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.
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 [
Mukherjea and Bamba [
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.
In this section, we describe our retrieval system using standard Information Retrieval (IR) techniques for indexing and retrieving patent documents.
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
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.
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).
2005 500 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
TFIDF 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 [
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. 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 [
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.
The results of our submitted runs using large and small topic set are shown on Table 5. 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.
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..