JP4759600B2 - Text search device, text search method, text search program and recording medium thereof - Google Patents

Description

  The present invention relates to a technique for searching a sentence on a site using an input query and an extended word related to the query.

  In order to improve the accuracy of the ranking of the text search system, one promising technique is called “query expansion”. This is a method of automatically obtaining, selecting, and assigning “something related” words (hereinafter referred to as extended words) to a query (search term) to derive better search results. is there.

As a document related to this, Non-Patent Document 1 uses a click log as acquisition source data of an extended word and a co-occurrence probability with a query as a selection criterion. The click log is a set of URLs of sites (clicked documents) that are actually selected by the web viewer in the search result for the query. Naturally, it is natural that an extended word suitable for a query changes from time to time. In Non-Patent Document 1, this is dealt with by using a click log.
“Probabilistic Query Expansion Usage Query Logs”. Hang Cui, et al. (2002) S. E. Robertson, “On term selection for query expansion”, Journal of Document, 46, pages 359-364, 1990.

  In Non-Patent Document 1, the URL of a site (clicked document) selected by a viewer is determined using a simple click count (absolute click count) obtained by analyzing a click log, and its title and summary sentence (snippet) ) To calculate the co-occurrence probabilities of keywords and queries included in Query expansion is performed by selecting the expanded words in order from the keyword with the highest measure.

  However, this absolute click count includes a tendency to click more frequently in the higher ranking sites in the search results, and many sites with this higher value are selected by the viewer and fit the query. It cannot be said that it was judged. Further, in Non-Patent Document 1, the number of extended words is 40 to 60, and the ranking accuracy is maximized, which requires a calculation cost.

  Therefore, in view of such problems, the present invention realizes query expansion that accurately identifies a site where access is concentrated by analyzing a click log, and can greatly improve search accuracy with a small number of expanded words. Is a solution issue.

  The present invention is a technical idea created in order to solve the above-mentioned problems, and obtains an extension word from a title and a summary sentence (snippet) of a site where access is concentrated that many viewers find useful. As a result, it is possible to obtain extended words with high suitability for queries.

Specifically, the invention described in claim 1 is a sentence search device that acquires an extended word related to an input query and searches a site using the extended word and the query . Click log analysis means for analyzing the click log for the query in which the actually selected site is ranked based on the number of clicks to identify the site where access is concentrated, the title and summary sentence of the site where access is concentrated, And an extended word acquisition unit that acquires the extended word, and the click log analysis unit calculates a difference in the number of clicks between sites existing in adjacent ranks in the click log as a relative number of clicks. and a first analyzing means for determining the access degree of concentration of the site in accordance with the relative number of clicks, click probability for each rank sites for each of the search There Referring to saved database, to calculate the actual number of clicks to the average click probability rank as occurrence probability, a second analyzing means for determining a candidate access concentration site by using a biological cause probability and threshold, the And an analysis result integration unit that integrates the analysis results of both analysis units and identifies a site where access is concentrated.

According to the second aspect of the present invention, the extended word acquisition means analyzes the title and summary sentence of the site where access is concentrated to obtain a candidate group of extended words; It is characterized by comprising means for ordering and means for selecting an extension word based on the ranking of each of the ordered candidates.

The invention according to claim 3 searches a site using the query and the extended word, and outputs search results, search execution means for outputting search results, and user click information for the search results of the search execution means. It further comprises click log feedback processing means for reflecting in the click log.

The invention according to claim 4 is a sentence search method for acquiring an extended word related to an input query and searching a site using the extended word and the query, wherein the click log analyzing means includes: A first step of analyzing a click log for the query in which a site actually selected by a viewer is ranked based on the number of clicks to identify a site where access is concentrated, and an expanded word acquisition unit concentrates the access. A second step of analyzing the title of the site and the summary sentence to obtain the extended word, wherein the first step is based on a relative difference in the number of clicks between sites existing in adjacent ranks in the click log. calculated as click-number, and obtaining access concentration degree of site in accordance with the relative number of clicks, for each rank of the site for each of the search With reference to database Rick probability is stored, calculates the actual number of clicks to the average click probability rank as occurrence probability, and determining a candidate access concentration site by using a biological cause probability and threshold, said both And integrating the analysis results of the steps to identify sites where access is concentrated.

The invention according to claim 5 is characterized in that the second step analyzes the title and summary sentence of the site where the access is concentrated to obtain an extended word candidate group, and for each of the extended word candidates, And ordering, and selecting expanded words based on the ranking of the ordered candidates.

The invention according to claim 6 is the step in which the search execution means searches the site using the query and the extended word and outputs the search result, and the click log feedback processing means responds to the search result in the step. A step of reflecting user click information in the click log.

The invention described in claim 7 is a text search program, which causes a computer to execute each step of the text search method according to any one of claims 4 to 6 .

The invention described in claim 8 is a computer-readable recording medium, wherein the sentence search program described in claim 7 is recorded.

According to the first to eighth aspects of the present invention, since the extended word is acquired from the title and summary sentence (snippet) of the site where access is concentrated, it becomes possible to acquire the extended word having high adaptability to the query. . Thereby, the number of extended words can be reduced, and a highly accurate search result can be obtained while suppressing the calculation cost.

In addition, since the analysis results of two independent measures are combined, it is possible to accurately identify the site where access is concentrated.

Further, according to the invention described in claim 3.6, it is possible to reflect the user's judgment (click) on the search result to the click log as needed.

  FIG. 1 shows a text search apparatus 1 according to an embodiment of the present invention. Here, an example in which the text search device 1 is configured by a computer will be described. However, the text search device is not limited to this, and includes, for example, an IC (Integrated Circuit) chip on which text search processing logic is mounted. It may be a computer or a mobile terminal such as a mobile phone.

  As shown in FIG. 1, the text search apparatus 1 includes a display unit 100 that displays four main functional blocks, that is, a query input screen 101 and a result display screen 102, an input query, and an extension related to the query. A search expression generation unit 110 that generates a search expression using words, an extended word selection unit 120 that acquires an expansion word for an input query, and a search execution unit 140 that executes the search expression are provided.

  The functions of the respective functional blocks 100.110.120.140 are realized by reading a text search program by a control unit (CPU: Central Processor Unit) of the text search device 1. The text search device 1 is used for normal components of a computer, for example, an input unit such as a keyboard and a mouse (not shown), a rewritable memory (RAM) for temporarily storing processing data, and a network connection. A communication device, a storage unit such as a hard disk drive, and a display unit such as a display are provided. Hereinafter, each functional block 100.110.120.140 will be described in detail with reference to FIG.

<Display unit 100>
On the display unit 100, a query input screen 101 for a user to input a query (search term) and a result display screen 102 for displaying a search result obtained from the search execution unit 140 are displayed via a browser. Here, the user inputs a query on the query input screen 101 using the keyboard or the like.

<Search Expression Generation Unit 110>
The search expression generation unit 110 receives the query input on the query input screen 101, and transmits the received query and an extended word request for the query to the extended word selection unit 120. In addition, when an extended word is received from the extended word selection unit 120, a search expression for performing a process of rearranging search results for an initial query is generated using the extended word, and is transmitted to the search execution unit 140. Using this search formula, the ranking of the initial search result for the input query is rearranged into a more accurate ranking using the extended word.

<Extended word selection unit 120>
The extended word selection unit 120 includes an instruction unit 121, an analysis result integration unit 122, an analysis unit A123, an analysis unit B124, an analysis unit C125, an information extraction unit 128, a collation unit 129, a search result acquisition unit 130, a function word extraction / noun. The phrase generation unit 133, the extended word weight / ordering unit 134, the click log feedback processing unit 150, three DBs (databases), that is, the click log DB 126, the click probability DB 127, the index DB 132, and the search engine 131 are included. Yes. Of these, each of the DBs 126.127.132 is constructed on the hard disk drive device.

  When the extended word selection unit 120 receives the query and the extended word request from the search expression generation unit 110, the extended word selection unit 120 transmits the query to the command unit 121.

  When receiving the query from the extended word selection unit 120, the command unit 121 transmits the received query and a click log analysis request to the analysis result integration unit 122.

  When the analysis result integration unit 122 receives the query and click log analysis request from the command unit 121, the analysis result integration unit 122 transmits them to the analysis unit A123 and the analysis unit B124.

  When the analysis unit A123 and the analysis unit B124 receive the query and click log analysis request from the analysis result integration unit 122, the analysis unit A123 and the analysis unit B124 read the click log corresponding to the query from the click log DB 126. In the click log DB 126, URLs of sites (clicked documents) actually selected by the viewer in the search result for the query are ranked based on the number of clicks, and stored as a click log. The click log DB 126 is constructed using click log data acquired in advance from a server (not shown) via the Internet.

  The analysis unit A123 analyzes the read click log by paying attention to the degree of access concentration for each site, and returns the analysis result to the analysis result integration unit 122.

  The analysis unit B124 further reads the click probability of the site for each rank stored in the click probability DB 127, analyzes the click log and the click probability read from both the DBs 126.127, and analyzes the analysis result as the analysis. It returns to the result integration unit 122. The click probability DB 127 is also constructed using the click log data on the server in the same manner as the click log DB 126.

  When the analysis result integration unit 122 receives the analysis results from the analysis unit A123 and the analysis unit B124, the analysis result integration unit 122 integrates them to specify an access concentration site (hereinafter abbreviated as ACS), and the URL Is sent back to the command unit 121. Details of the access concentration site (ACS) will be described later.

  Upon receiving the URL of the access concentration site (ACS) from the analysis result integration unit 122, the command unit 121 transmits this URL to the information extraction unit 128.

  When the information extraction unit 128 receives the query and the URL of the access concentration site (ACS) from the command unit 121, the information extraction unit 128 transmits them to the verification unit 129.

  When the collation unit 129 receives the query and the URL of the access concentrated site (ACS) from the information extraction unit 128, the collation unit 129 transmits the query to the search result acquisition unit 130.

  When the search result acquisition unit 130 receives the query from the collation unit 129, the search result acquisition unit 130 inputs the query to the search engine 131.

  When the search engine 131 receives a search result request from the search result acquisition unit 130 or the search execution unit 140, the search engine 131 returns a search result obtained by searching the index DB 132.

  In the index DB 132, “World Wide Web” or “Mobile Web” sites are indexed and stored. The index DB 132 is constructed using data acquired in advance from a server (not shown) via the Internet. Here, it is assumed that the search engine 131 includes the index DB 132.

  When the search result acquisition unit 130 receives a search result for the query from the search engine 131, the search result acquisition unit 130 returns the search result to the collation unit 129.

  Upon receiving the search result for the query from the search result acquisition unit 130, the collation unit 129 collates the URL of the search result with the URL of the access concentration site (ACS) received from the information extraction unit 128, and The access concentrated site (ACS) is identified from the URL of the search result. Then, information (URL, title, etc.) of the identified access concentration site (ACS) is returned to the information extraction unit 128.

  When the information extraction unit 128 receives the information (URL, title, etc.) of the access concentration site (ACS) from the collation unit 129, the information extraction unit 128 extracts a title and a snippet from the information and uses them as the function word. This is transmitted to the extraction / noun phrase generator 133.

  When the function word extraction / noun phrase generation unit 133 receives the title and snippet of the access concentration site (ACS) from the information extraction unit 128, the function word extraction / noun phrase generation unit 133 performs morphological analysis on the title and snippet to generate a function word extraction or noun phrase. Then, these are used as a candidate group of extension words to be used in the subsequent query extension, and transmitted to the extension word weight / ordering unit 134.

  When the extension word weight / ordering unit 134 receives the extension word candidate group from the function word extraction / noun phrase generation unit 133, each extension word is based on “Robertson Selection Value (RSV)” of Non-Patent Document 2. The candidate groups are weighted and ordered, and the result is transmitted to the extended word selection unit 120.

  When the extended word selection unit 120 receives the extended word candidate group from the extended word weight / ordering unit 134, the extended word selection unit 120 selects an extended word to be actually used from the candidate group, and supplies these to the search expression generation unit 110. Send back.

  The analysis unit C125 reads all click logs from the click log DB 126, calculates click probabilities for the sites of each rank, and updates the click probability DB 127 using the calculation results. This update process is executed at predetermined time intervals.

  The click log feedback processing unit 150 records information (click information) of a site clicked by a user on the search result displayed on the result display screen 102, and generates a new click log based on the information. Feedback to the click log DB 126 as needed.

<Search Execution Unit 140>
When the search execution unit 140 receives a search expression from the search expression generation unit 110, the search execution unit 140 inputs the search expression into the search engine 131 and receives a corresponding search result. Then, the search result is displayed on the result display screen 102.

<Operation example>
The text search apparatus 1 analyzes the click log based on the query input from the user, acquires an extended word, and obtains a more accurate search result by expanding the query using the extended word. Yes. This series of processing mainly includes four phases, that is, a query input phase, a click log analysis phase, an extended word acquisition phase, and a search execution phase. Hereinafter, the detailed processing content of each phase will be described with reference to FIGS.

(1) Query Input Phase FIG. 2 shows a processing flow of the query input phase. First, the user inputs a query on the query input screen 101. The input query is transmitted to the search expression generation unit 110. The search expression generation unit 110 transmits an extended word extraction / selection request together with the query to the extended word selection unit 120.

(2) Click log analysis phase In the click log analysis phase, click logs are analyzed using the input query. The purpose is to identify sites where access is concentrated among the sites in the click log. This is because when a viewer selects (clicks) a site in the search results, it is considered that the viewer looks at the title and snippet (summary) of the site. This is because it can be expected that there are keywords that the person has judged useful. And by expanding the query with this keyword, it can be expected that the accuracy of the search is greatly improved. Here, such a site is called an access concentrated site (ACS), and its title and snippet (Titles and Snippets) are called TS. In the click log analysis phase, an access concentrated site (ACS) is specified by analyzing the click log using the input query.

  FIG. 3 shows a processing flow of the click log analysis phase. When the command unit 121 receives the query from the extended word selection unit 120 (following the symbol A in FIG. 2), the command unit 121 transmits the received query and a click log analysis request thereto to the analysis result integration unit 122. . Upon receiving this request, the analysis result integration unit 122 transmits the query to the analysis unit A123 and the analysis unit B124, and an analysis phase is started.

  The analysis unit A123 calculates an access concentration degree (ACD) as an index for specifying an access concentration site (ACS). Specifically, the analysis unit A123 reads a click log for the query from the click log DB 126. Then, paying attention to the number of clicks of a site existing in a rank of the read click log and the sites existing in both adjacent ranks, the relative number of clicks is calculated by Expression (1) and Expression (2). These expressions (1) and (2) are defined in the program of the text search apparatus 1.

  Here, c (q, r) indicates the number of clicks of the site ranked r in the click log for the query q, and c (q, r-1) and c (q, r + 1) are the left of rank r. It shows the number of clicks on the site that has the next rank r−1 and the right next rank r + 1.

Further, slope _L and slope _R correspond to respective slopes for rank r-1 and rank r + 1 when a click log is analyzed for a certain query q and a curve is drawn based on the rank and the number of clicks. An example of this curve is shown in FIG.

  In FIG. 4, the horizontal axis indicates the rank of the site, and the vertical axis indicates the number of clicks. Here, the specific query curve is a curve drawn by analyzing the click log for a query, and the average query curve is a curve drawn by analyzing the click log for all queries and averaging the number of queries. To express.

  At this time, it is assumed that the site where the slope of the specific query curve is steep, that is, the rank where the specific query curve has a strong peak is a site where access is concentrated, and this is regarded as a candidate for an access concentrated site (ACS). . The degree of the slope of this curve is defined as the access concentration degree ACD (q, r) in Expression (3). This expression (3) is defined in the program of the text search apparatus 1.

Here, θ _L (r) and θ _R (r) indicate angles with respect to the slopes slope _L and slope _R, and the access concentration degree ACD (q, r) is characterized by these angles. The analysis unit A123 calculates the access concentration degree ACD (q, r) for each site, and returns it to the analysis result integration unit 122 as an analysis result.

  The analysis unit B124 statistically determines whether or not the number of clicks of the site in the click log is obviously large as another index for specifying the access concentrated site (ACS). That is, if the number of clicks of a site existing in a certain rank r greatly exceeds the number of clicks (average number of clicks) expected by that rank r, it can be said that the site was intentionally selected by the viewer, not by chance. .

  Specifically, when the analysis unit B124 receives the query q and click log analysis request from the analysis result integration unit 122, the analysis unit B124 reads the click log for the query q from the click log DB 126. At the same time, the click probability P (r) of the site existing in the rank r is read with reference to the click probability DB 127. It is assumed that the distribution of clicks of a site existing in a certain rank r follows a binomial distribution, and the actual number of clicks c (q, r) greatly exceeds the number of clicks (average clicks) that the rank expects. Is calculated as an occurrence probability p (q, r) by the equation (4). This equation (4) is defined in the program of the text search apparatus 1.

  Here, n (q, r) is the total number of times all viewers have viewed the site of rank r, and this unit number corresponds to “trial”. This total number n (q, r) is represented by the sum of the number of clicks c (q, r) of the site in this rank r and the number of times nc (q, r) that passed. The number of times of passing indicates the number of times a site having a rank lower than rank r has been clicked. If the same viewer clicks on a site of rank r and rank r or lower successively, it is identified by the viewer ID and counted once.

  Here, whether or not a site existing in the rank r is clicked is regarded as a “trial per time” expressed by the binomial distribution, and the occurrence probability p (q, r) at which the number of clicks on the site occurs is extremely high. If the number of clicks is small, the number of clicks is considered to be clearly high.

  For example, in the graph shown in FIG. 5, the horizontal axis is the number of clicks, the vertical axis is the occurrence probability, the click probability P (r) = 32% of the rank r site, and the number of trials n (q, r) = 100. The binomial distribution of the number of clicks is shown. In other words, when the click probability of the site having the rank r becomes 32% after 100 trials, it is assumed that the expected distribution of the number of clicks follows a binomial distribution as shown in the graph of FIG.

  At this time, if the number of clicks c (q, r) acquired by the site actually ranked r is included in the 2.5% region (hereinafter referred to as region S) on the right side of the graph, the site is clearly Considering that many clicks have been made, this site is regarded as a candidate for an access concentration site (ACS). Then, the analysis unit B124 returns the access concentration site (ACS) candidate group obtained in this way to the analysis result integration unit 122 as an analysis result.

  In FIG. 5, 2.5% on the right side, which is the threshold value of the region S, is a threshold value at which a significant difference generally appears before and after that, but this value can be appropriately changed according to the setting.

  When the analysis result integration unit 122 receives the analysis results from the analysis unit A123 and the analysis unit B124, the analysis result integration unit 122 specifies an access concentration site (ACS) based on these analysis results. That is, the top K cases of the sites whose click probability is within 2.5% on the right side of the binomial distribution and whose access concentration degree ACD (q, r) is high are regarded as access concentration sites (ACS). Get the URL. Then, this URL is returned to the command unit 121.

  At this time, if the occurrence probability of the number of clicks does not enter the region S in any site, the sites are ordered in descending order of the access concentration degree ACD (q, r), and the top K cases are regarded as the access concentration sites (ACS). . The value of the upper number K can be changed as appropriate according to the setting.

  The analysis unit C125 reads all click logs from the click log DB 126 and calculates a click probability P (r) for each rank site. Then, the click probability DB 127 is updated using this calculation result. This update process is performed at predetermined time intervals.

(3) Extended word acquisition phase In the extended word acquisition phase, an extended word for query expansion is acquired from the title and summary sentence (snippet) of the access concentrated site (ACS) specified in the click log analysis phase. This is because the titles and snippets of sites where access is concentrated have keywords that the viewer has determined to be useful, and by expanding the query with these keywords, a significant improvement in search accuracy can be expected.

  FIG. 6 shows the process flow of the extended word acquisition phase. After the information extraction unit 128 receives the URL of the access concentration site (ACS) and the query from the command unit 121 (continuing from the symbol B in FIG. 3), an extended word acquisition phase is started.

  The information extraction unit 128 transmits the received URL of the access concentration site (ACS) and the query to the collation unit 129. The collation unit 129 transmits the query to the search result acquisition unit 130, and then the search result acquisition unit 130 inputs the query to the search engine 131.

  The search engine 131 executes a search for the index DB 132 using this query, acquires a URL and the number N of search results, and returns these to the search result acquisition unit 130. Upon receiving these, the search result acquisition unit 130 returns the search results (URL and the number N of search results) to the verification unit 129.

  The collation unit 129 collates the URL of the search result received from the search result acquisition unit 130 with the URL of the access concentration site (ACS) received from the information extraction unit 128 to obtain an access concentration site (ACS). Identify. Then, the information (URL, title, snippet, etc.) of the identified access concentration site (ACS) and the search result number N are returned to the information extraction unit 128.

  The information extraction unit 128 extracts a title and a snippet from the received information on the access concentration site (ACS), and transmits the extracted information and the search result number N to the function word extraction / noun phrase generation unit 133.

  The function word extraction / noun phrase generation unit 133 decomposes the received title and snippet into morphemes, acquires a function word, and generates a noun phrase. Then, the function word and the noun phrase are transmitted to the weight / ordering unit 134 of the expanded word together with the search result number N as a candidate group of expanded words used for query expansion performed later.

  The extended word weighting / ordering unit 134 receives each extended word candidate i, the number of access concentrated sites (ACS) including each extended word candidate i in the title and snippet an (i), and the number of search results. N, using the total number of access concentrated sites (ACS) K, the weights and ordering of extended words using the formulas (5) and (6) of “Robertson Selection Value (RSV)” of Non-Patent Document 2 . These expressions (5) and (6) are defined in the program of the text search apparatus 1.

  Here, n (i) is the number of sites that include the expansion word candidate i in the title and snippet in the search result number N. The extended word weight / ordering unit 134 transmits the extended word candidate group thus ordered to the extended word selection unit 120.

  The extended word selection unit 120 employs the top T of the received expanded word candidate groups as expanded words that are actually used in query expansion, and returns the employed expanded word group to the search expression generation unit 110. The value of the upper order number T can be changed as appropriate according to the setting.

(4) Search execution phase In the search execution phase, a search expression is generated using the extended word acquired in the extended word acquisition phase, and a search based on the search expression is executed, and the user's judgment (click) on the search result is performed. Reflect in the click log.

  FIG. 7 shows the processing flow of the search execution phase. When receiving the extended word group from the extended word selecting unit 120 (following the symbol C in FIG. 6), the search expression generating unit 110 generates a search expression for executing query expansion using the extended word group, Is transmitted to the search execution unit 140.

  The search execution unit 140 inputs the received search expression to the search engine 131. The search engine 131 uses the search formula to search each of the site title and body (text) stored in the index DB 132.

  In this search formula, first, a set of search results (ranked site group) for the initial query input by the user is determined, and then the site group is scored using the extended word, and this score is obtained. A method of rearranging the rankings with higher accuracy based on the above is defined.

  Specifically, for the sentence that is an element of each site of the search result for the initial query, “extension word weight rsv (i) × extension word tf · idf in each sentence” is calculated and scored, Sort sites by this score. Here, “tf · idf” represents the weight of a word obtained by a measure such as how much a certain word appears in one document. Then, the rearranged result is displayed on the result display screen 102. Therefore, the user can confirm the ranking arranged with higher accuracy as a search result with respect to the first input query.

  For this search result, the user makes a decision (click) to select a site. This click information (information on the site clicked by the user) is transmitted to the click log feedback processing unit 150. The click log feedback processing unit 150 generates a new click log from the click information, and updates the click log DB 126 as needed using the click log.

  As described above, according to the text search device 1 according to the present embodiment, when specifying an access concentration site (ACS), it is specified by combining two independent measures, the access concentration degree ACD and the occurrence probability of the number of clicks. Therefore, it is possible to accurately determine the access concentration site (ACS).

  In addition, since query expansion is performed using keywords extracted from the title and snippet of the access intensive site (ACS), it is possible to achieve a significant improvement in search accuracy with a small number of expanded words (1 to 5 words).

  In addition, the click log DB used to acquire extended words is updated as needed by automatic processing of feedback from users. Can be extracted.

<Other examples>
The click log DB 126 and the index DB 132 are not necessarily implemented in the text search apparatus 1, and are, for example, implemented in a server (not shown) connected to the text search apparatus 1 via a network. May be. In this case, the connection to both DBs 126.132 is made via the communication device.

  That is, in the click log analysis phase, the analysis unit A123 and the analysis unit B124 connect to the click log DB 126 via the communication device and acquire a click log corresponding to the query. The acquired click log data is temporarily stored in the memory (RAM), and both the analysis units 123.124 analyze this data by the above-described method.

  Similarly, the analysis unit C125 connects to the click log DB 126 via the communication device and acquires all click logs. Then, the click probability is calculated from the acquired click log, and the click probability DB 127 is updated.

  In the extended word acquisition phase and the search execution phase, the search engine 131 connects to the index DB 132 via the communication device and performs a search. Further, the click log feedback processing unit 150 similarly connects to the click log DB 126 via the communication device, and updates the click log DB 126 as needed.

  The present invention can also be provided as a text search program that causes a computer to function as each functional block 100.110.120.140 of the text search device 1. This program may cause the computer to execute all the processes of the functional blocks 100.110.120.140, or may execute a part of the processes.

  This program is provided to the computer by downloading from a website or the like. The program is stored in a recording medium such as a CD-ROM, DVD-ROM, CD-R, CD-RW, DVD-R, DVD-RW, MO, HDD, Blu-ray Disk (registered trademark). It may be provided to a computer. Since the program code read from the recording medium causes the computer to function as each functional block of the present embodiment, the recording medium also constitutes the present invention.

The lineblock diagram of the text search device concerning the embodiment of the present invention. The processing flow figure of the same query input phase. The processing flow diagram of the click log analysis phase. The graph which shows the same access concentration degree ACD. A graph showing the binomial distribution of the number of clicks. The processing flow figure of the same extended word acquisition phase. The processing flow figure of the same search execution phase.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... Text search device 100 ... Display part 101 ... Query input screen 102 ... Result display screen 110 ... Search formula production | generation part 120 ... Extended word selection part 121 ... Command part 122 ... Analysis result integration part 123 ... Analysis part A (1st analysis) means)
124... Analysis unit B (second analysis means)
125 ... analysis unit C
126 ... Click log DB
127 ... Click probability DB
128 ... Information extraction unit 129 ... Verification unit 130 ... Search result acquisition unit 131 ... Search engine 132 ... Index DB
133 ... Function word extraction / noun phrase generation unit 134 ... Extended word weight / ordering unit 140 ... Search execution unit 150 ... Click log feedback processing unit

Claims (8)

A sentence search device that acquires an extended word related to an input query and searches a site using the extended word and the query,
Click log analysis means for analyzing a click log for the query in which a site actually selected by a visitor is ranked based on the number of clicks, and identifying a site where access is concentrated;
Analyzing the title and summary sentence of the site where the access is concentrated to obtain the extended word, and an extended word acquisition means,
The click log analysis means calculates a difference in the number of clicks between sites existing in adjacent ranks in the click log as a relative number of clicks, and calculates a degree of access concentration of the site according to the relative number of clicks Means,
By referring to a database in which the click probability for each rank of the site is stored for each search, the actual number of clicks with respect to the average click probability of the rank is calculated as the occurrence probability, and the access concentration site using the occurrence probability and the threshold value A second analysis means for obtaining a candidate for
An analysis result integration unit that integrates the analysis results of the two analysis units and identifies a site where access is concentrated;
A sentence search device characterized by comprising: The extended word acquisition means includes
Means for analyzing the title and summary sentence of the site where the access is concentrated to obtain a candidate group of extended words;
Means for ordering each candidate for the extended word;
The sentence search apparatus according to claim 1, further comprising: means for selecting an extended word based on the ranking of the ordered candidates. Search execution means for searching a site using the query and the extended word and outputting a search result;
The text search apparatus according to claim 1, further comprising: a click log feedback processing unit that reflects user click information on the search result of the search execution unit in the click log. . A sentence search method for acquiring an extended word related to an input query and searching a site using the extended word and the query,
A first step in which a click log analyzing unit analyzes a click log for the query in which a site actually selected by a viewer is ranked based on the number of clicks, and identifies a site where access is concentrated;
An extended word acquisition means comprising: a second step of acquiring the extended word by analyzing a title and a summary sentence of the site where the access is concentrated,
The first step includes
Calculating the difference in the number of clicks between sites existing in adjacent ranks in the click log as the number of relative clicks, and determining the access concentration degree of the site according to the relative number of clicks ;
By referring to a database in which the click probability for each rank of the site is stored for each search, the actual number of clicks with respect to the average click probability of the rank is calculated as the occurrence probability, and the access concentration site using the occurrence probability and the threshold value Seeking a candidate for,
And a step of identifying the site where access is concentrated by integrating the analysis results of the two steps. The second step includes
Analyzing the title and summary sentence of the site where access is concentrated to obtain a candidate group of extended words;
Ordering each candidate for the extended word;
The sentence search method according to claim 4 , further comprising: selecting an extended word based on the ranking of each of the ordered candidates. A search execution means for searching a site using the query and the extension word and outputting a search result;
6. The text search according to claim 4 , further comprising a step of causing the click log feedback processing means to reflect user click information on the search result of the step in the click log. Method. A text search program that causes a computer to execute each step of the text search method according to claim 4 . A computer-readable recording medium on which the text search program according to claim 7 is recorded.

Images