The present invention relates in general to text mining and, in particular, to a computer-implemented system and method for identifying relevant documents.
Document warehousing extends data warehousing to content mining and retrieval. Document warehousing attempts to extract semantic information from collections of unstructured documents to provide conceptual information with a high degree of precision and recall. Documents in a document warehouse share several properties. First, the documents lack a common structure or shared type. Second, semantically-related documents are integrated through text mining. Third, essential document features are extracted and explicitly stored as part of the document warehouse. Finally, documents are often retrieved from multiple and disparate sources, such as over the Internet or as electronic messages.
Document warehouses are built in stages to deal with a wide range of information sources. First, document sources are identified and documents are retrieved into a repository. For example, the document sources could be electronic messaging folders or Web content retrieved over the Internet. Once retrieved, the documents are pre-processed to format and regularize the information into a consistent manner. Next, during text analysis, text mining is performed to extract semantic content, including identifying dominant themes, extracting key features and summarizing the content. Finally, metadata is compiled from the semantic context to explicate essential attributes. Preferably, the metadata is provided in a format amenable to normalized queries, such as database management tools. Document warehousing is described in D. Sullivan, “Document Warehousing and Text Mining, Techniques for Improving Business Operations, Marketing, and Sales,” Chs. 1-3, Wiley Computer Publishing (2001), the disclosure of which is incorporated by reference.
Text mining is at the core of the data warehousing process. Text mining involves the compiling, organizing and analyzing of document collections to support the delivery of targeted types of information and to discover relationships between relevant facts. However, identifying relevant content can be difficult. First, extracting relevant content requires a high degree of precision and recall. Precision is the measure of how well the documents returned in response to a query actually address the query criteria. Recall is the measure of what should have been returned by the query. Typically, the broader and less structured the documents, the lower the degree of precision and recall. Second, analyzing an unstructured document collection without the benefit of a priori knowledge in the form of keywords and indices can present a potentially intractable problem space. Finally, synonymy and polysemy can cloud and confuse extracted content. Synonymy refers to multiple words having the same meaning and polysemy refers to a single word with multiple meanings. Fine-grained text mining must reconcile synonymy and polysemy to yield meaningful results.
In the prior art, text mining is performed in two ways. First, syntactic searching provides a brute force approach to analyzing and extracting content based on literal textual attributes found in each document. Syntactic searching includes keyword and proximate keyword searching as well as rule-based searching through Boolean relationships. Syntactic searching relies on predefined indices of keywords and stop words to locate relevant information. However, there are several ways to express any given concept. Accordingly, syntactic searching can fail to yield satisfactory results due to incomplete indices and poorly structured search criteria.
A more advanced prior art approach uses a vector space model to search for underlying meanings in a document collection. The vector space model employs a geometric representation of documents using word vectors. Individual keywords are mapped into vectors in multi-dimensional space along axes representative of query search terms. Significant terms are assigned a relative weight and semantic content is extracted based on threshold filters. Although substantially overcoming the shortcomings of syntactic searching, the multivariant and multidimensional nature of the vector space model can lead to a computationally intractable problem space. As well, the vector space model fails to resolve the problems of synonymy and polysemy.
Therefore, there is a need for an approach to dynamically evaluating concepts inherent in a collection of documents. Such an approach would preferably dynamically discover the latent meanings without the use of a priori knowledge or indices. Rather, the approach would discover semantic relationships between individual terms given the presence of another item.
There is a further need for an approach to providing a graphical visualization of concepts extracted from a document set through semantic indexing. Preferably, such an approach would extract the underlying meanings of documents through statistics and linear algebraic techniques to find clusters of terms and phrases representative of the concepts.
The present invention provides a system and method for indexing and evaluating unstructured documents through analysis of dynamically extracted concepts. A set of unstructured documents is identified and retrieved into a document warehouse repository. Individual concepts are extracted from the documents and mapped as normalized data into a database. The frequencies of occurrence of each concept within each document and over all documents are determined and mapped. A corpus graph is generated to display a minimized set of concepts whereby each concept references at least two documents and no document in the corpus is unreferenced. A subset of documents occurring within predefined edge conditions of a median value are selected. Clusters of concepts are grouped into themes. Inner products of document concept frequency occurrences and cluster concept weightings are mapped into a multi-dimensional concept space for each theme and iteratively generated until the clusters settle. The resultant data minima indicates those documents having the most pertinence to the identified concepts.
An embodiment provides a computer-implemented system and method for identifying relevant documents. A set of documents each associated with one or more concepts is identified. At least a portion of the documents in the set are clustered based on the concepts. A matrix that provides a summary of the documents most relevant to one such concept is generated by determining for each document a measure of similarity between a concept frequency occurrence and concept weights of each cluster. The matrix is populated with the calculated measures of similarity. Those documents associated with a threshold measure of similarity are identified as the most relevant documents to one such concept.
In summary, the present invention semantically evaluates terms and phrases with the goal of creating meaningful themes. Document frequencies and co-occurrences of terms and phrases are used to select a minimal set of highly correlated terms and phrases that reference all documents in a corpus.
Still other embodiments of the present invention will become readily apparent to those skilled in the art from the following detailed description, wherein is described embodiments of the invention by way of illustrating the best mode contemplated for carrying out the invention. As will be realized, the invention is capable of other and different embodiments and its several details are capable of modifications in various obvious respects, all without departing from the spirit and the scope of the present invention. Accordingly, the drawings and detailed description are to be regarded as illustrative in nature and not as restrictive.
The document analyzer 12 analyzes documents retrieved from a plurality of local sources. The local sources include documents 17 maintained in a storage device 16 coupled to a local server 15 and documents 20 maintained in a storage device 19 coupled to a local client 18. The local server 15 and local client 18 are interconnected to the system 11 over an intranetwork 21. In addition, the document analyzer 12 can identify and retrieve documents from remote sources over an internetwork 22, including the Internet, through a gateway 23 interfaced to the intranetwork 21. The remote sources include documents 26 maintained in a storage device 25 coupled to a remote server 24 and documents 29 maintained in a storage device 28 coupled to a remote client 27.
The individual documents 17, 20, 26, 29 include all forms and types of unstructured data, including electronic message stores, such as electronic mail (email) folders, word processing documents or Hypertext documents, and could also include graphical or multimedia data. Notwithstanding, the documents could be in the form of structured data, such as stored in a spreadsheet or database. Content mined from these types of documents does not require preprocessing, as described below.
In the described embodiment, the individual documents 17, 20, 26, 29 include electronic message folders, such as maintained by the Outlook and Outlook Express products, licensed by Microsoft Corporation, Redmond, Wash. The database is an SQL-based relational database, such as the Oracle database management system, release 8, licensed by Oracle Corporation, Redwood Shores, Calif.
The individual computer systems, including system 11, server 15, client 18, remote server 24 and remote client 27, are general purpose, programmed digital computing devices consisting of a central processing unit (CPU), random access memory (RAM), non-volatile secondary storage, such as a hard drive or CD ROM drive, network interfaces, and peripheral devices, including user interfacing means, such as a keyboard and display. Program code, including software programs, and data are loaded into the RAM for execution and processing by the CPU and results are generated for display, output, transmittal, or storage.
During text analysis, the text analyzer 42 identifies terms and phrases and extracts concepts in the form of noun phrases that are stored in a lexicon 18 maintained in the database 30. After normalizing the extracted concepts, the text analyzer 42 generates a frequency table 46 of concept occurrences, as further described below with reference to
Each module is a computer program, procedure or module written as source code in a conventional programming language, such as the C++ programming language, and is presented for execution by the CPU as object or byte code, as is known in the art. The various implementations of the source code and object and byte codes can be held on a computer-readable storage medium or embodied on a transmission medium in a carrier wave. The document analyzer 12 operates in accordance with a sequence of process steps, as further described below with reference to
Once identified and retrieved, the set of documents 44 is analyzed (block 73), as further described below with reference to
Following preprocessing, a histogram 48 of the frequency of terms (shown in
Next, a document reference frequency (corpus) graph 49, as further described below with reference to
The selected set of terms and phrases falling within the thresholds are used to generate themes (and concepts) (block 85) based on correlations between normalized terms and phrases in the documents set. In the described embodiment, themes are primarily used, rather than individual concepts, as a single co-occurrence of terms or phrases carries less semantic meaning than multiple co-occurrences. As used herein, any reference to a “theme” or “concept” will be understood to include the other term, except as specifically indicated otherwise.
Next, clusters are created (block 86) from groups of highly-correlated concepts and themes. Individual concepts and themes are categorized based on, for example, Euclidean distances calculated between each pair of concepts and themes and defined within a pre-specified range of variance, such as described in commonly-assigned U.S. Pat. No. 6,778,995, issued Aug. 17, 2004, the disclosure of which is incorporated by reference.
A matrix 47 of the documents 44 is created (block 87), as further described below with reference to
Initially, noun phrases are extracted (block 91) from each document 44. In the described embodiment, concepts are defined on the basis of the extracted noun phrases, although individual nouns or tri-grams (word triples) could be used in lieu of noun phrases. In the described embodiment, the noun phrases are extracted using the LinguistX product licensed by Inxight Software, Inc., Santa Clara, Calif.
Once extracted, the individual terms or phrases are loaded into records stored in the database 30 (shown in
Referring back to
A median value 145 is selected and edge conditions 146a-b are established to discriminate between concepts which occur too frequently versus concepts which occur too infrequently. Those documents falling within the edge conditions 146a-b form a subset of documents containing latent concepts. In the described embodiment, the median value 145 is document-type dependent. For efficiency, the upper edge condition 146b is set to 70% and the 64 concepts immediately preceding the upper edge condition 146b are selected, although other forms of threshold discrimination could also be used.
For a set of n documents, the distance dcluster is calculated by taking the sum of products (inner product) by terms between document concept frequency occurrences and cluster concept weightings, using the following equation:
where docterm represents the frequency of occurrence for a given term i in the selected document and clusterterm represents the weight of a given cluster for a given term i. The weights of the individual inner products are iteratively updated until the clusters settle. The goal is to calculate the minimum distances between as few clusters as possible until the rate of change goes constant. The rate of change can be calculated, for example, by taking the first derivative of the inner products over successive iterations.
Satisfactory results are shown when a meaningful cluster of documents is found. Objectively, each document within a given theme will have an inner product falling within a pre-defined variance of other related documents, thereby reflecting a set amount of similarity. The cluster itself represents a larger grouping of document sets based on related, but not identical, themes.
If necessary, the results are re-run (block 182). One reason to re-run the results set would be to re-center the median value 145 of the corpus graph 140 (shown in
While the invention has been particularly shown and described as referenced to the embodiments thereof, those skilled in the art will understand that the foregoing and other changes in form and detail may be made therein without departing from the spirit and scope of the invention.
This patent application is a continuation of U.S. patent application Ser. No. 14/949,829, filed Nov. 23, 2015, pending; which is a continuation of U.S. Pat. No. 9,195,399, issued Nov. 24, 2015; which is a continuation of U.S. Pat. No. 8,725,736, issued May 13, 2014; which is a continuation of U.S. Pat. No. 8,380,718, issued Feb. 19, 2013; which is a continuation of U.S. Pat. No. 8,015,188, issued Sep. 6, 2011; which is a continuation of U.S. Pat. No. 7,809,727, issued Oct. 5, 2010; which is a continuation of U.S. Pat. No. 7,313,556, issued Dec. 25, 2007; which is a continuation of U.S. Pat. No. 6,978,274, issued Dec. 20, 2005, the priority filing dates of which are claimed and the disclosures of which are incorporated by reference.
Number | Date | Country | |
---|---|---|---|
Parent | 14949829 | Nov 2015 | US |
Child | 15484122 | US | |
Parent | 14275802 | May 2014 | US |
Child | 14949829 | US | |
Parent | 13767849 | Feb 2013 | US |
Child | 14275802 | US | |
Parent | 13225325 | Sep 2011 | US |
Child | 13767849 | US | |
Parent | 12897710 | Oct 2010 | US |
Child | 13225325 | US | |
Parent | 11964000 | Dec 2007 | US |
Child | 12897710 | US | |
Parent | 11304406 | Dec 2005 | US |
Child | 11964000 | US | |
Parent | 09944474 | Aug 2001 | US |
Child | 11304406 | US |