=Paper=
{{Paper
|id=Vol-1176/CLEF2010wn-CLEF-IP-WanagiriEt2010
|storemode=property
|title=Prior Art Retrieval Using Various Patent Document Fields Contents
|pdfUrl=https://ceur-ws.org/Vol-1176/CLEF2010wn-CLEF-IP-WanagiriEt2010.pdf
|volume=Vol-1176
}}
==Prior Art Retrieval Using Various Patent Document Fields Contents==
https://ceur-ws.org/Vol-1176/CLEF2010wn-CLEF-IP-WanagiriEt2010.pdf
Prior Art Retrieval Using Various Patent Document
Fields Contents
Metti Zakaria Wanagiri and Mirna Adriani
Fakultas Ilmu Komputer, Universitas Indonesia
Depok 16424, Indonesia
metti.zakaria@ui.edu, mirna@cs.ui.ac.id
Abstract. In this paper, we report our approach to retrieve patent documents
based on the prior art. We use the standard Information Retrieval (IR)
techniques which contain indexing and retrieval processes. We use various
combinations of document fields for the query formulation. Based on the
evaluation summary, we achieve the best result for the combinations of
invention-title, description and claims fields in terms of precision and recall.
Keywords: patent retrieval
1 Introduction
There are a lot of inventions that have been invented in the industry and sciences. The
number of inventions is growing from time to time as there is a high demand and need
from human to have better and easier life, such as the living environment, working
environment and so on. For example, around April 2010, Apple Inc. developed a new
portable tablet computer called iPad which is one of the latest patented inventions.
One of the functions is it can enable humans to read any e-book documents anytime
and anywhere.
An invention can be granted an exclusive right called patent by the national
government for a limited period of time in exchange for public disclosure of those
inventions. This exclusive right granted to an inventor is the right to prevent others
from making, using, selling or distributing the patented invention without permission.
So, with this exclusive right, an inventor can fully protect its patented invention from
any misuses in the given period of time.
According to World Intellectual Property Organization (WIPO) of United Nations,
patent inventions/applications consist of patent specifications, official forms and
correspondence relating to the applications. A patent specification is a document that
describes the invention which generally contains the invention/application title,
section detailing the background and overview of the invention, a description of the
invention and embodiments of the invention and claims, which set out the scope of
the protection. It also includes an abstract which provides a summary of the invention.
The claims of a patent specification define the scope of protection of a patent granted
by the patent and describe the invention in a specific legal style.
As the number of patent applications increases, the patent domain is considered
quite important. Since there are many new inventions that are being set out for patent
granting, then it should be a justification on those new inventions. A new invention
should be checked whether there are any existing patents which may invalidate them.
So a patent specification or patent document plays a vital role in differentiating any
inventions.
In 2009, the Cross Language Evaluation Forum (CLEF) launches a track called
CLEF-IP which focuses on Intellectual Property domain. It investigates the use of
Information Retrieval techniques for patent document retrieval. The main task in this
track is to find any existing patent documents that may invalidate a new invention
who apply for its patent. Jarvelin and Preben [1] use an automatic query generation
algorithm. They compare queries generated by human experts to those generated by
system and the automatic generated queries achieve the better performance. Lopez
and Romary [2] use multiple retrieval models for producing several sets of ranked
results. Then they apply Multiple SVM regression models to merge the results.
Toucedo and Losada [9] build queries by extracting terms from some textual patent
documents fields using inverse document frequency (idf) and give preference to the
title terms. BM25 retrieval model is used and the best result is achieved when the title
terms and the standard parameters of BM25 retrieval model are used.
Mukherjea and Bamba [5] also develop a retrieval system for biomedical patents
called BioPatentMiner. It integrates information from the patents with knowledge
from biomedical ontologies to create a Semantic Web. Takaki et al. [8] propose the
invalidity patent search by applying an associative document retrieval method, in
which a document is used as a query to search for other similar documents. They use
subtopics or compositional elements extraction to extract subtopics which correspond
to an element constituting the claim section. Then content words which mainly nouns
are extracted from each compositional element as query terms. Evaluation results
show that the method used was effective in the patent search. Mase et al. [3] use two
retrieval stages which consists of query term extraction from claim text, query term
weighting without term frequency (tf) and using measurement terms (terms that
accompanied by numerical values) and text retrieval using claims as targets.
Evaluation results show that the effectiveness of the method varies depending on the
test sets used.
In this paper, we report our participation in the 2nd CLEF-IP. We focused on the
Prior Art Candidate Search (PAC) task to find patent documents that are likely to
constitute prior art to a given patent application (patent topic).The remaining of this
paper is organized as follows: section 2 discusses our retrieval system for patent
documents, section 3 describes the experiments, section 4 describes the evaluation
summary and section 5 is the conclusion.
2 Patent Documents Retrieval System
In this section, we describe our retrieval system using standard Information Retrieval
(IR) techniques for indexing and retrieving patent documents.
2.1 Extracting Patent Fields
Before indexing process is carried out, we need to extract the patent fields in the
multilingual document collection. There are about 60 different fields in a patent
document, however their contents are not always informative and important. So we
need to figure out which fields that are considered important to a corresponding patent
application. First, we randomly take some patent documents from the CLEF-IP 2010
corpus. Then we extract all of the document fields by recognizing the associated tags.
Then we create a list of unique patent fields
2.2 Indexing Documents
We choose a number of informative fields from the list of unique patent fields in
which the contents considered to be valuable and represent all the information about
the corresponding patent application. There are 30 chosen patent fields (see Table 1)
that are used in the indexing process.
Table 1. Patent Fields for Indexing Process
abstract address agents
applicant application-reference claim
claim-text claims classification-ecla
classification-ipc classification-ipcr classification-symbol
classifications-ipcr colspec copyright
country date dates-of-public-availability
description designated-states doc-number
doc-page document-id invention-title
inventor inventors patent-citations
patent-document priority-claim priority-claims
2.3 Query Formulation
The CLEF-IP 2010 topic documents are categorized into two sets: the large topic set
and small topic set. Each topic document is a patent document in XML format which
has the same structured data as the patent documents in CLEF-IP 2010 corpus. Both
sets come in three different languages: English, French and German (see Table 2).
Table 2. Two Sets of Topic Documents
Topic Sets Number of Docs
Large Topic Set 2005
Small Topic Set 500
Total 2505
Our task in this track is to find all relevant patent documents in the collection that
invalidate a given topic documents. In this case, we build some appropriate and
effective queries from the topic documents.
In this query formulation process, we use the standard term weighting algorithm of
TF-IDF [6[. Essentially, TF-IDF works by determining the relative proportion of
words in a specific document compared to the inverse proportion of that word over
the entire document corpus. This calculation determines how relevant a given word is
in a particular document.
So, given a document collection D and a document d є D, the calculation of TF-
IDF for a word w is
wd = fw,d * log (N/fw,D) (1)
where fw,d is the number of times w appears in d, N is the number of documents in D
and fw,D is the number of documents in D in which w appears [6].
For each of the topic documents from both sets, we apply these steps of query
formulation:
1. Extracting the contents from three patent fields: invention-title,
description and claims,
2. Extracting words from the extracted contents by applying the standard
weighting algorithm of TF-IDF,
3. Retrieving top 10 words with high TF-IDF, and
4. Forming the 10 words as one query.
As there are three patent fields that we used for query formulation, we define three
possible combinations that will be used for our experiments. The combinations are:
1. claims
2. invention-title + description
3. invention-title + description + claims
So following the steps above, after we extract the contents from each patent field, we
combine the contents based on the combinations above and then extract the top 10
words as the query. Finally, there are three sets of query that we will use in the
retrieving process. Table 3 shows the details of the query sets.
Table 3. Details of Query Sets
Query Set Patent Fields Number of Queries
QS-1 invention-title + description 2505
QS-2 claims 2505
QS-3 invention-title + description + claims 2505
3 Experiments
For the experiments, we use CLEF-IP 2010 corpus which contains around 2 million
patent documents from European Patent Office (EPO). Each patent document is an
XML file containing structured data with different fields delimited by specified tags.
We index the documents using Indri1 which is part of the Lemur2 Toolkit. Indri
retrieval model is based on a combination of language model and inference network
frameworks [7]. We remove stopwords from the corpus but we don’t stem the words.
We don’t use any cross language technique in those runs therefore no language
specific methods are used.
We run three experiments based on three sets of query and retrieve top 1000 patent
documents which are relevant to each query from sets. In the experiments, we
combined the title, description, and claims that occurred o the documents. These three
experiments or runs are the submitted runs for CLEF-IP 2010. For all of the runs, we
use both large and small topic sets. Table 4 shows the details of the submitted runs.
Table 4. Details of Experiments
Run ID Run Name Query Set Topic Set
ui-1 ui_title&desc_Run1_PAC_all QS-1 Large + Small
ui-2 ui_claims_Run2_PAC_all QS-2 Large + Small
ui-3 ui_title-desc-claims_Run3_PAC_all QS-3 Large + Small
4 Evaluation
The results of our submitted runs using large and small topic set are shown on Table
5.
Table 5. The Performance of the Submitted Runs
Topic Set Run ID P R MAP NDCG
ui-1 0.0064 0.2937 0.052 0.1705
Large ui-2 0.0059 0.2827 0.0457 0.1592
ui-3 0.007 0.3301 0.0581 0.1898
ui-1 0.0062 0.2859 0.0425 0.1551
Small ui-2 0.0059 0.2756 0.0441 0.1559
ui-3 0.007 0.3332 0.0537 0.1846
We present our evaluation summary in Table 5 in four measures: precision (P),
recall (R), Mean Average Precision (MAP) and NDCG. The retrieval performance of
1 http://www.lemurproject.org/indri/
2 http://www.lemurproject.org/
all the topics sets show that the recall is much higher that the precision. The MAP of
the large topic set (0.0581) is higher than the small topic set (0.0537).
Our results have motivated us to explore more on the patent fields’ contents that
are valuable for retrieval process. Furthermore the query formulation process needs to
be improved using different approach.
5 Conclusion
This year we participate in the Patent Retrieval track in CLEF-IP 2010. We use
standard IR techniques for retrieving patent documents. We identify several fields that
are used in the indexing process. For the retrieval process, we combine several fields
such as title, description, and claims. The evaluation shows that the precision is much
lower than the recall.
There are still rooms for improvement such as adding more context to the query
using query expansion or relevance feedback and also using different term weighting
algorithm..
References
1. Jarvelin, A., Preben, H.: UTA and SICS at CLEF-IP. In: 1st CLEF-IP, Corfu, Greece
(2009)
2. Lopez, P., Romary, L.: Multiple Retrieval Models and Regression Models for Prior Art
Search. In: 1st CLEF-IP, Corfu, Greece (2009)
3. Mase, H., Matsubayashi, T., Ogawa, Y.: Proposal of Two Stage Patent Retrieval Method
Considering the Claim Structure. ACM Transactions on Asian Language Information
Processing 4(2) (2005)
4. Michel, J.: Considerations, challenges and methodologies for implementing best practices
in patent office and like patent information departments. World Patent Information 28:132-
135 (2006)
5. Mukherjea, S., Bamba, B.: BioPatentMiner: An Information Retrieval System for
Biomedical Patents. In: VLDB '04: Proceedings of the Thirtieth International Conference
on Very Large Data Bases, pp. 1066-1077 (2004)
6. Salton, G., Buckley, C.: Term-weighting approaches in automatic text retrieval. In:
Information Processing & Management, 24(5), pp. 513-523 (1988)
7. Strohman, T., Metzler, D., Turtle, H., Croft, B.: Indri: A language-model based search
engine for complex queries (extended version). Technical Report IR-407, Department of
Computer Science, University of Massachusetts (2005)
8. Takaki, T., Fujii, A., Ishikawa, T.: Associative Document Retrieval by Query Subtopic
Analysis and its Application to Invalidity Patent Search. In: Proceedings of CIKM (2004)
9. Toucedo, J.C., Losada, D.E.: University of Santiago de Compostela at CLEF-IP09. In: 1st
CLEF-IP, Corfu, Greece (2009)