JP2005250558A - Chained record search device - Google Patents

Abstract

The present invention provides a chained record search apparatus capable of taking out a record similar to user input key information on a global basis and taking out an unexpected record.
Means 101 for inputting key information, parameters, new recording, etc., and outputting the extracted recording, means 104 for calculating similarity between key information and recording, and similarity between recordings, based on similarity And means 105 for calculating a score indicating whether or not the record should be taken out, a database 103 for storing the record and the score, a means 102 for controlling the whole, and the like. In the score calculation means, the score of the record similar to the key information is increased as viewed globally, and in the record where the score exceeds the target average value, the average value of the scores of similar records in the vicinity To be close to.
[Selection] Figure 1

Description

  The present invention relates to a technology for supporting management and retrieval of information in a personal information terminal and the like, and in particular, a chain for retrieving related records according to user input from user experiences and events stored on a computer. The present invention relates to a type record search apparatus.

  With the development of information devices such as personal computers and ubiquitous computing that enables them to be used anytime and anywhere, individuals can easily record large amounts of information. These records are not only different in form, such as text, images, audio, etc., but also have various generation purposes, so the method of classifying them according to the content and purpose and managing them will reset the categories Must be performed frequently, which is a burdensome task for the user. Further, when a desired record is taken out, a method for taking out only a record that matches a keyword or the like needs to specify a new keyword each time, and is inefficient for a large amount of various records.

  Conventionally, for example, the following method has been proposed as a technique for dealing with the problem of management and retrieval of such a record. The first method is to manage the state of the computer desktop as it is in time order. Records that match the specified keyword and specified records (hereinafter referred to as key information) exist / create / disappear. By searching the desktop at the time, the work and work records performed at that time can be taken out as records related to the key information (see, for example, Non-Patent Document 1). The second method considers the similarity between records based on the appearance frequency of words included in each record in addition to the record management in time order. Records related to information can be taken out (see, for example, Non-Patent Document 2). In the second method, by using the extracted record as new key information, a record similar to the extracted record can be extracted as a related record.

  In the above prior art, by setting a measure to measure the closeness between records, such as the time when the record was generated and the frequency of words included in the record, it is related to the key information without assuming the category classification of the record. Record to be taken out.

Jun Rekimoto, Time-Machine Computing: A Time-centric Approach for the Information Environment, ACM UIST '99, 1999 Toshiyuki Masui Information retrieval using neighborhood relations, IPSJ Human Interface Research Report 2003-HI-104, pp. 53-58, 2003

  When a user retrieves past records, he / she may find similar experiences and events in the past with respect to problems and issues that he is currently facing, and use them to solve current problems. . At this time, records that match exactly the same conditions as the current problem may not exist in the past, so it is possible to find clues to solve the problem by sequentially referencing related records that satisfy some conditions. is there.

  In the above prior art, when a desired record is clear, it can be taken back to the point where the record exists, or a record existing at the same time as the desired record can be taken out as a related record. When there are a lot of time points to be processed or when there are few related records included in each time point, this work is burdensome for the user. In addition, there are various records depending on the purpose and activity at each time point, and it may not be said that there is a useful relationship between these records just because they existed at the same time point. It is not effective to pull back related records back in time.

  In order to solve such a problem, the applicant must not only record records directly related to key information but also indirectly related records based on the similarity between records and key information and between records. A technique has been proposed that enables retrieval as a related record (Japanese Patent Application No. 2003-296851 “Recollection type information storage device”). In this case, each record is taken out in a chain according to the history of taking it out in the past. Each record is given a degree as to whether or not it should be taken out, and this changes with time depending on the current value and the similarity between the records. This value becomes larger as the number of similar records increases, and as the records extracted in the past, so it is easy to extract the next time. On the contrary, there is no need for records with few similar records or records that were not extracted in the past. It is considered as a record and it becomes difficult to retrieve.

  By the way, when a user takes out past records, there are cases where a desired record is not clear or an unexpected record is desired. If there is such a user request, it is possible that all the records taken out as the records related to the key information may become similar and uniform in the previously proposed technique, so another key information is newly set. It is necessary to take out related records, and this work is burdensome for the user when there are various and large numbers of records. In addition, even when the extracted records are chained out as new key information, a record similar to the initially set key information becomes new key information, so a similar record is extracted again. It takes a lot of time to arrive at an unexpected record.

  It is an object of the present invention to determine not only records similar to key information but also related records by determining whether or not each record should be taken out based on the similarity between key information and records and between records. It is another object of the present invention to provide a chained record search device that can be chained out.

  It is another object of the present invention to provide a chained record search apparatus that has a history that makes it easier for a record taken out in the past to be taken out at the next time.

  Also, an object of the present invention is to provide a chained record search device that considers records that have no other similar records and records that are not input with similar key information as unnecessary records and are not extracted. It is to provide.

  It is another object of the present invention to provide a chained record search device that makes it easy to pull out records similar to key information globally by adjusting the average value of whether or not all records should be taken out. is there.

  In addition, the object of the present invention is to locally adjust the degree of the recording with a large degree of whether or not to be taken out based on the average value of the degree only in the recording similar to the recording, It is an object of the present invention to provide a chained record search apparatus that can take out only a record that is very similar to key information and can take out a surprising record.

  The chain-type record search apparatus of the present invention inputs key information for extracting related records from records on a record database, inputs parameters relating to how to extract related records, inputs for adding / deleting new records, and Input / output means for outputting the extracted records, similarity calculation means for calculating the similarity between the key information and the records, and the similarity between the records, based on the similarity between the key information and the records, It includes score calculation means for calculating a score, which is a value indicating whether or not to take a record, a record and a record database for storing the score, and a control means for controlling the operation of each part.

  In the present invention, it is assumed that attribute information (hereinafter referred to as meta information) for explaining itself is given to the key information and the record. The similarity calculation means calculates the similarity between the key information and the records and between the records using the meta information. The score calculation means calculates a score, which is a value indicating the degree of whether or not to take out the record, based on the similarity calculated by the similarity calculation means. Each score in the record database is given this score. The control means takes out a record that is stochastically related according to the score of each record in the record database and presents it to the user.

  The score of each record increases with time as the record is similar to the key information, and as there are other records that are similar to the record and have a large score. As a result, not only the records similar to the key information but also the records related to them can be taken out in a chain.

  Further, since the score is calculated so that the score at the next time point increases as the current score increases, the record drawn in the past has a history characteristic that is easy to draw at the next time point.

  In addition, the score is monotonously decreased with time separately from the relationship with the key information and other records so as to reflect the freshness and expiration date of each record. As a result, the score of a record that has no other similar record or a record in which similar key information has not been input is reduced, and is regarded as an unnecessary record and cannot be taken out.

  Further, in order to avoid that the scores of all the records increase in a chained manner, the increase in the score is suppressed in proportion to the product of the scores of other similar records and the similarity. By adjusting the strength of this suppression action so that the average value of the scores of all records (hereinafter referred to as the average value of all scores) approaches the preset target average value, the score of the records similar to the key information can be increased. It becomes large and it becomes easy to pull out the record.

  Further, for a record whose score exceeds the target average value, the score is locally adjusted so as to be close to the average value of scores in a record similar to the record (hereinafter referred to as a local score average value). As a result, the score of a record whose score has become too large is reduced and the score of a record having a small score is increased. Therefore, it is possible to avoid taking out only records that are very similar to key information and take out surprising records. It becomes possible.

  According to the present invention, not only records similar to the key information but also related records can be taken out in a chain, and the past taken out record has a history that can be easily taken out at the next time point. A record that does not have a similar record or a record that does not receive similar key information is regarded as an unnecessary record and should not be taken out. A record that is similar to the key information can be easily extracted globally. Thus, there is an effect that it is possible to take out a surprising record while avoiding taking out only a very similar record.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
FIG. 1 shows an overall configuration diagram of an embodiment of a chained record search apparatus according to the present invention. In FIG. 1, 101 is an input / output unit, 102 is a control unit, 103 is a recording database, 104 is a similarity calculation unit, and 105 is a score calculation unit. Actually, the chain type record search apparatus is constituted by a computer system including an input / output device, a central processing unit (CPU), an internal storage device (RAM, ROM, etc.), an external storage device (hard disk, etc.) and the like.

  The input / output unit 101 inputs key information for extracting related records from records on the recording database 103, inputs parameters related to how to extract related records, inputs records to be added / deleted, and records of the extracted records Display and so on. The control unit 102 performs recording input / output operations and other control of the entire apparatus. The recording database 103 stores a record of user experiences and events, meta information thereof, a score calculated by the score calculation unit 105, and a similarity between the records calculated by the similarity calculation unit 104. The recording database 103 also temporarily stores the similarity between key information and recording. The similarity calculation unit 104 calculates the similarity between the key information and the record, and the similarity between the records. The score calculation unit 105 calculates the score of each record from the similarity between the record and the key information and between the records, and changes it with time.

Next, the operation of the entire chained record search apparatus will be described.
FIG. 2 is a flowchart showing an outline of processing when a new record is added to the recording database 103. The user uses the input / output unit 101 to input a record to be newly registered, its meta information, and an initial score value (step 201). The control unit 102 registers the input record, meta information, and initial score value in the record database 103 (step 202). Next, the control unit 102 retrieves all existing records and meta information from the recording database 103 and outputs them together with the newly registered record and its meta information to the similarity calculation unit 104. The degree of similarity between the newly recorded record and the newly added record is calculated, and the control unit 102 registers the result in the record database 103 (step 203). Various methods for calculating the degree of similarity have been proposed and are not a feature of the present invention, and therefore the details thereof are omitted. Next, the control unit 102 outputs the scores of all the records and the similarity between the records from the recording database 103 to the score calculation unit 105, and the score calculation unit 105 calculates the score of each record, and the score calculation result Is registered in the recording database 103 (step 204). Here, it is assumed that the parameters relating to how to take out the records used for score calculation are already input by the user using the input / output unit 101 and set in the score calculation unit 105. The processing of the score calculation unit 105 will be described in detail later.

  FIG. 3 is a flowchart showing an outline of processing when key information is inputted and a record related to the key information is taken out from the record database 103. The user inputs key information using the input / output unit 101 (step 301). The control unit 102 extracts all records and their meta information from the recording database 103, outputs them together with the input key information to the similarity calculation unit 104, and calculates the similarity between the record and the key information by the similarity calculation unit 104. The result is registered in the recording database 103 by the control unit 102 (step 302). Next, the control unit 102 outputs the scores of all the records and the similarity between the records and the similarity between the key information and the records to the score calculation unit 105 from the recording database 103, and the score calculation unit 105 determines the score of each record. Then, the control unit 102 registers the score calculation result in the recording database 103 (step 303). Also in this case, it is assumed that a parameter relating to how to take out a record has already been input by the user using the input / output unit 101, and this parameter has been set in the score calculation unit 105. Next, for each record in the record database 103, the control unit 102 probabilistically determines whether the record is related to the key information according to the score, and outputs the determined record to the input / output unit 101. As a result, the input / output unit 101 presents it to the user (step 304). In determining the record related to the key information, a random number is generated for each determination for each record, and if the value is smaller than the score of the record, the record is determined as the associated record. Finally, the control unit 102 clears the input key information and records the key information and the recording from the recording database 103 so that the score calculation is not affected in the next input key information. Is deleted (step 305).

Next, the processing of the score calculation unit 105 will be described in detail. The score calculation unit 105 calculates the score x _i of the record i at a certain time t according to the following equation.

ここで、ｍ_ｊｉは記録ｉから見た記録ｊの類似度、ｍ_ｉｊは記録ｊから見た記録ｉの類似度であり、また、ｍ_ｉは記録ｉから見たキー情報の類似度であり、これらはいずれも０から１の値になるように正規化されている。ｍ_ｊｉ及びｍ_ｉｊは、新たに記録を登録する際に一度計算されると、それ以降は記録データベース１０３に保持されるものである（図２参照）。一方、ｍ_ｉはキー情報が入力された際に計算されると、記録データベース１０３に保持されるが、不要になると除去されるものである（図３参照）。α，β，γ_ｉは、記録の取り出し方に関するパラメータ（スコアの変化速度のパラメータ）であり、その初期値は、予めユーザが入出力部１０１を用いて入力し、制御部１０２によりスコア計算部１０６に保持される。なお、γ_ｉは、その記録ｉと対応づけて記録データベース１０３に保持することでもよい。

Here, m _ji similarity of recording j as viewed from the recording i, m _ij is the similarity of the recording i viewed from the recording j, also, I m _i is an similarity key information as seen from the recording i These are normalized so as to have a value of 0 to 1. When m _ji and m _ij are calculated once when newly registering a record, they are held in the record database 103 thereafter (see FIG. 2). On the other hand, _mi is stored in the recording database 103 when it is calculated when key information is input, but is removed when it becomes unnecessary (see FIG. 3). α, β, and γ _i are parameters (score change speed parameters) relating to how to take out records, and the initial values are input by the user using the input / output unit 101 in advance, and the control unit 102 calculates the score calculation unit. 106. Note that γ _i may be stored in the recording database 103 in association with the recording i.

The sum of the first term and the second term on the right side in the above equation (1) is calculated only for the record j that is similar to the record i. When certain key information is input through the input / output unit 101, the record score similar to the key information is increased by the third term on the right side in the equation (1), but more similar to the record similar to the key information. Since the score of the record also increases in a chain by the first term on the right side, these records can be easily extracted as records related to the key information. The second term on the right-hand side serves to suppress the score between similar records so that the scores of all the records do not become too large in a chain. β (> 0) is a parameter representing the weight of the suppression action, and by adjusting this value, the score of the recording similar to the key information can be increased globally. This global score adjustment will be described later. α (> 0) is a parameter representing the weight of score calculation based on the similarity between recordings and between key information and recordings. The fourth term on the right side serves to decrease the score with time. As a result, the score of a record that has no other similar record or a record in which similar key information has not been input decreases with time, and is regarded as an unnecessary record and cannot be taken out. γ _i is a parameter representing the weight of this term, and is set by the expiration date, freshness, etc. for each recording. The larger this value, the faster the recording score decreases. In addition, γ _i is adjusted so as to decrease the score with respect to a record having a high score. As a result, the score of a record whose score has become too large is reduced, and the record with a small score is less affected by the global score adjustment. Therefore, it is possible to take out an unexpected record. This local score adjustment will be described later.

  In the calculation of the above formula (1), as the current score is larger, the score is likely to be larger at the next time point. Therefore, when the similar key information is input, the past-drawn record is easily retrieved at the next time point. Has a history. At the same time, it has one aspect that can draw out unexpected records.

  Next, global adjustment of the score will be described. In order to increase the score of the recording similar to the key information globally, the value of β is controlled so that the average value of all recording scores (average value of all scores) approaches a preset target average value. Good. Here, it is assumed that the method described in the document “Kantaris, N., Howden, PF, The Universal Equation Solver by BASIC, SIGMA PRESS, 1983” is used.

  FIG. 4 is a flowchart showing an overview of the process of global adjustment of scores. The value of β at time t + 1 is calculated according to the following equation.

  Update β when the average value of all scores is smaller than the target average value and decreases with time at time t, and when the average value of all scores is larger than the target average value and increases with time To do. In these cases, if the difference between the average value of all scores and the target average value at the times t-1 and t is the same sign, p in the equation (3) is increased by 1, and the average value of all scores is quickly determined as the target average. The value is brought close to the value (step 401). On the other hand, if the signs are different, r in equation (3) is increased by 1, and the average value of all scores is prevented from oscillating near the target average value, and gradually approaches the target average value (step 402). . Next, β is updated according to equation (3) (step 403), and when β> 0 is satisfied, the difference between the average value of all scores and the target average value at time t−1 is adjusted for subsequent score adjustment. The sign is deleted, the sign of the difference between the average value of all scores and the target average value at time t is stored (step 404), and the process ends. On the other hand, if β> 0 is not satisfied, the update of p, r, β in step 401 or step 402 and step 403 is invalidated (step 405), and the process is terminated.

Next, local adjustment of the score will be described. In order to maintain diversity, the score is locally increased even in a record other than the record similar to the key information. This means that γ _i of the record _i of interest (record whose score exceeds the target average value) is made closer to the average value (local score average value) of the score of the record similar to the record. Control is sufficient. Again, suppose that the method published in the previous literature is used.

FIG. 5 is a flowchart showing an overview of the process of local adjustment of the score. Only for the record i whose score exceeds the target average value at time t, the value of γ _i at time t + 1 is calculated according to the following equation.

H in the equation (7) is defined by the above equation (4). At time t, when the score of record i is smaller than the local score average value of the record and decreases with time, and when the score is larger than the local score average value and increases with time Update γ _i . In these cases, if the difference between the score and the local score average value at time t-1 and t have the same reference numerals, close to p _i is incremented 1, quickly local score mean scores in (7) (Step 501). Conversely, if a different code, (7) the r _i is incremented by 1 in the equation, the score is to approach gradually the local score average prevents oscillates about local score average (step 502) . Next, γ _i is updated according to equation (7) (step 503), and γ _i is larger than a value initially input by the user (hereinafter referred to as γ _init ) and satisfies α> γ _i , that is, When α> γ _i > γ _init is satisfied, the sign of the difference between the score at time t-1 and the local score average value is deleted for subsequent score adjustment i, and the score and local score average value at time t The sign of the difference is stored (step 504), and the process ends. On the other hand, if α> γ _i > γ _init is not satisfied, the update of p _i , r _i , and γ _i in step 501 or 502 and 503 is invalidated (step 505), and the process ends.

Thus, by always satisfying α> γ _i , if very similar key information is input even for a record with no similar record (m _i ≈1), the score increases. To. Further, by always satisfying γ _i > γ _init , it is possible to prevent the score from decreasing faster than the speed specified by the user even when there is no similar record or when similar key information is not input. .

  Next, an example of processing of the score calculation unit 105 will be described. FIG. 6 shows the change over time in the average value of all scores when key information having attribute value 50 is input to 99 records having attribute values 1,..., 99 as meta information. FIG. 7 shows the score of each record after 200 unit time. However, the similarity between the records and between the key information and the records is assumed to be more similar as the attribute value is closer, and the target average value is set to 0.6.

  As shown in FIG. 6, the overall average score value approaches the target average value by adjusting the global score. Further, as shown in FIG. 7, the overall score adjustment increases the score of the recording near the key information, and the local score adjustment increases the score even when the recording is not near the key information. It is getting bigger. That is, a record similar to the key information can be easily extracted on a global basis, and it is also possible to take out a surprising record by avoiding taking out only a record very similar to the key information. .

  The chain type record search apparatus of the present invention has been specifically described based on the embodiments. However, the present invention is not limited to the above embodiments, and various modifications can be made without departing from the scope of the claims. Needless to say, you get.

  It should be noted that the processing functions of some or all of the components in the chained record search apparatus shown in FIG. 1 can be configured by a computer program and the program can be executed using the computer to implement the present invention, or Needless to say, the processing procedures shown in FIGS. 2 to 5 can be configured by a computer program, and the program can be executed by the computer. In addition, a computer-readable recording medium such as an FD, MO, ROM, memory card, CD, or the like is stored in the computer. In addition, the program can be recorded and stored on a DVD, a removable disk, etc., and the program can be distributed through a network such as the Internet.

It is a figure which shows the whole structure of one Example of the chain type | mold recording search apparatus of this invention. It is a flowchart which shows the outline | summary of the operation | movement at the time of the new recording input of one Example of this invention. It is a flowchart which shows the outline | summary of the operation | movement at the time of the key information input of one Example of this invention. It is a flowchart which shows the outline | summary of the global adjustment operation | movement of the score in the score calculation part of this invention. It is a flowchart which shows the outline | summary of the local adjustment operation | movement of the score in the score calculation part of this invention. It is a figure which shows an example of a process of the score calculation part of this invention. It is a figure which similarly shows an example of the process of the score calculation part of this invention.

Explanation of symbols

101 Input / Output Unit 102 Control Unit 103 Recording Database 104 Similarity Calculation Unit 105 Score Calculation Unit

Claims (5)

A chain-type record search device for taking out records related to user input information (hereinafter, key information) from records stored on a record database,
Input / output means for inputting key information, inputting parameters related to how to take out related records, inputting new records, and outputting extracted records;
Similarity calculation means for calculating similarity between key information and records, and similarity between records;
Score calculating means for calculating a value (hereinafter referred to as a score) indicating the degree of whether or not to take the record of each record based on the similarity between the key information and the record, and the similarity between records;
A record database for storing records and their scores,
A chained record search apparatus characterized in that, on the basis of a score assigned to the record, related records are chained out according to key information. In the chain type record search device according to claim 1,
The score calculation means calculates the score x _i of the record i at a certain time t,

m _ji is the similarity of record j viewed from record i m _ij is the similarity of record i viewed from record j m _i is the similarity of key information viewed from record i x _j is the score of record j α, β, γ _i is calculated according to an equation of a parameter relating to how to extract related records. In the chain type record search device according to claim 2,
The score calculation means calculates the sum of the first term and the second term on the right side in the equation (1) only for a record j having similarity to the record i. In the chain type record search device according to claim 3,
The score calculation means controls the parameter β in the equation (1) so that the average value of the scores of all records approaches a preset target average value. In the chain type record search device according to claim 4,
The score calculation means calculates the parameter γ _i in the equation (1) for the record i whose score exceeds the target average value, and the average value of the scores in the record in which the score of the record i is similar to the record i A chained record search apparatus, characterized in that control is performed so as to approach.

Images