Patent Grant
7363279
The described technology relates generally to identifying a block within a display page that represents the primary topic of the display page.
Many search engine services, such as Google and Overture, provide for searching for information that is accessible via the Internet. These search engine services allow users to search for display pages, such as web pages, that may be of interest to users. After a user submits a search request that includes search terms, the search engine service identifies web pages that may be related to those search terms. To quickly identify related web pages, the search engine services may maintain a mapping of keywords to web pages. This mapping may be generated by “crawling” the web (i.e., the World Wide Web) to identify the keywords of each web page. To crawl the web, a search engine service may use a list of root web pages to identify all web pages that are accessible through those root web pages. The keywords of any particular web page can be identified using various well-known information retrieval techniques, such as identifying the words of a headline, the words supplied in the metadata of the web page, the words that are highlighted, and so on. The search engine service then ranks the web pages of the search result based on the closeness of each match, web page popularity (e.g., Google's PageRank), and so on. The search engine service may also generate a relevance score to indicate how relevant the information of the web page may be to the search request. The search engine service then displays to the user links to those web pages in an order that is based on their rankings.
Whether the web pages of a search result are of interest to a user depends, in large part, on how well the keywords identified by the search engine service represent the primary topic of a web page. Because a web page may contain many different types of information, it may be difficult to discern the primary topic of a web page. For example, many web pages contain advertisements that are unrelated to the primary topic of the web page. A web page from a news web site may contain an article relating to an international political event and may contain “noise information” such as an advertisement for a popular diet, an area related to legal notices, and a navigation bar. It has been traditionally very difficult for a search engine service to identify what information on a web page is noise information and what information relates to the primary topic of the web page. As a result, a search engine service may select keywords based on noise information, rather than the primary topic of the web page. For example, a search engine service may map a web page that contains a diet advertisement to the keyword “diet,” even though the primary topic of the web page relates to an international political event. When a user then submits a search request that includes the search term “diet,” the search engine service may return the web page that contains the diet advertisement, which is unlikely to be of interest to the user.
It would be desirable to have a technique for calculating the importance of various information areas of a web page to the primary topic of the web page.
An importance system identifies the importance of information areas of a display page. The importance system identifies information areas or blocks of a web page that represent areas of the web page that appear to relate to a similar topic. After identifying the blocks of a web page, the importance system provides the characteristics or features of a block to an importance function which generates an indication of the importance of that block to its web page. The importance system may learn the importance function by generating a model based on the features of blocks and the user-specified importance of those blocks.
A method and system for identifying the importance of information areas of a display page is provided. In one embodiment, an importance system identifies information areas or blocks of a web page. A block of a web page represents an area of the web page that appears to relate to a similar topic. For example, a news article of a web page may represent one block, and a diet advertisement of the web page may represent another block. After identifying the blocks of a web page, the importance system provides the characteristics or features of a block to an importance function that generates an indication of the importance of that block to its web page. The importance system “learns” the importance function by generating a model based on the features of blocks and the user-specified importance of those blocks. To learn the importance function, the importance system asks users to provide an indication of the importance of blocks of web pages in a collection of web pages. For example, one user may identify a certain block as of primary importance to a web page, and another user may identify that same block as of secondary importance to the web page. The importance system also identifies a feature vector for each block that may represent various spatial and content features of the block. For example, a spatial feature may be the size of a block or the location of a block, and a content feature may be the number of links within a block or the number of words within the block. The importance system then learns an importance function, based on the user-specified importance and the feature vectors of the blocks, that inputs a feature vector of a block and outputs the importance of the block. The importance system may attempt to learn a function that minimizes the square of the differences between the user-specified importance of the blocks and the importance calculated by the importance function. In this way, the importance system can identify a block of a web page that may be related to the primary topic of the web page based on features of blocks that users think are important.
The importance function, once learned, can be used to improve the accuracy of a wide range of applications whose results depend on discerning topics of web pages. A search engine service may use the importance function to calculate relevance of web pages of a search result to the search request. After the search result is obtained, the search engine service may use the importance function to identify the most important block of each web page and calculate a relevance of the most important block (or group of importance blocks) to the search request. Because this relevance is based on the information of the most important block of the web page, it may be a more accurate measure of relevance than a relevance that is based on the overall information of the web page, which may include noise information.
A search engine service may also use the importance function to identify web pages that are related to a certain web page. The search engine service may use the importance function to identify the most important block of that web page and then formulate a search request based on the text of the identified block. The search engine service may perform a search using that formulated search request to identify web pages that are related to that certain web page.
A search engine service may also use the importance function to classify web pages. For example, web pages may be classified based on their primary topic. The search engine service may use the importance function to identify the most important block of a web page and then classify the web page based on the text of the identified block.
A web browser may also use the importance function to identify portions of a web page to be displayed when the entire content of the web page cannot fit on a display device. For example, devices such as cell phones or personal digital assistants may have very small displays on which most web pages cannot reasonably be displayed because of the amount, size, and complexity of the information of the web page. A web page is not reasonably displayed when the content is so small that it cannot be effectively viewed by a person. When such a device is used to browse web pages, the importance function may be used to identify a most important block of a web page and to display only the identified block on the small display. Alternatively, a web browser may rearrange the blocks of a web page, highlight blocks of a web page, remove blocks of a web page, and so on based on the importance of the blocks as indicated by the importance function.
The importance system can identify the information areas of a web page using various segmentation techniques such as a document object model based (“DOM-based”) segmentation algorithm, a location-based segmentation algorithm, a vision-based segmentation algorithm, and so on. A DOM-based segmentation algorithm may use the HTML hierarchy of a web page to identify its various blocks. A location-based segmentation algorithm attempts to identify areas of a web page that may be considered a unit based on physical characteristics of the areas. A vision-based segmentation algorithm is described in U.S. patent application Ser. No. 10/628,766, entitled “Vision-Based Document Segmentation,” filed on Jul. 28, 2003, which is hereby incorporated by reference. That vision-based segmentation algorithm identifies blocks based on the coherency of the information content of each block. For example, an advertisement for a diet may represent a single block because its content is a directed to a single topic and thus has a high coherency. An area of a display page that includes many different topics may be divided into many different blocks.
The importance system may use various techniques to learn the importance function from the feature vectors representing the blocks and the user-specified importance of the blocks. For example, as described below in more detail, the importance system may learn the importance function using a neural network or a support vector machine. In addition, the importance system may use various spatial and content features of a block in a feature vector representation of the block. For example, the spatial features may include the center location of the block and the dimensions of the block, which are referred to as “absolute spatial features.” These absolute spatial features may be normalized based on the size of the web page, which are referred to as “relative spatial features,” or normalized based on the size of the window, which are referred to as “window spatial features.” The content features may relate to the images of the block (e.g., number and size of images), links of the block (e.g., the number of links and the number of words in each link), text of the block (e.g., number of words in block), user interaction of the block (e.g., number and size of input fields), and forms of the block (e.g., number and size). Various content features may also be normalized. For example, the number of links of a block may be normalized by the total number of links on the web page, or the number of words of text in the block may be normalized by the total number of words in the text of all the blocks of the web page.
In one embodiment, the importance system may allow users to specify the importance of blocks using discrete or continuous values. Higher values may represent a higher importance of a block to a web page. When using discrete values of 1-4, a value of 1 may represent noise information such as an advertisement, a copyright notice, a decoration, and so on. A value of 2 may represent useful information that is not particularly relevant to the primary topic of the page such as navigation information, directory information, and so on. A value of 3 may represent information that is relevant to the primary topic of the page but not of prominent importance such as related topics, topic indexes, and so on. A value of 4 may indicate the most prominent part of the web pages such as a headline, main content, and so on.
The computing device on which the importance system can be implemented may include a central processing unit, memory, input devices (e.g., keyboard and pointing devices), output devices (e.g., display devices), and storage devices (e.g., disk drives). The memory and storage devices are computer-readable media that may contain instructions that implement the importance system. In addition, the data structures and message structures may be stored or transmitted via a data transmission medium, such as a signal on a communications link. Various communications links may be used, such as the Internet, a local area network, a wide area network, or a point-to-point dial-up connection.
The importance system may be described in the general context of computer-executable instructions, such as program modules, executed by one or more computers or other devices. Generally, program modules include routines, programs, objects, components, data structures, and so on that perform particular tasks or implement particular abstract data types. Typically, the functionality of the program modules may be combined or distributed as desired in various embodiments.
In one embodiment, the importance system learns a block importance function by minimizing a least squares objective function based on user-specified importance of blocks of a collection of web pages. Each block can be represented by a feature vector and importance pair (x, y) where x is the feature vector of the block and y is its importance. The set of feature vector and importance pairs of the collection of web pages is referred to as a training set T. The importance system attempts to learn an importance function ƒ that minimizes the following objective function:
Σ(x,y)∈T|ƒ(x)−y|2 (1)
The objective function can be solved using classification techniques such as a support vector machine if y is discrete and regression techniques such as a neural network if y is continuous.
When the importance is represented by continuous real numbers, the importance system may apply a neural network learning for learning the optimal ƒ* which is given by minimizing the following objective function:
where m is the number of blocks in the training set. This is a multivariate non-parametric regression problem, since there is no a priori knowledge about the form of the true regression function that is being estimated.
A neural network model has three major components: architecture, cost function, and search algorithm. The architecture defines the functional form relating the inputs to the outputs (in terms of network topology, unit connectivity, and activation functions). The search in weight space for a set of weights that minimizes the objective function is the training process. In one embodiment, the importance system uses a radial basis function (“RBF”) network and a standard gradient descent as the search technique.
The importance system constructs an RBF network with three layers having different roles. The input layer comprises source nodes (i.e., sensory units) that connect the network to its environment (i.e., low-level feature space). The hidden layer applies a nonlinear transformation from the input space to the hidden space. Generally, the hidden space is of high dimensionality. The hidden layer has RBF neurons, which calculate the input of the hidden layer by combining weighted inputs and biases. The output layer is linear and supplies the block importance given the low-level block representation applied to the input layer.
The function learned by RBF networks can be represented by the following:
where i is a block in the training set, h is the number of hidden layer neurons,
where ci is the center for Gi and σi is the basis function width. The k-dimensional mapping can be represented as follows:
x→ƒ(x)=(ƒ1(x), ƒ2(x), . . . , ƒk(x)) (5)
where ƒ=[ƒ1, ƒ2, . . . , ƒk] is the mapping function.
The RBF neural network approximates the optimal regression function from feature space to block importance. The importance system can train the RBF neural network off-line with the training samples {xi, yi} (i=1, . . . , m) of the training set T. For a new block previously unprocessed, the importance system can calculate its importance using the regression function ƒ given the feature vector of the block.
When the importance is represented by discrete numbers, the importance system applies a support vector machine for learning the importance function. A support vector machine attempts to minimize structural risk, rather than empirical risk. The support vector machine may use a binary classification such that the training set is defined by the following:
D={xi, yi}i=1t (6)
where D is the training set, t is the number of training samples, and yi∈{−1,1}. The importance system attempts to select, among the infinite number of linear classifiers that separate the data, the one with the minimum generalization error. A hyperplane with this property is the one that leaves the maximum margin between the two classes. The importance function may be represented by the following:
where αi associated with the training sample xi expresses the strength with which that point is embedded in the final function and b is the intercept also known as the bias in machine learning. A property of this representation is that often only a subset of the points will be associated with non-zero αi. These points are called support vectors and are the points that lie closest to the separating hyperplane. The nonlinear support vector machine maps the input variable into a high dimensional (often infinite dimensional) space, and applies the linear support vector machine in the space. Computationally, this can be achieved by the application of a (reproducing) kernel. The corresponding nonlinear decision function is represented as follows:
where K is the kernel function. Some typical kernel functions include a polynomial kernel, Gaussian RBF kernel, and sigmoid kernel. For a multi-class classification problem, the importance system can apply a one-against-all scheme.
where HeaderHeight and FooterHeight are predefined constant values relating to the heights of headers and footers of a page. In block 303, the component identifies the content features of the block. In block 304, the component normalizes the content features as appropriate and then returns the feature vector.
One skilled in the art will appreciate that although specific embodiments of the importance system have been described herein for purposes of illustration, various modifications may be made without deviating from the spirit and scope of the invention. The principles of the importance system may be used to determine the importance of information areas within various information sources other than web pages. These information sources may include documents represented using HTML, XML, or other markup languages. One skilled in the art will appreciate that some applications of block importance may use an importance function that is not “learned.” For example, a developer can simply define an importance function that reflects their concept of block importance and not rely on collecting empirical data of user ratings of block importance. Such an importance function could rate the importance of a block based on a combination of its size and location. Accordingly, the invention is not limited except by the appended claims.
| Number | Name | Date | Kind |
|---|---|---|---|
| 6647410 | Scimone et al. | Nov 2003 | B1 |
| 6701350 | Mitchell | Mar 2004 | B1 |
| 20020007379 | Wang et al. | Jan 2002 | A1 |
| 20020062302 | Oosta | May 2002 | A1 |
| 20020087221 | Keeler et al. | Jul 2002 | A1 |
| 20020129011 | Julien | Sep 2002 | A1 |
| 20030099395 | Wang et al. | May 2003 | A1 |
| 20030115189 | Srinivasa et al. | Jun 2003 | A1 |
| 20030128877 | Nicponski | Jul 2003 | A1 |
| 20030130993 | Mendelevitch et al. | Jul 2003 | A1 |
| 20030158855 | Farnham | Aug 2003 | A1 |
| 20040024598 | Srivastava et al. | Feb 2004 | A1 |
| 20040044747 | Trevor et al. | Mar 2004 | A1 |
| 20040049374 | Breslau et al. | Mar 2004 | A1 |
| 20040076345 | Olszak et al. | Apr 2004 | A1 |
| Number | Date | Country |
|---|---|---|
| 0 964 340 | Dec 1999 | EP |
| Number | Date | Country | |
|---|---|---|---|
| 20050246296 A1 | Nov 2005 | US |
