=Paper=
{{Paper
|id=Vol-1171/CLEF2005wn-ImageCLEF-GrubingerEt2005
|storemode=property
|title=Towards a Topic Complexity Measure for Cross Language Image Retrieval
|pdfUrl=https://ceur-ws.org/Vol-1171/CLEF2005wn-ImageCLEF-GrubingerEt2005.pdf
|volume=Vol-1171
|dblpUrl=https://dblp.org/rec/conf/clef/GrubingerLC05a
}}
==Towards a Topic Complexity Measure for Cross Language Image Retrieval==
https://ceur-ws.org/Vol-1171/CLEF2005wn-ImageCLEF-GrubingerEt2005.pdf
Towards a Topic Complexity Measure for
Cross-Language Image Retrieval
Michael Grubinger1, Clement Leung1, Paul Clough2
1
School of Computer Science and Mathematics, Victoria University, Melbourne, Australia
michael.grubinger@research.vu.edu.au, clement@matilda.vu.edu.au
2
Department of Information Studies, Sheffield University, Sheffield, UK
p.d.clough@sheffield.ac.uk
Abstract. Selecting suitable topics in order to assess system effectiveness is a
crucial part of any benchmark, particularly those for retrieval systems. This in-
cludes establishing a range of example search requests (or topics) in order to
test various aspects of the retrieval systems under evaluation. In order to assist
with selecting topics, we present a measure of topic complexity for cross-
language image retrieval. This measure has enabled us to ground the topic gen-
eration process within a methodical and reliable framework for ImageCLEF
2005. This document describes such a measure for topic complexity, providing
concrete examples for every aspect of topic complexity and an analysis of top-
ics used in the ImageCLEF 2003, 2004 and 2005 ad-hoc task.
1 Introduction
Benchmarks for image retrieval consist of four main elements: a collection of still
natural images like [1] or [2]; a representative set of search requests (called queries or
topics); a recommended set of performance measures carried out on ground truths
associated with topics [3], [4]; and benchmarking events like [5] and [6] that attract
participants to make use of the benchmark.
The topic selection process is a very important part of any benchmarking event. In
order to produce realistic results, the topics should not only be representative of the
(image) collection, but also reflect realistic user interests/needs [7]. This is achieved
by generating the topics against certain dimensions, including the estimated number
of relevant images for each topic, the variation of task parameters to test different
translation problems, its scope (e.g. broad or narrow, general or specific), and the
difficulty of the topic (topic complexity).
Hence, as the types of search request issued by users of visual information sys-
tems will vary in difficulty (or complexity), a dimension of complexity with respect
to linguistic complexity for translation would help to set the context. Thus, there is a
need for a measure of topic complexity that expresses the level of difficulty for re-
trieval systems to return relevant images in order to ground the topic generation proc-
ess within a methodical and reliable framework.
As image retrieval algorithms improve, it is necessary to increase the average
complexity level of topics each year in order to maintain the challenge for returning
participants. However, if topics are too difficult for current techniques the results are
not particularly meaningful. Furthermore, it may prove difficult for new participants
to obtain good results and prevent them from presenting results and taking part in
comparative evaluations (like ImageCLEF). Providing a good variation in topic com-
plexity is therefore very important as it allows both the organizers (and participants)
to observe retrieval effectiveness with respect to complexity level.
Quantification of task difficulty is not a new concept; on the contrary, it has been
applied to many areas including information retrieval [8], machine learning [9], pars-
ing and grammatical formalisms [10], and language learning in general [11]. More
recent papers include the discussion of syntactic complexity in multimedia informa-
tion retrieval [12] and a measure of semantic complexity for natural language systems
[13]. However, none of this research deals with the definition of a topic complexity
measure for cross-language image retrieval.
This paper describes such a measure for topic complexity. Section 2 gives a short
overview, examples for each aspect of topic complexity are given in sections 3
(nouns), 4 (verbs) and 5 (adjuncts). Section 6 classifies and analyses the topics used
at the ImageCLEF ad-hoc tasks from 2003 to 2005. Finally, section 7 outlines further
improvement of the complexity measure and other future work.
2 Overview of a Measure for Topic Complexity
The first version of the proposed scale for topic complexity starts at 0 and is unlim-
ited as far as query difficulty is concerned. Expressed as a positive integer, the higher
the value c the higher is the topic complexity.
0≤c<∞ (1)
A value of zero implies that no translation is necessary and a simple keyword
search would suffice for effective retrieval. An example for such a topic would be
"David Beckham, 2005" as David Beckham is the same in every language1, and so is
the number 2005.
Each of the following topic elements adds one point to topic complexity value for
(cross-language) retrieval of images with complex image contents using text-based
search requests:
• nouns (used as subject, direct object, indirect object or in other cases, for ex-
ample genitive)
• qualifying attributes of nouns (adjectives)
• noun cardinality (grammatical number)
• verbs and qualifying attributes of verbs (adverbs)
• time, place, manner and reason adjuncts
1 We consider only languages that use some sort of alphabet (Latin, Cyrillic, Greek, etc.) and
exclude sign-based languages like Chinese etc.
In cross-language image retrieval, points are just added to the complexity level if
a translation for that specific topic part is necessary (see examples hereinafter). No
points are added for:
• meta-data like authors/photographers or the date of the picture
• any other part of the sentence that does not require any translation
Complexity points are accumulative: each of the elements can occur more than once
and thus add more than one point (e.g. a topic can easily have two adjectives, like
"traditional Scottish dancers"). However, logical OR constructs do not increase the
complexity level if they could be expressed differently (for example: boys or girls is
the same as children). The next three sections describe each of the topic elements in
detail and provide example complexity scores.
3 Nouns
A word or phrase that refers to a person, place, thing, event, substance or quality is
referred to as a noun (or noun substantive). Nouns can be classified in concrete nouns
and abstract nouns. Concrete nouns refer to definite objects (e.g. racket, ball),
whereas abstract nouns refer to ideas or concepts (e.g. fairness, freedom). In cross-
language image retrieval, just concrete nouns should be used. Further, nouns can be
proper nouns (e.g. "Michael"), common nouns (e.g. "boy"), or collective nouns (e.g.
"team") and sees the use of all three types.
In topics, nouns can occur in several different cases: as subjects (performers of ac-
tion), direct and indirect objects (recipients of action) and in the genitive case (indi-
cates possession). Nouns in topics can further be described by the use of adjectives
and have certain cardinality.
3.1 Subjects (Nominative Case).
The subject of a verb is the argument which generally refers to the origin of the ac-
tion. In languages where a passive voice exists, the subject of a passive verb may be
the target or result of the action. Passive voice should not be used in topic sentences
(see also 4.1).
Each noun used as a subject (e.g. in the nominative case) increments the topic
complexity level by one point.
Turtle eating leaf.
Tortuga comiendo hoja.
Schildkröte frisst Blatt.
Topic Complexity c = 3
(Subject, verb, direct object)
Fig. 1. Topic complexity example for subjects. Note that articles are omitted in all examples as
they are usually omitted in typical user search requests too.
3.2 Direct Objects (Accusative Case):
Objects represent the target of the verb's action. In many languages, the accusative
case of a noun is, generally, the case used to mark the direct object of a verb. The
accusative case exists (or existed once) in all the Indo-European languages (including
Latin, Sanskrit, Greek, German, Russian), in the Finno-Ugric languages, and in Se-
mitic languages (such as Arabic). In modern English, which lacks declension in its
nouns, objects are marked by their position in the sentence or using appositions (like
"to" in "I gave a book to him").
Each noun used as a direct object (in the accusative case) increments the topic
complexity by one point:
Man riding bicycle.
Hombre yendo a bicicleta.
Mann fährt Fahrrad.
Topic Complexity c = 3
(Subject, verb, direct object)
Fig. 2. Topic complexity example for direct objects.
3.2 Indirect Objects (Dative Case):
The dative case is a grammatical case for nouns and generally marks the indirect
object of a verb. Languages that use the dative case include Czech, Dutch, German,
Hungarian, Icelandic, Latin, Latvian, Lithuanian, Polish, Romanian, Russian, Serbian,
Croatian, Slovak, and Slovenian. In current English usage, the indirect object of an
action is sometimes expressed with a prepositional phrase of "to" or "for", though an
objective pronoun can also be placed directly after the main verb and used in a dative
manner, provided that the verb has a direct object as well; for example, "the soccer
referee shows a red card to David Beckham" can also be phrased as "the soccer refe-
ree shows David Beckham a red card".
Each noun used as an indirect object (in the dative case) increments the topic
complexity by one point:
Soccer referee showing card to soccer player.
Arbitro mostrando una tarjeta al futbolista.
Schiedsrichter zeigt Fußballspieler eine Karte.
Topic Complexity c = 4
(Subject, verb, direct object, indirect object)
Fig. 3. Topic complexity example for indirect objects
3.4 Nouns in Genitive Case (or other cases)
The genitive case is a grammatical case that indicates a relationship, primarily one of
possession, between the noun in the genitive case and another noun. In English, this
relation can be expressed by the use of the preposition of ("Lord of the Rings") or by
the possessive -'s ending ("Schindler's List"). Several languages have real genitive
cases, including Arabic, Latin, Irish, Greek, German, Dutch, Russian, and Finnish.
Each noun used in the genitive case (or in any other grammatical case that was not
mentioned here) increments the topic complexity by one point:
Man kissing woman's hand.
Hombre besando la mano de una mujer.
Mann küsst Hand einer Frau.
Topic Complexity c = 4
(Subject, verb, genitive case, direct object)
Fig. 4. Topic complexity example for the genitive case.
3.5 Qualifying Attributes of Nouns (Adjectives)
An adjective is a part of a sentence which modifies a noun, making its meaning more
specific. Adjectives can be used in a predicative (the sky is blue) or attributive man-
ner (the blue sky). In some languages e.g. the Germanic languages (like German,
English, etc), attributive adjectives precede the noun. In other languages, e.g. the
Romance languages (like Spanish), the adjective follows the noun. Some languages
do not even have any adjectives, for example Chinese (all the words that are trans-
lated into English as adjectives are actually stative verbs).
Each adjective used in a topic sentence increases the topic complexity by one
point. Just attributive adjectives should be used.
Austrian soccer referee showing red card to
Portuguese soccer player.
Österreichischer Schiedsrichter zeigt por-
tugiesischem Fußballspieler die rote Karte.
Arbitro austríaco mostrando la tarjeta roja a un
futbolista portugués.
Topic Complexity c = 7
(adjective, noun, verb, adjective, direct object,
adjective, indirect object)
Fig. 5. Topic complexity example for adjectives (compare to complexity of Fig. 3).
This example shows that a simple keyword search would not deliver good results
anymore. Since keyword search does not associate attributes to the according nouns,
it might as well return images showing: Portuguese soccer referee shows card to red
Austrian soccer player or Austrian Soccer player shows card to red, Portuguese
soccer referee etc.
3.6 Noun Cardinality (Grammatical Number, Numerals)
In linguistics, the grammatical number specifies the quantity of a noun or affects the
form of a verb or other part of speech depending on the quantity of the noun to which
it refers. Grammatical number is distinct from the use of numerals to specify the exact
quantity of a noun.
Topics querying a specific number of a noun provide a special challenge and
therefore is awarded one point for the topic complexity level.
Seven zebras drinking water.
Siete cebras bebiendo agua.
Sieben Zebras trinken Wasser.
Topic Complexity c = 4
(cardinality, subject, verb, direct object)
Fig. 6. Topic complexity example for noun cardinality.
4 Verbs
A verb is a part of a sentence that usually denotes action ("kick"), occurrence ("to
shine"), or a state of being ("stand"). Depending on the language, a verb may vary in
form according to many factors, including its tense, aspect, mood and voice. Verbs
can further be described by the use of adverbs.
4.1 Topic Verbs
Topics should just use verbs that clearly describe the situation in an image (like run-
ning, jumping, painting, hitting, and so on). Verbs or composite verb groups that need
some level of interpretation (e.g. finding, forgetting, trying to hit, attempting to es-
cape, etc.) are not used (see also [14]).
Verbs should be used in active voice only as passive voice does not exist in all
languages. Further, since the captions describe an action that is happening in the
image (at that moment), the grammatically correct form for English is the present
continuous tense (the auxiliary verb to be is omitted). In Spanish, the appropriate
tense is "el presente continuo" (present continuous tense), whereas in the German
language, actions that are happening at the time are expressed with the "Präsens"
(present tense).
Man pushing car in winter.
Hombre empujando coche en invierno.
Mann schiebt Auto im Winter.
Topic Complexity c = 4
(subject, verb, direct object, time adjunct)
Fig. 7. Topic complexity example for verbs. Auxiliary verbs (to be, estar) are omitted as they
are also typically omitted in real user requests.
4.2 Qualifying Attributes of Verbs (Adverbs)
An adverb is a part of a sentence that serves to modify verbs, adjectives, other ad-
verbs, and clauses. Each adverb used in a topic sentence increases the topic complex-
ity level by one point:
Tennis player hitting ball hard.
Tenista golpeando la pelota fuerte.
Tennisspielerin schlägt Ball hart.
Topic Complexity c = 4
(subject, verb, direct object, adverb)
Fig. 8. Topic complexity example for adverbs.
Topics should just use adverbs that modify verbs. Adverbs that modify adjectives
or other adverbs (adverbs of degree) are felt to be too subjective and should not be
used for cross-language image retrieval of complex image contents. Adverbs of de-
gree tell us about the intensity or degree of an action, an adjective or another adverb.
Examples for adverbs of degree are: almost, nearly, quite, just, too, enough, hardly,
scarcely, completely, very, extremely. For example: "Soccer player kicking ball very
hard" or "extremely tall boy playing basketball".
One might argue that even the example given above may be very subjective and
that adverbs should not increase the topic complexity level at all. However, as long as
adverbs clearly influence the result set, they should be considered as a factor. In the
example above, relevant images would include any tennis player driving, serving or
smashing the ball or hitting the ball with topspin. Tennis players slicing the ball,
playing a drop-volley or a stop-ball would not be relevant. This can clearly be seen by
the technique of the tennis player. Hence, the adverb "hard" increases the topic com-
plexity by one point.
4.3 Valency
The number of arguments that a verb takes is called its valency. According to its
valency, a verb can be classified as:
• Intransitive (valency = 1): the verb only has a subject. For example: "people
marching".
• Transitive (valency = 2): the verb has a subject and a direct object. For example:
"golfers swinging their clubs".
• Ditransitive (valency = 3): the verb has a subject, a direct object and an indirect
or secondary object. For example: "referee showing red card to soccer player".
It is possible to have verbs with valency = 0. A few of these appear in Spanish,
Italian and other languages and are called impersonal verbs. For example: "Llueve"
(Spanish) or "Piove" (Italian), which both mean "it rains".
Further, all languages are generally assumed to have a basic word order. Table 1
shows all possible word orders for the subject, verb, and object (in the order of the
most common to the rarest).
Table 1: Word order catagories and examples.
Rk Word Order Example Languages
1 S-O-V languages Turkish, Japanese, Korean, Latin, most Indian languages
2 S-V-O languages English, Spanish, Italian, Kiswahili, Chinese, French
3 V-S-O languages Arabic, Welsh, Gaelic
4 V-O-S languages Fijian
5 O-S-V languages Xavante
6 O-V-S languages Guajiro, Hixkaryana, Klingon
For topics with valency 2 or higher, a simple keyword search is in some cases not
sufficient any more as it can't detect grammatical relationships between search words.
Searching for an image of "Boy chasing dog" or "Boy giving girl a candy" a simple
keyword search would also return "Dog chasing boy" or "Girl giving boy a candy".
In this case, due to the increased difficulty of actually having to distinguish between
subject, direct and indirect object (which can't be done with the position of the noun
since the word order can be different in many languages), the complexity level is
increased by one point for verbs with a valency higher than 2.
5 Adjuncts
An adjunct is a type of adverbial illustrating the circumstances of the action. It ex-
presses such relations as time, manner, place, and reason, i.e. it answers the questions:
where (place adjuncts), when (time adjuncts), how (manner adjuncts) and why (ad-
juncts of reason).
5.1 Time Adjuncts
Time adjuncts indicate when an action happened. Topics will use prepositional
phrases as time adjuncts in order to refer to time. Example:
Man riding bicycle at night.
Hombre yendo a bicicleta por la noche.
Mann fährt Fahrrad in der Nacht.
Topic Complexity c = 4
(subject, verb, object, time adjunct)
Fig. 9. Topic complexity example for time adjuncts.
If the time element does not need any form of translation, the complexity level is
not increased, for example:
Tay Bridge Rail Disaster, 1879
Desastre del tren Tay Bridge, 1879
Das Tay Bridge Zugsunglück, 1879
Topic Complexity c = 1
(abstract noun)
Fig. 10. Topic complexity example for time adjuncts that do not increase the topic complexity
level as no translation is necessary for 1879.
5.2 Place Adjuncts
Place adjuncts indicate the location where the image was taken or where the action
occurred respectively. Topics will use prepositional phrases as place adjuncts. As
most of the countries, cities or other places have a different name in different lan-
guages, there will always be some sort of translation involved, thus the complexity
level is incremented by one.
Boat in Northern Ireland.
Barco en Irlanda del Norte.
Boot in Nordirland.
Topic Complexity c = 2
(noun, place adjunct)
Fig. 11. Topic complexity example for place adjuncts
5.3 Manner Adjuncts
Manner adjuncts further describe nouns or how actions are performed in an image. In
comparison to general adverbs that modify verbs (see 3.2.2), now we just talk about
prepositional phrases that describe how actions were performed.
Each manner adjunct increases the topic complexity by one point:
Woman in bikini leading horse with reins.
Mujer en bikini llevando un caballo con las riendas.
Frau in Bikini führt Pferd an den Zügeln.
Topic Complexity c = 6
(subject, manner adjunct, verb, direct object, manner
adjunct)
Fig. 12. Topic complexity example for manner adjuncts.
5.4 Adjunct of Reason
Each topic sentence can contain a reason adjunct which describes why actions are
taken. Each reason adjunct increases the topic complexity level by one point:
Bomb damage due to World War II.
Bombenschäden durch den zweiten Weltkrieg.
Daños por las bombas en la Segunda Guerra
Mundial.
Topic Complexity c = 2
(noun, reason adjunct)
Fig. 13. Topic complexity example for adjuncts of reason.
Note that the English word why does not only refer to reason adjuncts (due to, be-
cause of) but also to purpose adjuncts (in order to). Reason adjuncts like "due to
World War II" can be determined with some prior knowledge of the image, but pur-
pose adjuncts refer to the future and can just be assumed.
Purpose adjuncts are felt to be too subjective as it involves too much interpreta-
tion of the picture and are therefore not considered for the topic complexity measure.
6 Topic Complexity at ImageCLEF
The ImageCLEF retrieval benchmark was established in 2003 with the aim of evalu-
ating image retrieval from multilingual document collections [5][6]. This section
presents the results of the new topic complexity measure applied to the 2003, 2004
and 2005 ad-hoc ImageCLEF tasks.
6.1 Topic Complexity at ImageCLEF 2005
In the ImageCLEF 2005 ad-hoc task [15], the participants were provided with 28
topics translated into 33 different languages. Table 2 shows the analysis of the topic
complexity for each of the topic titles in English.
Table 2: Topic complexity analysis for English topic titles
ID Topic Title Topic Analysis c
1 aircraft on the ground noun, place adjunct 2
2 people gathered at bandstand noun, verb, place adjunct 3
3 dog (in) sitting (position) noun, verb 2
4 steam ship docked noun, noun, verb 3
5 animal statue noun, noun 2
6 small sailing boat adjective, noun 2
7 fishermen in boat noun, place adjunct 2
8 building covered in snow noun, verb, manner adjunct 3
9 horse pulling cart or carriage noun, verb, direct object (or direct object) 3
10 sun pictures, Scotland noun, place adjunct 2
11 Swiss mountain (scenery) adjective, noun 2
12 postcard from Iona, Scotland noun, place adjunct 2
13 stone viaduct with several arches noun, manner adjunct 2
14 people at the marketplace noun, place adjunct 2
15 golfer putting on green noun, verb, place adjunct 3
16 waves (breaking) on beach noun, place adjunct 2
17 man or woman reading noun (or noun), verb 2
18 woman in white dress noun, adjective, manner adjunct 3
19 composite postcards of Northern Ireland adjective, noun, place adjunct, adjective 4
20 royal visit to Scotland (not Fife) adjective, noun, place adjunct, exclusion 4
21 monument to Robert Burns noun 1
22 building with waving flag noun, manner adjunct, adjective 3
23 tomb inside church or cathedral noun, place adjunct (or place adjunct) 2
24 close-up pictures of bird noun, genitive noun 2
25 arched gateway adjective, noun 2
26 portrait pictures of mixed-sex groups noun, adjective, genitive noun 3
27 woman or girl carrying basket noun (or noun), verb, direct object 3
28 colour pictures of woodland scenes around adjective, noun, genitive noun, place adjunct 4
St. Andrews
Likewise, the complexity levels have been calculated for all alphabetical lan-
guages (Romanic alphabet) with more than 10 submitted runs (e.g. European and
Latin-American Spanish, Italian, German, French, Dutch, Portuguese) as same con-
cepts are sometimes expressed with different topic complexities across various lan-
guages.
A total of 11 research groups submitted 349 runs and produced the following
Mean Average Precision scores for each topic (Table 3, next page).
Table 3: Average MAP (Mean Average Precision) values for alphabetical languages with more
than 10 submitted runs (with their topic complexity in parenthesis)
ID ENG GER SPA – L SPA - E ITA FRA POR NED ALL
1 0.26 (2) 0.00 (2) 0.04 (2) 0.11 (2) 0.12 (2) 0.28 (2) 0.00 (2) 0.20 (2) 0.13 (2.00)
2 0.46 (3) 0.03 (3) 0.00 (3) 0.02 (3) 0.00 (3) 0.07 (4) 0.24 (3) 0.00 (2) 0.12 (3.00)
3 0.43 (2) 0.39 (3) 0.26(2) 0.26 (2) 0.26 (2) 0.43 (2) 0.29 (2) 0.44 (2) 0.35 (2.13)
4 0.28 (3) 0.20 (2) 0.18 (3) 0.16 (3) 0.04 (3) 0.11 (3) 0.03 (3) 0.10 (2) 0.15 (2.63)
5 0.70 (2) 0.71 (1) 0.68 (2) 0.70 (2) 0.65 (2) 0.36 (2) 0.77 (2) 0.61 (2) 0.58 (1.75)
6 0.50 (2) 0.49 (2) 0.38 (2) 0.10 (2) 0.36 (2) 0.15 (2) 0.45 (2) 0.48 (2) 0.31 (2.00)
7 0.35 (2) 0.06 (2) 0.31 (2) 0.25 (2) 0.39 (2) 0.31 (2) 0.27 (2) 0.33 (2) 0.26 (2.00)
8 0.08 (3) 0.05 (2) 0.06 (3) 0.06 (3) 0.07 (3) 0.20 (3) 0.07 (3) 0.05 (3) 0.09 (2.88)
9 0.32 (3) 0.23 (3) 0.34 (3) 0.34 (3) 0.17 (3) 0.14 (2) 0.25 (3) 0.45 (3) 0.27 (2.88)
10 0.32 (2) 0.22 (2) 0.26 (3) 0.24 (3) 0.24 (3) 0.28 (3) 0.28 (3) 0.29 (2) 0.24 (2.63)
11 0.50 (2) 0.14 (2) 0.66 (2) 0.20 (2) 0.09 (2) 0.15 (2) 0.10 (2) 0.06 (2) 0.34 (2.00)
12 0.29 (2) 0.30 (2) 0.26 (3) 0.28 (3) 0.32 (3) 0.32 (3) 0.24 (3) 0.31 (2) 0.23 (2.50)
13 0.37 (2) 0.26 (2) 0.27 (3) 0.31 (3) 0.07 (3) 0.27 (3) 0.26 (3) 0.22 (2) 0.26 (2.50)
14 0.13 (2) 0.42 (2) 0.44 (2) 0.45 (2) 0.15 (2) 0.40 (2) 0.74 (2) 0.49 (2) 0.36 (2.00)
15 0.35 (3) 0.15 (3) 0.19 (3) 0.08 (3) 0.13 (3) 0.06 (3) 0.14 (3) 0.16 (3) 0.15 (3.13)
16 0.41 (3) 0.40 (3) 0.33 (3) 0.42 (3) 0.33 (3) 0.43 (3) 0.39 (3) 0.04 (2) 0.30 (2.75)
17 0.47 (2) 0.46 (2) 0.36 (2) 0.07 (2) 0.33 (2) 0.47 (2) 0.55 (2) 0.46 (2) 0.37 (2.00)
18 0.08 (3) 0.08 (3) 0.08 (3) 0.08 (3) 0.04 (3) 0.09 (3) 0.04 (2) 0.11 (3) 0.08 (2.88)
19 0.22 (4) 0.00 (4) 0.00 (4) 0.00 (4) 0.00 (4) 0.00 (4) 0.00 (4) 0.03 (4) 0.05 (4.00)
20 0.06 (4) 0.03 (4) 0.03 (4) 0.03 (4) 0.04 (4) 0.07 (4) 0.05 (4) 0.08 (4) 0.07 (4.00)
21 0.48 (1) 0.44 (1) 0.46 (1) 0.48 (1) 0.46 (1) 0.55 (1) 0.37 (1) 0.43 (1) 0.39 (1.00)
22 0.32 (3) 0.43 (3) 0.39 (3) 0.39 (3) 0.34 (3) 0.29 (3) 0.21 (3) 0.43 (3) 0.36 (3.00)
23 0.48 (2) 0.34 (2) 0.33 (2) 0.06 (2) 0.02 (2) 0.08 (2) 0.26 (2) 0.54 (2) 0.22 (2.00)
24 0.22 (2) 0.25 (2) 0.15 (2) 0.12 (2) 0.16 (2) 0.17 (2) 0.23 (2) 0.26 (2) 0.19 (2.00)
25 0.45 (2) 0.13 (2) 0.07 (2) 0.11 (2) 0.03 (2) 0.38 (2) 0.22 (2) 0.06 (2) 0.19 (2.00)
26 0.53 (3) 0.36 (3) 0.22 (3) 0.15 (3) 0.08 (3) 0.29 (2) 0.10 (3) 0.37 (3) 0.25 (2.88)
27 0.35 (3) 0.28 (3) 0.14 (3) 0.15 (3) 0.21 (3) 0.29 (3) 0.08 (3) 0.33 (3) 0.22 (3.00)
28 0.13 (4) 0.13 (3) 0.12 (4) 0.10 (4) 0.10 (3) 0.12 (4) 0.09 (4) 0.15 (3) 0.11 (3.63)
In order to establish the existence of a relation between the level of complexity
and results obtained from ImageCLEF submissions, the correlation coefficient ρx,y is
calculated for each of the languages, using the following formula:
(2)
where
(3)
and
(4)
where X corresponds to the complexity levels for each topic and Y to their respective
results.
Figure 14 shows that a strong negative correlation exists between the level of
topic complexity and MAP of submitted ImageCLEF results (the higher the topic
complexity score, the lower the MAP score).
Topic Complexity Correlation (ImageCLEF 2005)
ENG SPA-E SPA-L ITA GER FRA NED POR ALL
0.0000
-0.1000
-0.2000
-0.3000
Correlation
-0.4000
-0.5000
-0.6000
-0.7000
-0.8000
Language
Fig. 14. Correlation between topic complexity score and MAP for ImageCLEF 2005 submis-
sions.
6.2 Topic Complexity at ImageCLEF 2004
In ImageCLEF 2004 [6], twelve participating groups submitted 190 runs to the ad-
hoc task. Similar to results in section 6.1, the levels of topic complexity were calcu-
lated for all topics and compared with the average MAP results for languages with
more than 10 submissions.
Topic Complexity Correlation (ImageCLEF 2004)
ENG SPA ITA GER FRA NED ALL
0.0
-0.1
-0.2
Correlation
-0.3
-0.4
-0.5
-0.6
-0.7
-0.8
Language
Fig. 15. Topic complexity correlation for ImageCLEF 2004
Figures 15 shows, again, a strong negative correlation. The correlation factor ρx,y
is always stronger than -0.4, reaches more than -0.6 for Italian and German and even
more than -0.7 for French.
6.3 Topic Complexity at ImageCLEF 2003
In ImageCLEF 2003 [5], four participating groups submitted 45 runs to the ad-hoc
task. Similar to results in sections 6.1 and 6.2, the levels of topic complexity were
calculated for all topics and compared with the average MAP results for languages
with more than 5 submissions.
Topic Complexity Correlation (ImageCLEF 2003)
ENG SPA ITA GER FRA ALL
0.0
-0.1
-0.2
Correlation
-0.3
-0.4
-0.5
-0.6
-0.7
-0.8
Language
Fig. 16. Topic complexity correlation for ImageCLEF 2003
The results shown in Figure 16 show a very strong negative correlation again.
Like in 2004 and 2005, the correlation factor ρx,y is always stronger than -0.4 (except
for Italian which is due to a couple of translation problems that produced surprising
results).
7 Conclusion and Future Work
In this paper, we present a measure for the degree of topic complexity for search
requests of cross-language image retrieval. Establishing such a measure is beneficial
when creating benchmarks such as ImageCLEF in that it is possible to categorise
results according to a level of complexity for individual topics. This can help explain
results obtained when using the benchmark and provide some kind of control and
reasoning over topic generation.
Examples illustrating various aspects of the linguistic structure of the complexity
measure and motivating its creation have been presented. Comparing the level of
complexity for topics created in ImageCLEF 2003 to 2005 for the ad-hoc task with
MAP scores from submitted runs by participating groups have shown a strong nega-
tive correlation indicating that more linguistically complex topics result in much
lower MAP scores due to the requirement of more complex translation approaches.
Future work will involve the improvement and refinement of the complexity
measure and further verification by analysing results from the 2006 ImageCLEF ad-
hoc task.
References
1. Clough, P., Sanderson, M., Reid, N.: The Eurovision St Andrews Photographic Collection
(ESTA), Image CLEF Report, University of Sheffield, UK (February 2003).
2. Grubinger, M., Leung, C.: A Benchmark for Performance Calibration in Visual Informa-
tion Search, In Proceedings of 2003 Conference of Visual Information Systems, Miami,
Florida, USA (September 24-26, 2003) 414 – 419.
3. Leung, C., Ip, H.: Benchmarking for Content Based Visual Information Search. In Laurini
ed. Fourth International Conference on Visual Information Systems (VISUAL’2000),
Lecture Notes in Computer Science, Springer Verlag, Lyon, France (November 2000) 442
– 456.
4. Müller, H., Müller, W., Squire, D., Marchand-Millet, S., Pun, T.: Performance Evaluation
in Content Based Image Retrieval: Overview and Proposals. Pattern Recognition Letters,
22(5), (April 2001), 563 – 601.
5. Clough, P., Sanderson, M.: The CLEF 2003 cross language image retrieval task. In Pro-
ceedings of the Cross Language Evaluation Forum (CLEF) 2003, Springer Verlag, (2004).
6. Clough, P., Müller, H., Sanderson, M.: Overview of the CLEF cross-language image re-
trieval track (ImageCLEF) 2004. In C. Peters, P.D. Clough, G. J. F. Jones, J. Gonzalo, M.
Kluck, and B. Magnini, editors, Multilingual Information Access for Text, Speech and Im-
ages: Result of the fifth CLEF evaluation campaign, Lecture Notes in Computer Science,
Springer Verlag, Bath, England (2005).
7. Armitage, L., Enser, P.: Analysis of User Need in Image Archives. Journal of Information
Science, (23 (4) 1997), 287 – 299.
8. Bagga, A., Biermann, A.: Analysing the complexity of a domain with respect to an infor-
mation extraction task. In Proceedings of the tenth International Conference on Research
on Computational Linguistics (ROCLING X), (August 1997), 175 – 194.
9. Niyogi, P.: The Informational Complexity of Learning from Examples. PhD Thesis, MIT.
(1996).
10. Barton, E., Berwick, R., Ristad, E.: Computational Complexity and Natural Language. The
MIT Press, Cambridge, Massachusetts (1987).
11. Ristad, E.: The Language Complexity Games. MIT Press, Cambridge, MA, (1993).
12. Flank, S.: Sentences vs. Phrases: Syntactic Complexity in Multimedia Information Re-
trieval. NAACL-ANLP 2000 Workshop: Syntactic and Semantic Complexity in Natural
Language Processing Systems (2000).
13. Pollard, S., Biermann, A.: A Measure of Semantic Complexity for Natural Language Sys-
tems. NAACL-ANLP 2000 Workshop: Syntactic and Semantic Complexity in Natural
Language Processing Systems.
14. Grubinger, M., Leung, C.: Incremental Benchmark Development and Administration, In
Proceedings of 2004 Conference of Visual Information Systems, Hotel Sofitel Conference
Centre, San Francisco, California, USA (September 8-10, 2004).
15. Clough, P., Desealers, T., Grubinger, M., Hersh, W., Jensen, J., Lehmann, T., Müller, H.:
The 2005 Cross-Language Image Retrieval Track. In these proceedings (2005).