf
|volume=Vol-1391
|dblpUrl=https://dblp.org/rec/conf/clef/PalottiZGKHJLP15
}}
==CLEF eHealth Evaluation Lab 2015, Task 2: Retrieving Information About Medical Symptoms==
https://ceur-ws.org/Vol-1391/inv-pap9-CR.pdf
CLEF eHealth Evaluation Lab 2015, Task 2:
Retrieving Information About Medical
Symptoms
João Palotti1 , Guido Zuccon2 , Lorraine Goeuriot3 , Liadh Kelly4 , Allan
Hanbury1 , Gareth J.F. Jones5 , Mihai Lupu1 , and Pavel Pecina6?
1
Vienna University of Technology, Austria,
palotti,hanbury,lupu@ifs.tuwien.ac.at
2
Queensland University of Technology, Australia, g.zuccon@qut.edu.au
3
Université Grenoble Alpes, France, lorraine.goeuriot@imag.fr
4
Trinity College, Dublin, Ireland Liadh.Kelly@tcd.ie,
5
Dublin City University, Ireland, gareth.jones@computing.dcu.ie
6
Charles University in Prague, Czech Republic, pecina@ufal.mff.cuni.cz
Abstract. This paper details methods, results and analysis of the CLEF
2015 eHealth Evaluation Lab, Task 2. This task investigates the effec-
tiveness of web search engines in providing access to medical information
with the aim of fostering advances in the development of these technolo-
gies.
The problem considered in this year’s task was to retrieve web pages to
support information needs of health consumers that are confronted with
a sign, symptom or condition and that seek information through a search
engine, with the aim to understand which condition they may have. As
part of this evaluation exercise, 66 query topics were created by potential
users based on images and videos of conditions. Topics were first created
in English and then translated into a number of other languages. A total
of 97 runs by 12 different teams were submitted for the English query
topics; one team submitted 70 runs for the multilingual topics.
Key words: Medical Information Retrieval, Health Information Seeking
and Retrieval
1 Introduction
This document reports on the CLEF 2015 eHealth Evaluation Lab, Task 2. The
task investigated the problem of retrieving web pages to support information
needs of health consumers (including their next-of-kin) that are confronted with
a sign, symptom or condition and that use a search engine to seek understand-
ing about which condition they may have. Task 2 has been developed within
the CLEF 2015 eHealth Evaluation Lab, which aims to foster the development
?
In alphabetical order, JP, GZ led Task 2; LG, LK, AH, ML & PP were on the Task
2 organising committee.
of approaches to support patients, their next-of-kin, and clinical staff in under-
standing, accessing and authoring health information [1].
The use of the Web as source of health-related information is a wide-spread
phenomena. Search engines are commonly used as a means to access health infor-
mation available online [2]. Previous iterations of this task (i.e. the 2013 and 2014
CLEFeHealth Lab Task 3 [3, 4]) aimed at evaluating the effectiveness of search
engines to support people when searching for information about their conditions,
e.g. to answer queries like “thrombocytopenia treatment corticosteroids length”.
These past two evaluation exercises have provided valuable resources and an
evaluation framework for developing and testing new and existing techniques.
The fundamental contribution of these tasks to the improvement of search engine
technology aimed at answering this type of health information need is demon-
strated by the improvements in retrieval effectiveness provided by the best 2014
system [5] over the best 2013 system [6] (using different, but comparable, topic
sets).
Searching for self-diagnosis information is another important type of health
information seeking activity [2]; this seeking activity has not been considered
in the previous CLEF eHealth tasks, nor in other information retrieval evalua-
tion campaigns. These information needs often arise before attending a medical
professional (or to help the decision of attending). Previous research has shown
that exposing people with no or scarce medical knowledge to complex medical
language may lead to erroneous self-diagnosis and self-treatment and that access
to medical information on the Web can lead to the escalation of concerns about
common symptoms (e.g., cyberchondria) [7, 8]. Research has also shown that
current commercial search engines are yet far from being effective in answering
such queries [9]. This type of query is the subject of investigation in this CLEF
2015 eHealth Lab Task 2. We expected these queries to pose a new challenge
to the participating teams; a challenge that, if solved, would lead to significant
contributions towards improving how current commercial search engines answer
health queries.
The remainder of this paper is structured as follows: Section 2 details the
task, the document collection, topics, baselines, pooling strategy, and evaluation
metrics; Section 3 presents the participants’ approaches, while Section 4 presents
their results; Section 5 concludes the paper.
2 The CLEF 2015 eHealth Task 2
2.1 The Task
The goal of the task is to design systems which improve health search, especially
in the case of search for self-diagnosis information. The dataset provided to
participants is comprised of a document collection, topics in various languages,
and the corresponding relevance information. The collection was provided to
participants after signing an agreement, through the PhysioNet website7 .
7
http://physionet.org/
Participating teams were asked to submit up to ten runs for the English
queries, and an additional ten runs for each of the multilingual query languages.
Teams were required to number runs such as that run 1 was a baseline run for
the team; other runs were numbered from 2 to 10, with lower numbers indicating
higher priority for selection of documents to contribute to the assessment pool
(i.e. run 2 was considered of higher priority than run 3).
2.2 Document Collection
The document collection provided in the CLEF 2014 eHealth Lab Task 3 [4]
is also adopted in this year’s task. Documents in this collection have been ob-
tained through a large crawl of health resources on the Web; the collection con-
tains approximately one million documents and originated from the Khresmoi
project8 [10]. The crawled domains were predominantly health and medicine
sites, which were certified by the HON Foundation as adhering to the HON-
code principles (appr. 60–70% of the collection), as well as other commonly
used health and medicine sites such as Drugbank, Diagnosia and Trip Answers9 .
Documents consisted of web pages on a broad range of health topics and were
likely targeted at both the general public and healthcare professionals. They
were made available for download in their raw HTML format along with their
URLs to registered participants.
2.3 Topics
Queries were manually built with the following process: images and videos re-
lated to medical symptoms were shown to users, who were then asked which
queries they would issue to a web search engine if they, or their next-of-kin, were
exhibiting such symptoms. Thus, these queries aimed to simulate the situation
of health consumers seeking information to understand symptoms or conditions
they may be affected by; this is achieved using imaginary or video stimuli. This
methodology for eliciting circumlocutory, self-diagnosis queries was shown to be
effective by Stanton et al. [11]; Zuccon et al. [9] showed that current commercial
search engines are yet far from being effective in answering such queries.
Following the methodology in [9, 11], 23 symptoms or conditions that man-
ifest with visual or audible signs (e.g. ringworm or croup) were selected to be
presented to users to collect queries. A cohort of 12 volunteer university stu-
dents and researchers based in the organisers’ institutions generated the queries.
English was the mother-tongue for all volunteers and they had no particular
prior knowledge about the symptoms or conditions, nor they had any specific
medical background: this cohort was then somehow representative of the average
user of web search engines seeking health advice (although they had a higher
8
Medical Information Analysis and Retrieval, http://www.khresmoi.eu
9
Health on the Net, http://www.healthonnet.org, http://www.hon.ch/HONcode/
Patients-Conduct.html, http://www.drugbank.ca, http://www.diagnosia.com,
and http://www.tripanswers.org
education level than the average level). Each volunteer was given 10 conditions
for which they were asked to generate up to 3 queries per condition (thus each
condition/image pair was presented to more than one assessor10 ). An example
of images and instructions provided to the volunteers is given in Figure 111 .
Imagine you are experiencing the health problem shown below.
Please provide 3 search queries that you would issue to find out what is wrong.
Instructions:
* You must provide 3 distinct search queries.
* The search queries must relate to what you see below.
Fig. 1. An example of instructions and images provided to volunteers for generating
potential search queries.
A total of 266 possible unique queries were collected; of these, 67 queries
(22 conditions with 3 queries and 1 condition with 1 query) were selected to be
used in this year’s task. Queries were selected by randomly picking one query per
condition (we called this the pivot query), and then manually selecting the query
that appeared most similar (called most) and the one that appeared least similar
(called least) to the pivot query. Candidates for the most and least queries were
identified independently by three organisers and then majority voting was used
to establish which queries should be selected. This set of queries formed the
English query set distributed to participants to collect runs.
In addition, we developed translations of this query set into Arabic (AR),
Czech (CS), German (DE), Farsi (FA), French (FR), Italian (IT) and Portuguese
(PT); these formed the multilingual query sets which were made available to par-
ticipants for submission of multilingual runs. Queries were translated by medical
experts available at the organisers institutions.
After the query set was released, numbered qtest1-qtest67, one typo was
found in query qtest62, which could compromise the translations. In order to
keep consistency between the English query and all translations made by the
experts, qtest62 was excluded. Thus, the final query set used in the CLEF 2015
10
With exception of one condition, for which only one query could be generated.
11
Note that additional instructions were given to volunteers at the start and end of
the task, including training and de-briefing.
eHealth Lab Task 2 for both English and multilingual queries consisted of 66
queries.
An example of one of the query topics generated from the image shown in
Figure 1 is provided in Figure 2. To develop their submissions, participants were
only given the query field of each query topic, that is, teams were unaware of
the query type (pivot, most, least), the target condition and the image or video
that was shown to assessors to collect queries.
...
qtest.23
red bloodshot eyes
non-ulcerative sterile keratitis
most
22
...
Fig. 2. Example query topic generated from the image of Figure 1. This query is of
type most and refers to the image condition 22 (as indicated by the field query index).
2.4 Relevance Assessment
Relevance assessments were collected by pooling participants’ submitted runs
as well as baseline runs (see below for a description of pooling methodology
and baseline runs). Assessment was performed by five paid medical students
employed at the Medizinische Universität Graz (Austria); assessors used Rel-
evation! [12] to visualise and judge documents. For each document, assessors
had access to the query the document was retrieved for, as well as the target
symptom or condition that was used to obtained the query during the query
generation phase.
Target symptoms or conditions were used to provide the relevance criteria as-
sessors should judge against; for example for query qtest1 – “many red marks on
legs after traveling from US” (the condition used for generating the query was
“Rocky Mountain spotted fever (RMSF)”), the relevance criterion read “Rel-
evant documents should contain information allowing the user to understand
that they have Rocky Mountain spotted fever (RMSF).”. Relevance assessments
were provided on a three point scale: 0, Not Relevant; 1, Somewhat Relevant; 2,
Highly Relevant.
Along with relevance assessments, readability judgements were also collected
for the assessment pool. The notion of readability and understandability of in-
formation is of important concern when retrieving information for health con-
sumers [13]. It has been shown that if the readability of information is accounted
for in the evaluation framework, judgements of relative system effectiveness can
vary with respect to taking into account (topical) relevance only [14] (this was
the case also when considering the CLEF 2013 and 2014 eHealth Evaluation
Labs).
Readability assessments were collected by asking the assessors whether they
believed a patient would understand the retrieved document. Assessments were
provided on a four point scale: 0, “It is very technical and difficult to read and
understand”; 1, “It is somewhat technical and difficult to read and understand”;
2, “It is somewhat easy to read and understand”; 3, “It is very easy to read and
understand”.
2.5 Example Topics
A different set of 5 queries was released to participants as example queries (called
training) to help develop their systems (both in English and the other consid-
ered languages). These queries were released together with associated relevance
assessments, obtained by evaluating a pool of 112 documents retrieved by a set
of baseline retrieval systems (TF-IDF, BM25, Language Model with Dirichlet
smoothing as implemented in Terrier [15], with the associated default parame-
ter values); the pool was formed by sampling the top 10 retrieved documents
for each query. Note that, given the very limited pool and system sample sizes,
these example queries should not be used to evaluate, tune or train systems.
2.6 Baseline Systems
The organisers generated baseline runs using BM25, TF-IDF and Language
Model with Dirichlet smoothing, as well as a set of benchmark systems that
ranked documents by estimating both (topical) relevance and readability. Ta-
ble 1 shows the 13 baseline systems created, 7 of them took into consideration
some estimation of text readability. No baselines were created for the multilin-
gual queries.
The first 6 baselines, named baseline1 -6, were created using either Xapian or
Lucene as retrieval toolkit. We vary the retrieval model used, including BM25
(with parameters k1 = 1, k2 = 0, k3 = 1 and b = 0.5) in baseline1, Vector Space
Model (VSM) with TF-IDF weighting (the default Lucene implementation) in
baseline2, and Language Model (LM), with Dirichlet smoothing with µ = 2, 000
in baseline3. Our preliminary runs based on the 2014 topics showed that remov-
ing HTML tags from documents in this collection could lead to higher retrieval
effectiveness when using BM25 and LM. We used the python package Beauti-
fulSoap (BS4)12 to parse the HTML files and remove HTML tags. Note that it
12
https://pypi.python.org/pypi/beautifulsoup4
does not remove the boilerplate from the HTML (such as headers or navigation
menus), being one of the simplest approaches to clean a HTML page and prepare
it to serve as the input of readability formulas [16] (see below). Baselines 4, 5
and 6 implement the same methods as in baselines 1, 2 and 3, respectively, but
execute a query that has been enhanced by augmenting the original query with
the known target disease names. Note that the target disease names were only
known to the organisers, participants had no access to this information.
For the baseline runs that take into account readability estimations, we used
two well-known automatic readability measures: the Dale-Chall measure [17]
and Flesch-Kincaid readability index [18]. The python package ReadabilityCal-
culator13 was used to compute the readability measures from the cleansed web
documents. We also tested a readability measure based on the frequency of
words in a large collection such as Wikipedia; the intuition behind this measure
is that an easy text would contain a large number of common words with high
frequency in Wikipedia, while a technical and difficult text would have a large
number of rare words, characterised by a low frequency in Wikipedia. In order
to retrieving documents accounting for their readability levels, we first generate
a readability score Read(d) for each document d in the collection using one of
the three measures above. We then combine the readability score of a document
with its relevance score Rel(d) generated by some retrieval model. Three score
combination methods were considered:
1. Linear combination: Score(d) = α × Rel(d) + (1.0 − α) × Read(d), where α
is a hyperparameter and 0 ≤ α ≤ 1 (in readability1 α is 0.9)
2. Direct Multiplication: Score(d) = Rel(doc) × Read(d)
Rel(doc)
3. Inverse Logarithm: Score(d) = log(Read(d))
Table 1 shows the settings of retrieval model, HTML processing, readability
measure and query expansion or score combination method that were considered
to produce the 7 readability baselines used in the task.
2.7 Pooling Methodology
In Task 2, for each query, the top 10 documents returned in runs 1, 2 and 3
produced by the participants14 were pooled to form the relevance assessment
pool. In addition, the baseline runs developed by the organisers were also pooled
with the same methodology used for participants runs. A pool depth of 10 doc-
uments was chosen because this task resembles web-based search, where often
users consider only the first page of results (that is, the first 10 results). Thus,
this pooling methodology allowed a full evaluation of the top 10 results for the 3
submissions with top priority for each participating team. The pooling of more
submissions or a deeper pool, although preferable, was ruled out because of the
limited availability of resources for document relevance assessment.
13
https://pypi.python.org/pypi/ReadabilityCalculator/
14
With the exclusion of multilingual submissions, for which runs were not pooled due
to the larger assessment effort pooling these runs would have required. Note that
only one team submitted multilingual runs.
Table 1. Scheme showing the settings of retrieval model, HTML processing, readability
measure and query expansion or score combination used to generate the organisers
baselines.
System Index Model Cleaning Expansion/Combination Readability
baseline1 Xapian BM25 BS4 - -
baseline2 Lucene VSM - - -
baseline3 Lucene LM BS4 - -
baseline4 Xapian BM25 BS4 Disease Name added -
baseline5 Lucene VSM - Disease Name added -
baseline6 Lucene LM BS4 Disease Name added -
readability1 Xapian BM25 BS4 Linear Combination Dale-Chall
readability2 Xapian BM25 BS4 Direct Multiplication Wikipedia Frequency
readability3 Xapian BM25 BS4 Inverse Logarithm Dale-Chall
readability4 Xapian BM25 BS4 Inverse Logarithm Flesch-Kincaid
readability5 Lucene VSM - Direct Multiplication Wikipedia Frequency
readability6 Lucene VSM BS4 Inverse Logarithm Dale-Chall
readability7 Lucene VSM BS4 Inverse Logarithm Flesch-Kincaid
2.8 Multilingual Evaluation: Additional Pooling and Relevance
Assessments
Because only one team submitted runs for the multilingual queries and only lim-
ited relevance assessment capabilities were available through the paid medical
students that performed the assessment of submissions for the English queries,
multilingual runs were not considered when forming the pools for relevance as-
sessments. However, additional relevance assessments were sought through the
team that participated in the multilingual task: they were thus asked to perform
a self-assessment of the submissions they produced. A new pool of documents
was sampled with the same pooling methodology used for English runs (see the
previous section). Documents that were already judged by the official assessors
were excluded from the pool with the aim to limit the additional relevance as-
sessment effort required by the team.
Additional relevance assessments for the multilingual runs were then per-
formed by a medical doctor (native Czech speaker with fluent English) associ-
ated with Team CUNI [22]. The assessor was provided with the same instruc-
tions and assessment system that the official assessors used. Assessments were
collected and aggregated with those provided by the official relevance assessors
to form the multilingual merged qrels. These qrels should be used with caution:
at the moment of writing this paper, it is unknown whether these multilingual
assessments are comparable with those compiled by the original, also medically
trained, assessors. The collection of further assessments from the team to ver-
ify their agreement with the official assessors is left for future work. Another
limitation of these additional relevance assessments is that only one system that
considered multilingual queries, that developed by team CUNI, was sampled and
thus it may further bias the assessment of retrieval systems with respect to how
multilingual queries are coped with.
2.9 Evaluation Metrics
Evaluation was performed in terms of graded and binary assessments. Binary
assessments were formed by transforming the graded assessments such that label
0 was maintained (i.e. irrelevant) and labels 1 and 2 were converted to 1 (rele-
vant). Binary assessments for the readability measures were obtained similarly,
with labels 0 and 1 being converted into 0 (not readable) and labels 2 and 3
being converted into 1 (readable).
System evaluation was conducted using precision at 10 (p@10) and nor-
malised discounted cumulative gain [19] at 10 (nDCG@10) as the primary and
secondary measures, respectively. Precision was computed using the binary rele-
vance assessments; nDCG was computed using the graded relevance assessments.
These evaluation metrics were computed using trec eval with the following
commands:
./trec eval -c -M1000 qrels.clef2015.test.bin.txt runName
./trec eval -c -M1000 -m ndcg cut qrels.clef2015.test.graded.txt runName
respectively to compute precision and nDCG values.
A separate evaluation was conducted using both relevance assessments and
readability assessments following the methods in [14]. For all runs, Rank Biased
Precision (RBP) [20] was computed along with readability-biased modifications
of RBP, namely uRBP (using the binary readability assessments) and uRBPgr
(using the graded readability assessments).
The RBP parameter ρ which attempts to model user behaviour15 (RBP
persistence parameter) was set to 0.8 for all variations of this measure, following
the findings of Park and Zhang [21].
To compute uRBP, readability assessments were mapped to binary classes,
with assessments 0 and 1 (indicating low readability) mapped to value 0 and
assessments 2 and 3 (indicating high readability) mapped to value 1. Then,
uRBP (up to rank K) was calculated according to
K
X
uRBP = (1 − ρ) ρk−1 r(k)u(k) (1)
k=1
where r(k) is the standard RBP gain function that is 1 if the document at rank k
is relevant and 0 otherwise; u(k) is a similar gain function but for the readability
dimension and is 1 if the document at k is readable (binary class 1), and zero
otherwise (binary class 0).
To compute uRBPgr, i.e. the graded version of uRBP, each readability label
was mapped to a different gain value. Specifically, label 0 was assigned gain 0
(least readability, no gain), label 1 gain 0.4, label 2 gain 0.8 and label 3 gain 1
(highest readability, full gain). Thus, a document that is somewhat difficult to
read does still generate a gain, which is half the gain generated by a document
15
High values of ρ representing persistent users, low values representing impatient
users.
that is somewhat easy to read. These gains are then used to evaluate the function
u(k) in Equation 1 to obtain uRBPgr.
The readability-biased evaluation was performed using ubire16 , which is
publicly available for download.
3 Participants and Approaches
3.1 Participants
This year, 52 groups registered for the task on the web site, 27 got access to the
data and 12 submitted any run for task 2. The groups are from 9 countries in 4
continents as listed in Table 2. 7 out of the 12 participants had never participated
in this task before.
Table 2. Participants for task 2 and their total number of submissions.
Runs Submitted
Continent Country Team Name
English Multilingual
Botswana UBML 10 -
Africa
Tunisia Miracl 5 -
Canada GRIUM 7 -
America
Canada YORKU 10 -
China ECNU 10 -
China FDUSGInfo 10 -
China USST 10 -
Asia
South Korea KISTI 8 -
Thailand KU-CS 4 -
Vietnam HCMUS 8 -
Czech Republic CUNI 10 70
Europe
France LIMSI 5 -
Total 9 Countries 12 Teams 97 70
3.2 Participant Approaches
Team CUNI [22] used the Terrier toolkit to produce their submissions. Runs ex-
plored three different retrieval models: Bayesian smoothing with Dirichlet prior,
Per-field normalisation (PL2F) and LGD. Query expansion using the UMLS
metathesaurus was explored by considering terms assigned to the same con-
cept as synonymous and choosing the terms with the highest inverse document-
frequency. Blind relevance feedback was also used as contrasting technique. Fi-
nally, they also experimented with linear interpolations of the search results
produced by the above techniques.
16
https://github.com/ielab/ubire
Team ECNU [23] explored query expansion and learning to rank. For query
expansion, Google was queried and the titles and snippets associated with the
top ten web results were selected. Medical terms were then extracted from these
resources by matching them with terms contained in MeSH; the query was then
expanded using those medical terms that appeared more often than a threshold.
As Learning to Rank approach, Team ECNU combined scores and ranks from
BM25, PL2 and BB2 into a six-dimensional vector. The 2013 and 2014 CLEF
eHealth tasks were used to train the system and a Random Forest classifier was
use to calculate the new scores.
Team FDUSGInfo explored query expansion methods that use a range of knowl-
edge resources to improve the effectiveness of a statistical Language Model base-
line. The knowledge sources that have been considering for drawing expansion
terms are MeSH and Freebase. Different techniques were evaluated to select the
expansion terms, including manual term selection. Team FDUSGInfo, unfortu-
nately, did not submit their working notes and thus the details of their methods
are unknown.
Team GRIUM [25] explored the use of concept based query expansion. Their
query expansion mechanism exploited Wikipedia articles and UMLS Concept
definitions and were compared to a baseline method based on Dirichlet smooth-
ing.
Team KISTI [26] focused on re-ranking approaches. Lucene was used for index-
ing and initial search, and the baseline used the query likelihood model with
Dirichlet smoothing. They explored three approaches for re-ranking: explicit se-
mantic analysis (ESA), concept-based document centrality (CBDC), and cluster-
based external expansion model (CBEEM). Their submissions evaluated these
re-ranking approaches as well as their combinations.
Team KUCS [27] implemented an adaptive query expansion. Based on the results
returned by a query performance prediction approach, their method selected the
query expansion that is hypothesised to be the most suitable for improving
effectiveness. An additional process was responsible for re-ranking results based
on readability estimations.
Team LIMSI [28] explored query expansion approaches that exploit external
resources. Their first approach used MetaMap to identify UMLS concepts from
which to extract medical terms to expand the original queries. Their second ap-
proach used a selected number of Wikipedia articles describing the most common
diseases and conditions along with a selection of MedlinePlus; for each query the
most relevant articles from these corpora are retrieved and their titles used to ex-
pand the original queries, which are in turn used to retrieve relevant documents
from the task collection.
Team Miracl [29]’s submissions were based on blind relevance feedback combined
with term selection using their previous work on modeling semantic relations
between words. Their baseline run was based on the traditional Vector Space
Model and the Terrier toolkit. The other runs employed the investigated method
by varying settings of two method parameters: the first controlling the number of
highly ranked documents from the initial retrieval step and the second controlling
the degree of semantic relationship of the expansion terms.
Team HCMUS [30] experimented with two approaches. The first was based on
concept-based retrieval where only medical terminological expressions in docu-
ments were retained, while other words were filtered-out. The second was based
on query expansion with blind relevance feedback. Common to all their ap-
proaches was the use of Apache Lucene and a bag-of-word baseline based on
Language Modelling with Dirichlet smoothing and standard stemming and stop-
word removal.
Team UBML [31] investigated the empirical merits of query expansion based
on KL divergence and the Bose-Einstein 1 model for improving a BM25 base-
line. The query expansion process selected terms from the local collection or
two external collections. Learning to rank was also investigated along a Markov
Random Fields approach.
Team USST [32] used BM25 as a baseline system and explored query expansion
approaches. They investigated pseudo relevance feedback approaches based on
Kullback-Liebler Divergence and Bose-Einstein models.
Team YorkU [33] explored BM25 and Divergence from Randomness methods
as provided by the Terrier toolkit, along with the associated relevance feedback
retrieval approaches.
4 Results and Findings
4.1 Pooling and Coverage of Relevance Assessments
A total of 8,713 documents were assessed. Of these, 6,741 (77.4%) were assessed
as irrelevant (0), 1,515 (17.4%) as somewhat relevant (1), 457 (5.2%) as highly
relevant (2). For readability assessments, the recorded distribution was: 1,145
(13.1%) documents assessed as difficult (0), 1,568 (18.0%) as somewhat difficult
(1), 2,769 (31.8%) as somewhat easy (2), and 3,231 (37.1%) as easy (3).
Table 3 details the coverage of the relevance assessments with respect to the
participant submissions, averaged over the whole query set. While in theory all
runs 1-3 should have full coverage (100%), in practice a small portion of docu-
ments included in the relevance assessment pool were left unjudged because the
documents were not in the collection (participants provided an invalid document
identifier) or the page failed to render in the relevance assessment toolkit (for
example because the page contained redirect scripts or other scripts that were
Table 3. Coverage of the relevance assessments for the top 10 results submitted by
participants in the task: 100% means that all top 10 results for all queries have been
assessed; 90% means that, on average, 9 out of 10 documents in the top 10 results have
been assessed, with one document being left unjudged.
Run Baseline Readab. CUNI ENUC FDUSG. GRIUM KISTI KUCS LIMSI Miracl HCMUS UBML USST YorkU Mean
1 99.98 100.0 100.0 99.98 98.77 100.0 100.0 99.64 99.83 99.98 99.92 99.92 100.0 99.62 99.83
2 99.82 99.98 100.0 99.88 98.77 100.0 99.98 98.77 99.92 99.98 99.89 100.0 100.0 100.0 99.79
3 99.98 99.95 99.94 99.95 98.77 100.0 100.0 92.61 99.85 100.0 99.79 100.0 100.0 100.0 99.35
4 93.64 94.65 99.95 99.86 98.08 99.65 99.80 91.58 92.00 96.82 97.65 98.38 98.64 99.98 97.19
5 92.61 99.15 99.58 96.00 97.91 99.94 99.58 - 92.00 99.15 94.67 98.42 98.30 99.85 97.47
6 93.74 98.89 99.23 98.11 91.65 99.98 99.73 - - - 93.12 98.58 97.91 99.68 97.33
7 - 97.33 99.79 96.56 91.65 99.98 99.70 - - - 94.65 99.48 96.24 99.53 97.49
8 - - 99.98 98.76 91.65 - 99.73 - - - 93.14 98.29 95.85 99.23 97.08
9 - - 99.61 99.79 91.33 - - - - - - 98.45 95.24 98.83 97.21
10 - - 97.94 98.70 91.33 - - - - - - 97.70 95.06 98.33 96.51
Mean 96.63 98.57 99.60 98.76 94.99 99.94 99.81 95.65 96.72 99.19 96.60 98.92 97.72 99.51 98
not executed within Relevation17 ). Overall, the mean coverage for runs 1-3 was
above 99%, with only run 3 from team KUCS being sensibly below this value.
This suggests that the retrieval effectiveness for runs 1-3 can be reliably mea-
sured. The coverage beyond submissions 3 is lower but always above 90% (and
the mean above 95%); this suggest that the evaluation of runs 4-10 in terms of
precision at 10 may be underestimated of an average maximum of 0.05 points
over the whole query set.
Table 4 details the coverage of relevance assessment for the multilingual runs.
As mentioned in Section 2.8, due to limited relevance assessment availability, only
the English runs were considered when forming the pool for relevance assessment.
The coverage of these relevance assessments with respect to the top 10 documents
ranked by each participants’ submissions is shown in the columns marked as Eng.
in Table 4. An additional document pool, made using only documents in runs1-
3 of multilingual submissions, was created to further increase the coverage of
multilingual submissions; the coverage of the union of the original assessments
and these additional ones (referred to as merged ) is shown in the columns marked
as Merged in Table 4 for the multilingual runs. The merged set of relevance
assessments was enough to provide a fairly high coverage for all runs, including
those not in the pool (i.e., runs beyond number 3), with a minimal coverage of
97%; this is likely because only one team submitted runs for the multilingual
challenge, thus providing only minimal variation in terms of top retrieval results.
4.2 Evaluation Results and Findings
Table 5 reports the evaluation of the participants submissions and the organisers
baselines based on P@10 and nDCG@10 for English queries. The evaluation
based on RBP and the readability measures is reported in Table 6.
Most of the approaches developed by team ECNU obtain significantly higher
values of P@10 and nDCG@10 compared to the other participants, demonstrat-
17
Note that before the relevance assessment exercise started, we removed the majority
of scripts from the pooled pages to avoid this problem.
Table 4. Coverage of the relevance assessments for the top 10 results submitted by
CUNI in the multilingual evaluation. As described in Section 2.8, two set of qrels were
used: those for the English task (Eng.), and those produced by merging the assessments
for English queries and the ones for multilingual queries (Merged.)
AR CS DE FA FR IT PT
Run
Eng. Merged Eng. Merged Eng. Merged Eng. Merged Eng. Merged Eng. Merged Eng. Merged
1 95.32 99.97 94.52 99.94 95.21 99.80 95.59 99.91 95.14 99.91 95.48 99.95 95.76 99.94
2 94.95 99.91 93.88 99.82 94.85 99.82 95.36 99.89 94.59 99.92 95.35 99.85 95.56 99.91
3 94.64 99.91 93.74 99.86 94.70 99.91 95.11 99.91 94.65 99.92 94.89 99.89 95.18 99.83
4 95.20 99.77 94.09 99.79 95.11 99.77 95.62 99.79 94.88 99.88 95.58 99.89 95.83 99.85
5 95.00 98.03 94.02 97.32 94.98 97.47 95.29 97.70 94.67 98.56 95.14 98.97 95.65 98.33
6 95.03 98.03 94.11 98.56 94.35 98.26 95.42 97.71 94.59 98.56 95.15 99.05 95.91 98.30
7 94.73 97.73 94.29 97.36 96.47 98.91 94.85 97.30 96.00 99.29 94.76 98.98 95.42 97.88
8 95.03 98.12 94.45 97.21 96.11 98.21 95.42 97.68 95.89 98.73 95.18 98.94 95.94 98.27
9 94.64 98.29 94.00 96.52 95.47 97.58 94.62 97.73 95.33 98.29 95.00 98.48 94.80 98.18
10 95.33 99.55 94.38 97.29 96.62 98.76 95.70 99.48 96.17 99.24 95.59 99.70 95.94 99.48
Mean 94.99 98.93 94.15 98.37 95.39 98.85 95.30 98.71 95.19 99.23 95.21 99.37 95.60 99.00
ing about 40% increase in effectiveness in their best run compared to the runner-
up team (KISTI). The best submission developed by the organisers and based
on both relevance and readability estimates has been proved difficult to out-
perform by most teams (only 4 out of 12 teams obtained higher effectiveness).
The pooling methodology does not appear to have significantly influenced the
evaluation of non-pooled submissions, as demonstrated by the fact that the best
runs of some teams are not those that were fully pooled (e.g. team KISTI, team
CUNI, team GRIUM).
There are no large differences between system rankings produced using P@10
or nDCG@10 as evaluation measure (Kendall τ = 0.88). This is unlike when
readability is also considered in the evaluation (the Kendall τ between system
rankings obtained with P@10 or uRBP is 0.76). In this latter case, while ECNU’s
submissions are confirmed to be the most effective, there are large variations in
system rankings when compared to those obtained considering relevance judge-
ments only. In particular, runs from team KISTI, which in the relevance-based
evaluation were ranked among the top 20 runs, are not performing as well when
considering also readability, with their top run (KISTI EN RUN.7) being ranked
only 37th according to uRBP.
The following considerations could be drawn when comparing the different
methods employed by the participating teams. Query expansion is found to of-
ten improve results. In particular, team ECNU obtained the highest effectiveness
among the systems that took part in this task; this was achieved when query
expansion terms are mined from Google search results returned for the original
queries (ECNU EN Run.3). This approach indeed obtained higher effectiveness
compared to learning-to-rank alternatives (ECNU EN Run.10). The results of
team UBML show that query expansion using the Bose-Einstein model 1 and
the local collection works better than other query expansion methods and ex-
ternal collections. Team USST also found that query expansion was effective
to improve results, however they found that the Bose-Einstein models did not
provide improvements over their baseline, while the Kullback-Liebler Divergence
based query expansion provided minor improvements. Health-specific query ex-
pansion methods based on the UMLS were shown to be effective above common
baselines and other considered query expansion methods by Team LIMSI and
GRIUM (this form of query expansion was the only one that delivered higher
effectiveness than their baseline).Team KISTI found that the combination of
concept-based document centrality (CBDC) and cluster- based external expan-
sion model (CBEEM) improved the results best. Few teams did not observe
improvements over their baselines; this was the case for teams KUCS, Miracl,
FDUSGInfo and HCMUS.
Tables 7 and 8 report the evaluation of the multilingual submissions based
on P@10 and nDCG@10; results are reported with respect to both the original
qrels (obtained by sampling English runs only) and the additional qrels (obtained
by sampling also multilingual runs, but using a different set of assessors); see
Section 2.8 for details about the difference between these relevance assessments.
Only one team (CUNI) participated in the multilingual task; they also submitted
to the English-based task and thus it is possible to discuss the effectiveness of
their retrieval system when answering multilingual queries compared to that
achieved when answering English queries.
The evaluation based on the original qrels allows us to compare multilingual
runs directly with English runs. Note that the original relevance assessments ex-
hibit a level of coverage for the multilingual runs that is similar to those obtained
for English submissions numbered 4-10. The evaluation based on the additional
qrels (merged) allows analysis of the multilingual runs using the same pooling
method used for English runs; thus submissions 1-3 for the multilingual runs
can be directly compared to the corresponding English ones, at the net of dif-
ferences in expertise, sensibility and systematic errors between the paid medical
assessors and the volunteer, student self-assessor used to gather judgements for
the multilingual runs.
When only multilingual submissions are considered, it can be observed that
there is not a language in which CUNI’s system is more effective: e.g. submis-
sions that considered Italian queries are among the best performing with original
assessments and are the best performing with the additional assessments, but
differences in effectiveness among top runs for different languages are not statis-
tically significant. However, it can be observed that none of CUNI’s submissions
that addressed queries expressed in not European languages (Farsi and Arabic)
are among the top ranked systems, regardless of the type of relevance assess-
ments.
The use of the additional relevance assessments naturally translates in observ-
ing increased retrieval effectiveness across all multilingual runs (because some of
the documents in the top 10 ranks that were not assessed, and thus irrelevant, in
the original assessments may have been marked as relevant in the additional as-
sessments). However, a noteworthy observation is that the majority of the most
effective runs according to the additional assessments are those that were not
fully sampled to form the relevance assessment pools (i.e. runs 4-10, as opposed
to the pooled runs 1-3).
When the submissions of team CUNI are compared across English and mul-
tilingual queries, it is possible to observe that the best multilingual runs do not
outperform English runs (unlike when the same comparison was instructed in
the 2014 task [4]), regardless of the type of relevance assessments. This result
does not come as unexpected and it indicates that the translation from a foreign
language to English as part of the retrieval process does degrade the quality of
queries (in terms of retrieval effectiveness), suggesting that more work is needed
to bridge the gap in effectiveness between English and multilingual queries when
these are used to retrieve English content.
5 Conclusions
This paper has described methods, results and analysis of the CLEF 2015 eHealth
Evaluation Lab, Task 2. The task considered the problem of retrieving web pages
for people seeking health information regarding unknown conditions or symp-
toms. 12 teams participated in the task; the results have shown that query
expansion plays an important role in improving search effectiveness. The best
results were achieved by a query expansion method that mined the top results
from the Google search engine. Despite the improvements over the organisers’
baselines achieved by some teams, further work is needed to sensibly improve
search in this context, as only about half of the top 10 results retrieved by the
best system were found to be relevant.
As a by-product of this evaluation exercise, the task contributes to the re-
search community a collection with associated assessments and evaluation frame-
work (including readability evaluation) that can be used to evaluate the effec-
tiveness of retrieval methods for health information seeking on the web. Queries,
assessments and participants runs are publicly available at http://github.com/
CLEFeHealth/CLEFeHealth2015Task2.
Acknowledgement
This task has been supported in part by the European Union Seventh Frame-
work Programme (FP7/2007-2013) under grant agreement no 257528 (KHRES-
MOI), by Horizon 2020 program (H2020-ICT-2014-1) under grant agreement
no 644753 (KCONNECT), by the Austrian Science Fund (FWF) project no
I1094-N23 (MUCKE), and by the Czech Science Foundation (grant number
P103/12/G084). We acknowledge the time of the people involved in the transla-
tion and relevance assessment tasks, in special we want to thank Dr. Johannes
Bernhardt-Melischnig (Medizinische Universitat Graz) for coordinating the re-
cruitment and management of the paid medical students that participated in
the relevance assessment exercise.
References
1. Goeuriot, L., Kelly, L., Suominen, H., Hanlen, L., Névéol, A., Grouin, C., Palotti,
J., Zuccon, G.: Overview of the clef ehealth evaluation lab 2015. In: CLEF 2015
- 6th Conference and Labs of the Evaluation Forum, Lecture Notes in Computer
Science (LNCS), Springer (September 2015)
2. Fox, S.: Health topics: 80% of internet users look for health information online.
Pew Internet & American Life Project (2011)
3. Goeuriot, L., Jones, G., Kelly, L., Leveling, J., Hanbury, A., Müller, H., Salanterä,
S., Suominen, H., Zuccon, G.: ShARe/CLEF eHealth Evaluation Lab 2013, Task 3:
Information retrieval to address patients’ questions when reading clinical reports.
In: Online Working Notes of CLEF, CLEF (2013)
4. Goeuriot, L., Kelly, L., Lee, W., Palotti, J., Pecina, P., Zuccon, G., Hanbury, A.,
Gareth J.F. Jones, H.M.: ShARe/CLEF eHealth Evaluation Lab 2014, Task 3:
User-centred health information retrieval. In: CLEF 2014 Evaluation Labs and
Workshop: Online Working Notes, Sheffield, UK (2014)
5. Shen, W., Nie, J.Y., Liu, X., Liui, X.: An investigation of the effec-
tiveness of concept-based approach in medical information retrieval grium@
clef2014ehealthtask 3. Proceedings of the ShARe/CLEF eHealth Evaluation Lab
(2014)
6. Zhu, D., Wu, S., James, M., Carterette, B., Liu, H.: Using discharge summaries to
improve information retrieval in clinical domain. Proceedings of the ShARe/-CLEF
eHealth Evaluation Lab (2013)
7. Benigeri, M., Pluye, P.: Shortcomings of health information on the internet. Health
promotion international 18(4) (2003) 381–386
8. White, R.W., Horvitz, E.: Cyberchondria: studies of the escalation of medical
concerns in web search. ACM TOIS 27(4) (2009) 23
9. Zuccon, G., Koopman, B., Palotti, J.: Diagnose this if you can: On the effectiveness
of search engines in finding medical self-diagnosis information. In: Advances in
Information Retrieval. Springer (2015) 562–567
10. Hanbury, A., Müller, H.: Khresmoi – multimodal multilingual medical information
search. In: MIE village of the future. (2012)
11. Stanton, I., Ieong, S., Mishra, N.: Circumlocution in diagnostic medical queries.
In: Proceedings of the 37th international ACM SIGIR conference on Research &
development in information retrieval, ACM (2014) 133–142
12. Koopman, B., Zuccon, G.: Relevation!: an open source system for information
retrieval relevance assessment. In: Proceedings of the 37th international ACM
SIGIR conference on Research & development in information retrieval, ACM (2014)
1243–1244
13. Walsh, T.M., Volsko, T.A.: Readability assessment of internet-based consumer
health information. Respiratory care 53(10) (2008) 1310–1315
14. Zuccon, G., Koopman, B.: Integrating understandability in the evaluation of con-
sumer health search engines. In: Medical Information Retrieval Workshop at SIGIR
2014. (2014) 32
15. Ounis, I., Lioma, C., Macdonald, C., Plachouras, V.: Research directions in terrier.
Novatica UPGRADE Special Issue on Web Information Access (2007)
16. Palotti, J., Zuccon, G., Hanbury, A.: The influence of pre-processing on the esti-
mation of readability of web documents. In: Proceedings of the 24th ACM Inter-
national Conference on Conference on Information and Knowledge Management
(CIKM). (2015)
17. Kincaid, J., Fishburne, R., Rogers, R., Chissom, B.: Derivation of New Readability
Formulas for Navy Enlisted Personnel. Technical report (1975)
18. Kincaid, J., Fishburne, R., Rogers, R., Chissom, B.: Derivation of New Readability
Formulas for Navy Enlisted Personnel. Technical report (1975)
19. Järvelin, K., Kekäläinen, J.: Cumulated gain-based evaluation of IR techniques.
ACM Transactions on Information Systems 20(4) (2002) 422–446
20. Moffat, A., Zobel, J.: Rank-biased precision for measurement of retrieval effective-
ness. ACM Transactions on Information Systems (TOIS) 27(1) (2008) 2
21. Park, L.A., Zhang, Y.: On the distribution of user persistence for rank-biased pre-
cision. In: Proceedings of the 12th Australasian document computing symposium.
(2007) 17–24
22. Saleh, S., Bibyna, F., Pecina, P.: CUNI at the CLEF 2015 eHealth Lab Task 2.
In: Proceedings of the ShARe/CLEF eHealth Evaluation Lab. (2015)
23. Song, Y., He, Y., Hu, Q., He, L.: ECNU at 2015 eHealth Task 2: User-centred
Health Information Retrieval. In: Proceedings of the ShARe/CLEF eHealth Eval-
uation Lab. (2015)
24. received, N.: Missing. In: Proceedings of the ShARe/CLEF eHealth Evaluation
Lab. (2015)
25. Liu, X.J., Nie, J.Y.: Bridging Layperson’s Queries with Medical Concepts -
GRIUM@CLEF2015 eHealth Task 2. In: Proceedings of the ShARe/CLEF eHealth
Evaluation Lab. (2015)
26. Oh, H.S., Jung, Y., Kim, K.Y.: KISTI at CLEF eHealth 2015 Task 2. In: Pro-
ceedings of the ShARe/CLEF eHealth Evaluation Lab. (2015)
27. Thesprasith, O., Jaruskulchai, C.: Task 2a: Team KU-CS: Query Coherence Anal-
ysis for PRF and Genomics Expansion. In: Proceedings of the ShARe/CLEF
eHealth Evaluation Lab. (2015)
28. D’hondt, E., Grau, B., Zweigenbaum, P.: LIMSI @ CLEF eHealth 2015 - task 2.
In: Proceedings of the ShARe/CLEF eHealth Evaluation Lab. (2015)
29. Ksentini, N., Tmar, M., Boughanem, M., Gargouri, F.: Miracl at Clef 2015 : User-
Centred Health Information Retrieval Task. In: Proceedings of the ShARe/CLEF
eHealth Evaluation Lab. (2015)
30. Huynh, N., Nguyen, T.T., Ho, Q.: TeamHCMUS: A Concept-Based Information
Retrieval Approach for Web Medical Documents. In: Proceedings of the ShARe/-
CLEF eHealth Evaluation Lab. (2015)
31. Thuma, E., Anderson, G., Mosweunyane, G.: UBML participation to CLEF
eHealth IR challenge 2015: Task 2. In: Proceedings of the ShARe/CLEF eHealth
Evaluation Lab. (2015)
32. Lu, F.: Employing query expansion models to help patients diagnose themselves.
In: Proceedings of the ShARe/CLEF eHealth Evaluation Lab. (2015)
33. Ghoddousi, A., Huang, J.X., Feng, T.: York University at CLEF eHealth 2015:
Medical Document Retrieval. In: Proceedings of the ShARe/CLEF eHealth Eval-
uation Lab. (2015)
Table 5. Participants and baseline results sorted by p@10.
R Run Name p@10 nDCG@10 R Run Name p@10 nDCG@10
1 ECNU EN Run.3 0.5394 0.5086 55 readability run.6 0.2970 0.2456
2 ECNU EN Run.10 0.4667 0.4525 57 Miracl EN Run.5 0.2939 0.2465
3 ECNU EN Run.8 0.4530 0.4226 57 YorkU EN Run.8 0.2939 0.2729
4 ECNU EN Run.6 0.4227 0.3978 59 YorkU EN Run.2 0.2924 0.2714
5 KISTI EN RUN.6 0.3864 0.3464 59 YorkU EN Run.4 0.2924 0.2717
5 KISTI EN RUN.8 0.3864 0.3464 59 YorkU EN Run.6 0.2924 0.2694
7 CUNI EN Run.7 0.3803 0.3465 62 FDUSGInfo EN Run.4 0.2848 0.2687
8 KISTI EN RUN.4 0.3788 0.3424 62 baseline run.4 0.2848 0.3483
9 CUNI EN Run.4 0.3742 0.3409 64 FDUSGInfo EN Run.5 0.2803 0.2665
10 KISTI EN RUN.7 0.3727 0.3459 64 YorkU EN Run.3 0.2803 0.2719
11 CUNI EN Run.1 0.3712 0.3423 66 YorkU EN Run.9 0.2788 0.2637
11 CUNI EN Run.2 0.3712 0.3351 67 UBML EN Run.5 0.2773 0.2500
13 ECNU EN Run.4 0.3652 0.3168 68 UBML EN Run.4 0.2742 0.2460
14 HCMUS EN Run.1 0.3636 0.3323 69 USST EN Run.4 0.2727 0.2305
15 CUNI EN Run.8 0.3621 0.3383 70 UBML EN Run.9 0.2697 0.2538
16 CUNI EN Run.6 0.3606 0.3364 70 YorkU EN Run.10 0.2667 0.2546
16 ECNU EN Run.2 0.3606 0.3220 72 UBML EN Run.8 0.2652 0.2533
16 ECNU EN Run.9 0.3606 0.3203 73 LIMSI EN run.3 0.2621 0.1960
16 KISTI EN RUN.1 0.3606 0.3352 73 UBML EN Run.6 0.2621 0.2265
16 KISTI EN RUN.5 0.3606 0.3362 73 baseline run.6 0.2621 0.3123
16 readability run.2 0.3606 0.3299 76 FDUSGInfo EN Run.2 0.2606 0.2488
22 KISTI EN RUN.3 0.3591 0.3395 76 HCMUS EN Run.3 0.2606 0.2341
23 CUNI EN Run.5 0.3530 0.3217 78 KUCS EN Run.1 0.2545 0.2205
23 CUNI EN Run.9 0.3530 0.3215 79 Miracl EN Run.3 0.2515 0.1833
25 CUNI EN Run.3 0.3485 0.3138 80 UBML EN Run.10 0.2485 0.2294
26 ECNU EN Run.1 0.3470 0.3144 81 USST EN Run.5 0.2470 0.2082
27 KISTI EN RUN.2 0.3455 0.3223 81 USST EN Run.6 0.2470 0.2056
28 readability run.1 0.3424 0.3226 83 USST EN Run.7 0.2439 0.2220
29 USST EN Run.2 0.3379 0.3000 84 Miracl EN Run.2 0.2424 0.1965
30 readability run.3 0.3364 0.2890 85 FDUSGInfo EN Run.3 0.2348 0.2234
31 HCMUS EN Run.2 0.3348 0.3137 86 LIMSI EN run.1 0.2318 0.1801
32 baseline run.1 0.3333 0.3151 87 LIMSI EN run.2 0.2303 0.1675
33 baseline run.3 0.3242 0.2960 88 KUCS EN Run.2 0.2288 0.1980
34 ECNU EN Run.7 0.3227 0.3004 88 readability run.7 0.2288 0.1834
35 Miracl EN Run.1 0.3212 0.2787 90 USST EN Run.8 0.1985 0.1757
36 UBML EN Run.2 0.3197 0.2909 91 HCMUS EN Run.4 0.1955 0.1866
37 GRIUM EN Run.6 0.3182 0.2944 91 baseline run.5 0.1955 0.2417
37 UBML EN Run.3 0.3182 0.2919 93 Miracl EN Run.4 0.1894 0.1572
39 GRIUM EN Run.3 0.3167 0.2913 93 YorkU EN Run.1 0.1894 0.1718
40 ECNU EN Run.5 0.3152 0.3006 95 HCMUS EN Run.5 0.1545 0.1574
41 GRIUM EN Run.1 0.3136 0.2875 96 HCMUS EN Run.7 0.1470 0.1550
42 UBML EN Run.1 0.3106 0.2897 97 USST EN Run.9 0.1439 0.1241
43 GRIUM EN Run.2 0.3091 0.2850 98 USST EN Run.10 0.1348 0.1145
43 UBML EN Run.7 0.3091 0.2887 99 readability run.4 0.1227 0.0958
45 readability run.5 0.3076 0.2595 100 HCMUS EN Run.6 0.1045 0.1139
46 GRIUM EN Run.7 0.3061 0.2798 101 HCMUS EN Run.8 0.0970 0.1078
47 GRIUM EN Run.5 0.3045 0.2803 102 FDUSGInfo EN Run.6 0.0773 0.0708
47 USST EN Run.1 0.3045 0.2841 102 FDUSGInfo EN Run.7 0.0773 0.0708
49 GRIUM EN Run.4 0.3030 0.2788 102 FDUSGInfo EN Run.8 0.0773 0.0708
49 USST EN Run.3 0.3030 0.2627 105 FDUSGInfo EN Run.9 0.0682 0.0602
51 YorkU EN Run.7 0.3015 0.2766 105 FDUSGInfo EN Run.10 0.0682 0.0602
51 baseline run.2 0.3015 0.2479 107 LIMSI EN run.4 0.0561 0.0378
53 CUNI EN Run.10 0.3000 0.2597 107 LIMSI EN run.5 0.0561 0.0378
53 YorkU EN Run.5 0.3000 0.2752 109 KUCS EN Run.3 0.0364 0.0299
55 FDUSGInfo EN Run.1 0.2970 0.2718 110 KUCS EN Run.4 0.0182 0.0163
Table 6. Participants and baseline results sorted by uRBP.
R Run Name RBP uRBP uRBPgr R Run Name RBP uRBP uRBPgr
1 ECNU EN Run.3 0.5339 0.3877 0.4046 56 FDUSGInfo EN Run.4 0.3019 0.2373 0.2393
2 ECNU EN Run.10 0.4955 0.3768 0.3873 57 YorkU EN Run.6 0.3081 0.2365 0.2431
3 CUNI EN Run.7 0.3946 0.3422 0.3312 58 YorkU EN Run.5 0.3109 0.2357 0.2416
4 ECNU EN Run.6 0.4459 0.3374 0.3453 59 UBML EN Run.8 0.2978 0.2352 0.2368
5 CUNI EN Run.2 0.3796 0.3354 0.3239 60 UBML EN Run.6 0.2766 0.2348 0.2310
6 CUNI EN Run.5 0.3736 0.3295 0.3169 61 FDUSGInfo EN Run.5 0.2989 0.2340 0.2356
7 CUNI EN Run.9 0.3727 0.3287 0.3163 62 YorkU EN Run.2 0.3151 0.2334 0.2404
8 CUNI EN Run.4 0.3894 0.3284 0.3256 63 UBML EN Run.9 0.2993 0.2332 0.2362
9 ECNU EN Run.8 0.4472 0.3273 0.3373 64 YorkU EN Run.4 0.3152 0.2319 0.2397
10 ECNU EN Run.9 0.3730 0.3249 0.3107 65 KUCS EN Run.1 0.2785 0.2312 0.2251
11 CUNI EN Run.6 0.3779 0.3224 0.3152 66 baseline run.4 0.3196 0.2291 0.2323
12 CUNI EN Run.3 0.3650 0.3218 0.3110 67 Miracl EN Run.5 0.2982 0.2262 0.2357
13 readability run.2 0.3756 0.3154 0.3117 68 UBML EN Run.4 0.2953 0.2255 0.2300
14 readability run.1 0.3675 0.3140 0.3064 69 FDUSGInfo EN Run.2 0.2757 0.2237 0.2252
15 ECNU EN Run.4 0.3638 0.3103 0.2990 70 UBML EN Run.5 0.2960 0.2220 0.2279
16 ECNU EN Run.1 0.3549 0.3080 0.2971 71 YorkU EN Run.3 0.3074 0.2216 0.2300
17 readability run.3 0.3390 0.3067 0.2929 72 USST EN Run.3 0.3148 0.2181 0.2336
18 CUNI EN Run.8 0.3842 0.3060 0.3102 73 UBML EN Run.10 0.2658 0.2125 0.2159
19 CUNI EN Run.1 0.3824 0.3027 0.3081 74 FDUSGInfo EN Run.3 0.2518 0.2114 0.2087
20 HCMUS EN Run.1 0.3715 0.3017 0.3062 75 USST EN Run.7 0.2726 0.2055 0.2102
21 baseline run.1 0.3567 0.2990 0.2933 76 LIMSI EN run.3 0.2417 0.2036 0.2060
22 ECNU EN Run.2 0.3527 0.2917 0.2830 77 baseline run.6 0.2843 0.2035 0.2143
23 ECNU EN Run.7 0.3548 0.2841 0.2869 78 HCMUS EN Run.3 0.2700 0.2012 0.2089
24 GRIUM EN Run.2 0.3305 0.2809 0.2768 79 USST EN Run.4 0.2815 0.1978 0.2110
25 UBML EN Run.7 0.3339 0.2795 0.2772 80 LIMSI EN run.1 0.2296 0.1929 0.1889
26 GRIUM EN Run.6 0.3306 0.2791 0.2761 81 KUCS EN Run.2 0.2562 0.1818 0.1906
27 GRIUM EN Run.5 0.3278 0.2780 0.2744 82 LIMSI EN run.2 0.2163 0.1815 0.1774
28 GRIUM EN Run.4 0.3244 0.2778 0.2719 83 USST EN Run.5 0.2540 0.1746 0.1890
29 GRIUM EN Run.3 0.3296 0.2775 0.2745 84 Miracl EN Run.3 0.2200 0.1698 0.1698
30 GRIUM EN Run.7 0.3272 0.2774 0.2739 85 USST EN Run.6 0.2410 0.1633 0.1771
31 ECNU EN Run.5 0.3531 0.2771 0.2804 86 Miracl EN Run.2 0.2291 0.1589 0.1626
32 UBML EN Run.3 0.3358 0.2757 0.2789 87 baseline run.5 0.2226 0.1530 0.1610
33 UBML EN Run.1 0.3294 0.2745 0.2771 88 KUCS EN Run.3 0.1679 0.1514 0.1425
34 baseline run.3 0.3369 0.2736 0.2751 89 Miracl EN Run.4 0.2001 0.1507 0.1570
35 GRIUM EN Run.1 0.3249 0.2725 0.2700 90 USST EN Run.8 0.2246 0.1492 0.1595
36 UBML EN Run.2 0.3305 0.2709 0.2735 91 HCMUS EN Run.4 0.2099 0.1467 0.1582
37 KISTI EN RUN.7 0.3299 0.2703 0.2739 92 HCMUS EN Run.5 0.1861 0.1299 0.1386
38 KISTI EN RUN.5 0.3203 0.2702 0.2725 93 HCMUS EN Run.7 0.1853 0.1266 0.1348
39 USST EN Run.2 0.3557 0.2659 0.2727 94 YorkU EN Run.1 0.1798 0.1127 0.1195
40 KISTI EN RUN.4 0.3306 0.2644 0.2709 95 USST EN Run.9 0.1629 0.1115 0.1195
41 baseline run.2 0.3150 0.2633 0.2587 96 readability run.4 0.1143 0.1080 0.1000
42 KISTI EN RUN.6 0.3332 0.2607 0.2695 97 USST EN Run.10 0.1467 0.0947 0.1039
42 KISTI EN RUN.8 0.3332 0.2607 0.2695 98 HCMUS EN Run.6 0.1257 0.0746 0.0861
42 KISTI EN RUN.2 0.3038 0.2607 0.2614 99 HCMUS EN Run.8 0.1210 0.0698 0.0808
45 KISTI EN RUN.3 0.3295 0.2596 0.2666 100 FDUSGInfo EN Run.6 0.0805 0.0609 0.0577
46 KISTI EN RUN.1 0.3222 0.2593 0.2646 100 FDUSGInfo EN Run.7 0.0805 0.0609 0.0577
47 FDUSGInfo EN Run.1 0.3134 0.2572 0.2568 100 FDUSGInfo EN Run.8 0.0805 0.0609 0.0577
48 USST EN Run.1 0.3342 0.2564 0.2639 103 KUCS EN Run.4 0.0656 0.0600 0.0567
49 HCMUS EN Run.2 0.3483 0.2556 0.2698 104 LIMSI EN run.4 0.0562 0.0476 0.0462
50 Miracl EN Run.1 0.3287 0.2546 0.2631 104 LIMSI EN run.5 0.0562 0.0476 0.0462
51 YorkU EN Run.8 0.3072 0.2504 0.2533 106 FDUSGInfo EN Run.9 0.0646 0.0473 0.0473
52 YorkU EN Run.7 0.3125 0.2470 0.2523 107 FDUSGInfo EN Run.10 0.0646 0.0473 0.0473
52 YorkU EN Run.9 0.2962 0.2470 0.2485 108 readability run.5 0.0362 0.0160 0.0227
54 CUNI EN Run.10 0.3060 0.2442 0.2459 109 readability run.6 0.0194 0.0117 0.0134
55 YorkU EN Run.10 0.2853 0.2415 0.2420 109 readability run.7 0.0194 0.0117 0.0134
Table 7. Results for multilingual submissions, sorted by p@10, obtained using the
original qrels.
R Run Name p@10 nDCG@10 R Run Name p@10 nDCG@10
1 CUNI DE Run10 0.2985 0.2825 34 CUNI IT Run5 0.2182 0.1856
2 CUNI DE Run7 0.2970 0.2757 37 CUNI AR Run1 0.2167 0.2117
3 CUNI FR Run10 0.2833 0.2615 37 CUNI AR Run7 0.2167 0.2133
4 CUNI FR Run7 0.2773 0.2568 39 CUNI CS Run8 0.2152 0.2137
5 CUNI IT Run10 0.2758 0.2369 39 CUNI FR Run1 0.2152 0.2056
6 CUNI IT Run1 0.2652 0.2278 39 CUNI PT Run2 0.2152 0.2227
7 CUNI IT Run4 0.2621 0.2221 42 CUNI FA Run6 0.2136 0.2107
8 CUNI PT Run6 0.2530 0.2492 43 CUNI CS Run1 0.2121 0.1924
9 CUNI PT Run8 0.2515 0.2382 43 CUNI DE Run1 0.2121 0.1969
10 CUNI DE Run8 0.2500 0.2413 43 CUNI IT Run7 0.2121 0.1812
10 CUNI FR Run9 0.2500 0.2188 43 CUNI PT Run3 0.2121 0.2253
12 CUNI FR Run8 0.2455 0.2271 47 CUNI FA Run7 0.2091 0.1806
12 CUNI IT Run6 0.2455 0.2142 48 CUNI CS Run5 0.2076 0.1958
14 CUNI DE Run9 0.2409 0.2107 48 CUNI FR Run3 0.2076 0.1943
14 CUNI PT Run10 0.2409 0.2451 48 CUNI FR Run5 0.2076 0.2017
16 CUNI IT Run2 0.2394 0.1913 51 CUNI FR Run4 0.2061 0.2074
17 CUNI IT Run3 0.2348 0.1952 52 CUNI AR Run8 0.2045 0.2026
17 CUNI PT Run7 0.2348 0.2266 52 CUNI DE Run5 0.2045 0.1940
19 CUNI IT Run8 0.2333 0.2105 54 CUNI AR Run4 0.2030 0.1966
20 CUNI CS Run10 0.2303 0.1926 54 CUNI AR Run9 0.2030 0.1768
20 CUNI FA Run10 0.2303 0.2277 54 CUNI CS Run6 0.2030 0.1605
20 CUNI PT Run1 0.2303 0.2338 57 CUNI DE Run4 0.2015 0.1869
20 CUNI PT Run5 0.2303 0.2180 58 CUNI DE Run3 0.2000 0.1652
24 CUNI PT Run4 0.2288 0.2352 59 CUNI FA Run9 0.1985 0.1735
25 CUNI AR Run10 0.2273 0.2202 60 CUNI FR Run6 0.1970 0.1661
25 CUNI FA Run4 0.2273 0.2267 61 CUNI CS Run9 0.1924 0.1530
25 CUNI IT Run9 0.2273 0.1856 62 CUNI CS Run4 0.1894 0.1721
28 CUNI FA Run1 0.2258 0.2227 63 CUNI AR Run2 0.1879 0.1831
29 CUNI CS Run7 0.2242 0.1897 63 CUNI FR Run2 0.1879 0.1854
30 CUNI FA Run3 0.2227 0.2049 63 CUNI PT Run9 0.1879 0.1719
30 CUNI FA Run5 0.2227 0.1991 66 CUNI AR Run3 0.1864 0.1894
32 CUNI AR Run5 0.2197 0.2148 67 CUNI CS Run3 0.1848 0.1609
32 CUNI AR Run6 0.2197 0.2017 68 CUNI DE Run6 0.1818 0.1485
34 CUNI FA Run2 0.2182 0.2087 69 CUNI CS Run2 0.1697 0.1470
34 CUNI FA Run8 0.2182 0.2201 69 CUNI DE Run2 0.1697 0.1517
Table 8. Results for multilingual submissions, sorted by p@10, obtained using addi-
tional qrels (merged ).
R Run Name p@10 nDCG@10 R Run Name p@10 nDCG@10
1 CUNI IT Run10 0.3727 0.3094 36 CUNI FR Run2 0.3061 0.2498
1 CUNI IT Run4 0.3727 0.3045 37 CUNI FA Run5 0.3045 0.2539
3 CUNI IT Run1 0.3712 0.3065 38 CUNI AR Run5 0.3030 0.2661
4 CUNI FR Run10 0.3682 0.3111 38 CUNI CS Run3 0.3030 0.2259
4 CUNI FR Run7 0.3682 0.3093 38 CUNI CS Run4 0.3030 0.2343
6 CUNI IT Run6 0.3606 0.2981 41 CUNI CS Run6 0.3000 0.2206
7 CUNI PT Run2 0.3576 0.3009 41 CUNI FA Run8 0.3000 0.2669
8 CUNI DE Run10 0.3561 0.3182 43 CUNI DE Run8 0.2985 0.2672
9 CUNI DE Run7 0.3545 0.3092 43 CUNI PT Run7 0.2985 0.2613
10 CUNI IT Run8 0.3515 0.2966 45 CUNI AR Run10 0.2924 0.2556
11 CUNI PT Run1 0.3500 0.2936 45 CUNI AR Run7 0.2924 0.2569
12 CUNI PT Run4 0.3485 0.2976 45 CUNI CS Run8 0.2924 0.2544
13 CUNI IT Run2 0.3424 0.2683 45 CUNI DE Run9 0.2924 0.2397
14 CUNI IT Run3 0.3394 0.2694 45 CUNI PT Run9 0.2924 0.2255
15 CUNI PT Run10 0.3379 0.2936 50 CUNI CS Run5 0.2909 0.2426
16 CUNI PT Run3 0.3364 0.2893 51 CUNI AR Run6 0.2894 0.2392
17 CUNI FA Run10 0.3333 0.2807 52 CUNI AR Run8 0.2879 0.2493
17 CUNI FR Run9 0.3333 0.2677 52 CUNI FR Run5 0.2879 0.2446
17 CUNI PT Run6 0.3333 0.2893 54 CUNI CS Run9 0.2864 0.2122
20 CUNI CS Run1 0.3318 0.2633 54 CUNI FA Run6 0.2864 0.2504
20 CUNI CS Run7 0.3318 0.2571 56 CUNI FA Run7 0.2803 0.2256
20 CUNI IT Run5 0.3318 0.2666 56 CUNI FR Run6 0.2803 0.2070
20 CUNI IT Run7 0.3318 0.2654 58 CUNI DE Run1 0.2773 0.2327
20 CUNI PT Run8 0.3318 0.2838 59 CUNI DE Run4 0.2742 0.2255
25 CUNI CS Run10 0.3288 0.2567 60 CUNI AR Run1 0.2727 0.2403
26 CUNI FA Run4 0.3273 0.2788 60 CUNI CS Run2 0.2727 0.2058
26 CUNI FR Run3 0.3273 0.2612 60 CUNI FA Run9 0.2727 0.2101
28 CUNI FA Run1 0.3258 0.2720 63 CUNI DE Run3 0.2682 0.2039
29 CUNI FA Run3 0.3242 0.2660 64 CUNI AR Run2 0.2621 0.2178
30 CUNI FA Run2 0.3227 0.2674 64 CUNI AR Run4 0.2621 0.2237
31 CUNI FR Run4 0.3182 0.2661 64 CUNI DE Run5 0.2621 0.2211
31 CUNI FR Run8 0.3182 0.2659 67 CUNI AR Run3 0.2591 0.2261
31 CUNI PT Run5 0.3182 0.2717 68 CUNI DE Run2 0.2485 0.1984
34 CUNI FR Run1 0.3121 0.2557 69 CUNI AR Run9 0.2439 0.1954
35 CUNI IT Run9 0.3106 0.2417 70 CUNI DE Run6 0.2364 0.1811