CN103136219A - Method and device for requirement mining and based on timeliness - Google Patents

Abstract

The invention provides a method and a device for requirement mining based on timeliness. The method comprises the following steps: obtaining clicking data in a selected time range from a searched log, the clicking data at least comprises queries of uses and clicked web titles of the searched results which correspond to the queries; obtaining queries in the clicked data, wherein the queries correspond to the web tiles and meet modes of preset requirement types; respectively calculating click rates of all the obtained queries, selecting queries which are larger or equal to a preset ration threshold value in the click rate, wherein the click ratio is the ratio that the click times, on the web titles with the modes of the preset requirement types, of the queries accounts for click times, on all the searched results, of the queries; and obtaining timeliness queries with the preset requirement type through the selected queries. According to the method and the device for the requirement mining and based on the timeliness, accuracy rate and recall rate of requirement query mining can be improved, so that an effect of requirement identification is improved.

Description

A method and device for demand mining based on timeliness

【Technical field】

The invention relates to the field of computer technology, in particular to a time-based demand mining method and device.

【Background technique】

With the rapid development and maturity of the Internet on a global scale, the information resources on the network are constantly enriched, and the amount of information data is also expanding rapidly. Obtaining information through search engines has become the main way for modern people to obtain information. In order to provide users with more convenient and accurate query services is the current and future development direction of search engine technology.

In search engine technology, identifying the user's search needs is an important part of improving search accuracy and effectiveness, especially in structured search (ie, vertical search). There is a way to mine the query with certain requirements in the search log in advance. When the user enters the query, directly match the query with the pre-mined query of various requirements to identify the requirements of the query entered by the user. .

When mining queries of various types of requirements, a keyword-based or template-based method is usually used to identify queries that contain keywords of a certain type of requirements as having that type of requirement, and identify queries that meet a template of a certain type of requirement as having that type of requirement. class requirements etc. However, this method is often unable to dig out time-sensitive queries in various types of requirements, such as the query "home-cooked food". Prior knowledge shows that this query only requires recipes. If the method is based on keywords or templates, the query It will be recognized that there is a demand for recipes even when the TV series "Home Cooking" is on the air. However, when the user enters "Home Cooking" during this TV show, the main demand may be video instead of recipes. Obviously, this kind of time-sensitive query cannot be recalled in demand mining, which will affect the accuracy rate in the process of demand identification.

【Content of invention】

The present invention provides a time-sensitive query identification method and device, so as to improve the accuracy rate and recall rate of demand mining, thereby improving the effect of demand identification.

The specific technical scheme is as follows:

A timeliness-based demand mining method, the method includes:

S1. Obtain the click data of the selected time period from the search log, the click data at least includes the user's search item query and the title of the clicked webpage in the search results corresponding to the query;

S2. Acquiring from the click data a query corresponding to a mode in which the title of the webpage satisfies a preset demand type;

S3. Calculate the click-through rate of each query obtained in the step S2 respectively, and select the query whose click-through rate is greater than or equal to the preset ratio threshold, wherein the click-through rate of the query is: the title of the web page of the query in the mode with the preset demand type The proportion of clicks on the query to the number of clicks on all search results;

S4. Obtain a time-sensitive query with the preset demand type from the selected query.

According to a preferred embodiment of the present invention, the pattern of the requirement type is composed of one or any combination of phrases, phrases, words, attribute identifiers, and segmentation symbols.

According to a preferred embodiment of the present invention, the mining of the pattern of the requirement type specifically includes:

A1. Obtain time-sensitive web page titles as corpus;

A2, clustering the acquired corpus;

A3. Perform steps A31 to A33 for the web page titles in each category of the clustering results:

A31. Segment the title of the web page into words;

A32, replace the named entity in the word segmentation result with the corresponding named entity type mark;

A33. Determine the n-gram n-gram of the word segmentation result, n is one or more preset positive integers, count the number of occurrences of each n-gram in the category of the webpage title, and extract the number of occurrences to meet the preset selected words The required n-grams serve as the schema for the type of requirement.

According to a preferred embodiment of the present invention, the step A1 specifically includes:

Obtain webpage titles as corpus from news websites of preset demand types; or,

The title of the clicked webpage corresponding to the time-sensitive seed query of the preset demand type is obtained from the search log as the corpus.

According to a preferred embodiment of the present invention, between the step A32 and the step A33, it also includes:

Look up the synonym vocabulary and normalize the words in the word segmentation results into synonym roots.

According to a preferred embodiment of the present invention, after the step A33, it also includes:

A34. Verify the pattern of the requirement type, and keep the verified pattern, wherein the verification process specifically includes: taking the time-sensitive webpage title as a positive example set, and taking the non-time-sensitive webpage title as a negative example set; for each pattern, calculate the ratio of the number of samples in the positive example set matched by the pattern to the sum of the number of samples in the positive example set matched by the pattern and the number of samples in the negative example set, and the calculated ratio As the score of the mode; if the score is greater than the preset score threshold, the verification is passed.

According to a preferred embodiment of the present invention, the step S4 includes:

Statistically count the website quantity corresponding to each query corresponding to the webpage title with the pattern of the preset demand type, and filter out the query corresponding to the webpage title whose source website quantity is lower than the preset quantity threshold; and / or,

Filter out queries containing preset blacklist words from the selected queries.

According to a preferred embodiment of the present invention, the step S4 includes:

Obtaining a query whose search times are greater than a preset threshold in the vertical search log of the preset demand type, and intersecting the acquired query with the selected query to obtain a time-sensitive query with the preset demand type; or,

The search times of the selected query in the vertical search log of the preset demand type are obtained respectively, and the queries whose search times are greater than a preset threshold are reserved as time-sensitive queries with the preset demand type.

According to a preferred embodiment of the present invention, after the step S4, it also includes:

The time-sensitive query with the preset demand type is combined with the query with the preset demand type mined in other ways to obtain the finally mined query with the preset demand type.

A demand excavation device based on timeliness, the device comprising:

A data acquisition unit, configured to acquire click data for a selected time period from the search log, the click data at least including the user's search item query and the title of the clicked webpage in the search results corresponding to the query;

A query acquisition unit, configured to acquire a query corresponding to a web page title that meets a preset demand type from the click data;

The query selection unit is used to separately calculate the click-through rate of each query acquired by the query acquisition unit, and select the query whose click-through rate is greater than or equal to a preset ratio threshold, wherein the click-through rate of the query is: the query has the preset demand type The proportion of the number of clicks on the title of the web page of the pattern to the number of clicks of the query on all search results;

The query determination unit is configured to obtain the time-sensitive query with the preset demand type from the query selected by the query selection unit.

According to a preferred embodiment of the present invention, the pattern of the requirement type is composed of one or any combination of phrases, phrases, words, attribute identifiers, and segmentation symbols.

According to a preferred embodiment of the present invention, the device further includes: a pattern mining unit;

The pattern mining unit specifically includes:

The corpus acquisition subunit is used to acquire time-sensitive web page titles as corpus;

The clustering subunit is used to cluster the acquired corpus;

The pattern extraction sub-unit is used to execute separately for the webpage titles in each category of the clustering results: segment the webpage titles, replace the named entities in the word segmentation results with the corresponding named entity type tags, and determine the n of the word segmentation results Metaphrase n-gram, n is one or more preset positive integers, count the number of occurrences of each n-gram in the category of the webpage title, and extract the n-gram whose occurrence times meet the preset word selection requirements as the described The schema for the requirement type.

According to a preferred embodiment of the present invention, the corpus acquisition subunit acquires web page titles as corpus from a news website of a preset demand type; or,

The title of the clicked webpage corresponding to the time-sensitive seed query of the preset demand type is obtained from the search log as the corpus.

According to a preferred embodiment of the present invention, the pattern extraction subunit is also used to find synonyms after replacing the named entity in the word segmentation result with the corresponding named entity type tag and before determining the n-gram of the word segmentation result Vocabulary, which normalizes the words in the word segmentation results into synonym roots.

According to a preferred embodiment of the present invention, the pattern extraction subunit is further configured to verify the pattern of the requirement type, and retain the pattern that passes the verification, wherein the verification process specifically includes: taking the time-sensitive webpage title as Positive example set, the non-time-sensitive web page title is used as a negative example set; for each pattern, the number of samples in the positive example set matched by the pattern and the number of samples in the positive example set matched by the pattern and the negative example set are calculated separately. The ratio of the sum of the number of samples in the example set, and the calculated ratio is used as the score of the mode; if the score is greater than the preset score threshold, the verification is passed.

According to a preferred embodiment of the present invention, the query determination unit includes: a first filtering subunit and/or a second filtering subunit;

The first filtering subunit is used to separately count the number of websites from which the webpage titles corresponding to each query have the pattern of the preset demand type, and count the query numbers corresponding to the webpage titles whose source website quantity is lower than the preset quantity threshold Filter out from the selected query;

The second filtering subunit is configured to filter out queries containing preset blacklist words from the selected queries.

According to a preferred embodiment of the present invention, the query determination unit includes: a first determination subunit or a second determination subunit;

The first determining subunit is configured to acquire a query whose search times are greater than a preset times threshold in the vertical search log of the preset demand type, and intersect the acquired query with the selected query to obtain the query with the preset query type. Set the timeliness query of the demand type;

The second determining subunit is configured to respectively obtain the search times of the selected query in the vertical search log of the preset demand type, and retain the query whose search times are greater than the preset number threshold as having the preset demand Type of timeliness query.

According to a preferred embodiment of the present invention, the device further includes: a mining and merging unit, configured to merge the time-sensitive query with the preset demand type with the query with the preset demand type excavated by other demand mining devices , to obtain the finally excavated query with the preset requirement type.

As can be seen from the above technical solutions, the present invention obtains the query of the webpage title satisfying the pattern of the preset demand type from the click data, and selects the query based on the click-through rate of the query on the webpage title of the pattern of the preset demand type to obtain the query with the preset demand type. The timeliness query of the preset demand type. That is to say, through the present invention, time-sensitive queries can be automatically mined in demand mining, so that the time-sensitive queries can be identified during demand identification, and the accuracy and recall rate of demand identification can be improved.

【Description of drawings】

FIG. 1 is a flowchart of a method provided by Embodiment 1 of the present invention;

FIG. 2 is a flow chart of mining short pattern strings provided by Embodiment 2 of the present invention;

FIG. 3 is a structural diagram of a demand digging device provided by Embodiment 3 of the present invention;

FIG. 4 is a structural diagram of a pattern mining unit provided by Embodiment 4 of the present invention.

【Detailed ways】

In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments.

Through observation, it is found that the titles of time-sensitive web pages often conform to some special patterns. If users click more on these web pages and less on other web pages when searching, it indicates that the user's search query has time-sensitive requirements. For example, when a user searches for "home cooking" query, if a TV series named "home cooking" is on the air for a period of time, the user may click on some news webpages. These webpages usually have similar patterns, and the user clicks on these There are much more webpages with the pattern than other webpages, indicating that the query has time-sensitive requirements. The time-sensitive query identification method and device provided by the present invention are based on this consideration, and the method provided by the present invention will be described in detail below through embodiments.

Embodiment one,

Fig. 1 is a flow chart of the method provided by Embodiment 1 of the present invention. As shown in Fig. 1, the method may include:

Step 101: Obtain the click data of the selected time period from the search log.

If it is necessary to analyze whether the query has timeliness requirements within a certain period of time, the click data within the period of time is obtained. The selected period of time can be the most recent period of time, for example, the click data of the latest day. Since demand mining is usually carried out periodically, the time period involved in this step can also be consistent with the time period adopted by ordinary demand mining. For example, if ordinary demand mining is conducted once a month, the time period involved in this step The length can also be one month, and the click data of the past month is obtained.

The acquired click data at least includes: the query searched by the user and the title of the clicked webpage in the search results corresponding to the query.

Step 102: Obtain the query corresponding to the pattern of the webpage title satisfying the preset demand type from the click data.

The pattern of the preset demand type can be represented by a short string of patterns. If the short string of the pattern is included in the title of the webpage, it means that the title of the webpage meets the pattern of the preset demand type. The composition of short pattern strings may include but not limited to: phrases, phrases, words, attribute identifiers, etc., and may also include some segmentation symbols, such as: [], [], (), {}, ~, ! ,, #, ¥, %, ..., &, *, _ and so on.

Taking pictures as an example, the pattern short strings include "shock feeling (picture)", "collision accident [number] people were killed", "accident [number] people were injured (picture)", where [number] is the attribute identification of the number .

The above pattern short strings can be manually summarized by observing the titles of the retrieval results, or can be automatically mined by machine learning. The automatic mining of the pattern short strings will be described in detail in Embodiment 2.

Step 103: Calculate the ratio of the number of clicks of each query on the title of the web page with the preset demand type to the number of clicks of the query on all search results, and select the query whose ratio is greater than or equal to the preset ratio threshold.

The preset ratio threshold in this step can be set according to specific demand types and actual conditions, and can be set to 0.3 in an embodiment.

Assume that the query obtained after step 102, the corresponding web page title and the number of clicks are shown in Table 1.

Table 1

After calculating the ratio in step 103, it is determined that the ratios calculated by "Shanghai Auto Show" and "Baidu Company" are both less than 0.3, while the ratio calculated by "Supernova Explosion" is greater than 0.3, so "Supernova Explosion" is selected.

Step 104: Count the number of websites from which the webpage titles corresponding to the preset demand types correspond to each query, and filter out the queries corresponding to the webpage titles whose number of source websites is lower than the preset number threshold.

If the number of websites that can indicate the timeliness of a certain query comes from the page title is very small, it is possible that its appearance and clicks are caused by the sudden behavior of the website, which does not reflect the timeliness of the query, so in this article In the step, the number of websites from which the webpage title comes is limited, and only the query corresponding to the webpage title whose number of source websites is greater than or equal to the preset number threshold is considered time-sensitive.

For example, the query obtained after step 103 and the corresponding number of source websites are shown in Table 2.

Table 2

query The number of corresponding source websites supernova explosion 8 US stealth helicopter crashes 12 Libyan Street Fighting 2 US satellite hits Earth 4

Assuming that the preset number threshold of the source website is 3, then "street fighting in Libya" is filtered out, and the remaining queries are: "supernova explosion", "US stealth helicopter crashed" and "US satellite crashed into the earth".

Step 105: Filter out queries containing preset blacklist words.

In this step, some obvious errors are removed by using the blacklist rules, and the blacklist words involved include but are not limited to: interrogative words, yellow anti-words, etc.

It should be noted that the filtering operations involved in the above step 104 and step 105 are not mandatory, and the two steps can be performed one by one, or can be performed sequentially in any order.

Step 106: Obtain the query whose search frequency is greater than the preset threshold in the vertical search log of the preset requirement type, and intersect the acquired query with the query obtained after filtering in step 105 to obtain a time-sensitive query with the preset requirement type.

The execution of this step is to further ensure that the mined query has a significant demand on the preset demand type, and it is also not a necessary step of the present invention. If the number of times the query is searched in the vertical search logs of the preset demand type is greater than the preset times threshold, it means that the query has a strong demand for the demand type.

In addition to the implementation of this step, another method can also be used, that is, to obtain the search times of each query obtained after filtering in step 105 in the vertical search log of the preset demand type, and keep the search times greater than the preset times Threshold queries are time-sensitive queries with preset demand types.

Continuing from the previous example, assume that the search times of "supernova explosion", "US stealth helicopter crashed" and "US satellite crashed into the earth" in the image vertical search log are 612, 5630 and 126 respectively. If the preset times threshold is 200, the final Determine "supernova explosion" and "US crashed stealth helicopter" as time-sensitive queries that require pictures.

After this step, the obtained time-sensitive query with the preset demand type can be merged with the query with the preset demand type mined in other ways to obtain the finally mined query with the preset demand type, thus It makes up for the time-sensitive query that cannot be recalled by the existing demand mining method.

Embodiment two,

Fig. 2 is the mining flowchart of the pattern short string provided by the second embodiment of the present invention, as shown in Fig. 2, the mining process specifically includes:

Step 201: Obtain time-sensitive web page titles as corpus.

In this step, the acquisition of the time-sensitive web page title may include but not limited to the following two methods:

Method 1: Acquiring the title of a web page from a news website of a preset demand type. Usually news websites report time-sensitive news. In order to attract readers' attention, time-sensitive news headlines usually describe the core of the event in the same documentary way and use some vocabulary to express the particularity of the news. In the title of the webpage of news websites using pictures, documentary descriptions such as "accident X people were injured", "accident X people were killed", etc., and words expressing the particularity of the news such as "[high-definition large picture]" and "(picture)" , "Group Map", etc. These titles are more suitable for mining short strings of patterns.

Method 2: Obtain the title of the clicked web page corresponding to the time-sensitive seed query of the preset demand type from the search log. Here, several time-sensitive seed queries can be artificially set, and the titles of clicked webpages corresponding to these seed queries also usually reflect time-effectiveness and are similar in patterns, so they are more suitable for mining short pattern strings.

Step 202: Clustering the acquired corpus.

In this step, when performing clustering, clustering may be performed based on events, sites, or channels. The purpose of clustering webpage titles is that after clustering, the titles in each clustering result describe the same or similar expressions. For example, the titles of webpages describing the same or similar events are more likely to have the same pattern, and the same It is more likely that the webpage titles of the sites have the same pattern, and the possibility of the webpage titles of the same channel has the same pattern. The clustering algorithm can adopt the existing text clustering method, which will not be repeated here.

For example, assume that in step 201, the news titles obtained from the picture news website are shown in Table 3.

table 3

serial number Headlines source website 1 A 6.0-magnitude and a 4.1-magnitude earthquake occurred within 2 minutes in Yili Prefecture, Xinjiang Baidu News 2 A 5.4-magnitude earthquake occurred at the junction of Sichuan and Gansu. The earthquake was felt in many places in Sichuan Province (Figure) Baidu News 3 Two people were injured in a collision between a train and a truck in Anqing City Baidu News 4 5 injured in train-truck collision in Japan (photo) Sina News 5 A train and a truck collided in Kunming and 5 carriages derailed Sina News 6 The 5.4-magnitude earthquake at the junction of Sichuan and Gansu caused no casualties and house collapses Sina News 7 A magnitude 5.4 earthquake occurred at the junction of Longnan, Gansu and Qingchuan County, Sichuan (Figure) Sina News 8 8 killed in train-car collision in Indonesia Tencent News 9 5 dead, dozens injured in Greece train truck collision Tencent News 10 The cause of the death of 4 people in an explosion in Fuquan, Guizhou is under investigation Tencent News 11 An explosion occurred at the Machangping toll station in Fuquan, Guizhou, and nearly a hundred officers and soldiers rescued Tencent News 12 Group pictures: Explosion in Fuquan, Guizhou, causing many casualties Tencent News 13 [People's Express] Two earthquakes of magnitude 6.0 and magnitude 4.1 occurred within 2 minutes in Yili, Xinjiang People's Daily Online 14 A 5.4-magnitude earthquake struck the border between Gansu and Sichuan, and Longnan residents were shaken awake People's Daily Online 15 In 2010, a magnitude 8.8 earthquake occurred in Chile People's Daily Online 16 A magnitude 7.1 earthquake hit Yushu County, Qinghai People's Daily Online 17 7.3-magnitude earthquake triggers tsunami, typhoon in Indonesia People's Daily Online 18 A 6.1-magnitude earthquake hit Zhaosu County, Xinjiang People's Daily Online

After clustering according to events, the clustering results of three categories are obtained as shown in Table 4.

Table 4

Cluster description News headline number earthquake 1, 2, 6, 7, 13, 14, 15, 16, 17, 18 traffic accident 3, 4, 5, 8, 9 explode 10, 11, 12

Step 203: Execute Step 203_1 to Step 203_5 for each web page title in each category of the clustering result.

Step 203_1: Segment the title of the webpage into words.

Taking the title of the webpage as "7.3 magnitude earthquake in Indonesia triggered tsunami and typhoon" as an example, the segmentation result is: "Indonesia/occurred/7.3/magnitude/earthquake/triggered/tsunami/and/typhoon".

Step 203_2: Replace the named entity in the word segmentation result with the corresponding named entity type tag.

The named entities here include but are not limited to: person names, place names, organization names, numbers, dates, currencies, addresses, etc.

The segmentation result of the above example is executed after this step to obtain “[place]/occurrence/[number]/level/earthquake/triggered/tsunami/and/typhoon”.

Step 203_3: Look up the synonym vocabulary, and normalize the words in the word segmentation result into synonym roots.

In the synonym vocabulary, each group of synonyms has a root, and the group of synonyms can be represented by the root. The purpose of this step is to convert more synonymous expressions into a unified template. Of course, this step is not a necessary operation.

Assuming that the root of the synonym of "和" is "and", then after performing this step in the above example, "[place]/occurrence/[number]/level/earthquake/initiation/tsunami/and/typhoon" is obtained.

Step 203_4: Determine the n-grams (n-gram) of the word segmentation result, count the number of occurrences of each n-gram in the category of the webpage title, and extract the n-gram whose number of occurrences meets the preset word selection requirements as a short string of patterns .

The so-called n-gram is a combination of n words with the smallest granularity appearing in sequence, where n is one or more preset positive integers. In the embodiment of the present invention, since 1-gram and 2-gram have a higher risk of escaping, n is usually selected as one or more positive integers greater than or equal to 3.

The preset requirement for word selection may be that the number of occurrences ranks in the top N1, where N1 is a preset positive integer, or the number of occurrences is greater than a preset threshold.

Assuming that n is selected as 3 and 4, the 3-grams obtained following the above example are: [place] occurrence of [number], occurrence of [number] magnitude, [number] magnitude earthquake, magnitude earthquake triggers, earthquake triggers tsunami, triggers tsunami and , tsunami and typhoon; 4-grams are: [place] occurrence of [number] level earthquake, occurrence of [number] level earthquake, [number] level earthquake triggers, level earthquake triggers tsunami, earthquake triggers tsunami and, triggers tsunami and typhoon.

After performing n-gram statistics on the 10 titles of the earthquake category, the top 3 n-grams are obtained, as shown in Table 5.

table 5

n-gram The number of occurrences Earthquake of magnitude [number] 10 An earthquake of magnitude [number] occurred 7 An earthquake of magnitude [number] occurred in [place] 7

The extracted n-grams can be directly used as the pattern string to end the pattern string mining process. In order to have better accuracy of the mined pattern string, subsequent steps can be further performed.

Step 203_5: Verify the extracted pattern strings, and keep the pattern strings that pass the verification.

The specific verification process is as follows: time-sensitive web page titles are used as positive examples, and non-time-sensitive web page titles are used as negative examples. For example, web page titles of news websites can be used as positive examples, and community websites (such as Baidu Tieba, Baidu Zhizhi, etc.) as the negative example set; for each pattern short string, calculate the number of samples in the positive example set matched by the pattern short string and the samples in the positive example set matched by the pattern short string The ratio of the sum of the number of samples and the number of samples in the negative example set, the calculated ratio is used as the score of the short string of the pattern; if the score is greater than the preset score threshold, the verification is passed, otherwise the verification is not passed. Wherein, the matching method between the pattern short string and the sample is as follows: if the title of the webpage contains the pattern short string, it means matching, otherwise it does not match.

The above is a detailed description of the method provided by the present invention, and the device provided by the present invention will be described in detail below in conjunction with Embodiment 3.

Embodiment three,

FIG. 3 is a structural diagram of a demand mining device provided by Embodiment 3 of the present invention. As shown in FIG. 3 , the device includes: a data acquisition unit 300 , a query acquisition unit 310 , a query selection unit 320 and a query determination unit 330 .

The data obtaining unit 300 obtains the click data of the selected time period from the search log, and the click data includes at least the user's search item query and the title of the clicked web page in the search result corresponding to the query.

The selected time period may be a recent period of time, for example, click data of the latest day. Since demand mining is usually carried out periodically, the time period involved in this step can also be consistent with the time period adopted by ordinary demand mining. For example, if ordinary demand mining is conducted once a month, the time period involved in this step The length can also be one month, and the click data of the past month is obtained.

The query obtaining unit 310 is configured to obtain, from the click data, a query corresponding to a pattern in which the title of the webpage satisfies a preset requirement type.

Wherein, the pattern of the requirement type is composed of one or any combination of phrases, phrases, words, attribute identifiers, and segmentation symbols. The pattern of the preset demand type can be represented by a short string of patterns. If the short string of the pattern is included in the title of the webpage, it means that the title of the webpage meets the pattern of the preset demand type.

The patterns of the above preset demand types can be manually summarized by observing the titles of the retrieval results, or can be automatically mined by machine learning. In order to realize the automatic mining of the patterns of the preset demand types, the device also includes: a pattern mining unit 340, the The specific structure of the unit will be described in detail in Embodiment 4.

The query selection unit 320 calculates the click-through rate of each query acquired by the query acquisition unit 310 respectively, and selects the query whose click-through rate is greater than or equal to the preset ratio threshold, wherein the click-through rate of the query is: the title of the web page of the query in the mode of the preset demand type The proportion of clicks on the query to the number of clicks on all search results for this query.

The query determination unit 330 is configured to obtain a time-sensitive query with a preset demand type from the query selected by the query selection unit 320 .

In order to further improve the accuracy of mining timeliness query, the query determination unit 330 may include: a first filtering subunit 331 and/or a second filtering subunit 332. In FIG. 3, two subunits are included as an example at the same time. When two subunits are included, the query selected by the query selection unit 320 can be processed in any order.

The first filtering subunit 331 counts the number of websites from which the webpage titles of the preset demand type patterns corresponding to each query come from, and selects the query corresponding to the webpage titles whose source website quantity is lower than the preset quantity threshold. filter out.

The second filtering subunit 332 filters out queries containing preset blacklist words from the selected queries. The blacklist words involved include but are not limited to: interrogative words, yellow anti-words, etc.

In addition, in order to further ensure that the mined query has a significant demand in the preset demand type, the query determining unit 330 may include: a first determining subunit 333 or a second determining subunit (not shown in FIG. 3 ).

The first determination subunit 333 obtains the query whose search frequency is greater than the preset threshold in the vertical search log of the preset requirement type, and intersects the obtained query with the selected query to obtain a time-sensitive query with the preset requirement type.

The second determining subunit respectively acquires the search times of the selected query in the vertical search log of the preset demand type, and retains the queries whose search times are greater than the preset number threshold as time-sensitive queries with the preset demand type.

Furthermore, the device further includes: a mining and merging unit 350, configured to merge the time-sensitive query with the preset demand type with the query with the preset demand type mined by other demand mining devices, to obtain the finally mined query with the preset demand type The query of the preset requirement type.

Embodiment four,

FIG. 4 is a structural diagram of a pattern mining unit provided in Embodiment 4 of the present invention. As shown in FIG. 4 , the pattern mining unit may specifically include: a corpus acquisition subunit 341 , a clustering subunit 342 and a pattern extraction subunit 343 .

The corpus acquisition subunit 341 acquires time-sensitive webpage titles as corpus.

Wherein, the method of acquiring corpus includes but is not limited to: obtaining webpage titles from news websites of preset demand types as corpus; or obtaining clicked webpage titles corresponding to time-sensitive seed queries of preset demand types as corpus from search logs.

The clustering subunit 342 clusters the acquired corpus.

When performing clustering, clustering can be performed based on events, websites, or channels. The clustering algorithm can adopt the existing text clustering method, which will not be repeated here.

The pattern extraction subunit 343 executes respectively for the webpage titles in each category of the clustering results: segmenting the webpage titles, replacing the named entities in the word segmentation results with corresponding named entity type tags, and determining the n-grams of the word segmentation results n-gram, n is one or more preset positive integers, count the number of occurrences of each n-gram in the category of the webpage title, and extract n-grams whose occurrences meet the preset word selection requirements as the mode of the demand type .

The named entities mentioned above include but are not limited to: person names, place names, organization names, numbers, dates, currencies, addresses, etc.

In the synonym vocabulary, usually there is a root for each group of synonyms, in order to convert more synonymous expressions into a unified template, the pattern extraction subunit 343 replaces the named entities in the word segmentation results with corresponding named entities After the type tag, and before determining the n-gram of the word segmentation result, it is also used to look up the synonym vocabulary, and normalize the words in the word segmentation result into synonym roots.

In order to further improve the accuracy of pattern mining, the pattern extraction subunit 343 is also used to verify the pattern of the requirement type, and retain the pattern that passes the verification.

The verification process specifically includes: taking time-sensitive web page titles as positive example sets, and non-time-sensitive web page titles as negative example sets; calculating the number of samples in the positive example set matched by the pattern for each pattern The ratio of the sum of the number of samples in the positive example set and the number of samples in the negative example set matched by the pattern is used as the score of the pattern; if the score is greater than the preset score threshold, the verification is passed. Wherein, the pattern matching method is: if the title of the web page contains the pattern, it means matching, otherwise it does not match.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the present invention. within the scope of protection.

Claims (18)

1. A demand mining method based on timeliness, characterized in that the method comprises: S1. Obtain the click data of the selected time period from the search log, the click data at least includes the user's search item query and the title of the clicked webpage in the search results corresponding to the query; S2. Acquiring from the click data a query corresponding to a mode in which the title of the webpage satisfies a preset demand type; S3. Calculate the click-through rate of each query obtained in the step S2 respectively, and select the query whose click-through rate is greater than or equal to the preset ratio threshold, wherein the click-through rate of the query is: the title of the web page of the query in the mode with the preset demand type The proportion of clicks on the query to the number of clicks on all search results; S4. Obtain a time-sensitive query with the preset demand type from the selected query. 2. The method according to claim 1, wherein the pattern of the requirement type is composed of one or any combination of phrases, phrases, words, attribute identifiers, and segmentation symbols. 3. The method according to claim 1 or 2, wherein the mining of the pattern of the demand type specifically comprises: A1. Obtain time-sensitive web page titles as corpus; A2, clustering the acquired corpus; A3. Perform steps A31 to A33 for the web page titles in each category of the clustering results: A31. Segment the title of the web page into words; A32, replace the named entity in the word segmentation result with the corresponding named entity type mark; A33. Determine the n-gram n-gram of the word segmentation result, n is one or more preset positive integers, count the number of occurrences of each n-gram in the category of the webpage title, and extract the number of occurrences to meet the preset selected words The required n-grams serve as the schema for the type of requirement. 4. The method according to claim 3, wherein the step A1 specifically comprises: Obtain webpage titles as corpus from news websites of preset demand types; or, The title of the clicked webpage corresponding to the time-sensitive seed query of the preset demand type is obtained from the search log as the corpus. 5. The method according to claim 3, characterized in that, between said step A32 and step A33, further comprising: Look up the synonym vocabulary and normalize the words in the word segmentation results into synonym roots. 6. The method according to claim 3, characterized in that, after the step A33, further comprising: A34. Verify the pattern of the requirement type, and keep the verified pattern, wherein the verification process specifically includes: taking the time-sensitive webpage title as a positive example set, and taking the non-time-sensitive webpage title as a negative example set; for each pattern, calculate the ratio of the number of samples in the positive example set matched by the pattern to the sum of the number of samples in the positive example set matched by the pattern and the number of samples in the negative example set, and the calculated ratio As the score of the mode; if the score is greater than the preset score threshold, the verification is passed. 7. The method according to claim 1, wherein said step S4 comprises: Statistically count the website quantity corresponding to each query corresponding to the webpage title with the pattern of the preset demand type, and filter out the query corresponding to the webpage title whose source website quantity is lower than the preset quantity threshold; and / or, Filter out queries containing preset blacklist words from the selected queries. 8. The method according to claim 1, wherein said step S4 comprises: Obtaining a query whose search times are greater than a preset threshold in the vertical search log of the preset demand type, and intersecting the acquired query with the selected query to obtain a time-sensitive query with the preset demand type; or, The search times of the selected query in the vertical search log of the preset demand type are obtained respectively, and the queries whose search times are greater than a preset threshold are reserved as time-sensitive queries with the preset demand type. 9. The method according to claim 1, further comprising after the step S4: The time-sensitive query with the preset demand type is combined with the query with the preset demand type mined in other ways to obtain the finally mined query with the preset demand type. 10. A demand excavation device based on timeliness, characterized in that the device comprises: A data acquisition unit, configured to acquire click data for a selected time period from the search log, the click data at least including the user's search item query and the title of the clicked webpage in the search results corresponding to the query; A query acquisition unit, configured to acquire a query corresponding to a web page title that meets a preset demand type from the click data; The query selection unit is used to separately calculate the click-through rate of each query acquired by the query acquisition unit, and select the query whose click-through rate is greater than or equal to a preset ratio threshold, wherein the click-through rate of the query is: the query has the preset demand type The proportion of the number of clicks on the title of the web page of the pattern to the number of clicks of the query on all search results; The query determination unit is configured to obtain the time-sensitive query with the preset demand type from the query selected by the query selection unit. 11. The device according to claim 10, wherein the pattern of the requirement type is composed of one or any combination of phrases, phrases, words, attribute identifiers, and segmentation symbols. 12. The device according to claim 10 or 11, further comprising: a pattern mining unit; The pattern mining unit specifically includes: The corpus acquisition subunit is used to acquire time-sensitive web page titles as corpus; The clustering subunit is used to cluster the acquired corpus; The pattern extraction sub-unit is used to execute separately for the webpage titles in each category of the clustering results: segment the webpage titles, replace the named entities in the word segmentation results with the corresponding named entity type tags, and determine the n of the word segmentation results Metaphrase n-gram, n is one or more preset positive integers, count the number of occurrences of each n-gram in the category of the webpage title, and extract the n-gram whose occurrence times meet the preset word selection requirements as the described The schema of the requirement type. 13. The device according to claim 12, wherein the corpus acquisition subunit acquires a web page title as a corpus from a news website of a preset demand type; or, The title of the clicked webpage corresponding to the time-sensitive seed query of the preset demand type is obtained from the search log as the corpus. 14. The device according to claim 12, wherein the pattern extraction subunit replaces the named entity in the word segmentation result with the corresponding named entity type mark, and determines the n-gram of the word segmentation result Previously, it was also used to look up the synonym vocabulary and normalize the words in the word segmentation results into synonym roots. 15. The device according to claim 12, wherein the pattern extraction subunit is further configured to verify the pattern of the requirement type, and retain the pattern that passes the verification, wherein the verification process specifically includes: Time-sensitive webpage titles are used as a positive example set, and non-time-sensitive webpage titles are used as a negative example set; for each pattern, the number of samples in the positive example set matched by the pattern and the positive examples matched by the pattern are calculated separately The ratio of the number of samples in the set to the sum of the number of samples in the negative set is used as the score of the pattern; if the score is greater than the preset score threshold, the verification is passed. 16. The device according to claim 10, wherein the query determining unit comprises: a first filtering subunit and/or a second filtering subunit; The first filtering subunit is used to separately count the number of websites from which the webpage titles corresponding to each query have the pattern of the preset demand type, and count the query numbers corresponding to the webpage titles whose source website quantity is lower than the preset quantity threshold Filter out from the selected query; The second filtering subunit is configured to filter out queries containing preset blacklist words from the selected queries. 17. The device according to claim 10, wherein the query determining unit comprises: a first determining subunit or a second determining subunit; The first determining subunit is configured to acquire a query whose search times are greater than a preset times threshold in the vertical search log of the preset demand type, and intersect the acquired query with the selected query to obtain the query with the preset query type. Set the timeliness query of the demand type; The second determining subunit is configured to respectively obtain the search times of the selected query in the vertical search log of the preset demand type, and retain the query whose search times are greater than the preset number threshold as having the preset demand Type of timeliness query. 18. The device according to claim 10, further comprising: a mining and merging unit, configured to combine the time-sensitive query with the preset demand type with the query with the preset demand mined by other demand mining devices. The query of the preset requirement type is merged to obtain the finally excavated query with the preset requirement type.

Images