CN102364473B - Netnews search system and method based on geographic information and visual information - Google Patents

Abstract

The invention proposes a network news retrieval system and method that integrate geographic information and visual information. The system includes: a data preprocessing module for crawling news data and performing text analysis and information data extraction, the news data including characters, places, time and text information; location correlation analysis module for performing news events and Correlation analysis of news locations; news map module, used to select appropriate images for news; retrieval result display module, used to sort and display retrieved news based on retrieval relevance. The system and method of the present invention comprehensively utilize geographic location information and visual information to describe and display network news, provide multimedia news retrieval based on geographic location for network users, and simultaneously integrate news location-news event relationship and news location correlation Sex and the relationship between news events, so as to provide users with a more vivid and informative news search results.

Description

Network news retrieval system and method integrating geographic information and visual information

technical field

The invention relates to the field of network news retrieval, in particular to a network news retrieval system and method that integrates geographical information and visual information.

Background technique

With the development of information technology and the globalization of the network, online news is becoming more and more popular, and it has become an important way for people to obtain information in their daily lives. People can obtain and browse news through some major Internet portals such as Yahoo, MSN or large news websites such as CNN, AOL and MSNBC.

However, the news display method in the prior art has several deficiencies.

For example, existing approaches to news presentation lack geographic-based organization. Studies have shown that users often give priority to news about a few specific locations, such as hometown and work. Most of the major news sites can organize news by relevant country. Users can submit a location as a search term to retrieve news. However, the geographic terms contained in the documents are often noisy, which reduces the retrieval performance.

In addition, existing news display methods do not contain comprehensive visual information.

Fig. 1 shows the distribution of the number of pictures contained in a news document in the prior art.

It can be seen from FIG. 1 that most news documents in the prior art have no pictures or contain very few pictures. For example, less than 5% of news documents contain more than one image.

Generally speaking, pictures are worth a thousand words. As a supplement to news texts, news pictures can enable users to obtain information faster. However, as shown in Figure 1, the number of pictures contained in the existing news documents is very small, so it is far from meeting the comprehensive needs of users for obtaining information.

Contents of the invention

The purpose of the present invention is to provide a network news retrieval system and method that integrates geographic and visual information. According to the system and method of the present invention, it is possible to provide users with news organized based on geographic information, so that users can quickly browse news events invented in the area they care about; furthermore, the present invention uses image information to supplement text information, so that users can quickly Accurately grasp the content of news events.

According to one aspect of the present invention, a network news retrieval system that integrates geographical information and visual information is provided, the system includes: a data preprocessing module, used to crawl news data and perform text analysis and information data extraction, the news The data includes person, location, time and text information; the location correlation analysis module is used to perform correlation analysis between news events and news locations; the news picture matching module is used to select appropriate images for news; the retrieval result display module uses Display the retrieved news in order based on retrieval relevance.

Wherein, the data preprocessing module includes: a news data crawling module, which is used to crawl news documents and corresponding news images from news websites; a text analysis module, which is used to extract the title, time, website, abstract of news data According to the text and the corresponding URL, the URL of the news image and the text information corresponding to the image are extracted; the news entity extraction module extracts the person, place and time from the news data.

The location correlation analysis module includes: geographic noun filtering and expansion module, used to obtain geographic location information of geographic nouns; matrix decomposition-based correlation analysis module, used to analyze news locations and news by using consistency constraint probability matrix decomposition method relationship between events.

Described news matching picture module comprises: retrieval term generation module, is used for extracting one or more key words from news data, it is combined into retrieval term and is submitted to image search engine and carries out image retrieval; Image sorting and selection module, uses It is used to sort and deduplicate the retrieved images, and select images that can express the content of news documents.

The retrieval result display module includes: a map view module, used to display the distribution position of the selected news on the map; a news event list module, used to sort and display the list of retrieved news events according to predetermined rules.

In the location correlation analysis module, the consistency constraint probability matrix decomposition method analyzes the relationship between news locations and news events based on the following rules: news events with higher similarity are likely to occur in the same place, and the correlation Higher multiple locations are closely related to the same news event.

In the news picture matching module, the search term generation module extracts search terms from multiple parts of the news data for image retrieval; the image sorting and selection module sorts the retrieved images using a method based on hierarchical aggregation.

In the retrieval result display module, the map view module displays the titles and corresponding images of the most relevant news events in response to the search words input by the user or clicking on any location on the map; the predetermined rule sorting includes the following One or more of: the correlation between news events, the correlation between news events and retrieval locations, and the time information of news occurrence.

According to another aspect of the present invention, there is also provided a network news retrieval method that integrates geographic information and visual information, the method includes: a data preprocessing step for crawling news data and performing text analysis and information data extraction, so The described news data includes people, place, time and text information; the place correlation analysis step is used to perform the correlation analysis between news events and news places; the news matching step is used to select suitable images for news; the retrieval result display step , used to sort and display the retrieved news based on retrieval relevance.

As mentioned above, the system and method of the present invention provide news retrieval based on geographical location, estimate and refine the relationship between news documents and geographical location, and comprehensively consider the initial relationship between news documents and locations in the process, Relationships between locations and similarities between news documents. In addition, by proposing a method for matching pictures to news documents, users can obtain news more directly and quickly. In addition, a retrieval module that supports news geographic terms, and a user retrieval and browsing interface that can be retrieved by clicking on the corresponding geographic location on the map are also proposed.

According to the consistency constraint probabilistic matrix decomposition model proposed by the present invention, the location-event relationship, the correlation between locations and the similarity between events can be fused together, the correlation between locations and events can be estimated and refined, and the correlation between locations and events can be removed. Noise and estimated potential relationships.

According to the method for extracting search words from texts to perform network image retrieval and image sorting proposed by the present invention, accurate picture matching of news documents can be realized according to various rules.

According to the method for combining retrieval words proposed by the present invention, the problems that current network search engines cannot handle complex retrieval and the problem that a single word cannot express document content as retrieval can be avoided.

In addition, for different image lists obtained from network search engines, the present invention also proposes a method based on level aggregation to fuse and sort these image lists, so as to select the image that best expresses the content of the news document.

According to the news document sorting method proposed by the present invention, the timeliness, importance and retrieval relevance of news are considered comprehensively. This method is based on the traditional Markov random walk model, which linearly fuses the news event-news location correlation obtained from the previous analysis and the timeliness of news documents into the initial state of the model, and based on the similarity between news documents , to achieve the retrieval relevance sorting of the news document collection.

The invention also provides an interactive interface for users to retrieve and browse news. Users can search by submitting a search or clicking on the map. At the same time, a search result is provided with title, news pictures and content abstracts, and users can quickly and vividly obtain the information they need. If users want to know more detailed information, they can get it by clicking the "More" button on the interface.

In summary, the present invention uses the name of the place where the news occurs as the search keyword or clicks on the geographical location of interest on the map to provide the user with a more vivid and informative news search result. The result display interface includes two parts : First, display the most relevant news titles and picture information on the real map according to the geographical location of the event; second, a multi-modal search result list with news titles, pictures and short descriptions.

Description of drawings

Figure 1 shows the distribution of the number of pictures contained in a news document in the prior art;

Fig. 2 has shown the schematic diagram of network news retrieval system of the present invention;

Fig. 3 is the matrix decomposition model figure of consistency constraint probability that the present invention proposes;

Fig. 4 has shown the example of a news matching picture among the present invention;

Fig. 5 has shown the network news retrieval and browsing interface of an embodiment of the present invention;

Figure 6 shows the retrieval performance evaluation results of the BM25 ranking model, the probability matrix factorization model and the consistency constraint probability matrix factorization model;

Figure 7 shows the results of changing parameters under the NDCG50 rule;

Fig. 8 is the performance comparison result of the news map matching method of the present invention and the prior art;

Fig. 9 has shown the comparison result of retrieval relevancy between the retrieval result sorting method of the present invention and the prior art sorting method;

Fig. 10 shows the comparison results of the retrieval result sorting method of the present invention and the sorting method of the prior art in terms of timeliness.

Detailed ways

In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in combination with specific examples and with reference to the accompanying drawings. Although the examples of the present invention are provided based on English news, the method of the present invention is not limited by the language category.

The invention proposes a computer-based news retrieval system based on multimedia analysis, which comprehensively utilizes geographic information and visual information. Firstly, the news location candidate set is extracted from the text, filtered and expanded using network information, and its geographic location information (latitude and longitude) is obtained. Discover the relationship between potential news locations and news events through the consistency constraint probability matrix decomposition relationship mining technology proposed by the present invention, comprehensively considering the correlation between news locations, the similarity of news events and news locations-news events the initial relationship between. Then, in order to enable users to obtain news quickly and vividly, the present invention proposes a method for matching pictures to news.

Although the current news documents also contain news pictures, there are too few corresponding news pictures, and even more than half of the documents still have no pictures, as shown in Figure 1. The related method proposed by the present invention can provide a document with multiple expressive pictures. For retrieval results, the present invention proposes a sorting method based on webpage sorting considering time information. A user-friendly interface for news retrieval and browsing is designed.

Fig. 2 shows a schematic diagram of the network news retrieval system of the present invention.

As shown in FIG. 2 , the news retrieval system integrating geographic information and visual information of the present invention includes a data preprocessing module, a location correlation analysis module, a news map module and a retrieval result display module.

The data preprocessing module is used to crawl news data and perform text analysis and information data extraction. The news data includes people, places, time and text information. Described data preprocessing module comprises submodules such as news data crawling module, text analysis module and news entity extraction module, wherein:

The news data crawling module uses web crawlers to crawl news documents and corresponding news images from news websites (for example, news websites such as ABC, BBC, CNN, and Google).

The text analysis module uses natural language processing technology to extract the title, time, website, abstract and text of the news document and the corresponding URL, and extracts the URL of the news image and the text information corresponding to the image.

The news entity extraction module uses natural language processing technology to remove duplicate documents and extract people, places and time from news documents.

The location correlation analysis module is used to perform a correlation analysis between news events and news locations. The location correlation analysis module includes geographic term filtering and expansion modules, matrix decomposition-based correlation analysis modules and other sub-modules, among which:

The geographic noun filtering and expansion module is used to obtain geographic location information (such as geographic longitude and latitude) of geographic nouns.

The correlation analysis module based on matrix decomposition adopts the consistency constraint probability matrix decomposition method of the present invention to analyze the relationship between news locations and news events.

The news image matching module is used to select images that can explain news content for news. The news map module includes sub-modules such as search term generation module, image sorting and selection module, among which:

The search term generation module is used to extract one or more keywords from the news data, combine them into search terms and submit them to the image search engine for image retrieval, that is, extract one or more keywords from the news documents, and combine them into search terms Combine them into search terms of different lengths and submit them to image search engines (such as Google) for image retrieval.

In this embodiment, it is possible to utilize the different importance of each part of the news data (title, abstract and body text, etc.) to extract search terms from the news for network image retrieval, thereby solving the problem that the current image search engine cannot handle long search terms and The problem that a single search term cannot express the content of the document.

Image sorting and selection module to sort and deduplicate the retrieved images and select the appropriate one. This module performs image sorting and selection, that is, considering the position of the image in the returned list and the similarity with the image contained in the source document, using the hierarchical aggregation method to learn the weights of the corresponding list of search words of different lengths, and then using these weights to sort the images. And remove duplicate images, and choose images that express the content of the news.

The retrieval result display module is used to sort and display retrieved news based on retrieval relevance. The invention provides a user interface for displaying retrieval results. As shown in Figure 2, the retrieval result display module includes submodules such as a map view module and a news event list module, among which:

The map view module displays the distribution position of the selected news on the map.

The news event list module is used to sort and display the retrieved news event list according to predetermined rules.

As shown in Figure 2, the user can enter search words in the search box to search, or browse the map and click the corresponding position on the map of the place to be searched, and the system will automatically return relevant results.

The retrieval result display module of the present invention comprehensively considers the timeliness of the news, the correlation with the retrieval and the importance of the news.

As shown in Figure 2, the headlines and top two images of the most relevant news are shown on the map. In the list on the right, each story shows the headline, related image and a short summary. More information is available by clicking the "More" button.

Introduced above the structural composition of the news retrieval system of the present invention, as shown in Figure 2, corresponding to each module of this news retrieval system, the present invention also proposes the network news retrieval method of fusing geographic information and visual information, and this method comprises following Said steps: a data preprocessing step for crawling news data and performing text analysis and information data extraction, said news data including characters, places, time and text information; a place correlation analysis step for performing news events and news The correlation analysis of the location; the step of news map matching, which is used to select a suitable image for the news; the retrieval result display step, which is used to sort and display the retrieved news based on the retrieval correlation.

The location correlation analysis step includes: geographic noun filtering and expansion steps, used to obtain geographic location information of geographic nouns; matrix decomposition-based correlation analysis step, used to analyze news locations and news using the consistency constraint probability matrix decomposition method relationship between events.

Preferably, the consistency constraint probability matrix decomposition method analyzes the relationship between news locations and news events based on the following rules: news events with high similarity are likely to occur in the same place, and multiple news events with high correlation The relationship between location and the same news event is similar.

Described news matching step comprises: search word generation step, is used to extract one or more key words from news data, it is combined into search word and is submitted to image search engine and carries out image retrieval; Image sorting and selection step, uses It is used to sort and deduplicate the retrieved images and select the appropriate one.

Preferably, the step of generating search words extracts search words from multiple parts of the news data for image retrieval; the step of sorting and selecting images sorts the retrieved images using a method based on hierarchical aggregation.

The step of displaying the retrieval results includes: a map view step for displaying the distribution position of the selected news on the map; a news event list step for sorting and displaying the list of retrieved news events according to predetermined rules.

Wherein, the predetermined ordering of rules includes one or more of the following: correlation between news events, correlation between news events and retrieval locations, and time information of news occurrence.

Wherein, in the map view step, in response to the search words input by the user or clicking on any location on the map, the titles and corresponding images of the most relevant news events are displayed.

As mentioned above, the news retrieval system adopts the following four main processing procedures: (1) geographical location correlation analysis based on the probability matrix factorization model of consistency constraints; (2) news map matching; (3) sorting of retrieval results; (4) Search result browsing interface.

The following takes English retrieval as an example to introduce the above sub-processes, mainly including the process of locating news documents and matching pictures, and the process of sorting user search results. But obviously, the present invention is not limited thereto, but can be reasonably applied to other languages, such as Chinese.

<Correlation Analysis of Geographical Location-News Events>

The process includes four steps: (1) extraction of candidate place names; (2) extracting the singularity of candidate place names; (3) obtaining the initial relationship between place names and documents; (4) detailed analysis of the relationship between place names and documents.

First of all, according to the use of natural language processing technology to extract news place names from the news documents in the database, a list of candidate place names is obtained; then, the obtained list is submitted to Wikipedia, and if there is no geographical information in the returned page, the candidate place name is considered is noise and removed.

Then, for the situation that different places correspond to the same name, the filtered list can be submitted to the geographic information system (GeoNames) for expansion and the corresponding geographic information (latitude and longitude) can be crawled.

Then, count the occurrence frequency of place names in the database news documents, so as to obtain the initial relationship between place names and documents.

The correspondence between place names and documents obtained according to the above method contains noise, for example, a news about a celebrity wedding will include the wedding venue and the hometowns of the bride and groom. The wedding location is the true location of the news and is the most relevant, other locations are irrelevant. In addition, the location related to the news may not appear in the article, such as the news about the Beijing Olympics, which introduces related sports events but does not mention Beijing, but the most relevant location for this news is Beijing.

Therefore, in order to better mine the relationship between news locations and news events (news events and news documents are one-to-one correspondence, that is, a news document describes a news event), the present invention is based on traditional probability matrix decomposition ( The Probabilistic Matrix Factorization (PMF) model (see Ruslan Salakhutdinov and Andriy Mnih. "Probabilistic Matrix Factorization", NIPS 2008.) proposed a Consistent Constraints Probabilistic Matrix Factorization model (CCPMF for short), for news location and new Analyze the relevance of documents. Compared with the traditional PMF model, the CCPMF model of the present invention introduces the consistent correlation between the news document and the place as the constraint condition of the optimal solution (that is, the relevant news document needs to correspond to the relevant place, and vice versa), so that it can be more Effectively analyze the true correlation between the two. Among them, the location correlation is to use search engines (such as Google distance) to calculate the statistical co-occurrence correlation between place names; the correlation between news documents is to calculate the linear combination by considering the different importance of news titles, abstracts and texts The resulting textual similarity. The fourth step above is described in detail below: detailed analysis of the relationship between place names and documents.

Fig. 3 is a model diagram of the consistency constraint probability matrix decomposition model proposed by the present invention.

As shown in Figure 3, considering that very related events are likely to occur in the same place and the same news may be related to several very related place names, the consistency constraint probability matrix decomposition model of the present invention comprehensively considers the relationship between place names. , the similarity between documents, and the relationship between place names and documents.

和

分别表示地点-事件关系矩阵、地点之间的相关性矩阵和事件相似性矩阵。采用矩阵分解的思想挖掘潜在的高质量特征空间，即采用P^TE近似R，其中

和

表示潜在的H维地名和事件特征矩阵。

是初始的地点-事件的0-1关系矩阵。在概率矩阵分解模型中，假设对地点-事件的关系估计误差从均值为0、方差为

的高斯分布，则有：Suppose there are M locations and N events.

and

represent the place-event relationship matrix, the correlation matrix between places and the event similarity matrix, respectively. Using the idea of matrix decomposition to mine the potential high-quality feature space, that is, using P ^T E to approximate R, where

and

Represents the underlying H-dimensional place-name and event feature matrix.

is the initial location-event 0-1 relationship matrix. In the probability matrix factorization model, it is assumed that the estimation error of the location-event relationship starts from a mean of 0 and a variance of

Gaussian distribution of , then:

表示均值为0、方差为σ²的高斯分布函数。p_i和e_j分别是矩阵P和E的第i列和第j列。δ是标识矩阵，如果i和j的关系大于零，则δ_ij＝1，否则δ_ij＝0。另外，假设潜在特征空间和系数矩阵服从球形高斯分布，即：in

Represents a Gaussian distribution function with mean 0 and variance ^σ2 . p _i and e _j are the ith and jth columns of matrices P and E, respectively. δ is an identity matrix, if the relationship between i and j is greater than zero, then δ _ij =1, otherwise δ _ij =0. In addition, it is assumed that the latent feature space and coefficient matrix obey a spherical Gaussian distribution, namely:

where I is the identity matrix. After simple Bayesian derivation and logarithmic operation, the objective function is obtained:

in ${\mathrm{\&lambda;}}_P={\mathrm{\&sigma;}}_R^2/{\mathrm{\&sigma;}}_P^2,$ as well as ${\mathrm{\&lambda;}}_{\mathrm{E.}}={\mathrm{\&sigma;}}_R^2/{\mathrm{\&sigma;}}_{\mathrm{E.}}^2.$

Considering the relationship between place names and the relationship between documents, the consistency constraint probabilistic matrix factorization model adds two consistency constraints on the basis of the probability matrix factorization model, and obtains the corresponding objective function:

$<span\; class="notranslate">\; +</span>\frac{{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; E.</span>}}}{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; Tr</span>}<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; E.</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; +</span>\frac{{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; C</span>}}}{\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; C</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\frac{{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; S</span>}}}{\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; S</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; )</span>$

和

分别考虑了地名的相互关系和文档的相互关系，定义为：Among them, λ _C and λ _S are two non-negative weight coefficients, which represent the balance parameters between place name correlation and document similarity.

and

Considering the mutual relationship of place names and the mutual relationship of documents respectively, it is defined as:

${\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; C</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; ik</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; jk</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; Tr</span>}<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; R</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}{\mathrm{<span\; class="notranslate">\; L</span>}}^{\mathrm{<span\; class="notranslate">\; C</span>}}\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; ]</span>$

${\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; S</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; the\; ki</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; kj</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; Tr</span>}<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; RL</span>}}^{\mathrm{<span\; class="notranslate">\; S</span>}}{\mathrm{<span\; class="notranslate">\; R</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; ]</span>$

D^S是对角矩阵，定义为Tr[]是矩阵求迹运算。于是，目标函数为：where L ^C =D ^C -C and L ^S =D ^S -S are Laplace matrices and D ^C is a diagonal matrix defined as

D ^S is a diagonal matrix defined as Tr[] is the matrix trace operation. Then, the objective function is:

$<span\; class="notranslate">\; +</span>\frac{{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; C</span>}}}{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; Tr</span>}<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; R</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}{\mathrm{<span\; class="notranslate">\; L</span>}}^{\mathrm{<span\; class="notranslate">\; C</span>}}\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; +</span>\frac{{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; S</span>}}}{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; Tr</span>}<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; RL</span>}}^{\mathrm{<span\; class="notranslate">\; S</span>}}{\mathrm{<span\; class="notranslate">\; R</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; .</span>$

Using the gradient descent method to solve the objective function, a local optimal solution can be obtained.

<News photo>

In order to enable users to quickly and vividly understand news content, the present invention provides a method for illustrating news documents with pictures. The process consists of two steps: image term generation and ranking of retrieved images.

Fig. 4 has shown the example of a news picture in the present invention.

As shown in Figure 4, the user wants to know what she looks like when seeing the news of American movie star Julia Roberts, and the information needed can be easily obtained by the pictures provided by the present invention; the user wants to know when seeing the news of the curling competition What curling looks like will quickly become apparent through the results provided by the present invention.

First, image terms are extracted from news documents. Since the current network image search engine cannot handle complex retrieval, and a single-word search term cannot express document content, the invention provides an effective image search term generation method.

Because an article is too long, it is more complicated to extract and retrieve from the text, and the manually edited document title is a good summary of the content of the document, so the present invention first extracts the items that form the retrieval from the title, and when the title is too short situation, and then extracted from the text.

个检索，提交给图像检索引擎检索并保存相关图像The above extraction method needs to score the words in the title and the text, and the present invention uses the word frequency-inverse word frequency model to score. For example, for each document, select c retrieval items. Usually, if too many search items are used to retrieve images, the search engine returns few or no results; and if a single item is used to search, the returned results cannot represent the content of the document. Therefore, better results will be obtained by combining items to form retrievals of different lengths. Accordingly, the present invention proposes to combine these retrieval items to form retrieval words of different lengths for image retrieval, and then fuse and sort the returned result lists. A total of c item combinations

Retrieval, submitted to the image retrieval engine to retrieve and save related images

Then, sort the saved images. Each retrieval crawls the first h images to form a list, and there are L lists in total. The present invention adopts the method of level aggregation to fuse and sort the L lists, so as to select suitable images for news documents. Since some documents contain pictures that are manually edited and screened, these pictures can well reflect the content of the document, so the pictures that are more visually similar to the pictures in the document should be ranked first. Additionally, the image's position in the list reflects its relevance to the retrieved text. Therefore, the proposed method of the present invention assigns an initial relation score to a single image considering its position in the list and its similarity to the image contained in the document:

是文档中图像集合。本发明采用1000维的视觉词袋特征以及余弦相似度度量图像之间的相似性。where x _i is the image at the kth position in the jth list,

is the collection of images in the document. The present invention uses 1000-dimensional visual word bag feature and cosine similarity to measure the similarity between images.

In order to adjust the effect of retrieval with different lengths, the present invention assigns the same weight to retrievals with the same length, that is,

个检索的权重。因此，x_i的打分是where η _k is the length k

search weight. Therefore, the score of _xi is

$\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; L</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}{\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>$

The c weights are obtained from the training data set. Use grid search to maximize the normalized nDCG (normalized discounted Cumulative Gain) at the 10th position on the training set.

According to the scores, a sorted list can be obtained. First, duplicate images are removed by using a duplicate detection algorithm, and then r images are selected for each document from the sorted list after deduplication. Among them, the images that come with the document need to be selected first.

<Sort of search results>

According to the search submitted by the user, the system returns a series of relevant results. For the news field, users focus on new, important and retrieval-related news. The invention proposes a news document sorting method which comprehensively considers the timeliness, importance and retrieval relevance of news. The retrieval result sorting process includes the following steps: time information quantification, place name document correlation normalization, and correlation sorting (sorting initialization and sorting).

Firstly, quantify the time information of news documents. Time is an important factor in news. First of all, the time of the news should be quantified, and the time should be expressed in the form of "year, month and day", for example, "September 12, 2010" should be expressed as "20100912". date _k represents the time quantization value of the kth document, and normalizes it:

${\mathrm{<span\; class="notranslate">\; date</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; date</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; min</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; date</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; max</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; date</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; min</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; date</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>}$

${\mathrm{<span\; class="notranslate">\; date</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; date</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}{\mathrm{<span\; class="notranslate">\; date</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}}$

A normalized document-to-place correlation is then performed. Through the consistency constraint probability matrix decomposition, the correlation value between the document and the place name has been obtained and normalized:

${\mathrm{<span\; class="notranslate">\; score</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; score</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}{\mathrm{<span\; class="notranslate">\; score</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}}$

Finally, the correlation sorting of news documents and locations is performed. In order to present the latest, hot and most relevant news to users, this invention proposes a news document sorting method based on the Markov random walk model, which comprehensively considers news time information, importance and relevance to retrieval . This model can be expressed as:

${\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}^{\mathrm{<span\; class="notranslate">\; iter</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; \&times;</span>{\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}^{\mathrm{<span\; class="notranslate">\; iter</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; +</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}^{\mathrm{<span\; class="notranslate">\; 0</span>}}$

是第k个文档在第iter次迭代的值，是第k个文档的初始排序值。y是权重系数，是非负常数。in,

is the value of the kth document at the iter iteration, is the initial ranking value of the kth document. y is the weight coefficient, which is a non-negative constant.

In the random walk model, an initial ranking value needs to be given. The present invention comprehensively considers the time information of the news and the correlation with retrieval, and sets the initial ranking value as

${\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}^{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; date</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; score</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}}{\mathrm{<span\; class="notranslate">\; 2</span>}}$

In the iterative process, considering the importance of news documents, the iterative formula proposed by the present invention is as follows:

${\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}^{\mathrm{<span\; class="notranslate">\; iter</span>}}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}^{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}\underset{\mathrm{<span\; class="notranslate">\; j</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}\frac{{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; kj</span>}}}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; mj</span>}}}{\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}^{\mathrm{<span\; class="notranslate">\; iter</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}$

Among them, S _kj represents the similarity of two documents, and y is set to 0.85. The above iterative process is repeated until a steady state is finally reached, and the sorting result is obtained.

The following describes the user interface in detail.

<Search result browsing interface>

Fig. 5 shows a network news retrieval and browsing interface of an embodiment of the present invention.

In order to provide users with a vivid and fast retrieval and browsing interface, the present invention provides a novel user retrieval and browsing interface.

Referring to FIG. 5 , the interface includes two modules, a map view and a news event list, respectively corresponding to the left and right parts in FIG. 5 . The left half is a map view based on Google Maps. Users can enter search terms in the above search box to search, or double-click the place of interest by browsing the map to search. The system returns the search results, and a window pops up at the corresponding location on the map, displaying the title of the first-ranked news and the first two most relevant images. Users can click the "More" button to get more information. The right half is arranged in descending order according to the correlation between news documents and retrieval, giving a list of all events related to retrieval. Each list entry corresponds to a news event, and 5 news pictures, titles and brief descriptions corresponding to the news event are given. If users want to know more details, they can click the "More" button to get the full text of the news. Users can also click on the headline of the news to visit the original web page of the news.

<preferred embodiment>

The technical effect of the algorithm and the interface provided by the present invention is illustrated below through a specific embodiment. In this embodiment, all the data are crawled from news websites such as ABC, BBC, CNN and Google. In total, there are 48,429 news documents and 20,862 news pictures. After filtering and expansion, a total of 4,742 place names and some parameter settings were obtained. For the consistency constrained probabilistic matrix factorization model, H=100, λ _P =λ _E =0.001, λ _C =2 ⁻³ and λ _S =2 ⁻⁴ . In the news picture, h=20, c=5, and 5 retrieval items are extracted from each document.

In order to evaluate the user's satisfaction with the whole system, this embodiment sets multiple users to score according to the set rules, and then uses the nDCG criterion to measure the evaluation results.

In the evaluation, it is necessary to manually calibrate some data, including the correlation between place names and documents and the correlation between pictures and documents. In this embodiment, three standards are set: very relevant, relevant and irrelevant, and quantified as 2, 1, 0 respectively. A user study was also conducted with 30 users between the ages of 20 and 30, who came from two countries and frequently read English news online.

The site correlation analysis is first evaluated experimentally. 500 documents were randomly selected to evaluate the relationship before and after the consistency constraint probability matrix decomposition and refinement, and two average values were obtained, which were 0.492 and 0.954, respectively. This shows that consistency-constrained probabilistic matrix factorization has a significant effect on denoising and refining place-event relations.

Second, the correlation analysis of place names is tested by retrieval. 100 place names were randomly selected for retrieval, and the consistency-constrained probabilistic matrix factorization, BM25 ranking model and traditional probabilistic matrix factorization model were compared.

Figure 6 shows the retrieval performance evaluation results of the BM25 ranking model, the probability matrix factorization model and the consistency constraint probability matrix factorization model.

In Fig. 6, the abscissa indicates the number of top-ranked documents selected during evaluation, and the ordinate uses the measurement value of the nDCG criterion. As shown in Figure 6, the nDCG measurement value of the analysis method of the Consensus Constrained Probabilistic Matrix Factorization Model (CCPMF) proposed by the present invention is obviously higher than the corresponding value of other methods, which shows that the analysis method of the present invention has significant technical effects.

In order to further observe the influence of the parameters in the model on the system, the present invention evaluates the results by changing λ _C (fixed λ _S ) and changing λ _S (fixed λ _C ).

Figure 7 shows the results of changing parameters under the NDCG50 rule. The results shown in Fig. 7 illustrate that the Consistency Constrained Probabilistic Matrix Factorization Model (CCPMF) proposed by the present invention is significantly better than the other two methods when the parameters vary in a large range.

Fig. 8 is the result of performance comparison between the news image matching method of the present invention and the prior art.

In the comparison test in Figure 8, in order to evaluate news pictures, 300 documents are randomly selected and the correlation between their images and documents is marked to form a training set to learn weight coefficients. Then 1,000 documents were randomly selected for testing, and the news matching method of the present invention was compared with simple search (using the title as a retrieval image) and simple fusion (using each word in the title as a retrieval image and fused) Compare. As shown in Figure 8, the method for matching news pictures of the present invention is obviously superior to the other two methods, which illustrates the effectiveness of the method of the present invention.

In addition, this embodiment also adopts user research to compare the performance before and after news with pictures. Each user is free to browse and compare the two versions. Score 1 for the poorer version, and 2, 3, and 1 if the other version is good, better, or equivalent, respectively. A two-way analysis of variance was also performed. The results of this comparison are shown in Table 1 below.

Table 1: User research results before and after news with pictures

Table 1 shows the mean, standard deviation and variance analysis before and after the news map, where the left part shows the mean score and standard deviation; the right part shows the result of variance analysis. It can be seen from Table 1 that users prefer to add pictures to news, and the performance of the news picture matching method of the present invention has obvious beneficial effects on statistical results.

Fig. 9 shows the comparison results of the retrieval relevancy between the retrieval result ranking method of the present invention and the prior art ranking method.

In the test shown in Figure 9, the method for sorting retrieval results of the present invention, the PRT (using time information as a static sorting web page sorting method) method, the PRR (using place name correlation value as a static sorting web page sorting method) method and BM25 were compared. Sort method. 100 searches were randomly selected for correlation evaluation, and the evaluation results are shown in Figure 9. As shown in Fig. 9, BM25 achieves the worst performance and PRR achieves the best performance. The search result sorting method of the present invention is only slightly inferior to PRR, but is obviously superior to the PRR method in terms of timeliness (see FIG. 10 ), that is, the present invention can preferentially display the latest news to users.

Fig. 10 shows the comparison results of the retrieval result sorting method of the present invention and the sorting method of the prior art in terms of timeliness.

In Figure 10, the vertical axis represents the percentage of news that occurred in the latest week. In this test, the average percentage of news that occurred in the last week among the first d (d=5, 10, 20, 50, 100) news in the 100 retrieved return lists was counted. It can be seen from FIG. 10 that the retrieval result sorting method of the present invention is only slightly inferior to the PRT method, but as shown in FIG. 9 , it is obviously superior to the PRT method in terms of retrieval relevance.

Combining the test results shown in Figure 9 and Figure 10, it can be seen that the comprehensive effect of the retrieval result sorting method of the present invention on relevance and timeliness is obviously better than the sorting method in the prior art, so satisfactory performance can be achieved .

It should be understood that the above specific embodiments of the present invention are only used to illustrate or explain the principle of the present invention, and not to limit the present invention. Therefore, any modification, equivalent replacement, improvement, etc. made without departing from the spirit and scope of the present invention shall fall within the protection scope of the present invention. Furthermore, it is intended that the appended claims of the present invention embrace all changes and modifications that come within the scope and metesques of the appended claims, or equivalents of such scope and metes and bounds.

Claims (12)

1. A network news retrieval system fusing geographic information and visual information, the system comprising:

the data preprocessing module is used for crawling news data and performing text analysis and information data extraction, wherein the news data comprises people, places, time and text information;

a location relevance analysis module for performing a relevance analysis of the news event to the news location;

the news matching module is used for selecting images capable of explaining news contents for news;

the retrieval result display module is used for displaying the retrieved news in a sequencing mode based on the retrieval relevance;

the location correlation analysis module includes:

the geographic noun filtering and expanding module is used for acquiring geographic position information of geographic nouns;

a matrix decomposition based correlation analysis module for analyzing the relationship between news locations and news events using a consistency constraint probability matrix decomposition method that analyzes the relationship between news locations and news events based on the following rules: the news events with high similarity are likely to occur in the same place, the relationship between a plurality of places with high relevance and the same news event is similar, and the objective function of the consistency constraint probability matrix decomposition method is as follows:

wherein M is the number of sites; n is the number of events; delta is an identification matrix, delta if the relationship of i and j is greater than zero_ij= l, otherwise δ_ij=0；

Is a 0-1 relationship matrix of initial location-events;

representing a desired place-event relationship matrix;

estimating a variance of a gaussian distribution to which the error obeys for the location-event relationship;

obeying the variance of Gaussian distribution for the potential H-dimensional place name feature matrix;obeying the variance of the Gaussian distribution to the potential H-dimensional event feature matrix;

representing a potential H-dimensional place name feature matrix;

representing a potential H-dimensional event feature matrix; lambda [ alpha ]_CAnd λ_SAre two non-negative weight coefficients; l is^C=D^C-C and L^S=D^SS is a Laplacian matrix, D^CIs a diagonal matrix, defined as

D^SIs a diagonal matrix, defined as

Representing a correlation matrix between the sites;

representing an event similarity matrix; tr [ 2 ]]Performing matrix tracing operation; solving to obtain P and E based on the model, and then adopting P^TE approximates R;

the news matching module comprises:

the search word generation module is used for extracting one or more key words from the news data, combining the key words into a search word and submitting the search word to an image search engine for image search;

and the image sorting and selecting module is used for sorting and removing the duplication of the retrieved images and selecting the images capable of explaining the news content.

2. The system of claim 1, the data pre-processing module comprising:

the news data crawling module is used for crawling news documents and corresponding news images from a news website;

the text analysis module is used for extracting the title, time, website, abstract and text of the news data and corresponding websites, and extracting the websites of the news images and text information corresponding to the images;

and the news entity extraction module is used for extracting people, places and time from the news data.

3. The system of claim 1, the search result presentation module comprising:

the map view module is used for displaying the distribution position of the selected news on a map;

and the news event list module is used for sequencing and displaying the list of the retrieved news events according to a preset rule.

4. The system of claim 1, wherein

The search term generating module extracts search terms from a plurality of parts of the news data to carry out image search;

the image sorting and selecting module sorts the retrieved images by a method based on rank aggregation.

5. The system of claim 3, wherein the predetermined rule ordering includes one or more of: the correlation between news events, the correlation between news events and retrieval locations, and the time information of news occurrences.

6. The system of claim 3, wherein the map view module displays the titles of the most relevant news events and the corresponding images in response to a search word input by a user or clicking on any one of the locations on the map.

7. A network news retrieval method fusing geographic information and visual information comprises the following steps:

the method comprises the steps of data preprocessing, wherein the data preprocessing is used for crawling news data and carrying out text analysis and information data extraction, and the news data comprises people, places, time and text information;

a location correlation analysis step of performing correlation analysis of the news event with a news location;

a news matching step, which is used for selecting images capable of explaining news contents for news;

a retrieval result display step for displaying the retrieved news in a sorted manner based on the retrieval relevance;

the location correlation analyzing step includes:

a geographic noun filtering and expanding step, which is used for acquiring the geographic position information of the geographic noun;

a correlation analysis step based on matrix decomposition for analyzing a relationship between the news site and the news event using a consistency constraint probability matrix decomposition method that analyzes a relationship between the news site and the news event based on the following rules: the news events with high similarity are likely to occur in the same place, the relationship between a plurality of places with high relevance and the same news event is similar, and the objective function of the consistency constraint probability matrix decomposition method is as follows:

wherein M is the number of sites(ii) a N is the number of events; delta is an identification matrix, delta if the relationship of i and j is greater than zero_ij= l, otherwise δ_ij=0;

Is a 0-1 relationship matrix of initial location-events;

representing a desired place-event relationship matrix;

estimating a variance of a gaussian distribution to which the error obeys for the location-event relationship;

obeying the variance of Gaussian distribution for the potential H-dimensional place name feature matrix;

obeying the variance of the Gaussian distribution to the potential H-dimensional event feature matrix;

representing a potential H-dimensional place name feature matrix;

representing a potential H-dimensional event feature matrix; lambda [ alpha ]_CAnd λ_SQi is two non-negative weight coefficients; LC = D^C-C and L^S=D^SS is a Laplacian matrix, D^CIs a diagonal matrix, defined as

D^SIs a diagonal matrix, defined as

Representing a correlation matrix between the sites;representing an event similarity matrix; tr [ 2 ]]Performing matrix tracing operation; solving to obtain P and E based on the model, and then adopting P^TE approximates R;

the news matching step comprises the following steps:

a search word generation step, which is used for extracting one or more key words from news data, combining the key words into a search word and submitting the search word to an image search engine for image search;

and an image sorting and selecting step for sorting and de-duplicating the retrieved images and selecting images capable of explaining news contents.

8. The method of claim 7, the data preprocessing step comprising:

a news data crawling step, which is used for crawling news documents and corresponding news images from a news website;

a text analysis step, which is used for extracting the title, time, website, abstract and text of the news data and corresponding websites, and extracting the websites of the news images and text information corresponding to the images;

and a news entity extraction step, namely extracting people, places and time from the news data.

9. The method of claim 7, wherein the step of displaying the search result comprises:

a map view step for displaying the distribution position of the selected news on a map;

and a news event listing step for sorting and displaying the retrieved news event list according to a predetermined rule.

10. The method of claim 7, wherein

The search word generating step extracts search words from a plurality of parts of the news data for image search;

the image sorting and selecting step sorts the retrieved images using a rank aggregation based approach.

11. The method of claim 9, wherein the predetermined rule ordering includes one or more of: the correlation between news events, the correlation between news events and retrieval locations, and the time information of news occurrences.

12. The method of claim 9, wherein the map view step displays a title of a most relevant news event and a corresponding image in response to a search word input by a user or clicking any one place on a map.

Images