JP5458640B2 - Rule processing method and apparatus - Google Patents

Description

  The present technology relates to a processing technology for rules obtained in advance by learning.

  For example, when a proper noun such as a person's name or place is extracted from a word string, conventionally, the following processing is necessary. Here, in order to simplify the explanation, it is assumed that the name of the person (denoted as “person”) and the other name (denoted as “O”) are distinguished. That is, a correct combination of a word and a proper noun type as shown in FIG. 1A is prepared, and a combination of a proper noun type and a feature set as shown in FIG. 1B is generated therefrom. The feature represents a clue for discrimination. In the feature, if the attribute type is the word itself, it is represented by W, and (0), (1), etc. after the symbol represent the relative appearance position based on the position (0) of the word of interest. For example, when focusing on the first word “Miyazaki” in FIG. 1A, it can be expressed as W (0) = Miyazaki. Further, as a clue for discrimination, the word at the current position and the previous and subsequent words are used. Therefore, assuming that the first word is the current position, the word “origin” moved in the positive direction is expressed as W (1) = origin. Further, when the current position is shifted to the second word, the word “from” at the current position is represented as W (0) = origin, and the word “Miyazaki” immediately before the current position is represented by W ( -1) = Miyazaki is represented, and the word “san” after the current position is represented by W (1) = san. Similarly, data as shown in FIG. 1B is generated from a word string as shown in FIG. 1A and its proper noun type. In this application, it is generalized and expressed as W (p) = wi. wi is the word at the i-th position. p represents the relative position of appearance from the position of interest. Although simplified here, other attributes such as a character type (for example, kanji or hiragana) or a part of speech may be used as the attribute type. In some cases, multiple attribute types may be combined.

  Furthermore, a known method (Iwakura, Tomoya and Okamoto, Seishi, “A Fast Boosting-based Learner for Feature-Rich Tagging and Chunking”, Proc. Of CoNLL 2008, pages : See {17-24}.) When machine learning is performed, rules as shown in FIG. 2 are generated. Each line in FIG. 2 is a rule, and each rule includes a condition, a proper noun type, and a score. Conditions are represented by features. In the example of FIG. 2, one condition includes one feature, but a combination of a plurality of features may be included. The higher the score, the higher the possibility that it is a corresponding proper noun type. The condition “W (0) = Miyazaki”, proper noun type “people”, and score “10” means that if W (0) = Miyazaki is satisfied, the word “Miyazaki” at the current position is “people”. Means “10”.

  The process for specifying a person name from the word strings “Miyazaki” “san” “to” “play” as shown in FIG. 3 using the rules as shown in FIG. 2 is as shown in FIG. 4 and FIG. It is. First, when applying the word “Miyazaki” at the current position to the rule, first, from the word “Miyazaki” at the current position and the next word “Ms.”, feature sets “W (0) = Miyazaki” and “W (1 ) = San ”is generated (step (1)). This is because the feature is generated from the word at the current position and the words before and after the current word. The rule condition is searched with the generated feature (step (2)), and the proper noun type and score corresponding to the matching condition are extracted. Since there is a rule with a condition that matches the feature W (0) = Miyazaki, the proper noun type “people” and the score “10” are obtained. Similarly, since there is a rule of a condition that matches the feature W (1) =, the proper noun type “person” and the score “30” are obtained. In summary, for the word “Miyazaki” at the current position, the proper noun type “people” becomes a score “40 (= 10 + 30)” and is registered (step (3)).

  Next, the word at the current position is shifted from “Miyazaki” to “san”, and from the words before and after the word “san”, the feature sets “W (−1) = Miyazaki” “W (0) = san” “ W (1) = to "is generated (step (4)). Then, the rule condition is searched with the generated feature (step (5)), and the proper noun type and score corresponding to the matching condition are extracted. Since there is a rule with a condition that matches the feature W (−1) = Miyazaki, the proper noun type “O” and the score “5” are obtained. Similarly, since there is a rule of a condition that matches the feature W (0) =, the proper noun type “O” and the score “20” are obtained. In summary, the proper noun type “O” becomes the score “25 (= 5 + 20)” and is registered for the word “san” at the current position (step (6)).

  As can be seen from FIGS. 4 and 5, words such as “Miyazaki” and “Mr.” appear as different objects at different positions such as relative positions “0”, “−1”, “1” in the generated features, You must match the rules each time. As a result, the number of searches increases, which causes a problem that the processing speed becomes slow.

Iwakura, Tomoya and Okamoto, Seishi, "A Fast Boosting-based Learner for Feature-Rich Tagging and Chunking", Proc. Of CoNLL 2008, pages: {17-24}

  As described above, in the prior art, since the same word is applied to the rule as different features at different positions many times, the number of searches for the rule increases and the processing speed is slowed down. In the above example, only the word itself is used as the attribute type of the word. However, the rule condition is defined by using other parts of speech or character types as the attribute type of the word, or by combining multiple features. If this is done, the number of searches will increase.

  Accordingly, an object of the present technology is to provide a technology for increasing the processing speed when applying a rule to determine a word type.

  This rule processing method includes a score, a relative appearance position and a type to which the score should be assigned, in association with each condition or combination of conditions represented by a combination of the attribute type, appearance relative position, and attribute value of the word. Access to a rule data storage unit in which one or a plurality of score setting rules are registered, and an element column data storage unit that stores an element column in which elements including attribute values of each attribute type of words are arranged in the order of appearance Executed by a computer. Then, this rule processing method uses the reference position of the appearance relative position included in the score setting rule associated with the maximum number of conditions included in the combination of conditions and the condition or combination of conditions from the element string data storage unit. In accordance with the candidate extraction condition specified by the maximum distance from the application candidate, an application candidate that is an element condition or a combination of element conditions represented by a combination of the attribute type of word, relative position of appearance, and attribute value is extracted for each element. When the rule data storage unit is searched for the extraction step and the application candidate and there is a corresponding condition or combination of conditions, the rule data storage unit is associated with the corresponding condition or combination of conditions 1 or A plurality of score setting rules are extracted, and the above-mentioned types and scores included in the extracted one or more score setting rules are extracted. Or the appearance relative position included in the plurality of scores set rules score for the types of elements identified from the appearance position of the applied candidates updated, and storing the element sequence data storage unit.

  The processing speed when applying the rule to determine the word type is increased.

FIG. 1 is a diagram for explaining the prior art. FIG. 2 is a diagram showing rule data of the prior art. FIG. 3 is a diagram illustrating an example of discrimination target data. FIG. 4 is a schematic diagram for explaining the processing of the prior art. FIG. 5 is a schematic diagram for explaining the processing of the prior art. FIG. 6 is a diagram illustrating rule data conversion according to the embodiment of the present technology. FIG. 7 is a schematic diagram showing application of rules in the embodiment. FIG. 8 is a schematic diagram showing application of rules in the embodiment. FIG. 9 is a functional block diagram of the rule processing device. FIG. 10 is a diagram showing a main processing flow in the embodiment. FIG. 11 is a diagram illustrating a processing flow of rule conversion processing. FIG. 12 is a diagram illustrating data representing the rule conversion process. FIG. 13 is a diagram illustrating data representing the rule conversion process. FIG. 14 is a diagram illustrating an example of generalized discrimination target data. FIG. 15 is a diagram illustrating a processing flow of the discrimination processing. FIG. 16 is a diagram illustrating an example of data stored in the determination result storage unit. FIG. 17 is a diagram illustrating a processing flow of the second rule conversion processing. FIG. 18 is a schematic diagram for explaining the contents of the second rule conversion process. FIG. 19 is a diagram illustrating a processing flow of the second rule conversion processing. FIG. 20 is a diagram illustrating a processing flow of the second determination processing. FIG. 21 is a diagram illustrating an example of discrimination target data. FIG. 22 is a diagram illustrating an example of check candidates. FIG. 23 is a diagram illustrating a rule application example in the second determination process. FIG. 24 is a diagram illustrating a processing flow of the second determination processing. FIG. 25 is a functional block diagram of a computer.

[Outline of this embodiment]
In the present embodiment, the rule as shown in FIG. 2 is converted into a new format rule as shown in FIG. Specifically, the new rule includes a condition, a word position to which a score is assigned, a proper noun type, and a score. However, one or more score setting rules including the position of the word to which the score is assigned, the proper noun type, and the score are associated with one condition. Therefore, when one condition is specified when applying the rule, the score setting rule can be extracted at a time.

  Then, as shown in FIG. 7, when processing a word string as shown in FIG. 3, the word “Miyazaki” at the current position is searched for the new rule shown in FIG. Confirm (step (10)). In the new rule of FIG. 6, the first rule is specified, and two score setting rules are extracted at once. Since the first score setting rule includes the position “0” of the word to which the score is assigned, the proper noun type “person”, and the score “10”, the word at the current position from the position “0” of the word to which the score is assigned. For “Miyazaki”, a score “10” is set for the proper noun type “people” (step (11)). Furthermore, since the second score setting rule includes the position “1” of the word to which the score is assigned, the proper noun type “O”, and the score “5”, the current position is determined from the position “1” of the word to which the score is assigned. A score “5” is set for the proper noun type “O” for the word “san” at the next position (step (12)).

  Also, as shown in FIG. 8, the current position is moved to the next word, and the new rule shown in FIG. 6 is searched for the word “Ms.” at the current position to check whether a matching condition exists (step ( 13)). In the new rule of FIG. 6, the second rule is specified, and two score setting rules are extracted at once. Since the first score setting rule includes the position “0” of the word to which the score is given, the proper noun type “O”, and the score “20”, the word at the current position from the position “0” of the word to which the score is given. For “san”, the score “20” is added to the current value “5” for the proper noun type “O”, and “25” is registered (step (14)). Similarly, since the second score setting rule includes the position “−1” of the word to which the score is assigned, the proper noun type “person”, and the score “30”, the position “−1” of the word to which the score is assigned. Is added to the current value “10” for the proper noun type “person”, and “40” is registered (step (15)).

  In this way, if a new rule is searched once for one word to be determined, the score setting rule to be applied can be extracted, the number of searches is reduced, and the processing speed is increased.

[Specific contents of this embodiment]
FIG. 9 shows a functional block diagram of the rule processing device. The rule processing device uses a learning data input unit 1, a learning data storage unit 3 that stores learning data input by the learning data input unit 1, and a rule learning process that uses data stored in the learning data storage unit 3. The rule learning unit 5 that implements the processing, the first rule data storage unit 7 that stores the processing result of the rule learning unit 5, and the first rule stored in the first rule data storage unit 7 as the second rule Input by the rule conversion unit 9 for conversion, the second rule data storage unit 11 for storing the data of the second rule generated by the rule conversion unit 9, the discrimination target data input unit 13, and the discrimination target data input unit 13 A determination target data storage unit 15 for storing the determined determination target data, a determination unit 17 for processing using the data stored in the second rule data storage unit 11 and the determination target data storage unit 15, Processing A determination result storing section 19 for storing, and an output unit 21 for outputting the data stored in the determination result storage unit 19.

  Next, processing contents of the rule processing device will be described with reference to FIGS. First, the learning data input unit 1 of the rule processing device accepts input of learning data from the user and stores it in the learning data storage unit 3 (FIG. 10: step S1). For example, data as shown in FIG. 1A, that is, a plurality of pairs of words and correct proper noun types are included. Then, the rule learning unit 5 generates a first rule by performing a known learning process on the learning data stored in the learning data storage unit 3 and stores the first rule in the first rule data storage unit 7 ( Step S3). For example, a data table as shown in FIG. 1A to FIG. 1B and FIG. 2 is generated. This process itself is well known and will not be described further.

  Next, the rule conversion unit 9 performs a rule conversion process (step S5). This rule conversion process will be described with reference to FIGS. However, at first, it is assumed that the word attribute type is only the word itself. The rule conversion unit 9 initializes r to 1 (step S21). Then, the condition of the r-th rule in the first rule table (FIG. 2) stored in the first rule data storage unit 7 is separated into feature appearance position information (= appearance relative position) p and value f. (Step S23). When the first line in FIG. 2 is processed, since the condition is W (0) = Miyazaki, p = 0 and f = Miyazaki are obtained.

  After that, the rule conversion unit 9 sets the second rule table in the second rule data storage unit 11 to the r-th rule in the first rule table with the value f as a condition and “−p” as the position of the word to which the score is assigned. The rule score and proper noun type are registered as they are (step S25). If it is the first row in FIG. 2, data as shown in FIG. 12 is registered in the second rule table in the second rule data storage unit 11. That is, the condition “Miyazaki” and the score setting rule including the position “0” of the word to which the score is assigned, the proper noun type “people”, and the score “10” are registered.

  Then, the rule conversion unit 9 increments r by 1, and determines whether r is less than m (m is the number of records in the first rule table of the first rule data storage unit 7) (step S29). If r is less than or equal to m, the process returns to step S23.

  When r = 2, since the condition is W (−1) = Miyazaki, p = −1 and f = Miyazaki are obtained. Then, as indicated by a bold line in FIG. 13, the score “5” and the proper noun type “O” of the r-th rule are used as they are, with f = Miyazaki as the condition, and −p = 1 as the position of the word to which the score is assigned. , Registered in the second rule table.

  On the other hand, if r exceeds m, the rule conversion unit 9 collects the rules by collecting the score setting rules for the same condition (step S30). Score setting rules associated with the same condition are collected and converted into a form in which the score setting rules collected in association with one condition are registered. This makes it possible to easily extract all score setting rules with a single search. Thereafter, the process returns to the original process.

  When such processing is performed, the conversion as shown in FIG. 6 is completed, and the second rule table is stored in the second rule data storage unit 11.

  Returning to the description of the processing in FIG. 10, separately from steps S <b> 1 to S <b> 3 or after step S <b> 5, the discrimination target data input unit 13 accepts input of discrimination target data from the user and stores it in the discrimination target data storage unit 15. (Step S7). The discrimination target data is, for example, data as shown in FIG. In such a simple example, the attribute type of the word is the word itself. However, the attribute type of the word is not necessarily limited to one type. For example, as shown in FIG. 14, the word itself, the part of speech, and the character type may be adopted as the attribute type. As described above, a combination of attribute values of all attribute types used for one word is referred to as an element w. If the input position of the word is “i” th input, it is expressed as wi.

  Then, the determination unit 17 performs a determination process using the data stored in the second rule data storage unit 11 and the determination target data storage unit 15 (step S9). This determination process will be described with reference to FIGS. 15 and 16. However, first, the process when there is only one type of condition will be described below. Specifically, a process when the second rule table on the right side of FIG. 6 is prepared will be described.

  First, the determination unit 17 initializes i to 1 (step S31). Then, one unprocessed attribute type of the element wi in the discrimination target data storage unit 15 is specified (step S33). As described above, the element wi includes attribute values of one or more attribute types of the i-th word in the discrimination target data storage unit 15. The example of FIG. 3 is a very simple example in which the attribute type is only the word itself, and the part of speech or character type described above may be combined.

  Then, the determination unit 17 determines whether the attribute value of the specified attribute type of the element w i satisfies any rule condition (step S35). In the example of FIG. 7, it is determined whether there is a rule having the attribute type of the word itself and the condition of the word “Miyazaki”, and two corresponding score setting rules are obtained. If the attribute value of the specified attribute type of the element wi does not match any rule condition, the process proceeds to step S39. On the other hand, when the attribute value of the specified attribute type of the element wi satisfies any rule condition, the score setting rule in the corresponding rule (that is, the position of the word to which the score is given (appearance relative position)) And a score is set for the element wi according to the type and score (step S37). In the example of FIG. 7, as in steps (11) and (12), a score “10” is set for the proper noun type “person” for the element of the word “Miyazaki” and a proper noun for the element of the word “san”. A score “5” is set for the type “O”. If a score has already been registered, the score value included in the current score setting rule is added. The process proceeds to step S39.

  In step S39, the determination unit 17 determines whether or not an unprocessed attribute exists for the element wi. If there is an unprocessed attribute type for the element wi, the process returns to step S33. On the other hand, if there is no unprocessed attribute type, i is incremented by 1 (step S41), and it is determined whether i is n (n is the number of elements) or less (step S43). If i is n or less, the process returns to step S33. On the other hand, if i exceeds n, for each element wi in the discrimination target data storage unit 15, the type with the highest score (in the example of FIG. 7, a person name or other) is specified and stored in the discrimination result storage unit 19. (Step S45). For example, as shown in FIG. 16, a discrimination result is registered in the discrimination result storage unit 19 for each word. That is, only “Miyazaki” is determined as “person”.

  As described above, in a simple example where the condition is only for one attribute type and there is no combination, if there are a plurality of attribute types, a rule is searched for the number of attribute types for one element. However, since one element is processed only once, the number of searches is reduced. If there is only one attribute type, one search is sufficient for one element.

  Returning to the description of the processing in FIG. 10, the output unit 21 outputs the determination result stored in the determination result storage unit 19 to an output device such as a display device or a printing device (step S <b> 21). When the rule processing device is connected to a network, it may be transmitted to another computer connected to the network.

  By performing the processing as described above, the most probable type is identified for each word of the discrimination target data while reducing the number of searches. In the example described above, whether or not the name is a personal name is specified, but other types can be discriminated. For example, part of speech can be determined. Furthermore, it is possible to determine noun phrases and verb phrases (basic phrase determination), phrase estimation, and the like.

  In the above, a simple example in which the rule condition is only for one attribute type and there is no combination has been described. However, in general, this is not the only simple example. Next, more generalized processing will be described. That is, the case where the condition is expressed by a combination of a plurality of features will be described with reference to FIGS. The rule converter 9 initializes r indicating the current position to 1 (step S51). Then, the minimum value MIN of the appearance relative position is initialized to “undefined” (step S53). Thereafter, one unprocessed feature is specified from the r-th rule in the first rule data storage unit 7 (step S55). For example, as shown in FIG. 18A, when the condition of each rule is expressed by a combination of two features, when processing the first rule, W (0) = Miyazaki and W (1) = For example, W (0) = Miyazaki is identified.

  And the rule conversion part 9 specifies the appearance relative position p of a feature from the specified feature (step S57). In the example specified above, p = 0. Thereafter, it is determined whether MIN is undefined or p <MIN (step S59). If MIN is undefined or p <MIN, MIN = p is set to minimize MIN (step S61). Then, control goes to a step S63. On the other hand, if MIN ≦ p, there is no need to update MIN, and the process proceeds to step S63.

  In step S63, the rule conversion unit 9 determines whether or not an unprocessed feature exists in the r-th rule (step S63). If an unprocessed feature exists, the process returns to step S55. In the example of FIG. 18A, W (1) = is identified as unprocessed, but since p = 1 and MIN = 0, it is determined that MIN = 0. In the case of the second rule, MIN = −1 because W (−1) = Miyazaki and W (0) = san.

  On the other hand, when there is no unprocessed feature in the r-th rule, the processing shifts to the processing in FIG. Next, the rule conversion unit 9 identifies one unprocessed feature again from the r-th rule (step S65). Then, the identified feature is separated into attribute type t, feature appearance relative position p, and value f (step S67). For W (0) = Miyazaki, the attribute type is t = W (ie, word), p = 0, f = Miyazaki, and so on. Then, NI = p-MIN is set as a new appearance relative position, and the feature expressed by the attribute type t, the new appearance relative position NI, and the value f is added to the new condition nr (step S69). For W (0) = Miyazaki in the first rule, since NI = 0-0 = 0, the content of the feature is not changed. There is no change in the content of the identity of W (1) =. On the other hand, for W (−1) = Miyazaki in the second rule, NI = −1 − (− 1) = 0, and W (0) = Miyazaki is changed. For W (0) = in the second rule, NI = 0 − (− 1) = 1 and is changed to W (1) =. Thereafter, it is determined whether an unprocessed feature exists in the r-th rule (step S71). If there is an unprocessed feature, the process returns to step S65.

  On the other hand, when there is no unprocessed feature in the r-th rule, the rule conversion unit 9 adds the condition “nr” and the word to which the score is assigned to the second rule table in the second rule data storage unit 11. A score setting rule including the position “−MIN” and the type and score of the r-th rule is registered (step S73). Then, r is incremented by 1 (step S75), and it is determined whether r is less than m (m is the number of rules) (step S77). If r is less than or equal to m, the process returns to step S53 via the terminal B. On the other hand, if r exceeds m, rules having the same condition are collected in the second rule table (step S79). The first rule and the second rule in FIG. 18A have the same conditions as shown in FIG. 18B, and therefore, association is performed so that two score setting rules can be extracted from one condition. . In the example of FIG. 18A, the conditions are the same for the third and fourth rules, so that two score setting rules can be extracted from one condition as shown in FIG. 18B. Make an association with. Then, the process returns to the original process.

  By performing the processing as described above, it is possible to cope with a case where the number of features included in the condition is plural or when features are configured for a plurality of attribute types.

  Next, more general determination processing will be described with reference to FIGS. Note that the second rule table in FIG. 18B is used. First, the number of elements w is set to n (step S81). For example, it is assumed that data as shown in FIG. 21 is stored in the discrimination target data storage unit 15. Here, the part of speech P is included in addition to the word itself W as the attribute type of the word. Since there are four elements here, n = 4. Further, the maximum distance from the reference position of the relative appearance position of the word in the rule condition is set in WR, and the maximum number of features included in the rule condition is set in MAX (step S83). This is set by searching the second rule table in the second rule data storage unit 11. In the case of the second rule table as shown in FIG. 18B, since the maximum value of p in “W (p) = wi” is 1, WR = 1. However, WR = “2” is assumed for the explanation of the following processing. Since each rule includes a maximum of two features, MAX = 2 is set.

Then, the determination unit 17 sets i = 1 (step S85), generates a check candidate according to WR and MAX with reference to the element w i, and stores it in the AR (step S87). For each element up to the element w _{i + WR} at the appearance position separated from the element w _i by the maximum distance WR, for each attribute type (for example, W or P), the attribute type (W or P) and the attribute value f of the attribute type And the appearance relative position pj from the element w i are generated. Then, check candidates are generated by combining them so as to include at least one of the features about the element w i within MAX. In the example of FIG. 21, if the element w1 is a processing target, the elements to be considered are the element w1 of the word “Miyazaki”, the element w2 of the word “san”, and the element w3 of the word “to”. . For element w1, W (0) = Miyazaki (ie, pj = 0, f = Miyazaki), P (0) = noun (ie, pj = 0, f = noun), W (1) = san (ie, pj = 1, f =), P (1) = suffix (ie, pj = 1, f = suffix), W (2) = and (ie, pj = 2, f = to), P (2) = particle (ie, pj) = 2 and f = particle). And since MAX = 2, when features are combined so as to include at least one of W (0) = Miyazaki and P (0) = noun with one or two features, as shown in FIG. Check candidates are generated.

Then, the determination unit 17 specifies one unprocessed check candidate cr from the AR (step S89). Thereafter, by searching the second rule table in the second rule data storage unit 11 using the check candidate cr, it is determined whether there is a rule for which the check candidate cr is an application condition (step S91). If there is no rule for which the check candidate cr is an application condition, the process proceeds to the process in FIG. On the other hand, if there is a rule for which the check candidate cr is an application condition, the corresponding score in the discrimination target data storage unit 15 is updated based on the score setting rule for the corresponding rule (step S93). As schematically shown in FIG. 23, when the second rule table stored in the second rule data storage unit 11 is searched for each of the check candidates generated in step S87, the fourth check candidate is the first check candidate. It matches the conditions of the rule. Then, two score setting rules are extracted, and the score of the corresponding part of the discrimination target data storage unit 15 is updated for each. In the case of the first score setting rule, since the position of the word to which the score is assigned is “0”, the score of the type “person” for the element w i is increased by “20”. In the case of the second score setting rule, since the position of the word to which the score is assigned is “1”, the score of the type “O” for the element w _{i + 1 is} increased by “10”. If it does so, the table of the discrimination | determination object data storage part 15 will be in the upper left state of FIG. The processing shifts to the processing in FIG.

  Shifting to the description of the processing in FIG. 24, the determination unit 17 determines whether or not unprocessed check candidates remain in the AR (step S97). If unprocessed check candidates remain, the process returns to step S89 in FIG. On the other hand, if all check candidates have been processed in the AR, i is incremented by 1 (step S99), and it is determined whether i is n or less (n is the number of elements) (step S101). If i is n or less, the process returns to step S87 in FIG. On the other hand, if i exceeds n, the maximum score type is specified for each element in the discrimination target data storage unit 15 and stored in the discrimination result storage unit 19 (step S103). Then, the process returns to the original process.

  By carrying out such processing, it is possible to cope with generalization of elements and rules.

  Although the embodiment of the present technology has been described above, the present technology is not limited to this. For example, although the example has only “person name” or “other” as the type, other types may be used, and it is possible to deal with a case where scores are set for many types.

  Furthermore, as long as the same result is obtained for the processing flow, the processing order may be changed or the processing flow may be performed in parallel. In particular, the processing as in step S87 may be modified so that it is generated every time the second rule table is searched, instead of generating all check candidates at once.

  Further, the functional block diagram of the rule processing device is an example, and may not necessarily match the actual program module configuration.

  The rule processing device described above is a computer device, and as shown in FIG. 25, a memory 2501, a CPU 2503, a hard disk drive (HDD) 2505, a display control unit 2507 connected to the display device 2509, and a removable device. A drive device 2513 for the disk 2511, an input device 2515, and a communication control unit 2517 for connecting to a network are connected by a bus 2519. An operating system (OS) and an application program for executing the processing in this embodiment are stored in the HDD 2505, and are read from the HDD 2505 to the memory 2501 when executed by the CPU 2503. If necessary, the CPU 2503 controls the display control unit 2507, the communication control unit 2517, and the drive device 2513 to perform necessary operations. Further, data in the middle of processing is stored in the memory 2501 and stored in the HDD 2505 if necessary. In an embodiment of the present technology, an application program for performing the above-described processing is stored in a computer-readable removable disk 2511 and distributed, and installed from the drive device 2513 to the HDD 2505. In some cases, the HDD 2505 may be installed via a network such as the Internet and the communication control unit 2517. Such a computer apparatus realizes various functions as described above by organically cooperating hardware such as the CPU 2503 and the memory 2501 described above, the OS, and necessary application programs.

  The present embodiment can be summarized as follows.

  This rule processing method includes a score, a relative appearance position and a type to which the score should be assigned, in association with each condition or combination of conditions represented by a combination of the attribute type, appearance relative position, and attribute value of the word. Access to a rule data storage unit in which one or a plurality of score setting rules are registered, and an element column data storage unit that stores an element column in which elements including attribute values of each attribute type of words are arranged in the order of appearance Executed by a computer. Then, this rule processing method uses the reference position of the appearance relative position included in the score setting rule associated with the maximum number of conditions included in the combination of conditions and the condition or combination of conditions from the element string data storage unit. In accordance with the candidate extraction condition specified by the maximum distance from the element, an element condition (feature in the embodiment) or a combination of element conditions represented by a combination of the attribute type, appearance relative position, and attribute value of each word for each element The extraction step for extracting the application candidate, and the rule data storage unit by searching for the application candidate, and if there is a corresponding condition or combination of conditions, the corresponding combination of the above condition or condition from the rule data storage unit 1 or a plurality of score setting rules associated with, and the types and types included in the extracted one or more score setting rules In the score, the score for the type of the element identified from the appearance relative position included in the extracted one or more score setting rules and the appearance position related to the application candidate is updated and stored in the element string data storage unit Including the step of.

  By preparing the rule data storage unit as described above, the rule data storage unit is searched for the application candidate, and if there is a condition or combination of conditions corresponding to the application candidate, 1 associated therewith Alternatively, since a plurality of score setting rules can be extracted at a time, the number of searches in the rule data storage unit can be reduced, and the overall processing speed is improved. There may be one or a plurality of attribute types. For example, a word itself, a part of speech, a character type, and the like are assumed.

  Further, the extraction step described above relates to the attribute type, the attribute value of the attribute type, and the processing for each attribute type for each of the elements up to the appearance position separated by the maximum distance from the appearance position of the element related to the processing. The element condition is included within the maximum number of the above conditions so as to include at least one of the element condition for the element related to the processing and the step of generating the element condition represented by the combination of the appearance relative position from the appearance position of the element And generating a candidate for application in combination. Since the application candidate is not generated again by using the data of the element before the element related to the process in the element string, the process is simplified and speeded up.

  Further, the rule processing method may further include a step of comparing the score for each type for each of the elements, and specifying the type for which the maximum score is set as the type of the element. For example, if the type is a proper noun or other type, it can be determined whether each word is a proper noun.

  In addition, the computer includes a score and a type to which the score should be assigned in association with each of the second condition or the combination of the second condition represented by a combination of the attribute type, appearance relative position, and attribute value of the word. The second rule data storage unit in which the second score setting rule is registered may be further accessible. In that case, for each of the second condition or the combination of the second condition stored in the second rule data storage unit, the smallest value at the appearance relative position included in the second condition or the combination of the second condition is used as the reference value. Identify and convert the appearance relative position of each second condition included in the second condition or combination of second conditions into a new appearance relative position from the reference value to generate the condition or combination of conditions, In the rule data storage unit, a step of storing a score setting rule including a value obtained by multiplying the reference value by (−1) as an appearance relative position in addition to the second setting rule in association with the combination of conditions in the rule data storage unit; And a step of extracting score setting rules having the same conditions or combinations of conditions and aggregating the conditions or combinations of conditions. For example, when the second rule data storage unit is prepared by the conventional technique, data of the rule data storage unit necessary for the present technique can be prepared in this way.

  It is possible to create a program for causing a computer to carry out the processes described above, and the program can be read by a computer such as a flexible disk, a CD-ROM, a magneto-optical disk, a semiconductor memory, and a hard disk. Stored in a storage medium or storage device. Note that data being processed is temporarily stored in a storage device such as a computer memory.

  The following supplementary notes are further disclosed with respect to the embodiments including the above examples.

(Appendix 1)
One or a plurality of score settings including a score and a relative position and type of appearance to which the score should be given in association with each condition or combination of conditions represented by a combination of the attribute type, appearance relative position, and attribute value of the word Rules executed by a computer that can access a rule data storage unit in which rules are registered and an element string data storage unit that stores an element sequence in which elements including attribute values of word attribute types are arranged in the order of appearance A processing method,
From the element string data storage unit, the maximum number of conditions included in the combination of conditions and the maximum distance from the reference position of the appearance relative position included in the score setting rule associated with the condition or combination of conditions; An extraction step of extracting an application candidate that is an element condition or a combination of element conditions represented by a combination of a word attribute type, an appearance relative position, and an attribute value for each element according to the candidate extraction condition specified by:
If the rule data storage unit is searched for the application candidate and the corresponding condition or combination of conditions exists, the rule data storage unit is associated with the corresponding condition or condition combination 1 or Extracting a plurality of score setting rules, the type and the score included in the extracted one or more score setting rules, and the appearance relative position included in the extracted one or more score setting rules Updating a score for the type of the element identified from the appearance position related to the application candidate, and storing the score in the element string data storage unit;
Rules processing method including.

(Appendix 2)
The extraction step comprises:
For each of the elements from the appearance position of the element related to the processing to the appearance position separated by the maximum distance, for each attribute type, the attribute type, the attribute value of the attribute type, and the appearance position of the element related to the processing Generating an element condition represented by a combination with an appearance relative position;
Generating the application candidate by combining the element conditions within a maximum number of the conditions so as to include at least any of the element conditions for the elements involved in the processing;
The rule processing method according to supplementary note 1 including:

(Appendix 3)
The rule processing method according to appendix 1 or 2, further comprising a step of comparing the score for each of the types for each of the elements and specifying the type for which the maximum score is set as the type of the element.

(Appendix 4)
The computer includes a score and a type to which the score should be assigned in association with each of the second condition or the combination of the second condition represented by a combination of the attribute type, appearance relative position, and attribute value of the word The second rule data storage unit in which the second score setting rule is registered is further accessible,
For each of the second condition or combination of second conditions stored in the second rule data storage unit, the smallest value in the appearance relative position included in the second condition or combination of second conditions is set as a reference value. And generating the condition or combination of conditions by converting the appearance relative position of each second condition included in the second condition or combination of second conditions into a new appearance relative position from the reference value. In addition to the second setting rule, the score setting rule including a value obtained by multiplying the reference value by (−1) as the appearance relative position in the rule data storage unit in association with the condition or the combination of conditions. Storing, and
In the rule data storage unit, extracting the score setting rule having the same condition or combination of conditions, and aggregating the conditions or combinations of conditions;
The rule processing method according to any one of appendices 1 to 3, further including:

(Appendix 5)
A program for causing a computer to execute the rule processing method according to any one of appendices 1 to 4.

(Appendix 6)
One or a plurality of score settings including a score and a relative position and type of appearance to which the score should be given in association with each condition or combination of conditions represented by a combination of the attribute type, appearance relative position, and attribute value of the word A rule data storage in which rules are registered;
An element string data storage unit that stores an element string in which elements including attribute values of each attribute type of words are arranged in the order of appearance;
From the element string data storage unit, the maximum number of conditions included in the combination of conditions and the maximum distance from the reference position of the appearance relative position included in the score setting rule associated with the condition or combination of conditions; An extraction means for extracting an application candidate that is an element condition or a combination of element conditions represented by a combination of a word attribute type, an appearance relative position, and an attribute value for each element according to the candidate extraction condition specified by:
If the rule data storage unit is searched for the application candidate and the corresponding condition or combination of conditions exists, the rule data storage unit is associated with the corresponding condition or condition combination 1 or Extracting a plurality of score setting rules, the type and the score included in the extracted one or more score setting rules, and the appearance relative position included in the extracted one or more score setting rules Means for updating a score for the type of the element identified from the appearance position related to the application candidate, and storing the score in the element string data storage unit;
A rule processing device.

DESCRIPTION OF SYMBOLS 1 Learning data input part 3 Learning data storage part 5 Rule learning part 7 1st rule data storage part 9 Rule conversion part 11 2nd rule data storage part 13 Discrimination object data input part 15 Discrimination object data storage part 17 Discrimination part 19 Discrimination result Storage unit 21 Output unit

Claims (6)

One or a plurality of score settings including a score and a relative position and type of appearance to which the score should be given in association with each condition or combination of conditions represented by a combination of the attribute type, appearance relative position, and attribute value of the word Rules executed by a computer that can access a rule data storage unit in which rules are registered and an element string data storage unit that stores an element sequence in which elements including attribute values of word attribute types are arranged in the order of appearance A processing method,
From the element string data storage unit, the maximum number of conditions included in the combination of conditions and the maximum distance from the reference position of the appearance relative position included in the score setting rule associated with the condition or combination of conditions; An extraction step of extracting an application candidate that is an element condition or a combination of element conditions represented by a combination of a word attribute type, an appearance relative position, and an attribute value for each element according to the candidate extraction condition specified by:
If the rule data storage unit is searched for the application candidate and the corresponding condition or combination of conditions exists, the rule data storage unit is associated with the corresponding condition or condition combination 1 or Extracting a plurality of score setting rules, the type and the score included in the extracted one or more score setting rules, and the appearance relative position included in the extracted one or more score setting rules Updating a score for the type of the element identified from the appearance position related to the application candidate, and storing the score in the element string data storage unit;
Rules processing method including. The extraction step comprises:
For each of the elements from the appearance position of the element related to the processing to the appearance position separated by the maximum distance, for each attribute type, the attribute type, the attribute value of the attribute type, and the appearance position of the element related to the processing Generating an element condition represented by a combination with an appearance relative position;
Generating the application candidate by combining the element conditions within a maximum number of the conditions so as to include at least any of the element conditions for the elements involved in the processing;
The rule processing method according to claim 1, comprising: The rule processing method according to claim 1, further comprising: comparing a score for each of the types for each of the elements and specifying a type for which the maximum score is set as the type of the element. One or more first and a single word first condition or each type to be given a score and the score associated with the combination of the first condition and the attribute type is expressed in combination with the appearance relative position and attribute values a rule processing method performed in the first rule data storage unit 1 score setting rule is registered by a computer accessible,
For each of the combinations of the first condition or the first condition is stored in the first rule data storage unit, a reference value the smallest value in the occurrence relative position included in the combination of the first condition or the first condition was identified as the said occurrence relative position of each of the first condition contained in the combination of the first condition or the first condition is converted into a new appearance relative position from the reference value, the word attribute type and the generating a combination of second condition or a second condition represented by the combination of the attribute value and the new appearance relative positions, in association with a combination of the second condition or the second condition, the first score setting rule In addition, storing a second score setting rule including a value obtained by multiplying the reference value by (−1) as an appearance relative position to which a score is to be assigned, in a second rule data storage unit;
In the second rule data storage unit, the method comprising the combination of the second condition or the second condition is to extract the same the score set rules, to aggregate the combination of the second condition or the second condition,
Rules processing methods, including.   A program for causing a computer to execute the rule processing method according to any one of claims 1 to 4. One or a plurality of score settings including a score and a relative position and type of appearance to which the score should be given in association with each condition or combination of conditions represented by a combination of the attribute type, appearance relative position, and attribute value of the word A rule data storage in which rules are registered;
An element string data storage unit that stores an element string in which elements including attribute values of each attribute type of words are arranged in the order of appearance;
From the element string data storage unit, the maximum number of conditions included in the combination of conditions and the maximum distance from the reference position of the appearance relative position included in the score setting rule associated with the condition or combination of conditions; An extraction means for extracting an application candidate that is an element condition or a combination of element conditions represented by a combination of a word attribute type, an appearance relative position, and an attribute value for each element according to the candidate extraction condition specified by:
If the rule data storage unit is searched for the application candidate and the corresponding condition or combination of conditions exists, the rule data storage unit is associated with the corresponding condition or condition combination 1 or Extracting a plurality of score setting rules, the type and the score included in the extracted one or more score setting rules, and the appearance relative position included in the extracted one or more score setting rules Means for updating a score for the type of the element identified from the appearance position related to the application candidate, and storing the score in the element string data storage unit;
A rule processing device.

Images