Understanding Discord NSFW Servers A Guide for Responsible Users.pdf
Comparing Cross-Language Retrieval Tools at CLEF-IP
1. Chances and Challenges in Comparing
Cross-Language Retrieval Tools
Giovanna Roda
Vienna, Austria
Irf Symposium 2010 / June 3, 2010
2. CLEF-IP: the Intellectual Property track at CLEF
CLEF-IP is an evaluation track within the Cross Language
Evaluation Forum (Clef). 1
1
http://www.clef-campaign.org
3. CLEF-IP: the Intellectual Property track at CLEF
CLEF-IP is an evaluation track within the Cross Language
Evaluation Forum (Clef). 1
organized by the IRF
1
http://www.clef-campaign.org
4. CLEF-IP: the Intellectual Property track at CLEF
CLEF-IP is an evaluation track within the Cross Language
Evaluation Forum (Clef). 1
organized by the IRF
first track ran in 2009
1
http://www.clef-campaign.org
5. CLEF-IP: the Intellectual Property track at CLEF
CLEF-IP is an evaluation track within the Cross Language
Evaluation Forum (Clef). 1
organized by the IRF
first track ran in 2009
running this year for the second time
1
http://www.clef-campaign.org
6. CLEF-IP: the Intellectual Property track at CLEF
CLEF-IP is an evaluation track within the Cross Language
Evaluation Forum (Clef). 1
organized by the IRF
first track ran in 2009
running this year for the second time
1
http://www.clef-campaign.org
7. What is an evaluation track?
An evaluation track in Information Retrieval is a cooperative action
aimed at comparing different techniques on a common retrieval
task.
8. What is an evaluation track?
An evaluation track in Information Retrieval is a cooperative action
aimed at comparing different techniques on a common retrieval
task.
produces experimental data that can be analyzed and used to
improve existing systems
9. What is an evaluation track?
An evaluation track in Information Retrieval is a cooperative action
aimed at comparing different techniques on a common retrieval
task.
produces experimental data that can be analyzed and used to
improve existing systems
fosters exchange of ideas and cooperation
10. What is an evaluation track?
An evaluation track in Information Retrieval is a cooperative action
aimed at comparing different techniques on a common retrieval
task.
produces experimental data that can be analyzed and used to
improve existing systems
fosters exchange of ideas and cooperation
produces a reusable test collection, sets milestones
11. What is an evaluation track?
An evaluation track in Information Retrieval is a cooperative action
aimed at comparing different techniques on a common retrieval
task.
produces experimental data that can be analyzed and used to
improve existing systems
fosters exchange of ideas and cooperation
produces a reusable test collection, sets milestones
Test collection
A test collection consists traditionally of target data, a set of
queries, and relevance assessments for each query.
12. Clef–Ip 2009: the task
The main task in the Clef–Ip track was to find prior art for a
given patent.
13. Clef–Ip 2009: the task
The main task in the Clef–Ip track was to find prior art for a
given patent.
Prior art search
Prior art search consists in identifying all information (including
non-patent literature) that might be relevant to a patent’s claim of
novelty.
16. Participants - 2009 track
1 Tech. Univ. Darmstadt, Dept. of CS,
Ubiquitous Knowledge Processing Lab (DE)
2 Univ. Neuchatel - Computer Science (CH)
3 Santiago de Compostela Univ. - Dept.
Electronica y Computacion (ES)
17. Participants - 2009 track
1 Tech. Univ. Darmstadt, Dept. of CS,
Ubiquitous Knowledge Processing Lab (DE)
2 Univ. Neuchatel - Computer Science (CH)
3 Santiago de Compostela Univ. - Dept.
Electronica y Computacion (ES)
4 University of Tampere - Info Studies (FI)
18. Participants - 2009 track
1 Tech. Univ. Darmstadt, Dept. of CS,
Ubiquitous Knowledge Processing Lab (DE)
2 Univ. Neuchatel - Computer Science (CH)
3 Santiago de Compostela Univ. - Dept.
Electronica y Computacion (ES)
4 University of Tampere - Info Studies (FI)
5 Interactive Media and Swedish Institute of
Computer Science (SE)
19. Participants - 2009 track
1 Tech. Univ. Darmstadt, Dept. of CS,
Ubiquitous Knowledge Processing Lab (DE)
2 Univ. Neuchatel - Computer Science (CH)
3 Santiago de Compostela Univ. - Dept.
Electronica y Computacion (ES)
4 University of Tampere - Info Studies (FI)
5 Interactive Media and Swedish Institute of
Computer Science (SE)
6 Geneva Univ. - Centre Universitaire
d’Informatique (CH)
20. Participants - 2009 track
1 Tech. Univ. Darmstadt, Dept. of CS,
Ubiquitous Knowledge Processing Lab (DE)
2 Univ. Neuchatel - Computer Science (CH)
3 Santiago de Compostela Univ. - Dept.
Electronica y Computacion (ES)
4 University of Tampere - Info Studies (FI)
5 Interactive Media and Swedish Institute of
Computer Science (SE)
6 Geneva Univ. - Centre Universitaire
d’Informatique (CH)
7 Glasgow Univ. - IR Group Keith (UK)
21. Participants - 2009 track
1 Tech. Univ. Darmstadt, Dept. of CS,
Ubiquitous Knowledge Processing Lab (DE)
2 Univ. Neuchatel - Computer Science (CH)
3 Santiago de Compostela Univ. - Dept.
Electronica y Computacion (ES)
4 University of Tampere - Info Studies (FI)
5 Interactive Media and Swedish Institute of
Computer Science (SE)
6 Geneva Univ. - Centre Universitaire
d’Informatique (CH)
7 Glasgow Univ. - IR Group Keith (UK)
8 Centrum Wiskunde & Informatica - Interactive
Information Access (NL)
22. Participants - 2009 track
9 Geneva Univ. Hospitals - Service of Medical
Informatics (CH)
23. Participants - 2009 track
9 Geneva Univ. Hospitals - Service of Medical
Informatics (CH)
10 Humboldt Univ. - Dept. of German Language
and Linguistics (DE)
24. Participants - 2009 track
9 Geneva Univ. Hospitals - Service of Medical
Informatics (CH)
10 Humboldt Univ. - Dept. of German Language
and Linguistics (DE)
11 Dublin City Univ. - School of Computing (IE)
25. Participants - 2009 track
9 Geneva Univ. Hospitals - Service of Medical
Informatics (CH)
10 Humboldt Univ. - Dept. of German Language
and Linguistics (DE)
11 Dublin City Univ. - School of Computing (IE)
12 Radboud Univ. Nijmegen - Centre for Language
Studies & Speech Technologies (NL)
26. Participants - 2009 track
9 Geneva Univ. Hospitals - Service of Medical
Informatics (CH)
10 Humboldt Univ. - Dept. of German Language
and Linguistics (DE)
11 Dublin City Univ. - School of Computing (IE)
12 Radboud Univ. Nijmegen - Centre for Language
Studies & Speech Technologies (NL)
13 Hildesheim Univ. - Information Systems &
Machine Learning Lab (DE)
27. Participants - 2009 track
9 Geneva Univ. Hospitals - Service of Medical
Informatics (CH)
10 Humboldt Univ. - Dept. of German Language
and Linguistics (DE)
11 Dublin City Univ. - School of Computing (IE)
12 Radboud Univ. Nijmegen - Centre for Language
Studies & Speech Technologies (NL)
13 Hildesheim Univ. - Information Systems &
Machine Learning Lab (DE)
14 Technical Univ. Valencia - Natural Language
Engineering (ES)
28. Participants - 2009 track
9 Geneva Univ. Hospitals - Service of Medical
Informatics (CH)
10 Humboldt Univ. - Dept. of German Language
and Linguistics (DE)
11 Dublin City Univ. - School of Computing (IE)
12 Radboud Univ. Nijmegen - Centre for Language
Studies & Speech Technologies (NL)
13 Hildesheim Univ. - Information Systems &
Machine Learning Lab (DE)
14 Technical Univ. Valencia - Natural Language
Engineering (ES)
15 Al. I. Cuza University of Iasi - Natural Language
Processing (RO)
34. 2009-2010: evolution of the CLEF-IP track
2009
1 task: prior art search
targeting granted patents
15 participants
all from academia
families and citations
manual assessments
standard evaluation mea-
sures
35. 2009-2010: evolution of the CLEF-IP track
2009
1 task: prior art search
targeting granted patents
15 participants
all from academia
families and citations
manual assessments
standard evaluation mea-
sures
36. 2009-2010: evolution of the CLEF-IP track
2009 2010
1 task: prior art search
targeting granted patents
15 participants
all from academia
families and citations
manual assessments
standard evaluation mea-
sures
37. 2009-2010: evolution of the CLEF-IP track
2009 2010
1 task: prior art search prior art candidate search
and classification task
targeting granted patents
15 participants
all from academia
families and citations
manual assessments
standard evaluation mea-
sures
38. 2009-2010: evolution of the CLEF-IP track
2009 2010
1 task: prior art search prior art candidate search
and classification task
targeting granted patents patent applications
15 participants
all from academia
families and citations
manual assessments
standard evaluation mea-
sures
39. 2009-2010: evolution of the CLEF-IP track
2009 2010
1 task: prior art search prior art candidate search
and classification task
targeting granted patents patent applications
15 participants 20 participants
all from academia
families and citations
manual assessments
standard evaluation mea-
sures
40. 2009-2010: evolution of the CLEF-IP track
2009 2010
1 task: prior art search prior art candidate search
and classification task
targeting granted patents patent applications
15 participants 20 participants
all from academia 4 industrial participants
families and citations
manual assessments
standard evaluation mea-
sures
41. 2009-2010: evolution of the CLEF-IP track
2009 2010
1 task: prior art search prior art candidate search
and classification task
targeting granted patents patent applications
15 participants 20 participants
all from academia 4 industrial participants
families and citations include forward citations
manual assessments
standard evaluation mea-
sures
42. 2009-2010: evolution of the CLEF-IP track
2009 2010
1 task: prior art search prior art candidate search
and classification task
targeting granted patents patent applications
15 participants 20 participants
all from academia 4 industrial participants
families and citations include forward citations
manual assessments expanded lists of relevant
docs
standard evaluation mea-
sures
43. 2009-2010: evolution of the CLEF-IP track
2009 2010
1 task: prior art search prior art candidate search
and classification task
targeting granted patents patent applications
15 participants 20 participants
all from academia 4 industrial participants
families and citations include forward citations
manual assessments expanded lists of relevant
docs
standard evaluation mea- new measure: pres, more
sures recall-oriented
44. What are relevance assessments
A test collection (also known as gold standard) consists of a target
dataset, a set of queries, and relevance assessments corresponding
to each query.
45. What are relevance assessments
A test collection (also known as gold standard) consists of a target
dataset, a set of queries, and relevance assessments corresponding
to each query.
The CLEF-IP test collection:
46. What are relevance assessments
A test collection (also known as gold standard) consists of a target
dataset, a set of queries, and relevance assessments corresponding
to each query.
The CLEF-IP test collection:
target data: 2 million EP patents
47. What are relevance assessments
A test collection (also known as gold standard) consists of a target
dataset, a set of queries, and relevance assessments corresponding
to each query.
The CLEF-IP test collection:
target data: 2 million EP patents
queries: full-text patents (without images)
48. What are relevance assessments
A test collection (also known as gold standard) consists of a target
dataset, a set of queries, and relevance assessments corresponding
to each query.
The CLEF-IP test collection:
target data: 2 million EP patents
queries: full-text patents (without images)
relevance assessments: extended citations
50. Relevance assessments
We used patents cited as prior art as relevance assessments.
Sources of citations:
51. Relevance assessments
We used patents cited as prior art as relevance assessments.
Sources of citations:
1 applicant’s disclosure: the Uspto requires applicants to
disclose all known relevant publications
52. Relevance assessments
We used patents cited as prior art as relevance assessments.
Sources of citations:
1 applicant’s disclosure: the Uspto requires applicants to
disclose all known relevant publications
2 patent office search report: each patent office will do a search
for prior art to judge the novelty of a patent
53. Relevance assessments
We used patents cited as prior art as relevance assessments.
Sources of citations:
1 applicant’s disclosure: the Uspto requires applicants to
disclose all known relevant publications
2 patent office search report: each patent office will do a search
for prior art to judge the novelty of a patent
3 opposition procedures: patents cited to prove that a granted
patent is not novel
57. Patent families
A patent family consists of patents granted by different patent
authorities but related to the same invention.
58. Patent families
A patent family consists of patents granted by different patent
authorities but related to the same invention.
simple family all family members share the same priority number
59. Patent families
A patent family consists of patents granted by different patent
authorities but related to the same invention.
simple family all family members share the same priority number
extended family there are several definitions, in the INPADOC
database all documents which are directly or
indirectly linked via a priority number belong to the
same family
64. Relevance assessments 2010
Expanding the 2009 extended citations:
1 include citations of forward citations ...
65. Relevance assessments 2010
Expanding the 2009 extended citations:
1 include citations of forward citations ...
2 ... and their families
66. Relevance assessments 2010
Expanding the 2009 extended citations:
1 include citations of forward citations ...
2 ... and their families
This is apparently a well-known method among patent searchers.
67. Relevance assessments 2010
Expanding the 2009 extended citations:
1 include citations of forward citations ...
2 ... and their families
This is apparently a well-known method among patent searchers.
Zig-zag search?
68. How good are the CLEF-IP relevance assessments?
CLEF-IP uses families + citations:
69. How good are the CLEF-IP relevance assessments?
how complete are extended
citations as a relevance
assessments?
70. How good are the CLEF-IP relevance assessments?
how complete are extended
citations as a relevance
assessments?
will every prior art patent be
included in this set?
71. How good are the CLEF-IP relevance assessments?
how complete are extended
citations as a relevance
assessments?
will every prior art patent be
included in this set?
and if not, what percentage
of prior art items are captured
by extended citations?
72. How good are the CLEF-IP relevance assessments?
how complete are extended
citations as a relevance
assessments?
will every prior art patent be
included in this set?
and if not, what percentage
of prior art items are captured
by extended citations?
when considering forward
citations, how good are
extended citations as a prior
art candidate set?
73. Feedback from patent experts needed
Quality of prior art candidate sets has to be assessed
75. Feedback from patent experts needed
at Clef–Ip 2009 7 patent search professionals assessed 12
search results
76. Feedback from patent experts needed
at Clef–Ip 2009 7 patent search professionals assessed 12
search results
the task was not well defined and there were
misunderstandings on the concept of relevance
77. Feedback from patent experts needed
at Clef–Ip 2009 7 patent search professionals assessed 12
search results
the task was not well defined and there were
misunderstandings on the concept of relevance
amount of data was not sufficient to draw conclusions
79. Some initiatives associated with Clef–Ip
The results of evaluation tracks are mostly useful for the research
community.
80. Some initiatives associated with Clef–Ip
The results of evaluation tracks are mostly useful for the research
community.
This community often produces prototypes that are of little
interest to the end-user.
81. Some initiatives associated with Clef–Ip
The results of evaluation tracks are mostly useful for the research
community.
This community often produces prototypes that are of little
interest to the end-user.
Next I’d like to present two concrete outcomes - not of Clef–Ip
directly but arising from work in patent retrieval evaluation
84. Soire
developed at Matrixware
service-oriented architecture - available as a a Web service
85. Soire
developed at Matrixware
service-oriented architecture - available as a a Web service
allows to replicate IR experiments based on classical
evaluation model
86. Soire
developed at Matrixware
service-oriented architecture - available as a a Web service
allows to replicate IR experiments based on classical
evaluation model
tested on the CLEF-IP data
87. Soire
developed at Matrixware
service-oriented architecture - available as a a Web service
allows to replicate IR experiments based on classical
evaluation model
tested on the CLEF-IP data
customized for the evaluation of machine translation
89. Spinque
a spin-off (2010) from CWI (the Dutch National Research
Center in Computer Science and Mathematics)
90. Spinque
a spin-off (2010) from CWI (the Dutch National Research
Center in Computer Science and Mathematics)
introduces search-by-strategy
91. Spinque
a spin-off (2010) from CWI (the Dutch National Research
Center in Computer Science and Mathematics)
introduces search-by-strategy
provides optimized strategies for patent search - tested on
CLEF-IP data
92. Spinque
a spin-off (2010) from CWI (the Dutch National Research
Center in Computer Science and Mathematics)
introduces search-by-strategy
provides optimized strategies for patent search - tested on
CLEF-IP data
transparency: understand your search results to improve
strategy
93. Clef–Ip 2009 learnings
The Humboldt University implemented a model for patent search
that produced the best results.
94. Clef–Ip 2009 learnings
The Humboldt University implemented a model for patent search
that produced the best results.
The model combined several strategies:
95. Clef–Ip 2009 learnings
The Humboldt University implemented a model for patent search
that produced the best results.
The model combined several strategies:
using metadata (IPC, ECLA)
96. Clef–Ip 2009 learnings
The Humboldt University implemented a model for patent search
that produced the best results.
The model combined several strategies:
using metadata (IPC, ECLA)
indexes built at lemma level
97. Clef–Ip 2009 learnings
The Humboldt University implemented a model for patent search
that produced the best results.
The model combined several strategies:
using metadata (IPC, ECLA)
indexes built at lemma level
an additional phrase index for English
98. Clef–Ip 2009 learnings
The Humboldt University implemented a model for patent search
that produced the best results.
The model combined several strategies:
using metadata (IPC, ECLA)
indexes built at lemma level
an additional phrase index for English
crosslingual concept index (multilingual terminological
database)
100. Some additional investigations
% runs class
≤5 hard
5 < x ≤ 10 very difficult
Some citations were hard to find
10 < x ≤ 50 difficult
50 < x ≤ 75 medium
75 < x ≤ 100 easy