JP5326781B2 - Extraction rule creation system, extraction rule creation method, and extraction rule creation program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an extraction rule creation system for efficiently creating a rule for extracting information requested by a user. <P>SOLUTION: When a text with tag and character string position information being information showing the position of a character string in the text with tag are applied, an extraction rule creation means 82 combines a word or tag corresponding to position shown by the character string position information with words or tags before and after the word or tag, and creates an extraction rule being a rule for extracting information from the text with tag. An adequate sentence position information extraction means 83 extracts adequate sentence position information being information showing the position of an adequate sentence including the word or the tag adequate to the extraction rule by text with tag stored in a text with tag storage means 81. An evaluation value calculation means 84 calculates an evaluation value higher according as the number of adequate sentences appearing in one text with tag becomes smaller, and calculates the evaluation value higher according as the adequate sentence appears more in texts with tag. <P>COPYRIGHT: (C)2011,JPO&amp;INPIT

Description

  The present invention relates to an extraction rule creation system, an extraction rule creation method, and an extraction rule creation program for creating an extraction rule for extracting information from a document.

  When extracting necessary information from a document, a user may want to extract information that matches some rule (pattern). If the pattern of information to be extracted is known, it is possible to extract information from other documents using the pattern.

For example, consider a case where a user extracts a person name from a document. If the user
When it is desired to collect suspect names from personal names, it can be assumed that the pattern of information that the user wants to extract is “person name” + suspect. In addition, when the user wants to collect personal names whose surname is “ab (ab is two kanji characters representing the surname)”, the information pattern that the user wants to extract is “ab +“ noun ””. I can guess. In this way, if the pattern of information extracted by the user is known, information that matches the pattern can be extracted from other documents.

  Patent Document 1 describes an information extraction rule generation device that easily generates an information extraction rule from a training corpus. In the information extraction rule generation device described in Patent Document 1, the tree structure display unit displays a syntax tree on the display unit. The worker inputs an annotation based on a mouse operation or a keyboard operation while referring to the syntax tree displayed on the display unit. The tree structure regular expression extraction unit extracts a tree structure expression displaying a corresponding rule based on the tree structure and the annotation.

  Non-Patent Document 1 describes an information extraction rule creation support method based on automatic rule generation and interactive selection. In the method described in Non-Patent Document 1, a plurality of extraction rules are automatically created in advance from one case, and extraction processing is executed based on each extraction rule. Then, after presenting the extraction result to the user, the user narrows down an appropriate extraction rule by interactively inputting whether the extraction result is correct or not. Thereby, the user can create an appropriate extraction rule only by inputting the correctness of the extraction result.

JP 2004-318809 (paragraphs 0028-0032)

Tsuyoshi Kawai, Shinichi Ando, "Proposal for Information Extraction Rule Creation Support Based on Automatic Rule Generation and Interactive Selection", Proc. Of the 13th Annual Conference of the Language Processing Society of Japan, D3-1, 2007 3 Moon

  When a user extracts a person name from a document, the user is considered to want information that matches some pattern related to the person name. However, as in the above-described example, when a user extracts a person name from a document, a plurality of patterns representing information that the user wants to extract can be estimated. For example, when the user wants to collect suspect names from personal names, the pattern of information that the user wants to extract is ““ person name ”+ suspect”. When the user wants to collect personal names whose surname is “ab (ab is two kanji characters representing surname)”, the pattern of information that the user wants to extract is “ab +“ noun ””. become. Thus, it is difficult to efficiently create a pattern for extracting information because multiple patterns for extracting information desired by the user are inferred only by the information “extract a person's name from a document”. In many cases.

  In the information extraction rule generation device described in Patent Document 1, the user must describe a complicated extraction rule based on the syntax tree displayed on the display unit. Therefore, in order to create an extraction rule using the apparatus described in Patent Document 1, there is a problem that it takes a lot of time. Moreover, since the input method is complicated in the apparatus described in Patent Document 1, there is a problem that the user must understand the operation method.

  Further, in the method described in Non-Patent Document 1, in order to create an appropriate extraction rule, the user only has to input correct / incorrect for the extraction result. However, until an appropriate extraction result is obtained, the user needs to repeatedly input the correctness of the extraction result. Therefore, there is still a problem that the user has a lot of trouble to generate the extraction rule.

  Therefore, an object of the present invention is to provide an extraction rule creation system, an extraction rule creation method, and an extraction rule creation program that can efficiently create rules for extracting information desired by a user.

  The extraction rule creation system according to the present invention is information added to an arbitrary position in a character string, and includes position information indicating the position of the character string to which the information is added and an attribute of a word corresponding to the position. Tagged text storage means for storing tagged text that is a document including a set of tags that is information indicating attribute information to be indicated, and characters that are information indicating the position of the tagged text and a character string in the tagged text When column position information is given, a word or tag corresponding to the position indicated by the character string position information is combined with words or tags before and after the word or tag to extract information from the tagged text For each tagged text stored in the tagged text storage means, an extraction rule creating means for creating an extraction rule that is a rule of The matching sentence position information extracting means for extracting the matching sentence position information that is information indicating the position of the matching sentence, and the evaluation value calculation for calculating the evaluation value that is a value obtained by evaluating the extraction rule based on the matching sentence position information The evaluation value calculation means calculates the evaluation value higher as the number of matching sentences appearing in one tagged text is smaller, and calculates the evaluation value as the matching sentence appears in more tagged texts. It is characterized by doing.

  The extraction rule creation method according to the present invention is information added to an arbitrary position in a character string, and includes position information indicating the position of the character string to which the information is added and an attribute of a word corresponding to the position. When a tagged text that is a document including a set of tags that is attribute information to be indicated and character string position information that is information indicating the position of the character string in the tagged text is given, the character string An extraction rule creating step for creating an extraction rule that is a rule for extracting information from tagged text by combining a word or tag corresponding to the position indicated by the position information and a word or tag before and after the word or tag. Then, for each tagged text stored in the tagged text storage means, the matching sentence position information which is the information indicating the position of the matching sentence including the word or the tag that matches the extraction rule is extracted. A matching sentence position information extraction step, and an evaluation value calculation step that calculates an evaluation value that is a value obtained by evaluating the extraction rule based on the matching sentence position information. In the evaluation value calculation step, within one tagged text The evaluation value is calculated to be higher as there are fewer conforming sentences appearing in, and the evaluation value is calculated to be higher as matching sentences appear in more tagged text.

  The extraction rule creation program according to the present invention is information added to an arbitrary position in a character string, and includes position information indicating the position of the character string to which the information is added and an attribute of a word corresponding to the position. An extraction rule creation program applied to a computer having tagged text storage means for storing tagged text, which is a document including a set of tags, which is information representing attribute information to be displayed. When the character string position information, which is information indicating the position of the character string in the tagged text, is given, the word or tag corresponding to the position indicated by the character string position information, and before and after the word or tag Extraction rule creation processing that creates extraction rules that are rules for extracting information from tagged text by combining words or tags, tagged text For each of the tagged text stored in the storage means, a matching sentence position information extraction process for extracting matching sentence position information, which is information indicating a position of a matching sentence including a word or tag that matches the extraction rule, and matching sentence position information Based on the evaluation value, the evaluation value calculation process for calculating the evaluation value, which is an evaluation value of the extraction rule, is executed. In the evaluation value calculation process, the lower the number of matching sentences that appear in one tagged text, the higher the evaluation value is calculated. In addition, the evaluation value is calculated to be higher as the matching sentence appears in more tagged text.

  According to the present invention, it is possible to efficiently create a rule for extracting information desired by a user.

It is explanatory drawing which shows the example of the text with a tag. It is explanatory drawing which shows the example of the syntax expressing the extraction rule. It is a block diagram which shows one Embodiment of the extraction rule preparation system by this invention. It is explanatory drawing which shows the example of the data format in the target document memory | storage part. It is explanatory drawing which shows the example of the pattern produced by the pattern synthetic | combination part. It is explanatory drawing which shows the example of the evaluation value which the pattern evaluation part 13 calculated. It is a flowchart which shows the example of a synthetic | combination step. It is a state transition diagram showing an example of a method for selecting a pattern. It is a flowchart which shows the example of the algorithm in step S10. It is a flowchart which shows the example of the algorithm in step S40. It is a flowchart which shows the example of an evaluation step. It is a flowchart which shows the example of a search process. It is a flowchart which shows the example of the determination process in step S52. It is explanatory drawing which shows the example of the pattern produced by step S10-S30. It is a block diagram which shows the minimum structure of this invention.

  Hereinafter, terms used in the description of the present invention will be defined. The tagged text is a document including at least a body that is a set of character strings and a set of tags added at arbitrary positions in the character strings. FIG. 1 is an explanatory diagram illustrating an example of tagged text. The tagged text illustrated in FIG. 1 is the main text that says “Nara Prefectural Police arrested suspected abcd (20) on 14th for murder of robbery.” (Where ab is a two-character kanji character representing a surname, cd Indicates a set of tags added to each character string in the text.

  The tag is information including a character string representing a word attribute (hereinafter referred to as a tag name) and a position (start position and end position) in the text of the character string to which the tag is added. The start position and the end position are represented by the number in the text of the character string to which the tag is added, for example, with the beginning of the sentence being “0” and the number of characters counted. In the tagged text illustrated in FIG. 1, the tag added to “abcd” can be expressed using tags with a tag name “person name”, a start position “9”, and an end position “13”. In the following description, a tag whose tag name is T may be referred to as a “T tag”.

  The designated extraction position is a position of a character string in the tagged text that designates a character string to be extracted by a user or an external program, and is information represented by a start position and an end position. For example, when the user designates the 9th character at the start position and the 13th character at the end position as the designated extraction position for the tagged text illustrated in FIG. 1, the user wants to extract the character string “abcd”. I understand that

  A case is information representing a set of one tagged text and a designated extraction position designated for the tagged text, and is created by a user or an external program.

  An extraction rule is a rule for extracting information from tagged text. By appropriately creating this extraction rule, information desired by the user can be extracted from the tagged text. The extraction rule is expressed by a combination of at least a character string, a tag name, and a wild card (sometimes referred to as a template), and the template includes information indicating a designated extraction position.

  FIG. 2 is an explanatory diagram showing an example of a syntax expressing the extraction rule in the present invention. The syntax illustrated in FIG. 2 indicates that the extraction rule R is defined by a character string including one or more columns each including the phrase PH and the extraction position pattern EP. The phrase PH is a character string composed of one or more condition keys. The condition KEY is expressed by any one of a character string itself, a tag name surrounded by “[” and “]”, a wild card (*), or a null character (φ). The syntax expressing the extraction rule may be described as a pattern to be extracted (or simply a pattern).

  The extraction position pattern EP is a pattern obtained by combining the components of the tagged text at the designated extraction position. Specifically, the extraction position pattern EP is a pattern including one or more tags and including a character string and a tag. The syntax illustrated in FIG. 2 indicates that the extraction position pattern EP is defined as a character string surrounded by the symbol “$” before and after the constituent element EPH. The component element EPH is a character string including one or more columns including a tag and a condition EKEY, or a character string obtained by combining the component element EPH with the condition EKEY. The condition EKEY is represented by either the character string itself, the tag name surrounded by “[” and “]”, or the empty character (φ).

  As an example of the extraction rule R, there is a rule of “$ [person name] $ suspect”. This extraction rule indicates that the character string of the part to which the personal name tag is added is extracted from the part where the personal name tag and the character string “suspect” are adjacent.

  Another example is a rule of “Nara * $ ab [noun] $” (where ab is two kanji characters representing a surname). This extraction rule is a sentence in which an arbitrary character string (*) exists after the character string “Nara”, followed by a character string “ab”, and immediately followed by a noun tag. Therefore, it is indicated that the rule is that “ab” and the character string of the noun tag part are extracted together.

  In the following description, a problem of creating an extraction rule based on a set of tagged text and one or more cases may be referred to as an extraction rule creation problem.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

  FIG. 3 is a block diagram showing an embodiment of the extraction rule creation system according to the present invention. The extraction rule creation system according to the present invention includes a target document storage unit 11, a pattern synthesis unit 12, a pattern evaluation unit 13, and a document search unit 14.

  The target document storage unit 11 is a storage device that stores a set of tagged text to be extracted. The target document storage unit 11 holds the tagged text illustrated in FIG. 1 in an arbitrary format. FIG. 4 is an explanatory diagram illustrating an example of a format in which the target document storage unit 11 stores tagged text. In the example illustrated in FIG. 4, the target document storage unit 11 stores the tagged text illustrated in FIG. 1 in two tables, a body table and a tag table.

  The body table is a table that stores tagged text for each sentence. The body table stores a document ID that is a unique identifier of tagged text and a sentence ID that is a unique identifier of a sentence in association with a character string of the body.

  The tag table is a table that stores all tags added to a certain tagged text. The tag table stores the tag name of a tag, the start position and end position of the text, the document ID of the document to which the tag is added, and the sentence ID of the sentence to which the tag is added in association with each other.

  For example, the tagged text may be registered in advance in the target document storage unit 11 by an administrator, or may be registered in the target document storage unit 11 by the document search unit 14 described later.

  When one or more cases (that is, a set of tagged text and a designated extraction position designated for the tagged text) are given by the method described later, the pattern synthesis unit 12 Extraction rule candidates are synthesized (created) based on the word or tag of the tagged text and the word or tag before and after the word or tag. FIG. 5 is an explanatory diagram illustrating an example of a pattern created by the pattern synthesis unit 12. For example, given the tagged text illustrated in FIG. 1 and a case in which the 9th to 13th characters of the tagged text are designated extraction positions, the pattern composition unit 12 determines the case and the vicinity of the designated extraction position. The pattern illustrated in FIG. 5 is created based on the word or tag.

  The pattern evaluation unit 13 examines the position where the pattern created by the pattern synthesis unit 12 occurs in the tagged text stored in the target document storage unit 11 and the distribution of the occurrence frequency, and evaluates the evaluation value based on this distribution. Is calculated. That is, for each tagged text stored in the target document storage unit 11, the pattern evaluation unit 13 extracts the position of a sentence including a word or tag that matches the pattern created by the pattern synthesis unit 12, and stores the position. And the evaluation value is calculated. Specifically, the pattern evaluation unit 13 determines, for each tagged text stored in the target document storage unit 11, the position of a sentence including a word or tag that matches the extraction rule (pattern) created by the pattern synthesis unit 12. Extract. Hereinafter, a sentence including a word or tag that conforms to the extraction rule is referred to as a conforming sentence. Then, the pattern evaluation unit 13 calculates the evaluation value of the extraction rule higher as the number of matching sentences appearing in one tagged text is smaller, and as the matching sentence appears in more tagged texts, High evaluation value is calculated.

  For example, the total number of occurrences of the pattern p in the tagged text in the target document storage unit 11 (that is, the number that includes a matching pattern) is pf (p), and the pattern in the tagged text in the target document storage unit 11 Assuming that df (p) is the number of times p appears (that is, the type of the document including the matching pattern), the evaluation value ipfdf (p) can be calculated by the following equation 1.

  In the following description, a case where the pattern evaluation unit 13 calculates an evaluation value using Equation 1 will be described. The evaluation value calculation method calculates the evaluation value of the extraction rule higher as the number of matching sentences appearing in one tagged text is smaller, and the extraction rule as the matching sentence appears in more tagged texts. It may be calculated using another formula that calculates a high evaluation value of.

  FIG. 6 is an explanatory diagram illustrating an example of evaluation values calculated by the pattern evaluation unit 13. For example, for each pattern illustrated in FIG. 5, the pattern evaluation unit 13 examines the distribution of the occurrence position and the occurrence frequency in the tagged text stored in the target document storage unit 11 and evaluates based on the distribution. Calculate the value.

  When the user or an external program or the like specifies an extraction condition (extraction target keyword) as a search query, the document search unit 14 extracts tagged text corresponding to the condition. The document search unit 14 is realized by, for example, a document search system that extracts tagged text corresponding to a specified condition. The tagged text to be extracted is stored in advance in a storage unit (not shown) in the document search unit 14, for example, in the format illustrated in FIG. The document search unit 14 extracts a list of document IDs and sentence IDs that satisfy a condition from the storage unit (not shown), reads the tagged text indicated by each document ID, and uses the tagged text as the target document storage unit 11. You may register with. Note that registering tagged text in the target document storage unit 11 means storing the tagged text in the target document storage unit 11.

  If the format stored in the storage unit is different from the format illustrated in FIG. 4, the document search unit 14 converts the tagged text corresponding to the extraction condition into the format illustrated in FIG. It may be registered in the document storage unit 11.

  Based on the above contents, the document search unit 14 performs a process of selectively collecting a set of tagged text to be extracted and inserting it into the target document storage unit 11 based on a search query expressed by a keyword. be able to.

  As described above, the document search unit 14 extracts the tagged text based on the search query designated by the user or an external program, and registers the extracted tagged text in the target document storage unit 11, thereby The set of tagged text in the document storage unit 11 can be changed. By changing the tagged text to be extracted, the evaluation value of the pattern desired by the user can be made higher, so that the extraction rule can be extracted efficiently.

  The pattern synthesis unit 12, the pattern evaluation unit 13, and the document search unit 14 are realized by a CPU of a computer that operates according to a program (extraction rule creation program). For example, the program is stored in a storage unit (not shown) included in an apparatus including the pattern synthesis unit 12, the pattern evaluation unit 13, and the document search unit 14, and the CPU reads the program and follows the program according to the program. The pattern evaluation unit 13 and the document search unit 14 may be operated. Further, each of the pattern synthesis unit 12, the pattern evaluation unit 13, and the document search unit 14 may be realized by dedicated hardware.

  Next, the operation will be described. In the following description, first, the pattern synthesis unit 12 synthesizes (creates) a pattern based on the input case, and the pattern evaluation unit 13 evaluates each pattern created by the pattern synthesis unit 12 (hereinafter referred to as “pattern”). These processes are collectively referred to as a case input process.). Thereafter, the document search unit 14 creates a set of tagged text including the keyword in the body from the tagged text stored in the internal storage unit (not shown) (hereinafter, this process is referred to as a search process). Will be described.).

  In the following description, a process in which the pattern synthesis unit 12 synthesizes (creates) a pattern based on an input case is referred to as a synthesis step, and the pattern evaluation unit 13 creates each pattern created by the pattern synthesis unit 12. The process for evaluating the pattern is referred to as an evaluation step.

  First, the case input process will be described. The case input process starts when a user or an external program inputs a case to the pattern synthesis unit 12.

  FIG. 7 is a flowchart showing an example of the synthesis step. When a certain example (that is, information including a tagged text and a specified extraction position) is input, the pattern synthesis unit 12 first extracts all the character strings and tags at the specified extraction position from the tagged text, and adds one tag. All combinations including the above are extracted to create a pattern (step S10).

  FIG. 8 is a state transition diagram illustrating an example of a method by which the pattern synthesis unit 12 selects a pattern. In the example illustrated in FIG. 8, as an example, a case where the tagged text illustrated in FIG. 1 and the designated extraction position “9th to 13th characters” are input will be described. The pattern synthesis unit 12 examines all the patterns of state transitions illustrated in FIG. 8 from the tagged text illustrated in FIG. 1, and displays “abcd”, “ab [noun]”, “[noun] cd”, “[ Noun] [noun] ”and“ [person name] ”are extracted. Among these, four patterns of “ab [noun]”, “[noun] cd”, “[noun] [noun]”, and “[person name]” are extracted as combinations including one or more tags.

  A pattern that does not include a tag can collect only a specific character string (for example, a character string “abcd”), and has a small effect as an extraction rule. Therefore, by extracting a combination including one or more tags, useless patterns can be omitted in advance, and the amount of calculation thereafter can be reduced.

  The process (algorithm) in step S10 will be described with reference to FIG. FIG. 9 is a flowchart illustrating an example of the algorithm in step S10. The algorithm illustrated in FIG. 9 creates a pattern by recursively calling a method (generate). Here, the generate method has three arguments: a pattern now created when the generate method is called, a position pos at that time, and an end position end.

  First, the pattern synthesis unit 12 receives, as arguments of the generate method, now = ”” (empty), pos = start position of the specified extraction position, end = end position of the specified extraction position, and performs the process indicated by the generate method. Start. In the generate method, the pattern synthesis unit 12 checks whether the position pos is larger than the end position end (step S11). If the position pos is larger than the end position end (YES in step S11), the pattern composition unit 12 ends the process (step S12). If the position pos is not larger than the end position end (NO in step S11) and the position pos is equal to the end position end (YES in step S13), the pattern composition unit 12 determines that the end position of the state transition has been reached, The pattern now at that time is examined. If even one tag is included in the pattern now, the value of the pattern now is extracted as a pattern, and the process is terminated (step S14).

  On the other hand, when the position pos is not equal to the end position end (NO in step S13), the pattern composition unit 12 extracts all the tags T having the position pos as the start position, and adds these tags T to the current pattern. In addition, the pattern synthesis unit 12 calls the generate method with information obtained by updating the current position pos to the end position of the tag T (step S15). Furthermore, the pattern composition unit 12 takes out the character c next to the current position pos (step S16). Then, the pattern synthesis unit 12 adds the character c to the pattern now and calls the generate method with information obtained by adding 1 to the current position pos (step S17).

  The generate method is a process for executing the above contents. That is, it can be said that the generate method is a process of creating a pattern while moving the designated extraction position to the right.

  Next, the pattern synthesizing unit 12 creates a pattern in which the character string of the right R character at the designated extraction position and the tag are combined, and a pattern in which the character string of the left L character at the designated extraction position and the tag are combined ( Step S20 in FIG. The values of R and L are, for example, arbitrary integers designated in advance by the user or developer.

  A pattern in which a character string of a right R character (left L character) at a designated extraction position and a tag are combined can be created by an algorithm similar to the algorithm illustrated in FIG. That is, when creating a pattern of R characters on the right side of the designated extraction position, the pattern synthesis unit 12 sets the generate method arguments as “now =” ”(empty), pos = end position of the designated extraction position, end = designated extraction position The processing may be performed as the end position + R. Further, when creating a pattern of L characters on the left side of the designated extraction position, the pattern synthesis unit 12 sets the generate method arguments as “now =” ”(empty), pos = start position of the designated extraction position−L, end = designation. Processing may be performed as the start position of the extraction position.

  Since the right R character (left L character) of the designated extraction position does not need to include a tag, in this case, the pattern synthesis unit 12 determines whether the pattern now includes a tag in step S14. It is not necessary to judge whether or not.

  Next, the pattern synthesis unit 12 creates a pattern based on the noun tag (step S30 in FIG. 7). Specifically, the pattern synthesis unit 12 extracts all nouns that appear in addition to the designated extraction position, the left L character of the designated extraction position, and the right R character of the designated extraction position. Then, the pattern synthesizing unit 12 extracts the noun on the left side of the designated extraction position as the left pattern word Plw and the noun on the right side of the designated extraction position among these nouns as the right pattern word Prw.

  In the following description, a case where the pattern composition unit 12 creates a pattern based on a noun tag (that is, a noun is extracted from the text) will be described. Is not limited to noun tags. For example, it may be an independent word such as a verb other than a noun, an adjective, or an adjective verb. By creating a pattern using such words, general and meaningless words such as particles can be excluded from the pattern.

  The pattern synthesis unit 12 includes a list of patterns created in step S10 (hereinafter referred to as list A), a list of right patterns created in step S20 (hereinafter referred to as list RP), and a list of left patterns. (Hereinafter referred to as a list LP) and a list of left pattern words (hereinafter referred to as a list LW) created in step S30 and a list of right pattern words (hereinafter referred to as a list RW) are synthesized. Then, extraction rule candidates are created (step S40).

  The process (algorithm) in step S40 will be described with reference to FIG. FIG. 10 is a flowchart illustrating an example of the algorithm in step S40. In the algorithm illustrated in FIG. 10, the pattern synthesis unit 12 first adds a null character “” to the list RP, list LP, list LW, and list RW (step S41). This empty character “” means that a pattern included in the list is not used. The pattern synthesis unit 12 extracts one pattern from each list (ie, list A, list RP, list LP, list LW, and list RW), and performs the following steps for all combinations of the extracted patterns of each list. The processing shown in S42 to S47 is performed.

  The pattern synthesis unit 12 creates a pattern R in which “$”, which is a symbol indicating the designated extraction position, is added to the pattern extracted from the list A (step S42). Next, the pattern synthesis unit 12 adds the pattern Pr extracted from the list RP to the right side of the pattern R (step S43). Similarly, the pattern synthesis unit 12 adds the pattern Pl extracted from the list LP to the left side of the pattern R (step S44). Next, the pattern synthesis unit 12 adds a pattern with a wild card “*” added to the left side of the pattern Prw extracted from the list RW to the right side of the pattern R (step S45). Similarly, the pattern composition unit 12 adds a pattern with a wild card “*” added to the right side of the pattern Plw extracted from the list LW to the left side of the pattern R (step S46). Finally, the pattern synthesis unit 12 notifies the pattern evaluation unit 13 of the created pattern R (step S47). As described above, the pattern synthesis unit 12 notifies the pattern evaluation unit 13 of a pattern created based on the given case.

  Next, the operation of the evaluation step will be described. FIG. 11 is a flowchart illustrating an example of the evaluation step. When the pattern evaluation unit 13 receives each pattern from the pattern synthesis unit 12, the pattern evaluation unit 13 searches for a sentence that matches the pattern and extracts a combination of the document ID and the sentence ID of the matching sentence (step S50). Then, the pattern evaluation unit 13 calculates the evaluation value of the received pattern from the extracted document ID and sentence ID (step S60). In the following description, the process in step S50 is referred to as a search process, and the process in step S60 is referred to as an evaluation value calculation process.

  The search process will be described with reference to FIG. FIG. 12 is a flowchart illustrating an example of search processing. The pattern evaluation unit 13 sequentially reads each record in the body table stored in the target document storage unit 11, and reads a tag corresponding to the document ID and sentence ID of each record from the tag table (step S51). Next, the pattern evaluation unit 13 compares each sentence and each tag with the input pattern (step S52), and when matching is established between them (that is, a sentence or tag that matches the pattern exists). Then, the document ID and sentence ID of the sentence are extracted (step S53).

  The method for performing the search process is not limited to the above method. Any other method may be used as long as it can extract the position of a sentence that matches the pattern created by the pattern synthesis unit 12.

  Processing for determining whether or not matching has been established between the two in step S52 will be described with reference to FIG. FIG. 13 is a flowchart illustrating an example of the determination process in step S52. The pattern evaluation unit 13 performs the following steps S71 to S85 based on the pattern P received from the pattern synthesis unit 12, the sentence S read in step S51, and the tag set TList added to the sentence S. Do.

  The pattern evaluation unit 13 divides the pattern P into tag name, character string, and wildcard conditions based on a predetermined syntax (the symbols “[”, “]”, “*”, “$”). Then, a list of separated conditions (hereinafter referred to as a condition column list List) is created (step S71). For example, in the case of a pattern of “$ [person name] $ suspect * arrest”, the pattern evaluation unit 13 may divide into four conditions of “[person name]”, “suspect”, “*”, and “arrest”. it can. The pattern evaluation unit 13 stores the separated conditions in the condition column list List in order from the top of the pattern.

  Next, the pattern evaluation unit 13 initializes a variable i representing a position in the sentence to 0 and a variable flag representing the presence / absence of a wild card to 1 (step S72). Then, the pattern evaluation unit 13 extracts the first condition C from the condition string list List (step S73). If the extracted condition C is a tag name (YES in step S74), the pattern evaluation unit 13 checks whether or not the variable flag is 1 (step S75). If the variable flag is 1 (YES in step S75), it can be said that this state is either that the wild card is valid or the beginning of the pattern. In this case, the pattern evaluation unit 13 checks whether or not the tag T specified by the condition C and the start position is larger than the variable i from the tag set TList (step S76). When the tag T specified by the condition C exists in the tag set TList (YES in step S76), the pattern evaluation unit 13 determines that the tag T matches the condition C, and ends the tag T in the variable i. The position is substituted (step S78).

  On the other hand, when the tag T specified by the condition C does not exist in the tag set TList in step S76 (NO in step S76), the pattern evaluation unit 13 outputs information indicating that the patterns do not match, and performs processing. finish.

  In step S75, if the variable flag is not 1 (NO in step S75), it is understood that the wild card is not valid. In this case, the pattern evaluation unit 13 checks whether or not the tag T specified by the condition C exists and the start position is equal to the variable i from the tag set TList (step S77). When the tag T specified by the condition C exists in the tag set TList (YES in step S77), the pattern evaluation unit 13 determines that the tag T matches the condition C, and ends the tag T in the variable i. The position is substituted (step S78). On the other hand, if the tag T specified by the condition C does not exist in the tag set TList in step S77 (NO in step S77), the pattern evaluation unit 13 outputs information indicating that the patterns do not match, and performs processing. finish.

  In step S74, when the condition extracted by the pattern evaluation unit 13 is not a tag name (NO in step S74), the pattern evaluation unit 13 checks whether the condition C is a character string (step S79). When the condition C is a character string (YES in step S79), the pattern evaluation unit 13 checks whether or not the variable flag is 1 (step S80). If the variable flag is 1 (YES in step S80), it can be seen that the wild card is valid. In this case, the pattern evaluation unit 13 checks whether or not the character string W specified as the condition C exists after the i-th character of the read sentence S (step S81). When the character string W exists (YES in step S81), the pattern evaluation unit 13 determines that the character string W matches the condition C, and adds the number of characters of the character string W to the variable i (step S83). On the other hand, if the character string W does not exist in step S81 (NO in step S81), the pattern evaluation unit 13 outputs information indicating that the patterns do not match, and ends the process.

  In step S80, if the variable flag is not 1 (NO in step S80), it is understood that the wild card is not valid. In this case, the pattern evaluation unit 13 checks whether or not the character string W designated as the condition C exists in the i character of the read sentence S (step S82). When the character string W exists (YES in step S82), the pattern evaluation unit 13 determines that the character string W matches the condition C, and adds the number of characters of the character string W to the variable i (step S83). On the other hand, if the character string W does not exist in step S82 (NO in step S82), the pattern evaluation unit 13 outputs information indicating that the patterns do not match, and ends the process.

  After substituting the end position of the tag T for the variable i in step S78, or after adding the number of characters of the character string W to the variable i in step S83, the pattern evaluation unit 13 determines whether the condition list List is empty. (That is, whether or not the extraction of the condition C has been completed) is determined (step S84). If the condition list List is empty (YES in step S84), the pattern evaluation unit 13 determines that all the conditions C are satisfied, outputs information indicating that the patterns match, and ends the process. On the other hand, when the condition list List is not empty (NO in step S84), the pattern evaluation unit 13 substitutes 0 for the variable flag (step S85), and repeats the processing after step S73.

  In step S79, if condition C is not a character string (NO in step S79), condition C is a wild card. Therefore, the pattern evaluation unit 13 substitutes 1 for flag (step S86), and repeats the processing after step S73.

  In this manner, the pattern evaluation unit 13 indicates whether or not the pattern P received from the pattern synthesis unit 12 matches the sentence S read in step S51 and the tag set TList added to the sentence S. Can be output.

  Next, the evaluation value calculation process will be described. In the evaluation value calculation process, the pattern evaluation unit 13 calculates pf (p) and df (p) from the set of document ID and sentence ID extracted in the search process. For example, the pattern evaluation unit 13 calculates the value of pf (p) by counting the number of pairs of document ID and sentence ID, and calculates the value of df (p) among the pair of document ID and sentence ID. Calculation is performed by counting the number of types of document IDs. For example, the pattern evaluation unit 13 calculates an evaluation value based on the equation shown in Equation 1.

  Note that the pattern synthesis unit 12 and the pattern evaluation unit 13 may perform the processes in the case input process in an arbitrary order as long as the logical context of each process does not change.

  Next, the search process will be described. The search process starts when a user or an external program inputs a query (that is, a condition) to the document search unit 14.

  When the keyword to be extracted is specified as a query, the document search unit 14 includes the specified keyword in the text from the tagged text stored in the storage unit (not shown) in the document search unit 14. Create a set of tagged text. Next, the document search unit 14 empties all the data in the target document storage unit 11 and then stores the created tagged text set in the target document storage unit 11 in the format illustrated in FIG. 4, for example. sign up.

  When the case input process has been processed (that is, when a pattern as an extraction rule has already been created), the pattern evaluation unit 13 causes the document search unit 14 to collect a set of tagged text in the target document storage unit 11. Immediately after registration, the evaluation step process may be started. Thereby, the change of the evaluation value accompanying the change of the tagged text set in the target document storage unit 11 can be reflected immediately.

  According to the present invention, when the tagged text and the extraction position information are given, the pattern synthesis unit 12 extracts a combination of the word or tag of the extraction position information and the word or tag before and after the word or tag. Create a rule. Then, the pattern evaluation unit 13 extracts, for each tagged text stored in the target document storage unit 11, a document ID and a sentence ID of a sentence (matching sentence) including a word or tag that conforms to the extraction rule, and the document An evaluation value is calculated based on the ID and sentence ID. At this time, the pattern evaluation unit 13 calculates a higher evaluation value as there are fewer matching sentences appearing in one tagged text, and calculates a higher evaluation value as more matching sentences appear in more tagged text. Therefore, it is possible to efficiently create a rule (extraction rule) for extracting information desired by the user.

  For example, as a case, consider a case where the tagged text illustrated in FIG. 1 and “abcd” from the ninth character to the thirteenth character are input as the designated extraction position. For example, if it is assumed that the user simply wants to collect personal names, the extraction rule should be "$ [person name] $". Also, if it is assumed that the user wants to collect suspect names, the extraction rule should be “$ [person name] $ suspects”. Further, if it is assumed that the user wants to collect names with the surname “ab”, the extraction rule should be “$ ab [noun] $”. Thus, when only a case is input, the extraction rule to be determined differs depending on what the user wants as described above.

  In general, it can be said that the difficulty of the extraction rule creation problem is that a user or an external program must infer what information the user wants to extract from an input example. However, according to the present invention, an extraction rule is created based on a set of tagged text and one or more cases, and an evaluation value is calculated for each extraction rule. Therefore, the extraction rule according to the user's extraction request can be created while reducing the user's trouble.

  Further, the document search unit 14 extracts the tagged text corresponding to the designated condition and registers it in the target document storage unit 11, and the pattern evaluation unit 13 performs the tagging text registered in the target document storage unit 11. The document ID and sentence ID of the conforming sentence may be extracted. In this case, since the evaluation value of the pattern desired by the user can be customized by changing the tagged text extracted by the pattern evaluation unit 13, the extraction rule that matches the information desired by the user can be efficiently extracted.

  The pattern composition unit 12 may select a combination pattern including one or more tags from the created extraction rules. In this case, since unnecessary patterns are omitted in advance, the amount of calculation thereafter can be reduced.

  In addition, the pattern synthesis unit 12 creates an extraction rule by combining predetermined types of independent words (such as nouns) among words or tags before and after the word or tag corresponding to the position indicated by the extraction position information. Also good. In this case, common meaningless words such as particles can be excluded from the pattern.

  Hereinafter, the present invention will be described with reference to specific examples, but the scope of the present invention is not limited to the contents described below. In the following description, when the newspaper article data illustrated in FIG. 1 is stored in the document search unit 14 as tagged text, the user wants to create a list of suspect names of murder cases. Will be described.

  In the search process, for example, when the user designates a keyword such as “suspected murderer”, the document search unit 14 creates a set of tagged text related to the arrest information of the murder case from among the tagged text stored therein, Register in the target document storage unit 11. Thereby, the tagged text related to the murder case is stored in the target document storage unit 11.

  Next, in the case input process, for example, when the user inputs the tagged text illustrated in FIG. 1 and the designated extraction position from the ninth character to the thirteenth character to the pattern composition unit 12, the pattern composition unit 12 performs the composition step. To start. Here, it is assumed that list A, list RP, list LP, list RW, and list LW illustrated in FIG. 14 are created by the processes of steps S10, S20, and S30. Furthermore, in step S <b> 40, the pattern synthesis unit 12 combines these patterns, creates a pattern illustrated in FIG. 5, and notifies the pattern evaluation unit 13 of the pattern.

  Next, in the evaluation step, the pattern evaluation unit 13 performs a search process and an evaluation value calculation process on each received pattern, and calculates the pattern illustrated in FIG. 5 and the evaluation value illustrated in FIG. . In the present embodiment, since a set of tagged text created based on the keyword “murder suspect” is stored in the target document storage unit 11, it is considered that there are many sentences regarding the arrest information of the murder case. Therefore, the evaluation value in a pattern such as “$ [person name] $ suspect”, “$ [person name] $ suspect * murder”, or “$ [person name] $ suspect * murder” increases.

  The evaluation value according to the present invention is a value that appears more frequently in more documents and that a less frequent pattern becomes higher. For this reason, it means that a pattern that appears once for every tagged text in the target document storage unit 11 is highly evaluated. For this reason, keywords such as “Nara” and “ab (where ab is a two-character kanji for surname)” cannot be said to appear in more documents, so the evaluation value is Lower. Further, since a pattern that matches an arbitrary person name such as “$ [person name] $” matches not only the suspect name but also the victim name, etc., the frequency increases, so the evaluation value decreases. This is why the evaluation values in the patterns such as “$ [person name] $ suspect”, “$ [person name] $ suspect * murder”, and “$ [person name] $ suspect * murder” are high.

  In addition, the present invention can be used as a search system that realizes a list-type search in which information desired by a user is listed from a text and output in units of words. It can also be used as a text mining system that visualizes the extracted values with a graph or the like.

  Next, the minimum configuration of the present invention will be described. FIG. 15 is a block diagram showing the minimum configuration of the present invention. The extraction rule creation system according to the present invention is information added to an arbitrary position in a character string, and position information indicating the position of the character string to which the information is added (for example, a start position and an end position), Tagged text storage means 81 for storing tagged text that is a document including attribute information (for example, a noun, a person name, etc.) indicating the attribute of a word corresponding to the position and a set of tags. When the target document storage unit 11) and character string position information (for example, extracted position information) which is information indicating the position of the tagged text and the character string in the tagged text are given, the character string position information An extraction rule (e.g., a rule for extracting information from tagged text by combining a word or tag corresponding to the position indicated by and a word or tag before and after the word or tag) For example, for each tagged text stored in the extraction rule creation means 82 (for example, the pattern synthesis unit 12) and the tagged text storage means 81 for creating a pattern, a matching sentence including a word or tag that matches the extraction rule Based on the matching sentence position information, the matching sentence position information extracting means 83 (for example, the pattern evaluation unit 13) that extracts matching sentence position information (for example, document ID and sentence ID) that is information indicating the position of Evaluation value calculation means 84 (for example, pattern evaluation unit 13) that calculates an evaluation value that is a value obtained by evaluating a rule is provided.

  The evaluation value calculation unit 84 calculates the evaluation value higher as the number of matching sentences appearing in one tagged text is smaller, and calculates the evaluation value as the matching text appears in more tagged texts (for example, (Evaluation value is calculated based on Equation 1).

  With such a configuration, it is possible to efficiently create a rule for extracting information desired by the user.

  Also, the above embodiment discloses an extraction rule creation system having the following configuration.

(1) Information added to an arbitrary position in a character string, and position information (for example, start position and end position) indicating the position of the character string to which the information is added, and a word corresponding to the position Tagged text storage means (for example, target document storage unit 11) that stores tagged text that is a document including attribute information (for example, nouns, personal names, etc.) indicating a set of tags. When a character string position information (for example, extracted position information) that is information indicating the position of a tagged text and a character string in the tagged text is given, a word corresponding to the position indicated by the character string position information Or, an extraction rule that creates an extraction rule (for example, a pattern) that is a rule for extracting information from a tagged text by combining a tag and the word or a word or tag before or after the tag. Relevant sentence position information which is information indicating the position of a relevant sentence including a word or tag that matches the extraction rule for each tagged text stored in the generating means (for example, the pattern synthesis unit 12) and the tagged text storage means. Based on the matching sentence position information extraction unit (for example, the pattern evaluation unit 13) that extracts (for example, document ID and sentence ID), an evaluation value that is a value obtained by evaluating the extraction rule is calculated based on the matching sentence position information. Evaluation value calculation means (for example, pattern evaluation unit 13), and the evaluation value calculation means calculates the evaluation value higher as the number of matching sentences appearing in one tagged text is smaller, and in more tagged text An extraction rule creation system that calculates a higher evaluation value as a matching sentence appears in (for example, calculates an evaluation value based on Equation 1).

(2) Tagged text that extracts tagged text corresponding to a specified condition from a plurality of tagged text and registers the tagged text extracted by the tagged text extracting means in the tagged text storage means. An extraction rule creation system comprising registration means (for example, a document search unit 14), wherein the matching sentence position information extraction means extracts matching sentence position information for each tagged text registered by the tagged text registration means.

(3) An extraction rule creation system in which the extraction rule creation means selects a combination including one or more tags from the created extraction rules.

(4) The extraction rule creating means combines the predetermined types of independent words (for example, nouns) out of the words or tags before and after the word or tag corresponding to the position indicated by the character string position information, and sets the extraction rule. Extraction rule creation system to be created.

(5) The extraction rule creating means combines the wild card (for example, “*”) between a word or tag corresponding to the position indicated by the character string position information and a predetermined type of self-supporting word, and sets the extraction rule. Extraction rule creation system to be created.

  The present invention is suitably applied to an extraction rule creation system that creates an extraction rule for extracting information from a document.

11 target document storage unit 12 pattern synthesis unit 13 pattern evaluation unit 14 document search unit

Claims (9)

Information added to an arbitrary position in the character string, which is position information indicating the position of the character string to which the information is added and attribute information indicating the attribute of the word corresponding to the position. Tagged text storage means for storing tagged text that is a document including a set of tags;
When character string position information, which is information indicating the position of the tagged text and the character string in the tagged text, is given, the word or tag corresponding to the position indicated by the character string position information, the word or An extraction rule creating means for creating an extraction rule that is a rule for extracting information from the tagged text by combining words or tags before and after the tag;
For each tagged text stored in the tagged text storage means, matched sentence position information extracting means for extracting matched sentence position information, which is information indicating the position of a matched sentence including a word or tag that matches the extraction rule; ,
Evaluation value calculation means for calculating an evaluation value, which is a value obtained by evaluating the extraction rule, based on the relevant sentence position information;
The evaluation value calculation means calculates the evaluation value higher as there are fewer matching sentences appearing in one tagged text, and calculates the evaluation value as the matching sentences appear in more tagged text. An extraction rule creation system characterized by Tagged text registration means for extracting the tagged text corresponding to the specified condition from the plurality of tagged text and registering the tagged text in the tagged text storage means,
The extraction rule creation system according to claim 1, wherein the matching sentence position information extraction unit extracts matching sentence position information for each tagged text registered by the tagged text registration unit. The extraction rule creation system according to claim 1 or 2, wherein the extraction rule creation means selects a combination including one or more tags from the created extraction rules. The extraction rule creating means creates an extraction rule by combining predetermined types of independent words among words or tags before and after the word or tag corresponding to the position indicated by the character string position information. 4. The extraction rule creation system according to any one of items 3. The extraction rule creating means according to claim 4, wherein the extraction rule creating means creates an extraction rule by combining a wild card between a word or tag corresponding to the position indicated by the character string position information and a predetermined type of independent word. system. Information added to an arbitrary position in the character string, which is position information indicating the position of the character string to which the information is added and attribute information indicating the attribute of the word corresponding to the position. When a tagged text that is a document including a set of tags and character string position information that is information indicating a position of a character string in the tagged text are given, a word corresponding to the position indicated by the character string position information Or an extraction rule creating step for creating an extraction rule that is a rule for extracting information from the tagged text by combining the tag and the word or the word or tag before and after the tag;
For each tagged text stored in the tagged text storage means, a matching sentence position information extracting step for extracting matching sentence position information, which is information indicating a position of a matching sentence including a word or tag that matches the extraction rule;
An evaluation value calculating step of calculating an evaluation value that is a value obtained by evaluating the extraction rule based on the relevant sentence position information;
In the evaluation value calculating step, the evaluation value is calculated to be higher as there are fewer matching sentences appearing in one tagged text, and the evaluation value is calculated to be higher as matching sentences appear in more tagged text. An extraction rule creation method characterized by A tagged text registration step of extracting tagged text corresponding to a specified condition from a plurality of tagged text and registering the tagged text in the tagged text storage means;
The extraction rule creation method according to claim 6, wherein in the matching sentence position information extraction step, matching sentence position information is extracted for each tagged text registered in the text registration step. Information added to an arbitrary position in the character string, which is position information indicating the position of the character string to which the information is added and attribute information indicating the attribute of the word corresponding to the position. An extraction rule creation program applied to a computer having tagged text storage means for storing tagged text that is a document including a set of tags,
In the computer,
When character string position information, which is information indicating the position of the tagged text and the character string in the tagged text, is given, the word or tag corresponding to the position indicated by the character string position information, the word or An extraction rule creation process for creating an extraction rule that is a rule for extracting information from the tagged text by combining words or tags before and after the tag,
For each tagged text stored in the tagged text storage means, a matching sentence position information extraction process for extracting matching sentence position information, which is information indicating a position of a matching sentence including a word or tag that matches the extraction rule;
Based on the relevant sentence position information, an evaluation value calculation process for calculating an evaluation value that is a value obtained by evaluating the extraction rule is executed,
In the evaluation value calculation process, the evaluation value is calculated to be higher as there are fewer matching sentences appearing in one tagged text, and the evaluation value is calculated to be higher as matching sentences appear in more tagged text. Extraction rule creation program. On the computer,
Extract the tagged text that meets the specified conditions from the tagged text, and execute the tagged text registration process for registering the tagged text in the tagged text storage means,
The extraction rule creation program according to claim 8, wherein in the matching sentence position information extraction process, the matching sentence position information is extracted for each tagged text registered in the text registration process.

Images