Patent Application
20050080781
The present invention relates to taxonomies for information resources, and in particular to a system and process for generating a taxonomy for a plurality of information resources in a communications network.
The enormous number of stored electronic documents and other information resources available in modem communications networks such as the Internet poses particular problems for classification and categorisation. For example, the world wide web provides access to an ever-increasing number of electronic documents, many of them generated dynamically, and it is often difficult to retrieve a document of interest without knowing in advance at least part of an identifier, address or locator for the resource. For this reason, search engines have been developed which attempt to generate lists of relevant documents in response to keywords typed in by a user. However, such searches are limited by the choice of keywords entered by the user. As an alternative, directories of web resources have been created by manual vetting and categorisation of web documents into hierarchical category structures known as web directories. These directories are extremely useful for locating relevant documents once a particular category has been chosen. However, the development of these directories is a challenge in itself. For example, companies such as Yahoo! have employed more than 300 people for maintaining the structure of their online directory. This level of expenditure is not justifiable for most companies. More recently, some solutions have appeared which replace the manual vetting with automatic classification based on a manually created taxonomy. Although this alleviates the problem to some extent, the manpower needed to create and maintain the appropriate taxonomy is still considerable. It is desired, therefore, to provide an improved system and process for generating a taxonomy for information resources in a communications network, or at least a useful alternative.
In accordance with the present invention there is provided a process for generating a taxonomy for a plurality of information resources in a communications network, including:
collecting said resources from said network;
generating cluster criteria from said resources; and
generating said taxonomy as a hierarchy of resource clusters based on said criteria, wherein the number of said resource clusters is determined by content of said resources.
The present invention also provides an information resource taxonomy system, including
a data collector for collecting information resources from a communications network; and
a taxonomy generator for generating a taxonomy represented by a hierarchy of resource clusters, using cluster criteria generated from said resources, wherein the number of said resource clusters is determined by content of said resources.
Preferred embodiments of the present invention are hereinafter described, by way of example only, with reference to the accompanying drawings, wherein:
As shown in
An administrator of the taxonomy system can login to the system from a terminal associated with the management system 16. The administrator can then submit to the taxonomy system a text file that defines the taxonomy specifications, i.e., the taxonomy creation tasks to be performed by the system. This file includes a list of universal resource indicators (URIs) and a corresponding list of ‘include’ specifications. The URIs indicate high-level domains that are to be clustered or categorised by the taxonomy system, and the ‘include’ specifications indicate the types of documents that are to be included in the taxonomy. For example, it may be desired to include only textual documents in one or more of the following formats: HTML, text, Microsoft Word®, FrameMaker, and StarOffice. The text file containing these specifications is sent to the data collector 10.
The components of the taxonomy system can be implemented using standard computer system hardware and adding unique software modules. For example, the data collector 10 and the renderer 4 are 850 MHz Pentium 3 and 1.5 GHz Pentium 4 personal computers, respectively, each running a Linux operating system. The data processing system 12 is a Sun Ultra Enterprise four-CPU server running a Solaris 8 operating system. The management system 16 is a 1.5 GHz Pentium 4 personal computer running a Windows XP operating system. The data processing system 12 includes a number of data processing modules 18 to 26, including a pre-processor 18, a sampler 20 a clusterer and classifier 22, a taxonomy database 24, and a post processor module 26. The data processing system 12 can further include parallel clusterers 28, and/or parallel classifiers 30. The renderer 14 includes a taxonomy rendering module 15 and a web server module 17. The management system 16 includes a process management component 19 and an editor module 21. Whilst these modules are preferably implemented by software code, at least some of the processing steps executed by the modules, described below, may be implemented by hardware circuits such as application-specific integrated circuits (ASICs).
The data collector 10 executes a data collection process, as shown in
HTML data sent to the data processing system 12 from the data collector 10 is received by the pre-processor 18. Alternatively, HTML data can be directly submitted to the pre-processor 18 by the administrator using the management system 16. The pre-processor 18 executes an HTML processing process, as shown in
The clusterer 22 partitions the documents based on their content. It does this by forming groups or clusters of documents based on their natural affinity rather than requiring a pre-specified number of categories. The clustering and feature selection processes are based upon processes described in the specification of International Patent Application No. PCT/AU01/00198 (“the TACT specification”), incorporated herein by reference. First, each document is represented by a word frequency vector including words from the document and their frequencies of occurrence, where some words are excluded using feature selection criteria. A numeric similarity measure is then determined as a function of any two word vectors to determine the similarity of any two documents. For example, a new cluster can be formed by two documents if their similarity falls within a threshold similarity value for clustering. Once formed, a cluster is characterised by a word frequency vector that is the average of the word frequency vectors of its constituent documents. This average word frequency vector is referred to as the cluster centroid. The similarity measure used is the cosine similarity function, described in the TACT specification. The clustering process uses this similarity measure to group similar documents into clusters by assigning each document to the most similar cluster. An optimal similarity threshold value for creating clusters from a given document set is determined by creating different groupings of the documents at different thresholds and then evaluating these to determine the best grouping, as described in An Evaluation of Criteria for Measuring the Quality of Clusters by B. Raskutti and C. Leckie, pp. 905-910, in Proceedings of the Sixteenth International Joint Conference on Artificial Intelligence, 1999. This evaluation is based on minimising a goodness value that is based on the similarity of documents within clusters, which tends to reduce the number of documents in each cluster, and the separation of cluster centroids from the global centroid, which encourages larger clusters. For example, a goodness value for a document set can be determined by simply summing these two values.
Hierarchical clustering is achieved by iterative clustering of larger, less coherent clusters. The coherence of a cluster is determined by the intra-cluster similarity value of the cluster. If the documents in a cluster are very similar, i.e., the similarity values of each document with the cluster centroid fall within a similarity threshold for coherence, then the cluster is deemed coherent. If this criterion is not met, then documents within the cluster are formed into sub-clusters of the original cluster. These sub-clusters are sub-nodes of the original parent cluster or node, thus forming a hierarchy of clusters or nodes. By performing this sub-clustering iteratively, a hierarchical tree structure of coherent clusters is formed, to provide the taxonomy. The computational complexity of this clustering process is proportional to n, the number of documents, K, the number of threshold evaluations and m, the average number of clusters per threshold.
The clustering process includes several steps for alleviating some of the scalability issues by reducing n and K. Whilst m is much smaller that n, it is proportional to n, therefore reducing n also reduces m. In one form, execution time is reduced by using percentage-based random sampled clustering of the document space whereby the sampler 20 provides a fixed fraction of the document space to the clusterer 22 for clustering.
A second form is provided by stopping the clustering process after a predefined time interval in order to generate a clustered sample of the document space. These two forms of optimisation can be used independently or in conjunction.
After the initial clustering has been performed on a subset or sample of the document set, the remaining documents are subsequently assigned to the clusters by one of three processes. The first process simply classifies documents into the existing clusters using the existing cluster centroids. That is, a new document is added to an existing cluster if its similarity to the cluster centroid falls within a fixed threshold similarity value. Any documents failing the threshold evaluation criteria for all clusters are set aside for later clustering.
The second process uses the sample document clusters as a training set for an alternative document classification system. In this case, a support vector machine (SVM) is used as an alternative classifier. The SVM is described in the specification of International Patent Application No. PCT/AU01/00415, incorporated herein by reference. As with the first process, any documents not classified are set aside for later clustering.
The third process simply continues to cluster, but using the optimal threshold similarity value determined whilst clustering the initial sample documents. This process forms new clusters for new documents that are not similar to the existing clusters.
Each of these three processes is an approximation and assumes that the original sample is representative of the complete (or future) document space. Consequently, errors are introduced over time as more documents are added to the clusters due to cluster centroid drift. Two processes are used to combat this effect In the first process, the coherence of the clusters is maintained as the number of documents n increases by reducing the similarity threshold with increasing n.
In the second process, a new random sample better representing the population is determined as the document collection grows. The new sample is used as a metric for evaluating the optimality of the existing clusters and/or as a means for determining a new quasi-optimal similarity threshold value for subsequent re-clustering of the document space to improve accuracy.
To reduce the time required by the search for an optimal or quasi-optimal similarity threshold value, cluster formation with different threshold values can be performed by different threads or on different processors in an SMP or distributed processing framework, such as the parallel clusterers 28. The time spent searching for the optimum threshold value is also reduced by using an efficient search process based on knowledge of the topography of the goodness vs threshold similarity curve. For example,
The taxonomy produced by clustering is stored in the taxonomy database 24. After clustering, the postprocessor module 26 augments the clustered data by extracting titles from metadata of each document, and adding summary text generated by the clustering process, as described in the TACT specification. In cases where access logs (i.e., web server or proxy cache logs) are available for each document, the clusters and/or documents within each cluster can be ranked using the access frequency of each document. For example, on a corporate web server, the most popular pages are listed near the top of each category listing, and/or the most popular categories are listed near the top of a listing of categories.
The management system 16 includes an editor 21 that allows the administrator to manually edit a taxonomy to create a new document hierarchy. This new structure can then be used as the training set for adding further documents to the database using the classifier function of the clusterer/classifier 22. The speed of document classification by the categorisation system can be improved by using the parallel classifiers 30 to classify many documents in parallel.
The editor 21 offers a number of editing functions, including moving branches of the hierarchical taxonomy to other branches, editing meta descriptions for documents and branches, and creating, deleting, and merging new branches in the taxonomy. The editor 12 presents information from the taxonomy database 24 using HTML forms. Changes can then be made to the taxonomy by modifying input fields in the forms and then submitting the changes via submit buttons of the forms.
The taxonomy rendering module 15 of the renderer 14 generates dynamic web pages using the taxonomy database 24 to provide structure to the original resource content These web pages can be accessed by providing to the web browser 32 a URI associated with the web server module 17. The visual presentation provided by these web pages is derived from a configuration file detailing the arrangement of the various fields on the rendered page. The pages represent a web ‘view’ into the hierarchy using a ‘directory’ style wherein the URI of the displayed page corresponds to the position or branch within the taxonomy that is being browsed. Each level in the ‘view’ can contain documents and/or categories, ie., deeper branches in the taxonomy. Browsing into a category produces a new view with a greater level of specificity. Each branch in the taxonomy is initially labelled automatically by extracting descriptive information from the data during taxonomy generation, as described above, and is manually editable by invoking the editor module 21 of the management system 16. Documents are presented using their titles and summaries. Browsing to the document opens the document or a representation of the document.
Many modifications will be apparent to those skilled in the art without departing from the scope of the present invention as herein described with reference to the accompanying drawings.
| Number | Date | Country | Kind |
|---|---|---|---|
| PR 9589 | Dec 2001 | AU | national |
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/AU02/01719 | 12/18/2002 | WO |
