Introduction
Given an entity in a source domain, finding its matched entities from another (target) domain is an important task in many applications. For example, a patent expert may be interested in finding related patents for a product. The problem can be generalized as a document matching problem across corpora from heterogeneous sources. In other words, given a document describing an entity (e.g., product) in one source, the goal is to find related documents describing other types of entities (e.g., patent) from a different source. One key challenge to solve the problem is that there could be only a small portion of text overlap or even no text overlap between two different sources (e.g., general terms in Wiki vs. special terminologies in patents).
In this paper, we propose a probabilistic model, Cross-Source Topic (CST) model, to solve the problem for a two-source case, which integrates the topic extraction and document matching into a unified model. By using this model, we can not only find the matched documents for a query document, but also explain why these documents are related by showing the common topics they share. Our model can successfully overcome the little-text-overlap problem across heterogeneous sources of corpus, by modeling a topic with terms coming from all the sources and utilizing the relevance labels for documents across different sources. Our experiments in two real-world applications show that the proposed model can extensively improve the matching performance (averagely +10.1%) compared with several alternative methods.
Code & Data Set
data labeling spec
INTRODUCTION
The following describes the specs that how we label and get the product-patent matching ground truth data. There four annotator label the data.
GENERAL PRINCIPLE
Label product-patent matching ground truth data.
Generally speaking, our ground truth data are defined as shown in Table 1,
| product | Feature | Wiki Page | Corresponding Patents |
| iPhone | visual keyboard | http://en.wikipedia.org/wiki/IPhone_4S | 7602378 |
Table 1
We label our ground truth data through the feature of a product, Company’s Patent Lawsuit, official document of a product.
Feature of a product
Let us take the example of how we find one of the IPhone’s patents (US7602378).
First, each annotator will read the Wiki articles which describes the “IPhone”. We get the information that IPhone has many unique features and each feature might have a certain patent describing and supporting it.
For example, the first paragraph of IPhone’s WIKI article is shown as follows.
The iPhone ) is a line of smartphones designed and marketed by Apple Inc. It runs Apple's iOS mobile operating system, known as the "iPhone OS" until mid-2010, shortly after the release of the iPad. The first iPhone was released on June 29, 2007;[15] the most recent iPhone, the sixth-generation iPhone 5, on September 21, 2012.[16] The user interface is built around the device's multi-touch screen, including a virtual keyboard. The iPhone has Wi-Fi and cellular connectivity (2G, 3G and 4G).
We can find that the “virtual keyboard” is one unique feature of IPhone, so we guess there might be a “virtual keyboard” patent which was granted by Apple Inc. Then, we use the “apple IPhone Virtual Keyboard patent” which is shown in figure 1 as the query to search this information in GOOGLE.
Figure1
After we get the searching result, we view them from top to bottom. Then, when we click the second search result, we will find a website which tells us “Apple Wins Key iPhone Virtual Keyboard Patent”. We will get the following sentence in this web-site.
Apple credits Bas Ording as the sole inventor of granted patent 7,614,008 officially titled: Operation of a computer with touch screen interface. The patent was originally filed in Q3 2005. Update: Apple was granted another patent for their virtual keyboard in October, under number 7,602,378.
According to this sentence, we can get our ground truth that the patent No. 7,602,378 and IPhone Wiki article is relevant from the sentence above.
Company’s Patent Lawsuit
News about Company’s Patent Lawsuit will tell us the information about that company’s patent and product. We can also use this to find one product’s related patent.
For example, we know that the Apple and Samsung have several patent lawsuit, so we use the “patent apple Samsung lawsuit” which is shown in figure 2 as the query to search this information in GOOGLE.
Figure2
After we get the searching result, we view them from top to bottom. Then, when we click the third search result, we will find this website tells us the lawsuits between Apple Inc. and Samsung Electronics. After reading this website, we get the following sentence.
Design Patent 504,889 (describing the ornamental design of the iPod) was one of the few patents the jury concluded Samsung had not infringed.
From this information, we can get our ground truth that the patent No. 504,889 and IPod Wiki article is relevant.
Official documents of the products
We know that some patent document has a field called “other references”. This field contains prior art reference in the form of books, journals, articles, abstracts, official documents of the products and other non-patent documents that are cited by the Patent Examiner during patent examination. Thus, it will tell us other documents which are related to this patent
For example, when we search the patent D639763 in the patent website (www.uspto.gov), we will find the following information.
According to this information, we can get our ground truth that the patent No. 7,602,378 and Samsung_SGH-F700 Wiki article is relevant.
Get product-patent matching ground truth data.
After we get the relationship between the product and patent, we are going to extract the useful information from the Wiki article and patent document
Extraction content and URL of Wiki article
First we get the source web page of one WIKI article. For example the source web page of IPhone Wiki article is view-source:http://en.wikipedia.org/wiki/Iphone.
Generally, speaking, this extraction is defined as shown in Table 2
| Section | Annotation |
| content | wiki article |
| URL | wiki url link |
Table 2
Each component in the right column of the Table 2 is defined as a property of a wiki page and may consist of words and url.
About tokenization
We should note that annotated tags should not destroy the structure of the html. For example, keep as what it is. Likewise for , . Especially, both font and bold can be nested, annotation should keep the nested relationship as what it be.
” and “
” in HTML. When extracting information from the Web, we cleaning all the other tags except the “” and “
” tag. We download the information from the Web so that we can get content of a certain wiki article which describes a product.Extraction the patent document
We get the patent number, title, abstract, claims, author from our Patentminer Database.
Generally speaking, a patent document profile is defined as shown in Table 3.
| patent document |
| number |
| titile |
| abstract |
| claim |
| author |
Table3
LINE BRAKER
One can conduct experiments by viewing each line in the web page as a unit. The line breaker is naturally the line break between two lines.
WORD BRAKER
A html tag is a unique word.
A return is a word
For the others, the breaker is defined as:
[\n\s\,\?\!\(\)\[\]\{\}]{1})|(?:\.\s+)|(?:\s*$)
EXAMPLES
Wiki article
| Examples of Wiki article |
| The iPhone is a line of smartphonesdesigned and marketed by Apple Inc. It runs Apple's iOS mobile operating system, known as the "iPhone OS" until mid-2010, shortly after the release of the iPad. The first iPhone was released on June 29, 2007;[15] the most recent iPhone, the sixth-generation iPhone 5, on September 21, 2012.[16] The user interface is built around the device's multi-touch screen, including a virtual keyboard. The iPhone has Wi-Fi and cellular connectivity (2G, 3G and 4G). |
| An iPhone can shoot video (though this was not a standard feature until the iPhone 3GS), take photos, play music, send and receive email, browse the web, send texts, and receive visual voicemail. Other functions—games, reference, GPS navigation, social networking, etc.—can be enabled by downloading apps; as of 2012, the App Store offered more than 775,000 apps by Apple and third parties.[17] |
| There are six generations of iPhone models, each accompanied by one of the six major releases of iOS. The original iPhone was a GSM phone, and established design precedents, such as a button placement that has persisted through all models and a screen size maintained until the launch of the iPhone 5 in 2012. The iPhone 3G added 3G cellular network capabilities and A-GPS location. The iPhone 3GS added a faster processor and a higher-resolution camera that could record video at 480p. The iPhone 4 featured a higher-resolution 960 × 640 "retina display", a higher-resolution rear-facing camera and a lower-resolution front-facing camera for video calling and other apps.[18] The iPhone 4S added an 8-megapixel camera with 1080p video recording, a dual-core processor, and a natural language voice control system called Siri.[19][20][21] iPhone 5 features the new A6 processor, holds a 4-inch Retina display that is larger than its predecessor's 3.5-inch display, and replaces the 30-pin connector with an all-digital Lightning connector. |
URL
| Examples of url |
| /wiki/IPhone |
| /wiki/IPhone_(disambiguation) |
| /wiki/IPhone_5 |
patent document
| patent document | content |
| number | 8086604 |
| titile | Universal interface for retrieval of information in a computer system |
| titile | The present invention provides convenient access to items of information that are related to various descriptors input by a user, by means of a unitary interface which is capable of accessing information in a variety of locations, through a number of different techniques. Using a plurality of heuristic algorithms to operate upon information descriptors input by the user, the present invention locates and displays candidate items of information for selection and/or retrieval. Thus, the advantages of a search engine can be exploited, while listing only relevant object candidate items of information. |
| claim | What is claimed is:1. A method for locating information in a network using a computer, comprising: receiving by the computer an inputted information descriptor from a user-input device; providing said information descriptor received from the user-input device to a plurality of heuristic modules, wherein: each heuristic module corresponds to a respective area of search and employs a different, predetermined heuristic algorithm corresponding to said respective area to search the area for information that corresponds to the received information descriptor, and the search areas include storage media accessible by the computer; searching by the heuristic modules, based on the received information descriptor, the respective areas of search using the predetermined heuristic algorithms corresponding to each respective area of search; providing at least one candidate item of information located by the heuristic modules as a result of said searching; and displaying by the computer a representation of said candidate item of information on a display device. |
| author | # Yan Arrouye# Keith Mortensen |
Product-patent matching.
In this task, given a Wiki article describing a specific product, we aim to find relevant patents, e.g., a Wiki article and a patent should be relevant if they are both talking about the topic of Siri. We collect 13,085 Wiki articles and 15,000 patents from Wikipedia and USPTO respectively. For some Wiki article that describes a product, we use it as a query to find patents related with the same product. One Wiki article may be matched with more than one patents, e.g., a Wiki article describing iPhone corresponds to patents that claim on touch screen, camera, soft keyboard, etc.. We sample 233Wiki articles as queries and find 1060 matching relations in total. We randomly choose 30% of the matching relations as known. The remaining relations are regarded as unknown and need to be inferred. The training set also contains all 13,085 Wiki articles and 15,000 patents.
Cross-lingual matching.
In this task, given an English Wiki article, we aim to find a Chinese article, which reports the same content, from a Chinese Wiki knowledge base. We collect the data set as follows: we randomly select an English article A with a crosslingual link to a Chinese article B from Wikipedia, we then use the B’s title to find another Chinese article C with the same title in Baidu Baike5. As A is cross-lingually linked with B in Wikipedia, and B has the same main idea with C (normally aWiki article uses its main idea as the title). It is reasonable to say there is a crosslingual matching relation between A and C. We totally collect 2,000 English articles from Wikipedia, and 2,000 Chinese articles from Baidu Baike. Thus in the data set, each English article corresponds to one Chinese article.
The code of the Cross-Source Topic Model and the data set are publicly available. They can be downloaded here.