Date post: | 28-Nov-2014 |
Category: |
Technology |
Upload: | mike-salampasis |
View: | 256 times |
Download: | 0 times |
Multilayer Collection Selection and Search of Topically Organized Patents
Michail Salampasis
Vienna University of Technology
Anastasia Giahanou
University of Macedonia
Giorgos Paltoglou
University of Wolverhampton
2
Contents
Overview:
Aim and Objectives of this work
Distributed Information Retrieval / Federated Search
Topically Organised Patents
Integration of DIR in patent search: Multilayer Source
Selection
Experiment Setup
Results
Conclusions
Aim of this work
3
To explore the thematic organization of patent documents
using the subdivision of patent data by International
Patent Classification (IPC) codes , and
if this organization can be used to build search tools that
could improve patent search effectiveness using DIR
methods
Which search tools and how should be integrated?
4
It is a mistake if we think the search tools which should be
integrated into patent search systems depend only on
existing IR or text processing technologies,
Probably it has more to do with the attitude that a patent
search is conducted.
Furthermore, it is also very important to deeply
understand a search process and how a specific tool can
attain a specific objective of this process and therefore
increase its efficiency.
If these parameters are not carefully considered
5
• Professional searchers will be skeptical and with a very conservative attitude towards adopting search methods, tools and technologies beyond the ones which dominated their domain.
• A typical example is patent search where professional search experts typically use the Boolean search syntax and quite complex intellectual classification schemes
Understanding Patent Search processes *
* Taken from Mihai Lupu and Allan Hanbury, Review Patent Retrieval
Objectives
7
•The improvement of our method relates to the very fundamental step in professional patent search (step 3 in the use case presented by Lupu and Hanbury) which is “defining a text query, potentially by Boolean operators and specific field filters”. • In prior art search probably the most important filter is based on the IPC (CPC now) classification
Objectives
8
•The method and tool which we present in this paper can support this step by automatically selecting IPCs given a query, make a filtered search based on the query and the automatically selected IPCs •The tool can be used for classification search which will be used as a starting point to identify and closer examine technical concepts as these are expressed in IPCs and to which a patent could be related
9
Distributed IR
Elements composing a Distributed Information Retrieval System
. . .
(1) Source
Representation
. . . . Collection 1 Collection 2 Collection 3 Collection 4 Collection Ν
(2) Source
Selection
…… ……
(3) Results
Merging
User
Topically Organised Patents based on IPC taxonomy
10
IPC is a standard taxonomy for classifying patents, and has currently
about 71,000 nodes which are organized into a five-level hierarchical
system which is also extended in greater levels of granularity.
Patent documents produced worldwide have manually-assigned
classification codes which in our experiments are used to topically
organize, distribute and index patents through hundreds or
thousands of sub-collections.
Topically Organised Patents
11
Topically Organised Patents
12
The patents in average have three IPC codes. In the experiments we
report here, we allocated a patent to each sub-collection specified by
at least one of its IPC code, i.e. a sub-collection might overlap with
others in terms of the patents it contains.
IPC are assigned by humans in a very detailed and purposeful
assignment process, something which is very different by the creation
of sub-collections using automated clustering algorithms or the naive
division method by chronological or source order, a division method
which has been extensively used in past DIR research
Topically Organised Patents
13
Analysis of IPC distribution of topics and their relevant documents
14
IPC Level
# of topics
# relevant docs per
topic (a)
# of IPC
classes of each topic
(b)
# of IPC classes of relevant
docs (c)
# of common IPC
classes between (b)
and (c)
Training
Split 3
300 8.22 2.08 4.8 1.76
Split 4
300 8.22 3.1 8.76 2.34
Split 5
300 8.22 5.82 19.84 3.63
Testing
Split 3
300 8.57 2.09 5.15 1.75
Split 4
300 8.57 2.95 9.02 2.21
Split 5
300 8.57 5.58 20.56 3.73
Experiment Setup
15
We indexed the collection with the Lemur toolkit.
The fields which have been indexed are: title, abstract, description (first 500 words), claims, inventor, applicant and IPC class information.
Patent documents have been pre-processed to produce a single (virtual) document representing a patent.
Our pre-processing involves also stop-word removal and stemming using the Porter stemmer. In the experiments reported here we use the Inquery algorithm implementation of Lemur
Two different types of Source Selection Algorithms were used
16
Hyper-document approach (CORI)
oThe main characteristic of CORI which is probably the most widely used and tested source selection method is that it creates a hyper-document representing all the documents-members of a sub-collection.
Source Selection as Voting
oThis is a shift of focus from estimating the relevancy of each remote collection to explicitly estimating the number of relevant documents in each.
Source Selection Results (level 3)
17
Source Selection Results (level 4)
18
Source Selection Results (level 5)
19
Discussion
• The superiority of CORI as source selection method is unquestionable
• best runs are those requesting fewer sub-collections 10 or 20 and more documents from each selected sub-collection
• This fact is probably the result of the small number of relevant documents which exist for each topic
20
Results of Retrieval Results
SPLIT4
10 Collections Selected 20 Collections Selected
Pres@100 MAP@100 Pres@100 MAP@100 Optimal 0.313 0.128 0.313 0.128
Centralised 0.257 0.105 0.257 0.105 CORI-CORI 0.203 0.081 0.213 0.086 CORI-SSL 0.221 0.091 0.231 0.097
BordaFuse-SSL 0.077 0.035 0.087 0.039 Multilayer 0.256 0.105 0.261 0.105
SPLIT5
10 Collections Selected 20 Collections Selected
Pres@100 MAP@100 Pres@100 MAP@100 Optimal 0.346 0.146 0.351 0.148
Centralised 0.257 0.105 0.257 0.105 CORI-CORI 0.267 0.107 0.259 0.105
CORI-SSL 0.27 0.11 0.263 0.107
BordaFuse-SSL 0.03 0.02 0.04 0.028 Multilayer 0.269 0.106 0.267 0.102
Conclusions
DIR approaches managed to perform better than the centralized index approaches, with 9 DIR combinations scoring better than the best centralized approach.
Much more work is required:
oWe plan to explore further this line of work with exploring modifications to state-of-the-art DIR methods which didn’t perform well enough in this set of experiments
oAlso, we would like to experiment with larger distribution levels based on IPC (subgroup level). We plan to report the runs using split-5 in a future paper.
22
23
Thank you…