CN100447793C - Extraction Method of Page Query Interface Based on Visual Feature - Google Patents

Abstract

The invention discloses a method for extracting a page query interface based on visual features. Firstly, a page document containing a query interface is obtained; a visual block tree is constructed for the above-mentioned page document by using a vision-based document segmentation method; the query interface area is positioned; The feature identifies the label block; and then uses the visual feature to complete the grouping of the control block and the label block, thereby determining the control in the query interface and its corresponding attribute label, and realizing the automatic extraction of the query interface. The present invention can realize the automatic extraction of the query interface, which provides a basis for the integrated search of deep web pages; the experiment proves that the automatic extraction method of the query interface based on visual features of the present invention is feasible and has higher precision; the present invention The integrated search applied to the deep web can improve the accuracy of the search, thereby improving people's work efficiency on a large scale.

Description

Extraction Method of Page Query Interface Based on Visual Feature

technical field

The invention relates to an information retrieval method, in particular to a search method based on information extraction technology, which is used to realize the automatic extraction of deep web page query interfaces.

Background technique

There are a large number of information pages on the Internet. Usually, search engines can search these pages through web crawlers (Crawlers), so that visitors can obtain the information pages they need according to keywords. However, with the wide application of Web databases, the Internet is accelerating its "deepening". A large number of pages are dynamically generated by the background database. The information on these pages cannot be obtained directly through static links, but can only be obtained by filling out forms and submitting queries. . Since traditional web crawlers do not have the ability to fill in forms and cannot search these pages, existing search engines cannot provide information on pages generated by such databases, resulting in this part of information being hidden and invisible to search engine users. Visible, can be called deep web (Deep Web, also known as Invisible Web, HiddenWeb). Deep web is a concept corresponding to Surface Web, which was first proposed by Dr. Jill Ellsworth in 1994, and refers to web pages whose information content is difficult to find by ordinary search engines. Deep web information is generally stored in a database. Compared with static web pages, it usually has more information, more specific topics, better information quality, better information structure, and faster growth rate. Research shows that there are 500 times more information on the deep web than on the surface, with nearly 450,000 deep web sites. Therefore, realizing large-scale deep web data integration is an effective way to facilitate users to use deep web information.

To realize large-scale deep web integrated search, it is necessary to solve: 1) Deep Web Discovery; 2) Query Interface Extraction; 3) Source Classification; 4) Query Transfer ); 5) Five key issues such as ResultMerging, among which, how to realize query interface extraction is the most basic and most important issue in order to obtain information from the Web database. Realizing the automatic extraction of query interface is one of the key issues (steps) to realize large-scale integrated retrieval, and it is the basis of Web database modeling, deep web page classification, query interface pattern matching, and construction of unified query interface.

In the prior art, the web page query interface is usually located in the searchable surface web page, which is manually established by people using HTML language, which causes different modes of query interface establishment on different web pages, and many semantically related contents are in the HTML text It appears scattered on the Internet, and the content, form, and query capabilities of the query interface are also different. Therefore, it is a very challenging task to automatically obtain the attributes on the query interface and understand the information content related to the attribute semantics.

Although automatic query interface extraction is an important issue for many applications, there are not many methods dedicated to this problem. For example the following citations, B. Chidlovskii and A. Bergholz. Crawling for domain-specific hidden web resources. In Proceedings of 4th International Conference on Web Information Systems Engineering, 2003; H. He, W. Meng, C. Yu, and Z. Wu. Wise-integrator: An automatic integrator of web search interfaces for e-commerce. In VLDB Conference, 2003; S. Raghavan and H. Garcia-Molina. Crawling the hidden web. In VLDB Conference, 2001. In these citations, some methods deal with the problem of automatic filling of forms, however they only deal with query forms with only simple keywords, or use only drop-down selection labels for simple interactions. For example, it is proposed to use simple heuristic rules such as proximity and alignment to associate elements and labels in the form. This method cannot handle a variety of query interfaces in HTML files, and it is easy to cause the extracted interfaces to be inconsistent. Integrity, which leads to low extraction accuracy of the query interface and affects the construction of the integrated query interface.

Document segmentation based on visual features can identify one or more parts of a document's semantic content. For example, Chinese invention patent application CN1577328A discloses a vision-based document segmentation method, which can segment a Web page document according to visual features, and identify one or more parts of its semantic content. A plurality of visual blocks are identified in the document, and one or more delimiters between the visual blocks of the plurality of visual blocks are detected for identification. A content structure is constructed for the document based at least in part on the plurality of visual blocks and the one or more separators, and the content structure identifies one or more portions of semantic content of the document. The content structure obtained using vision-based document segmentation can optionally be used in the document retrieval process. The technical solution can realize the division of semantic content, thereby improving the accuracy in the web page search process. However, the Web page search it targets is still the surface page search in the prior art, and cannot be directly applied to the deep page search.

Contents of the invention

The purpose of the present invention is to provide a method for extracting query interfaces on pages based on visual features, to be applied to deep page searches, to realize the automatic extraction of query interfaces for deep page searches, thereby improving the recall rate, degree of automation and query efficiency of deep page searches .

In order to achieve the above object, the technical solution adopted in the present invention is: a method for extracting page query interface based on visual features, comprising the following steps:

(1) Obtain a page document containing a query interface;

(2) Using a visual-based document segmentation method to construct a visual block tree for the above-mentioned page documents;

(3) Location query interface area;

(4) identification label blocks, including,

4-1) arrange the text blocks in the query interface area into a list, and get the first text block and classify it into the first class;

4-2) Take a text block and calculate the similarity between it and the existing class. The similarity formula between two text blocks is,

Sim(B1, B2)=w1×wfs(B1, B2)+w2×was(B1, B2)+w3×wcs(B1, B2)+w4×wss(B1, B2)

In the formula, wfs(B1, B2) represents whether the font and background color of B1 and B2 are the same, the same is 1, otherwise it is 0,

was(B1, B2) represents whether the text of B1 and B2 is left-aligned or right-aligned, the alignment is 1, otherwise it is 0,

wcs(B1, B2) represents whether B1 and B2 appear at the same time or not, and it is 1 at the same time, otherwise it is 0,

wss(B1, B2) represents whether the texts of B1 and B2 are in the same line, if they are not in the same line, it is 1, otherwise it is 0,

w1 is 3.5-4.5, w2 is 1.5-2.5, w3 is 1.5-2.5, w4 is 1.5-2.5, and w1+w2+w3+w4=10;

The similarity between a text block and a class is the average of the similarities between the text block and all text blocks in the class. If the similarity between a text block and a certain class is greater than the similarity threshold (Tas), the text block will be Be classified into this class; If its similarity with any existing class is not greater than the similarity threshold, then create a new class, and this text block is classified into the newly created class, and the similarity threshold is 6;

4-3) repeat step 4-2), until finishing the classification of text blocks;

4-4) According to the display feature, determine that the class with the highest degree of conformity in the obtained text block class is the label class; the display feature includes that the label is usually not in the same line, multiple text blocks appear in the same line, and the first block is Labels; Labels are usually left-aligned or right-aligned; Labels have the same font size, color, and background color;

(5) Grouping of control blocks and label blocks,

5-1) Create a list of control blocks, delete the submit, reset, and image control blocks;

5-2) Comparing each control block with the label block obtained in step (4), grouping the control blocks and label blocks displayed on the same row into one group;

5-3) According to the display characteristics, group the remaining control blocks and the adjacent label blocks above them into one group to complete the grouping of control blocks and label blocks;

In this way, the controls in the query interface and their corresponding attribute labels are determined, so as to realize the automatic extraction of the query interface.

In the above technical solution, the page document is an HTML format document.

The method of constructing a visual block tree for the above-mentioned page document using a vision-based document segmentation method is a prior art, for example, the document segmentation method disclosed in Chinese invention patent application CN1577328A can be used.

In the above technical solution, the query interface area is usually the smallest block containing the <form> tag. However, there are usually not only query forms on a webpage, but also forms that are meaningless for deep page searches, such as member login and email subscription. interference, therefore, in order to obtain better results, the further technical solution is: the method for positioning and querying the interface area is:

3-1) Set the control number threshold;

3-2) In the visual block tree obtained in step (2), take the smallest block containing the <form> tag as the preset query interface area, and calculate the number of controls contained therein;

3-3) If the number of controls is greater than the set control number threshold, mark the block as the query interface area, otherwise detect the next smallest block containing the <form> tag;

3-4) Repeat steps 3-2) and 3-3) to complete the positioning of the query interface area.

Wherein, the control number threshold is between 2 and 4.

Or, the method for locating the query interface area is: in the visual block tree obtained in step (2), take the smallest block containing the <form> tag as the preset query interface area, if the PASSWORD control is included in the area, then Delete the preset query interface area, and detect the next smallest block containing the <form> tag until the query interface area is determined or all visual blocks are detected.

In the above technical solution, in step 5-3), the display feature also includes that the control and the label are vertically aligned in different rows; the control and the label are adjacent to each other; if there is no label on the same line of the control, the control is closest to the top Label association.

Due to the use of the above-mentioned technical solutions, the present invention has the following advantages compared with the prior art:

1. The present invention combines the document segmentation based on visual features with the visual features of labels and controls, and realizes the grouping and combination of controls and their label attributes in the query interface area, so that automatic extraction of query interfaces can be realized. The integrated search provides the basis for;

2. Experiments have proved that the automatic extraction method of query interface based on visual features of the present invention is feasible and has higher precision;

3. Applying the present invention to the integrated search of deep web pages can improve the accuracy of search, thereby improving people's work efficiency in a large range.

Description of drawings

Accompanying drawing 1 is the automatic extraction system diagram of the query interface based on the visual feature of the embodiment of the present invention;

Accompanying drawing 2 is the query interface page display schematic diagram of the embodiment of the present invention;

Accompanying drawing 3 is the schematic diagram of page content structure of Fig. 2;

Accompanying drawing 4 is the schematic diagram of visual block tree corresponding to Fig. 3;

Accompanying drawing 5 is a schematic diagram of the automatic extraction process of the present invention.

Detailed ways

The present invention will be further described below in conjunction with accompanying drawing and embodiment:

Embodiment 1: Referring to accompanying drawings 1 to 5, a method for extracting a page query interface based on visual features comprises the following steps:

(1) Obtain a page document containing a query interface;

Referring to accompanying drawing 2, a query interface includes some form controls to allow users to input query information, such as controls such as text boxes (Textbox), radio buttons (Radio Button), check boxes (Check box) and drop-down lists (Selection List). Each control is usually associated with a label - a descriptive text, each control can have one or more values (value), for example a drop-down list has a list of values for the user to choose from, radio buttons and check boxes usually have a value . Logically speaking, a control and its associated label constitute an attribute, which corresponds to a field in the backend database of the Deep Web. Typically, a property contains a label and one or more form controls. For example, the Author attribute in Figure 2 has 4 form controls, including 1 text box and 3 radio buttons. The label in the attribute can be regarded as the name of the attribute (attribute name), and the form control in the attribute can be regarded as the domain of the attribute (attribute domain). If a property includes multiple form controls, there is some relationship between those form controls. For example, the four form controls of the author attribute in Figure 2, the text box can be regarded as a domain element (domain element), and the three radio buttons define the constraints of the domain control respectively. We can regard them as constraint elements (constraint elements) ).

To be able to extract the query interface from the above page document, first determine the query interface area, then extract the labels and form controls from the query interface area, and then reorganize them into attributes according to the logical relationship (a logic of the query condition unit). We can abstractly define a query interface QI as QI=(N, {A ₁ , A ₂ ,...,A _n }), where N is the name of the form, A _i represents the logical attributes contained in the query interface, A _i ={L _i , E _i...k }, where L _i is an attribute label, and E _i is a form control.

(2) Using a visual-based document segmentation method to construct a visual block tree for the above-mentioned page documents;

The goal of visual feature-based page segmentation algorithms (VIPs) is to extract the content structure of Web pages according to their visual presentation. These content structures are a tree structure, and each node in the tree corresponds to a rectangular area in the Web page. A leaf node is a block that cannot be further divided, and it represents a minimum semantic unit. The relationship between the parent node and the child node is an inclusion relationship. The rectangular area corresponding to the leaf node is included in the rectangular area corresponding to the parent node. This tree is called a visual block tree. We use the VIPs algorithm to convert the page where the query interface is located into a visual block tree. Attached Figure 3 shows the content structure of the page located in Figure 2. The dotted box is the area of the query interface. Figure 4 shows the corresponding visual block tree Schematic diagram, the actual visual block tree is more complex than that shown in Figure 4, which usually contains hundreds of blocks.

(3) Location query interface area;

Generally, the smallest Block containing the <form> tag can be used as the query interface area, but usually a web page not only has a query form but also forms such as member login, mail subscription, etc. that are meaningless to the present invention, forming interference. For this reason, in this embodiment, further optimization can be done, and the query interface area is determined according to some heuristic rules. For example, some web forms have TEXTAREA controls and PASSWORD controls. According to actual experience, we can directly determine that such web forms are not query interfaces. In addition, a threshold can be set for the number of controls in a web form. When the number of controls in a web form is lower than this threshold, it can be considered that the web form is not a query interface. There's a textbox and a submit button, and we don't get enough information about this type of web form, so we can classify them as non-query interfaces. The value of the threshold is an empirical value, generally, the value of 2, 3 or 4 is more appropriate.

(4) Identification label block

To identify label blocks, it is first necessary to understand the visual features of the query interface.

The query interface is used to allow users to input query information and submit queries. In order to make it easier for users to understand and use the query interface, designers usually incorporate various types of visual features, such as fonts, colors, layouts, etc. Therefore, visual features are very important for query interface extraction. Below we will describe some visual features that are mainly used by the method of the present invention.

Position Features (PF, Position Features): These features describe the relative position of labels and form controls in the form.

PF1: Labels are usually not in the same line, and there are multiple text blocks in the same line. The first block is a label, and the other blocks are explanatory text, such as: between...and

PF2: Form controls and their property labels are usually vertically aligned on the same line or on different lines.

Layout Features (LF, Layout Features): These features indicate the arrangement of attributes and form controls in the query interface.

LF1: Labels are usually left-aligned or right-aligned.

LF2: The label in the attribute is adjacent to the form control.

LF3: If there is no label on the same line as the form control, the control is associated with the nearest neighbor label above.

Appearance Features (AF, Appearance Features): These features describe the visual characteristics of labels and form controls.

AF1: The font size, font color, and background of the label are the same.

AF2: Tags usually end with a colon.

There are multiple text blocks in the query interface, and the label block needs to be determined according to the PF, LF, and AF features. The specific method is:

4-1) arrange the text blocks in the query interface area into a list in any order, and take the first text block and classify it into the first class;

4-2) Take a text block and calculate the similarity between it and the existing class. The similarity formula between two text blocks is,

Sim(B1, B2)=w1×wfs(B1, B2)+w2×was(B1, B2)+w3×wcs(B1, B2)+w4×wss(B1, B2)

In the formula, wfs(B1, B2) represents whether the font and background color of B1 and B2 are the same, the same is 1, otherwise it is 0,

was(B1, B2) represents whether the text of B1 and B2 is left-aligned or right-aligned, the alignment is 1, otherwise it is 0,

wcs(B1, B2) represents whether B1 and B2 appear at the same time or not, and it is 1 at the same time, otherwise it is 0,

wss(B1, B2) represents whether the texts of B1 and B2 are in the same line, if they are not in the same line, it is 1, otherwise it is 0,

w1 is 3.5-4.5, w2 is 1.5-2.5, w3 is 1.5-2.5, w4 is 1.5-2.5, and w1+w2+w3+w4=10; the preferred values are 4, 2, 2, 2 respectively;

The similarity between a text block and a class is the average value of the similarity between the text block and all text blocks in this class, if the similarity between a text block and a certain class is greater than 6, the text block is classified into this class; If its similarity with any existing class is not greater than 6, then create a new class, and classify the text block into the newly created class;

4-3) repeat step 4-2), until finishing the classification of text blocks;

4-4) According to the display feature, determine that the class with the highest degree of conformity in the obtained text block class is the label class; the display feature includes that the label is usually not in the same line, multiple text blocks appear in the same line, and the first block is Labels; labels are usually left-aligned or right-aligned; labels have the same font size, color, and background color.

(5) Grouping of control blocks and label blocks,

The previous step found and divided the label blocks into one category, which is equivalent to finding the name of the attribute and determining the number of attributes. Then through three steps, the control block and the related label block are composed of logic attributes.

5-1) Create a list of control blocks. Since the query interface contains form controls for users to submit queries, there are usually three types of submit, reset, and image. These controls do not have logical attributes, so these control blocks must be deleted first. ;

5-2) Comparing each control block with the label block obtained in step (4), grouping the control blocks and label blocks displayed on the same row into one group;

5-3) According to the display features PF2, LF2, and LF3, group the remaining control blocks and the adjacent label blocks above them into one group to complete the grouping of control blocks and label blocks.

In this way, the controls in the query interface and their corresponding attribute labels are determined, so as to realize the automatic extraction of the query interface.

Claims (5)

1. A method for extracting page query interfaces based on visual features, comprising the following steps: (1) Obtain a page document that includes a query interface, and the page document is an HTML format document; (2) Using a visual-based document segmentation method to construct a visual block tree for the above-mentioned page documents; (3) Location query interface area; (4) identification label blocks, including, 4-1) arrange the text blocks in the query interface area into a list, and get the first text block and classify it into the first class; 4-2) Take a text block and calculate the similarity between it and the existing class. The similarity formula between two text blocks is, Sim(B1, B2)=w1×wfs(B1, B2)+w2×was(B1, B2)+w3×wcs(B1, B2)+w4×wss(B1, B2) In the formula, wfs(B1, B2) represents whether the font and background color of B1 and B2 are the same, the same is 1, otherwise it is 0, was(B1, B2) represents whether the texts of B1 and B2 are left-aligned or right-aligned, and the alignment is 1, Otherwise, it is 0, wcs(B1, B2) represents whether B1 and B2 appear at the same time or not, and it is 1 at the same time, otherwise it is 0, wss(B1, B2) represents whether B1 and B2 text are in the same line, not in the same line 1 , otherwise it is 0, w1 is 3.5~4.5, w2 is 1.5~2.5, w3 is 1.5~2.5, w4 is 1.5~2.5, and w1+w2+w3+w4=10; The similarity between a text block and a class is the average of the similarities between the text block and all text blocks in the class. If the similarity between a text block and a class is greater than the similarity threshold, the text block will be included in the class. class; if its similarity with any existing class is not greater than the similarity threshold, then create a new class, and put the text block into the newly created class, and the similarity threshold is 6; 4-3) repeat step 4-2), until finishing the classification of text blocks; 4-4) According to the display feature, determine that the class with the highest degree of conformity in the obtained text block class is the label class; the display feature includes that the label is usually not in the same line, multiple text blocks appear in the same line, and the first block is Labels; Labels are usually left-aligned or right-aligned; Labels have the same font size, color, and background color; (5) Grouping of control blocks and label blocks, 5-1) Create a list of control blocks, delete the submit, reset, and image control blocks; 5-2) Comparing each control block with the label block obtained in step (4), grouping the control blocks and label blocks displayed on the same row into one group; 5-3) According to the display characteristics, group the remaining control blocks and the adjacent label blocks above them into one group to complete the grouping of control blocks and label blocks; In this way, the controls in the query interface and their corresponding attribute labels are determined, so as to realize the automatic extraction of the query interface. 2. the page query interface extraction method based on visual features according to claim 1, characterized in that: the method for the location query interface area is, 3-1) Set the control number threshold; 3-2) In the visual block tree obtained in step (2), take the smallest block containing the <form> tag as the preset query interface area, and calculate the number of controls contained therein; 3-3) If the number of controls is greater than the set control number threshold, mark the block as the query interface area, otherwise detect the next smallest block containing the <form> tag; 3-4) Repeat steps 3-2) and 3-3) to complete the positioning of the query interface area. 3 . The visual feature-based page query interface extraction method according to claim 2 , characterized in that: the control number threshold is between 2 and 4. 4 . 4. the method for extracting page query interface based on visual features according to claim 1, characterized in that: the method for the location query interface area is, in the visual block tree obtained in step (2), take <form> The smallest block of the label is used as the preset query interface area. If the area contains PASSWORD controls, delete the preset query interface area, and check the next smallest block containing the <form> tag until the query interface area is determined or all detections are completed. visual block. 5. The method for extracting page query interface based on visual features according to claim 1, characterized in that: in step 5-3), said display features also include that controls and labels are vertically aligned in different rows; controls and labels Adjacent to each other; if the control has no label on the same row, the control is grouped with the nearest label above.

Images