JP6251637B2 - Information retrieval method, apparatus and program - Google Patents

Description

  The present invention is based on ambiguous time expressions such as “in the past”, “a little before”, “soon”, “several years ago”, “before”, etc. included in the search query. The present invention relates to an information search method, apparatus, and program for setting or optimizing search conditions, and in particular, a model that predicts a quantitative or specific time expression from an ambiguous time expression is updated based on a user's response to a search result. The present invention relates to an information search method, apparatus, and program.

  Ambiguous time expressions such as “old days”, “a little before”, “soon”, “several years ago”, “previous”, etc. are specific time expressions that each expression represents (3 days ago) -5 days ago, 6 months ago-3 years ago, 1 day later-1 month later, etc.), it becomes an obstacle to information retrieval and communication of intention during conversation. Therefore, an apparatus that extracts an ambiguous time expression included in daily speech to acquire an individual's time sensation and supports information retrieval and intention sharing during conversation has been studied.

  Patent Document 1 discloses a function of assigning a quantitative time expression to an ambiguous time expression input by a user in an information retrieval scene. Here, the correspondence table for mapping ambiguous time expressions (yesterday, around 5 o'clock, etc.) and quantitative time expressions (● month • day, 4 o'clock to 6 o'clock, etc.) (“midnight” is 0 Time) etc. and a rule (such as “about” is a time ± 1) is prepared in advance, and the time expression is converted according to these correspondence tables and rules.

  Patent Document 2 discloses a technique in which a user designates an accurate date and time as a search range when searching for information.

  Non-Patent Document 1 discloses an ambiguity according to the size of a unit of time for an expression representing a degree such as “almost” or “a little” when converting an ambiguous time expression into a quantitative time expression. Is disclosed. For example, since the degree of ambiguity is different between “about 3 minutes” and “about 3 hours”, a width corresponding to the size (minute, time) of the unit is given.

Japanese Patent Laid-Open No. 10-254891 JP 2000-244958 A Japanese Patent Application No. 2014-101383

"Expression and Reasoning of Time Concept Including Ambiguity", Proceedings of the 39th Annual Conference (1), 357-358, 1989-10-16

  When searching for past memories or information, or when making a promise in the future, there are cases where time information cannot be input accurately, and as an ambiguous time expression, for example, "a little before" or "soon" May be used. However, the quantitative time expression expressed by the ambiguous time expression, such as whether “a little before” is 1 hour to 3 hours ago, or 1 hour to 5 hours ago, is different for each individual. Therefore, if an ambiguous time expression is used as it is, there arises a problem that the information retrieval system and the conversation partner cannot correctly understand the time information.

  In Patent Document 1, an attempt is made to solve with a conversion method using correspondence tables and rules prepared in advance. However, since the time sensation varies from person to person, (1) Administrators must create it manually for each person. (2) The quantitative time expression that sets the time range only needs to assume the risk of losing, so set the time range wider than necessary. In addition, there is a technical problem that it is necessary to further determine (3) what time probability within the range is high. For this reason, for example, in a scene used for information search, excessive information is acquired as a search result, and prioritization cannot be performed on the search result, so that the readability of the search result cannot be improved.

  In Patent Document 2, a specific date and time can be specified as a search range, but ambiguous time expressions such as “old”, “a little before”, “soon”, “several years ago”, “previous”, etc. The search range could not be specified with. In Non-Patent Document 1, a range can be given to an expression according to a certain rule, but it cannot cope with a flexible setting considering individual differences.

  In response to such a technical problem, the inventors of the present invention are able to identify the correspondence between the ambiguous time expression and the specific time expression as a probability distribution without providing a special learning opportunity or a specialized administrator. A method, apparatus, and program for learning in units have been invented and a patent application (Japanese Patent Application No. 2014-101383) has been filed.

  However, the correspondence between the ambiguous time expression and the specific time expression changes due to the influence of the user's time sensation that changes over time. Also, if the learning environment is different from the practical environment, there will be a subtle shift in the correspondence between the ambiguous time expression and the specific time expression, and the search range intended by the user will match the actual search range. It may disappear.

  An object of the present invention is to solve the above technical problems, quantitatively interpret a date and time condition intended by a user from an ambiguous time expression included in a search query based on a prediction model, and reflect this in the search condition. Information retrieval that enables the retrieval of highly accurate information and updates the prediction model based on the user's response to the search results, so that the date and time conditions intended by the user for ambiguous time expressions can be reflected in the prediction model. To provide a method, an apparatus, and a program.

  In order to achieve the above object, the present invention has the following configuration in an information search apparatus that responds to a search query including an ambiguous time expression with a search result based on a search condition corresponding to the time expression. There is a feature.

  (1) A prediction model that probabilistically predicts a quantitative date and time range for each ambiguous time expression, and an ambiguous time expression included in a search query is applied to the prediction model so that the probability value is within a predetermined setting range. A means for predicting a date and time range; a search function means for displaying a search query and a search result obtained by a search using the date and time range as a search condition; and a user reaction detecting means for detecting a user reaction to the search result; And a prediction model updating unit that corrects the probability value corresponding to the prediction result of the date and time range of the prediction model based on a user's reaction.

  (2) The prediction model update unit increases the probability value corresponding to the prediction result of the date and time range of the prediction model when the user reaction detection unit detects an input operation that the user is estimated to be satisfied with the search result. Correction is made, and when an input operation presumed that the user is not satisfied with the search result is detected, a reduction correction is made.

  (3) The increase / decrease of the increased or decreased probability value is gradually decreased with time.

  (4) Learning data extracting means for extracting learning data including ambiguous time expressions and event expressions that can specify date and time from learning data candidates, and searching the history date and time related to the event expressions from the history information And a means for calculating the actual value corresponding to the ambiguous time expression and a means for constructing a prediction model based on the correspondence between the ambiguous time expression and the actual value.

  (5) Equipped with an evaluation means that evaluates the search result of the date and time related to the event expression, so that the correspondence relationship where the evaluation value does not satisfy the predetermined reference value is not used for the construction of the prediction model.

  (6) It comprises means for detecting a speech signal, means for converting the speech signal into a time series of character data, and means for generating a learning data candidate by dividing the time series of character data according to a predetermined rule.

According to the present invention, the following effects are achieved.
(1) A model that predicts a quantitative date and time range from an ambiguous time expression is given as a probability distribution, and the date and time range of the prediction model is based on the user's reaction and evaluation to search results that use this date and time range as a search condition. The probability value corresponding to is increased or decreased. Therefore, the user's evaluation can be reflected in the update of the prediction model, and the accuracy in predicting the quantitative date and time range from the ambiguous time expression, and hence the search accuracy can be expected to be improved over time.

  (2) Since the user's reaction to the search result is detected based on the user's input operation, the user's subjectivity can be physically detected.

  (3) When reflecting the user evaluation in the update of the prediction model, the increase / decrease correction based on the newer search result is more strongly reflected in the prediction model, so the relationship between the user's ambiguous time expression and the quantitative date and time range Even if a change over time occurs, it is possible to accurately reflect the change over time in the prediction of the date and time range, so that an accurate information search can be realized.

  (4) Based on the history information of the user who uttered the ambiguous time expression, the actual value corresponding to the ambiguous time expression used by the user in the past, that is, the corresponding actual period is searched and associated. Since learning is performed, a prediction model capable of accurately predicting a quantitative time expression from an ambiguous time expression can be constructed for each user.

  (5) The search result of the actual value corresponding to the ambiguous time expression is evaluated based on the matching accuracy and variance, and the search result with low evaluation value is not used for the construction of the prediction model, so the accuracy is low It is possible to prevent a prediction model from being built based on a search result or an incorrect search result.

  (6) Since learning data candidates are generated by detecting voice data uttered by users on a daily basis, it is easy to learn considering individual differences without creating special learning opportunities or specialized administrators. You can do it.

It is a functional block diagram of the information search device concerning a 1st embodiment of the present invention. It is the figure which showed an example of the prediction model accumulate | stored in prediction model DB. It is the figure which showed the prediction method of the date-time range based on the probability value of a prediction model. It is the flowchart which showed the operation | movement of 1st Embodiment of this invention. It is the figure which showed the update method of a prediction model. It is a functional block diagram of the time expression learning device concerning a 2nd embodiment of the present invention. It is the figure which showed the method of extracting learning data from a learning data candidate. It is the flowchart which showed the operation | movement of 2nd Embodiment of this invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. FIG. 1 is a functional block diagram showing a configuration of a main part of an information search apparatus according to an embodiment of the present invention. Each function described later is provided to a general-purpose computer or server, or a mobile terminal such as a smartphone or tablet terminal. It can be configured by implementing an application (program) to be realized. Alternatively, it can be configured as a dedicated machine or a single-function machine in which a part of the application is hardware or ROMized. Moreover, when it comprises with a computer and a server, it is applicable also to the cloud service which several people share on a network. Here, application to a smartphone will be described as an example, and illustrations of components unnecessary for the description of the present invention are omitted.

  The user interface (I / F) 1 is connected to a microphone MC that detects a voice signal of a smartphone user (hereinafter referred to as a smartphone user) and a touch screen TS that receives text input and various touch operations.

  The search request accepting unit 5 accepts a search request including a search query that is input from the touch screen TS by a smartphone user or a search query that is input from the microphone MC and converted into text data by an appropriate speech recognition algorithm.

  The prediction model database (DB) 4 includes ambiguous time expressions such as “old days”, “a little before”, “soon”, “several years ago”, “previous”, etc. A prediction model for predicting a quantitative or specific time expression is registered in advance.

  FIG. 2 is a diagram showing an example of a prediction model stored in the prediction model DB 4. In this embodiment, a quantitative time expression is registered as a probability distribution for each ambiguous time expression. The probability distribution of each prediction model is actually a discrete value, but is shown here as a continuous value for convenience.

  Figure (a) shows the probability distribution for the ambiguous time expression `` Recent '', Figure (b) shows the probability distribution for the ambiguous time expression `` Old '', and Figure (c) shows the ambiguous time expression `` Probability distribution for `` several years ago '', figure (d) shows the probability distribution for the ambiguous time expression `` soon '', figure (e) shows the probability distribution for the ambiguous time expression `` previous '' .

  Returning to FIG. 1, the date / time range prediction unit 6 extracts an ambiguous time expression included in the search query and applies it to the corresponding prediction model, whereby the date / time range in which the probability value satisfies a predetermined setting condition. As a temporal search condition corresponding to the ambiguous temporal expression.

  FIG. 3 is a diagram schematically representing a date / time range prediction method based on the probability value of the prediction model. When the probability value setting condition is set to “0.8 or more”, for example, the probability value is 0.8 or more. The date range Δt1 is predicted as a temporal search condition corresponding to the ambiguous time expression. Similarly, when “0.6 or more and less than 0.8” is set as the condition of the probability value, for example, the date / time range Δt2 corresponding thereto is predicted as a temporal search condition corresponding to the ambiguous time expression.

  Returning to FIG. 1, the Web search function unit 7 transmits a search request including the search query and the date / time range to the appropriate search server on the network from the communication interface 2, and matches a plurality of search results that are responded to the search request. A list is displayed on the display 3 in an order corresponding to the accuracy and presented to the smartphone user.

  The user response detection unit 8 is a touch operation for accessing any of the information sources of a large number of search results or a touch for displaying the next page of the search results as a smartphone user's response to the search results displayed in the list. An operation is detected and used as an index as to whether or not the user is satisfied with the current search result. The prediction model update unit 9 updates the probability distribution of the prediction model used for the prediction of the date / time range based on the detected smartphone user reaction.

  Next, the operation of the embodiment of the present invention will be described with reference to the flowchart of FIG. In step S1, a search request including a search query is received by the search request receiving unit 5. In step S2, it is determined whether or not an ambiguous time expression is included in the search query. If not included, the process proceeds to a normal search process, and if included, the process proceeds to step S3.

  In step S3, the probability value range of the prediction model is set to the highest level as a condition for predicting a quantitative time expression corresponding to the ambiguous time expression. Here, the description will be continued assuming that the probability value range is set to “0.8 or more”.

  In step S4, the date and time range prediction unit 6 applies the ambiguous time expression to the corresponding prediction model in the prediction model DB 4, and the date and time with a probability value of 0.8 or more is obtained as a quantitative time expression corresponding to the time expression. A range is extracted. For example, the date / time range Δt1 shown in FIG. 3 is extracted.

  In step S5, a search request including the date and time range Δt1 as an optional search condition together with the search query is provided by the web search function unit 7 to a search engine on the network. For example, if the search query is "What was the title of the comedy program you saw recently?", The search keyword was "I saw", "Comedy program", "Title", "What", and ambiguous time expression A search request using a specific date range Δt1 predicted from “recent” as a search condition is provided to the search engine.

  In step S6, a number of search results responded by the search engine are displayed in a list on the display 3. In steps S7 and S10, the user response to the search result is detected by the user response detection unit 8. For example, if the smartphone user cannot find a desired search result from the displayed search result list and performs a transition operation to the next page in order to obtain a further search result, this is detected in step S10 and step S11. Proceed to

  In step S11, the probability value range set when predicting the date / time range corresponding to the ambiguous time expression is updated to a value lower than the initial value set in step S3. In the present embodiment, the probability value range is updated from “0.8 or more” to, for example, “0.6 or more and less than 0.8”.

  In step S12, the probability value of the prediction model corresponding to the predicted date / time range is updated by the reduction correction. FIG. 5A is a diagram schematically showing a method for reducing and correcting the probability value of the prediction model. The probability value (frequency) corresponding to the date / time range Δt1 is uniformly reduced and corrected.

  Thereafter, the processing returns to step S4, and the information search is continued using the date / time range Δt2 (see FIG. 4) corresponding to the probability value range of 0.6 or more and less than 0.8 as a new search condition.

  On the other hand, when the smartphone user finds a desired search result from the displayed search result list and tries to access the search result, this is detected in step 7 and proceeds to step S8.

  In step S8, the probability value of the prediction model corresponding to the predicted date / time range Δt1 is updated by increasing correction. FIG. 5 (b) is a diagram schematically showing a method of increasing and correcting the probability value of the prediction model, and the probability value (frequency) corresponding to the date / time range Δt1 is uniformly increased and corrected. In step S9, a communication session is established with the access destination, and search result information is acquired.

  Note that it is desirable that the increase / decrease correction amount of the probability value be gradually decreased with the passage of time since the correction. In this way, the increase / decrease correction based on newer search results can be reflected more strongly in each prediction model, so the relationship between the vague time expression and quantitative time expression of the smartphone user changes over time. Even in such a case, accurate date / time range prediction is possible.

  According to the present embodiment, a model that predicts a quantitative date and time range from an ambiguous time expression is given as a probability distribution, and a prediction model based on a user's reaction and evaluation to a search result using the date and time range as a search condition The probability value of the date / time range is corrected for increase / decrease. Therefore, the user's evaluation can be reflected in the update of the prediction model, and as a result, the accuracy when predicting the quantitative date and time range from the ambiguous time expression and thus the search accuracy can be expected to be improved over time. Become.

  FIG. 6 is a functional block diagram showing the configuration of the second embodiment of the present invention. The same reference numerals as those described above represent the same or equivalent parts. Here again, an explanation will be given taking adaptation to a smartphone as an example.

  In the first embodiment described above, it is assumed that the prediction model is registered in the database 4 in advance. However, in this embodiment, a prediction model unique to each smartphone user is constructed by the smartphone and registered in the prediction model DB 4. Is done. Therefore, in this embodiment, in addition to the practical part corresponding to the first embodiment, a learning part and various databases 10, 11, and 12 for constructing the prediction model DB 4 are provided.

  The voice signal recognition unit 13 applies appropriate voice recognition processing to the voice signal of the smartphone user detected by the microphone MC to convert it into character data, and outputs the time series. The learning data candidate generation unit 14 decomposes the time series of the character data, for example, in units of sentences based on the contents, the presence / absence of unvoiced portions, and the like, and outputs each sentence as a learning data candidate.

  The learning data extraction unit 15 extracts only learning data candidates suitable as learning data from among a large number of learning data candidates, and outputs this as learning data. In the present embodiment, a list of ambiguous time expressions is registered in advance in the time expression DB 10, and a list of event expressions that can specify the date and time is registered in the event expression DB 11 in advance. The learning data extraction unit 15 extracts each of the learning data candidates as learning data if both of the ambiguous time expression and the event expression capable of specifying the date and time are included.

  FIG. 7 is a diagram schematically showing a method of extracting learning data from a large number of learning data candidates. The first learning data candidate “Recently seen ● ● (entertainer name) came out "The program was interesting" includes "Recent" as an ambiguous time expression and "Comedy program that I saw, ●●" as an event expression that can specify the date and time. Extracted as learning data.

  The second candidate for the learning data “I want to go to the live of ● ● once again” includes “old” as an ambiguous time expression, and “go ●” as an event expression that can specify the date and time. ● "Live" is included, so it is extracted as learning data.

  The third candidate for learning data, “Shibuya's ●● company is not growing rapidly” visited several years ago, includes “several years ago” as an ambiguous time expression, and the date and time can be specified Since “Shibuya ●● company visited” is included as an expression, it is extracted as learning data.

  The fourth candidate for learning data “I'm looking forward to a drinking party to be held soon” includes “coming soon” as an ambiguous time expression, and “event to be held” as an event expression that can specify the date and time. Since it is included, it is extracted as learning data.

  The fifth learning data candidate “A previously purchased desk lamp has been broken” includes “Previous” as an ambiguous time expression, and includes “Bought desk lamp” as an event expression that can specify the date and time. Therefore, it is extracted as learning data.

  On the other hand, the sixth learning data candidate “What was the title of the TV program that I watched recently” includes “Recent” as an ambiguous time expression, but does not include an event expression that can identify the date and time. It is not extracted as learning data.

  On the other hand, the seventh learning data candidate “It seems to be raining” includes “It seems to rain” as an event expression that can specify the date and time, but does not include ambiguous time expression, so it is not extracted as learning data .

  Returning to FIG. 6, the history information management unit 12 uses the online purchase or the purchase history using the wallet function, the telephone conversation history, the schedule registered by the user as the history information that can identify the date and time specific to the smartphone user of the speaker. It manages history information such as history, tracking history based on GPS and wireless base station location information, document history related to user-created document files, web browsing history using browser.

  The date search unit 16 searches all the history information under the history information management unit 12 using the event expression that can be specified as the date included in the extracted learning data as a search key, and finds corresponding history information. Then, the actual value is calculated from the date and time information, and this is associated with the ambiguous time expression included in the learning data candidate.

  For example, in the case of the first learning data candidate “Recently watched ●● (entertainer name) comedy program, it was interesting”, search key “●● (entertainer name)” and “comic program” All the history information is searched, and the search key is registered in the viewing history of 1Seg broadcasting, and when this is answered as a search result, the time difference between the viewing date and time of the comedy program and the current time is ambiguous. Corresponding as a “recent” actual value.

  Similarly, if the second candidate for the learning data is “I want to go to ●● live again in the past”, the history information is searched using “●● live” as a search key. When a search key is registered in the history and this is returned as a search result, the time difference between the date and time of the schedule book and the current time is associated with the actual value of the ambiguous time expression “old”.

  The evaluation unit 17 evaluates the search result, extracts a correspondence relationship between an ambiguous time expression that satisfies a predetermined evaluation condition and its actual value, and registers it in the prediction model DB 4. In the present embodiment, it is assumed that the date search unit 16 responds with the matching accuracy together with the search result, and the correspondence relationship in which the matching accuracy is lower than a predetermined reference value is not registered in the prediction model DB 4.

  In addition, a plurality of search results are obtained for an event expression that can specify one date and time, and there is a large variation in each date and time, in other words, when there is a large variation in the associated actual value, it corresponds. Since the actual value cannot be specified, the evaluation is low, and registration in the prediction model DB 4 is postponed.

  FIG. 8 is a flowchart showing a learning procedure of a prediction model according to the present invention. Here, a case where offline learning is performed based on user's utterance data stored in advance will be described as an example. You may learn online if you have enough ability to handle the process.

  In step S21, a known voice recognition process is performed on the voice signal detected by the microphone MC by the voice signal recognition unit 13, and a time series of character data is output. In step S22, the learning data candidate generation unit 14 decomposes the time series of the character data into a plurality of sentences based on the contents and silent portions, and sets each sentence as a learning data candidate.

  In step S23, attention is paid to one of the learning data candidates, and in step S24, the learning data extraction unit 15 determines whether the focused learning data candidate includes the ambiguous time expression. If an ambiguous time expression is included, the process proceeds to step S25.

  In step S25, the learning data extraction unit 15 determines whether or not the learning data candidate includes the event expression that can specify the date and time. If the event expression that can specify the date and time is included, the current learning data candidate is adopted as the regular learning data, and the process proceeds to step S26. The date and time search unit 16 uses the event expression that can specify the date and time as a search key to manage history information. Part 12 is searched.

  In step S27, the evaluation unit 17 evaluates all search results. In the present embodiment, a higher evaluation value is given to a search result with higher matching accuracy based on the matching accuracy of each search result.

  Furthermore, when a plurality of search results (actual values) are obtained for an event expression that can specify one date and time, the variation of each search result is calculated as variance, for example. As a result, a low evaluation value is given if there is a large variation in search results, for example, the date and time of one search result is one month ago while the other search result is one year ago. . On the other hand, a high evaluation value is given if the variation is small, for example, the primary of one search result is 40 days ago and the date and time of the other search result is 42 days ago.

  In step S28, the evaluation value of each search result is compared with a predetermined reference value. For a search result whose evaluation value exceeds the predetermined reference value, the actual value is associated with the ambiguous time expression in step S29. Are registered in the prediction model DB 4. That is, the frequency is added to the corresponding actual value of the corresponding prediction model.

  If the evaluation value does not satisfy the predetermined reference value, the process proceeds to step S30, and it is determined whether or not the above processing has been completed for all learning data candidates. If not completed, the process returns to step S23, and the prediction model described with reference to FIG. 2 is sequentially constructed by switching the learning data candidate to be noticed and repeating the above processes.

  According to this embodiment, based on the history information of the user who spoke the ambiguous time expression, the actual value corresponding to the ambiguous time expression used by the user in the past, that is, the corresponding actual date and time is searched. Since learning is performed in association with each other, a prediction model capable of accurately predicting a quantitative time expression from an ambiguous time expression can be constructed.

  Further, according to the present embodiment, the search result of the actual value corresponding to the ambiguous time expression is evaluated based on the matching accuracy and variance, and the search result having a low evaluation value is not used for the construction of the prediction model. Therefore, it is possible to prevent the prediction model from being constructed based on a search result with low accuracy or an incorrect search result.

  Furthermore, according to the present embodiment, learning data candidates can be generated by detecting voice data uttered by the user on a daily basis, so that learning that takes into account individual differences can be achieved without providing special learning opportunities or specialized managers. Can be done easily.

  In each of the above embodiments, the prediction model is described as being constructed for each user. However, the present invention is not limited to this. For example, a learning function is provided on the cloud, and each user The learning data and history information may be uploaded from its own terminal, the learning data and history information uploaded from each user on the cloud may be classified based on user attributes, and a prediction model may be constructed for each user attribute. .

  DESCRIPTION OF SYMBOLS 1 ... User I / F, 2 ... Communication I / F, 3 ... Display, 4 ... Prediction model DB, 5 ... Search request reception part, 6 ... Date / time range prediction part, 7 ... Web search function part, 8 ... User reaction detection , 9 ... Prediction model update part, 10 ... Time expression DB, 11 ... Event expression DB, 12 ... History information management part, 13 ... Speech signal recognition part, 14 ... Learning data candidate generation part, 15 ... Learning data extraction part, 16 ... Date search part, 17 ... Evaluation part

Claims (16)

In an information search apparatus that responds to a search query including an ambiguous time expression with a search result based on a search condition corresponding to the time expression,
A prediction model that probabilistically predicts a quantitative date and time range for each ambiguous time expression;
Means for applying an ambiguous time expression included in a search query to the prediction model to predict a date and time range within which a probability value is within a predetermined setting range;
Search function means for displaying a list of search results obtained by the search using the search query and the date and time range as search conditions;
User response detection means for detecting a user response to the search results;
An information search apparatus comprising: a prediction model update unit that increases or decreases a probability value corresponding to a prediction result of the date and time range of the prediction model based on a reaction of the user.   The prediction model update unit increases a probability value corresponding to the prediction result of the date and time range of the prediction model when the user reaction detection unit detects an operation that the user is estimated to be satisfied with the search result. The information search apparatus according to claim 1, wherein correction is performed.   The information search apparatus according to claim 2, wherein the operation that the user is estimated to be satisfied with the search result is an access request for any of the search results.   The prediction model update unit is configured to detect a probability value corresponding to a prediction result of the date and time range of the prediction model when an operation that the user is estimated not to be satisfied with the search result is detected by the user reaction detection unit. 4. The information search apparatus according to claim 1, wherein the correction is performed by reducing the correction.   The information search apparatus according to claim 4, wherein the operation that the user is estimated not to be satisfied with the search result is an operation that requests a list display of the next search result.   The means for predicting the date and time range, when an operation presumed that the user is not satisfied with the search result is detected, lowers the probability value setting range and re-predicts the date and time range. The information search device according to claim 4 or 5.   7. The information search apparatus according to claim 1, wherein the increase / decrease of the probability value that has been increased or decreased is gradually decreased as time elapses. Learning data extraction means for extracting learning data including ambiguous time expressions and event expressions that can specify date and time from learning data candidates;
Means for searching a history date and time related to the event expression from history information, and calculating a performance value corresponding to the ambiguous time expression;
8. The information search apparatus according to claim 1, further comprising means for constructing the prediction model based on a correspondence relationship between the ambiguous time expression and the actual value. Comprising an evaluation means for evaluating a search result of date and time related to the event expression,
9. The information search apparatus according to claim 8, wherein a correspondence relationship whose evaluation value does not satisfy a predetermined reference value is not used for construction of a prediction model.   The information search apparatus according to claim 9, wherein the evaluation unit evaluates the search result higher as the matching accuracy is higher.   11. The information search apparatus according to claim 9, wherein the evaluation means evaluates the lower the greater the variance of each date and time when a plurality of the search results are obtained. Means for detecting an audio signal;
Means for converting the voice signal into a time series of character data;
12. The information search apparatus according to claim 8, further comprising means for dividing the time series of the character data according to a predetermined rule to generate learning data candidates. A time expression database that stores a number of ambiguous time expressions;
An event expression database for storing a large number of date and time identifiable event expressions,
9. The learning data extracting unit extracts a time expression stored in the time expression database and an event expression stored in the event expression database from the learning data candidates. 12. The information search device according to any one of 12 above. In an information search apparatus that responds to a search query including an ambiguous time expression with a search result based on a search condition corresponding to the time expression,
A prediction model that probabilistically predicts a quantitative date and time range for each ambiguous time expression;
Means for applying an ambiguous time expression included in a search query to the prediction model to predict a date and time range within which a probability value is within a predetermined setting range;
Search function means for displaying a list of search results using the search query and the prediction result of the date and time range as a search condition;
User response detection means for detecting a user response to the search results;
An information search apparatus comprising: a prediction model update unit that increases or decreases a probability value corresponding to a prediction result of the date and time range of the prediction model based on a reaction of the user. In an information search method for responding to a search query including an ambiguous time expression with a search result based on a search condition corresponding to the time expression,
Applying ambiguous time expressions included in the search query to a prediction model that probabilistically predicts a quantitative date and time range for each ambiguous time expression and predicting a date and time range in which the probability value is within a predetermined setting range; ,
A procedure for displaying a list of search results obtained by a search using the search query and the date and time range as search conditions;
Detecting a user's response to the search results;
And a procedure for increasing or decreasing the probability value corresponding to the prediction result of the date / time range of the prediction model based on the reaction of the user. In an information search program for responding to a search query including an ambiguous time expression with a search result based on a search condition corresponding to the time expression,
A procedure for accepting a search request including a search query via an interface;
A procedure for extracting an ambiguous temporal expression included in the search query;
Predict the date and time range in which the probability value is within the predetermined setting range by applying the ambiguous time representation to the corresponding prediction model in the database that stores the prediction model that stochastically predicts the quantitative date and time range for each ambiguous time expression And the steps to
A procedure for displaying a list of search results obtained by a search using the search query and the date / time range as a search condition;
A procedure for detecting a user response to the search result via an interface;
An information search program for causing a computer to execute a procedure for increasing or decreasing a probability value corresponding to a prediction result of the date and time range of the prediction model based on a reaction of the user.