JP2010165041A - Correct answer determining apparatus, correct answer determining system, correct answer determining method, and correct answer determining program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a correct answer determining apparatus for determining a combination of words useful for replacing an unknown word included in text information with a known word. <P>SOLUTION: The correct answer determining apparatus segments first text information composed of a plurality of words and second document information in which any word of the plurality of words included in the first text information is replaced with another word into a plurality of words, compares a word segmented by an analysis unit with a previously stored known word, determines the segmented word as the known word if it is stored previously, determines the segmented word as an unknown word if the divided word is not stored previously, and on the basis of the determination result, detects a combination of words such that a word based on the first text information is an unknown word and a word based on the second text information corresponding to a word based on the first text information is a known word. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a correct answer determination apparatus, a correct answer determination system, a correct answer determination method, and a correct answer determination program for converting a word of a sentence included in an electronic document into a known word.

  2. Description of the Related Art Conventionally, morpheme analyzers that analyze text-based sentence information such as blogs published on the Internet as analysis objects (text corpus) and divide them into morphemes that are the smallest units constituting sentences have been used. For example, such a morphological analyzer stores in advance dictionary information in which a predetermined word (known word) and a part of speech of the word are associated with each other. Then, the sentence information to be analyzed is read to extract words constituting the sentence, and each of the extracted words is compared with words included in dictionary information stored in advance. The morphological analysis device discriminates a boundary between a word included in the sentence information to be analyzed and another word based on a word included in the dictionary information and a connection relationship between the parts of speech, and divides the sentence into morphemes.

Here, in many cases, a word expressed in a sentence form is used as a word included in dictionary information serving as a reference for analyzing text information for morphological analysis. For this reason, if you want to analyze text information that contains unknown words with unique text expressions such as blogs and news published on the Internet, and electronic bulletin board services, the corresponding words cannot be detected from the dictionary information. Morpheme division may not be possible.
Therefore, Patent Document 1 proposes a technique in which a user registers a phrase specific to a colloquial according to sentence information to be analyzed, and performs morphological analysis based on the registered phrase. Here, among the words included in the sentence information to be analyzed, the emphasis expression in which the words that are normally written in hiragana are intentionally written in katakana is detected based on the defined correction rules, The morphological analysis is performed after correcting the notation. Patent Document 2 calculates a probability distribution of character strings before and after a character string in a sentence to be analyzed, and performs part-of-speech estimation based on the assumption that the probability distributions of words belonging to the same part of speech are similar. Has been proposed.

Japanese Patent No. 2870375 JP 2006-031295 A

  However, when performing morphological analysis using words specific to colloquial input from the user as in Patent Document 1, it is difficult for the user to input all words specific to colloquial and the user who inputs the word The analysis varies depending on experience and preferences, and general-purpose morphological analysis cannot be performed. In addition, as in Patent Document 2, the technique of estimating the part of speech based on the probability distribution is useful for learning technical terms and new words, but for analyzing sentences such as blogs in which unique expressions appear irregularly. Not suitable.

  The present invention has been made in view of such circumstances, and a correct answer determination apparatus, a correct answer determination system, a correct answer determination method, and a correct answer for determining a combination of useful words that replace unknown words included in sentence information with known words. A judgment program is provided.

  In order to solve the above-described problem, the present invention provides a known word storage unit that stores a known word that is predetermined as a known word, first sentence information that includes a plurality of words, The input of the first document information and the second document information is received by receiving the input of the second document information in which any one of the plurality of words included in the first sentence information is replaced with another word. Is divided into a plurality of words, and the words divided by the analysis unit are compared with the known words stored in the known word storage unit. The word based on the first sentence information is an unknown word based on the determination result of the correct answer determining unit that determines that there is an unknown word if it is not stored in the known word storage unit, and the correct answer determining unit A set of words in which the word based on the second text information corresponding to the word based on the first text information is a known word A detection unit for detecting a combined, characterized in that it comprises a.

  In addition, the present invention provides an analysis target information storage unit in which first sentence information is stored, a first word to be corrected among a plurality of words included in the first sentence information, and a first word. A correction rule storage unit that stores a correction rule associated with a second word to be replaced, a first word included in the first sentence information, and the extracted first word as a second A correction rule applying unit that performs correction processing for generating second sentence information replaced with words and inputs the first sentence information and the second sentence information to the analysis unit is further provided.

  Further, the present invention associates a word based on the first sentence information as a first word and a word based on the detected second sentence information as a second word among the combinations of words detected by the detection unit. And a correction rule registration unit that generates the correction rule and causes the correction rule storage unit to generate the correction rule.

  In the present invention, the correction rule registration unit associates a word obtained by dividing a word based on the first sentence information detected by the detection unit and a word based on the second sentence information into words having a shorter number of characters. The modified rule is generated and stored in the modified rule storage unit.

  Further, according to the present invention, the correction rule registration unit concatenates a word based on the first text information detected by the detection unit and a word based on the second text information to another word adjacent to each other in each text. Then, the associated correction rule is generated and stored in the correction rule storage unit.

  In the present invention, the correction rule application unit corrects the correction when the correct answer determination unit determines that both the word based on the first sentence information and the word based on the second sentence information are unknown words. The correction process is performed again based on another correction rule stored in the rule, and the correction rule registration unit detects the first detected by the detection unit based on the processing result of the correction process performed again by the correction rule application unit. A correction rule that associates a word based on the sentence information as the first word and a word based on the detected second sentence information as the second word is generated and stored in the correction rule storage unit.

  Moreover, this invention respond | corresponds to a 1st word, when the correct answer determination part determines that both the word based on 1st text information and the word based on 2nd text information are unknown words. The correction rule registration unit further includes an input unit that receives an input of the second word, and the correction rule registration unit sets the word based on the first sentence information detected by the detection unit as the first word, and the word input to the input unit after correction A correction rule associated with the second word is generated and stored in the correction rule storage unit.

  The present invention also provides an analysis object information storage device that stores first sentence information composed of a plurality of words, a first word to be corrected, and a second word that replaces the first word. Is stored in a correction rule storage device, a known word storage unit storing a known word that is predetermined as a known word, and a first sentence information included in the first sentence information. A first rule composed of a plurality of words and a correction rule applying unit that performs a correction process for generating second sentence information by extracting one word and replacing the extracted first word with a second word The information and the second document information in which any one of the plurality of words included in the first sentence information is replaced with another word are received, and the input first sentence information and second The document information is divided into a plurality of words, an analysis unit, a word divided by the analysis unit, and a known word storage unit A correct answer determination unit that determines that the stored word is a known word if it is stored in the known word storage unit, and that it is an unknown word if it is not stored in the known word storage unit; Based on the determination result of the determination unit, a combination of words in which the word based on the first sentence information is an unknown word and the word based on the second sentence information corresponding to the word based on the first sentence information is a known word And a correction rule registration device that generates a correction rule based on the detection result of the detection unit of the correct answer determination device and stores the correction rule in a correction rule storage device. Is a correct answer determination system.

  In addition, the present invention is a correct answer determination method of a correct answer determination apparatus including a known word storage unit in which a known word that is predetermined as a known word is stored. The input of the sentence information and the second document information in which any one of the plurality of words included in the first sentence information is replaced with another word is received. Dividing the document information of 2 into a plurality of words, comparing the divided words with the known words stored in the known word storage unit, and if the words are stored in the known word storage unit, A step of determining that there is an unknown word if it is not stored in the known word storage unit, and based on the determination result, the word based on the first sentence information is an unknown word, and the first sentence information A combination of words in which the word based on the second sentence information corresponding to the word based on is a known word Detecting allowed, characterized by comprising a.

  In addition, the present invention provides a computer of a correct answer determination apparatus including a known word storage unit that stores a known word that is predetermined as a known word, first sentence information composed of a plurality of words, The input of the first document information and the second document information is received by receiving the input of the second document information in which any one of the plurality of words included in the first sentence information is replaced with another word. Are divided into a plurality of words, and the divided words are compared with the known words stored in the known word storage unit, and determined to be known words if stored in the known word storage unit. The word based on the first sentence information is an unknown word based on the determination result based on the determination result if the word is not stored in the known word storage unit, and the word based on the first sentence information A combination of words whose words based on the corresponding second sentence information are known words Is correct determination program to be executed a step of leaving, the.

  As described above, according to the present invention, any one of the first sentence information composed of a plurality of words and the plurality of words included in the first sentence information is replaced with another word. The second document information is divided into a plurality of words, the word divided by the analysis unit is compared with a known word stored in advance, and if it is stored in advance, it is determined as a known word, If it is not stored in advance, it is determined as an unknown word, and based on the determination result, the word based on the first text information is an unknown word, and the second text information corresponding to the word based on the first text information Since a combination of words whose known words are known words is detected, unknown words are replaced with known words by replacing words included in the first sentence information to be analyzed with other words. Text information containing unknown words is placed in a sentence composed of known words. It is possible to detect a useful word for.

It is a block diagram which shows the structure of the correct answer determination system by one Embodiment of this invention. It is a figure which shows the example of data of the correction rule by one Embodiment of this invention. It is a figure which shows the operation example of the label provision apparatus by one Embodiment of this invention. It is a figure which shows the operation example of the scoring apparatus by one Embodiment of this invention. It is a figure which shows the operation example of the rule specialization apparatus by one Embodiment of this invention. It is a figure which shows the operation example of the rule coupling | bonding apparatus by one Embodiment of this invention. It is a figure which shows the operation example of the rule generalization apparatus by one Embodiment of this invention. It is a figure which shows the operation example of the active learning apparatus by one Embodiment of this invention. It is a figure which shows the operation example of the correct answer determination system by one Embodiment of this invention.

Hereinafter, an embodiment of the present invention will be described with reference to the drawings.
FIG. 1 is a block diagram showing a configuration of a correct answer determination system 1 according to the present embodiment. The correct answer determination system 1 includes a text corpus storage device 10, a correction rule storage device 20, a label assignment device 30, a scoring unit 41, a rule specialization device 50, a rule combination device 60, and a rule generalization device 70. And an active learning device 80 and an input terminal 89.

  The text corpus storage device 10 is a computer device that stores natural language text information to be analyzed by this device. The sentence information is, for example, sentence information included in web contents such as news sites and blogs acquired via the Internet. The sentence information stored in the text corpus storage device 10 is composed of a plurality of words and sentences, and includes various sentence expressions such as colloquial expressions, secret words, abbreviations, slang words, youth words, and slang.

  The correction rule storage device 20 includes a first word to be corrected among a plurality of words included in the sentence information, and a sentence word that has the same semantic content as the first word and can be analyzed by the morphological analysis unit 34. A computer device that stores a plurality of correction rules associated with a second word. The correction rule storage device 20 stores in advance some basic correction rules in the initial state. For example, if the example of the correction rule is expressed as “(first word) → (second word)”, “Cha → Teha”, “Ka → Ka”, “ゎ → Ha”, “ゎ → wa” Such correction rules are stored in advance. The correction rule storage device 20 stores new correction rules generated by the scoring device 40, the rule specializing device 50, the rule combining device 60, the rule generalizing device 70, and the active learning device 80.

  FIG. 2 is a diagram illustrating an example of correction rule data stored in the correction rule storage device 20. Information such as a registration type, a score, and a label is stored in association with the correction rule. The registration type is a type indicating the cause of the correction rule stored in the correction rule storage device 20, “initial” indicating that it is a primitive rule stored in advance, or specialization processing by the rule specializing device 50 There is a “specialization” or the like indicating that the correction rule is generated and stored by. The score is a value indicating the degree of effectiveness of the correction rule calculated by the scoring device 40. The higher the numerical value, the higher the effectiveness. In this example, the score indicates the probability that a known word is converted to an unknown word by applying the correction rule. The label is information provided by the label attaching device 30 and is information indicating a change in the state between the known word and the unknown word by applying the correction rule. In the present embodiment, it is assumed that all such information is stored in the correction rule storage device 20, but the registration type, score, label, and the like are stored and managed separately by providing different storage devices. You may do it.

  The label assigning device 30 is a computer device that replaces words included in the sentence information stored in the text corpus storage device 10 based on the correction rules stored in the correction rule storage device 20 and determines the usefulness of the replacement. . The label assigning device 30 includes an applied sentence extraction unit 31, a correction rule application unit 32, a known word storage unit 33, a morpheme analysis unit 34, and a label determination unit 35.

  The application sentence extraction unit 31 reads out the sentence information to be analyzed stored in the text corpus storage device 10 and extracts a pre-application sentence that is a sentence including the first word of the correction rule stored in the correction rule storage device 20. To do. For example, when the correction rule to be applied to the analysis target sentence information (TX) is “X → Y”, the application sentence extraction unit 31 uses the first word “X” from the analysis target sentence information (TX). Is extracted (grep X from TX), and the pre-application sentence (TX1) is extracted. The application sentence extraction unit 31 inputs the extracted pre-application sentence to the correction rule application unit 32 and the morpheme analysis unit 34.

  The modification rule application unit 32 is an application in which the first sentence included in the pre-application sentence extracted from the sentence information to be analyzed by the application sentence extraction unit 31 based on the modification rule is replaced with a corresponding second word. Generate postscript. The correction rule application unit 32 replaces the first word (X) included in the pre-application sentence (TX1) extracted by the application sentence extraction unit 31 with the second word (Y) based on the correction rule (sed X / Y TX1) and post-application sentence (TX2) are generated. For example, the correction rule applying unit 32 applies the correction rule “Cha → Teha” to the sentence before applying the colloquial sentence “Hurry must be”, and generates the sentence “After the sentence hastened” after applying the sentence. The correction rule application unit 32 inputs the generated post-application sentence to the morpheme analysis unit 34.

The known word storage unit 33 stores dictionary information including a list of a plurality of known words associated with information such as parts of speech. Here, the known word is a predetermined word used for morphological analysis or the like. Known words are generally expressed in sentence language.
The morpheme analysis unit 34 receives input of the pre-application sentence input from the application sentence extraction unit 31 and the post-application sentence input from the correction rule application unit 32, and the input pre-application sentence and post-application sentence are input, Morphological analysis based on dictionary information stored in the known word storage unit 33 is performed, and each sentence is divided into morphemes. Morphological analysis, for example, divides a sentence in natural language into known words included in the dictionary information based on part-of-speech connection restrictions according to natural language rules, determines word boundaries, and composes sentences This is a process of extracting a morpheme that is a minimum unit. For example, the sentence “There are two chickens in the garden” can be obtained by performing morphological analysis to obtain “garden” (noun), “ni” (particle), “ha” (particle), “two” (noun). ), “Feather” (noun), “chicken” (noun), “ga” (particle), “is” (verb), and so on.

  The label determination unit 35 compares a word based on each morpheme divided by the morpheme analysis unit 34 with a word stored in the known word storage unit 33, and the same word as the divided word is the morpheme analysis unit 34. If it is not stored in the known word storage unit 33, it is determined that it is an unknown word. For example, the label determination unit 35 uses the morpheme analysis result sentence (IN1) for the pre-application sentence to which the correction rule “X → Y” is applied and the morpheme for the post-application sentence to which the correction rule “X → Y” is similarly applied. Using the sentence (IN2) as the analysis result as an input, the word (P1) including the first word (X) of the correction rule is extracted from the morphological analysis result (IN1) for the pre-application sentence.

  In addition, the label determination unit 35 determines whether, for each morpheme included in the extracted word (P1), the morpheme or the same word as the word including the morpheme is stored in the known word storage unit 33 as a known word. Is determined to determine whether the morpheme or the word including the morpheme is a known word or an unknown word. Similarly, the label determination unit 35, for each morpheme of the word (P2) corresponding to the word (P1) included in the morpheme analysis result sentence (IN1) among the morpheme analysis result sentence (IN2) for the post-application sentence. Next, it is determined whether it is a known word or an unknown word. Then, the label determination unit 35 assigns a label indicating a change in state between the known word and the unknown word for the word combination “P1 → P2”.

  Here, the label is given a symbol “O”, “□”, “×”, or “Δ”, and “O” indicates that P1 is an unknown word and P2 is a known word. Indicates. “□” indicates that both P1 and P2 are known words. “X” indicates that P1 is a known word and P2 is an unknown word. “Δ” indicates that both P1 and P2 are unknown words. For example, when the correction rule “I → I” is applied to “Kawaii Koinu” (P1) and “Cute Koinu” (P2) is obtained, the label determination unit 35 stores the known word. With reference to the information stored in the unit 33, “Kaiikoinu” is determined to be an unknown word, and “Kawaii Koinu” is determined to be a known word. In this case, the label determination unit 35 assigns the label “◯” to “P1 → P2”. Here, the label determination unit 35 associates the replaced word combination “P1 → P2” with the label given to the word combination, and temporarily stores them in the correction rule storage device 20. Anyway.

  The scoring device 40 evaluates the correction rule based on the processing result of the label assigning device 30 and assigns a score, and causes the correction rule storage device 20 to store the correction rule that calculates a score exceeding a predetermined threshold. The scoring device 40 includes a scoring unit 41 and a correction rule registration unit 42. The scoring unit 41 is effective for each correction rule stored in the correction rule storage device 20 based on the labels assigned by the label assignment device 30 for each of the plurality of pieces of text information stored in the text corpus storage device 10. A score indicating the evaluation of is calculated (scoring). For example, the scoring unit 41 calculates the number of pre-application sentences to which the correction rule is applied and the number of correct answers whose label is “O” by performing the replacement by applying the correction rule to the pre-application sentence. Then, the ratio of the number of correct answers to the number of pre-application sentences (correct answer rate) is calculated as a score. For example, when the correction rule “i → i” is applied to two pre-application sentences and the number of correct answers is one, the correct answer rate is 50%.

  The correction rule registration unit 42 registers the correction rule in the correction rule storage device 20 according to the score calculated by the scoring unit 41. For example, the correction rule registration unit 42 stores a correction rule in which the score calculated by the scoring unit 41 exceeds a predetermined threshold (for example, 50%) in the correction rule storage device 20 as a new correction rule. Further, the correction rule registration unit 42 may have a function of detecting an unnecessary rule out of the correction rules stored in the correction rule storage device 20 and deleting it from the correction rule storage device 20. For example, the correction rule registration unit 42 makes the score calculated again by the label assigning device 30 out of the correction rules already stored in the correction rule storage device 20 fall below a predetermined threshold (for example, 20%). In such a case, the correction rule may be deleted from the correction rule storage device 20.

  The rule specializing device 50 performs specialization processing for the correction rules stored in the correction rule storage device 20. In the correction rule specialization process, the rule specializing device 50 performs a new correction in which another adjacent character (string) is connected to the first word included in the correction rule stored in the correction rule storage device 20. A rule is generated, and the labeling device 30 and the scoring device 40 are operated using the generated new correction rule, and the correction rule for which a score exceeding a predetermined threshold is calculated is corrected as a new correction rule It is stored in the rule storage device 20.

  For example, the rule specializing device 50 adds the first word (X) included in the correction rule “X → Y” from the pre-application sentence extracted by the label assigning device 30 before or after being adjacent in the pre-application sentence. A word (wX or Xw) obtained by concatenating characters (w) (w is one or more characters (strings)) is detected, and the label determination unit 35 and the scoring device 40 are input using the detected word (wX or Xw) as an input. And make it work. For example, the rule specializing apparatus 50 uses a sentence including “cha” in the pre-application sentence (for example, “does not go”) for the correction rule “cha → teha” from the result sentence obtained by the morphological analysis by the label assigning apparatus 30. ‘Chat’, ‘Chan’, ‘Must see’, ‘Baby’, ‘Ochawan’, ‘Shakucha’, ‘Kakucha’, etc. ”,“ Taneha ”,“ Need to see ”,“ Red Tane ”,“ Otewa Wan ”,“ Need to do ”,“ Kutehaku Hate ”) The label determination unit 35 gives labels to the detected pre-application sentence and post-application sentence. As a result, the rule specializing apparatus 50 obtains the label “◯” for the sentence “I have to go → I have to go”, for example. Similarly, label “×” for “Chanto → Taneto”, label “○” for “I have to see → I have to see”, and label “×” for “Baby → Red” , “Ochawan → Otewawan” is labeled “×”, “Shakucha → I have to do” is labeled “○”, and “Kakukucha → Kutehakuhaha” is labeled “×” Get.

  The rule specializing device 50 causes the scoring device 40 to calculate a score based on the label thus obtained. In the above example, the scoring device 40 applies the correction number “4” to the correction rule “Kaku → Kateha” in which “ku” is connected before the correction rule “Cha → Teha”. Therefore, the score is calculated as 75 (%). Similarly, the number of applications for “Nakucha → Nenai”, in which “Naku” is connected before “Chaya → Teha”, is 3, and the number of correct answers is 3, so the score is 100. For “Chan → Tehan” in which “Chan → Teha” is connected to “Chan → Tehan”, the number of applications is 2 and the number of correct answers is 0, so the score is 0. For “Ocha → Oteha”, in which “O” is connected before “Cha → Teha”, the number of applications is 1 and the number of correct answers is 0, so the score is 0. With regard to “Chawa → Tehawa” in which “Cha → Teha” is connected to “Wa”, the number of applications is 2 and the number of correct answers is 0, so the score is 0.

  The rule specializing device 50 stores, in the correction rule storage device 20, a correction rule that exceeds a predetermined threshold among the correction rules whose scores are calculated in this way by the scoring device 40 as a new correction rule. For example, if the threshold value of the score is 70, the correction rule storage device 20 stores “Kaku → Kaku wa” and “Kaku kan → Kaku” among the above correction rules. Here, the rule specializing device 50 may not perform the specialization process on the correction rule for which a score equal to or higher than a predetermined threshold is calculated. For example, when the threshold value is 100, in the above example, the specialization process based on “necessary → necessary” may not be performed. In order to prevent the correction rule from explosively increasing, a restriction may be provided such that a correction rule having a character string length equal to or greater than a predetermined threshold is not generated. Further, the correction rule to be specialized processing is performed based only on the correction rule stored in advance in the correction rule storage device 20 in the initial state and the correction rule generated by the rule generalization device 70 described later, You may not perform based on the correction rule produced | generated by the other method. For example, in the above example, the specialization process based on “Cha → Teha” is performed, but the specialization process based on the correction rule “Kaku → Cute” generated thereby is not performed. This is because when “Cha → Teha” is specialized, a rule that specializes “Cha → Teha” can be obtained.

  The rule combining device 60 performs a process of combining the correction rules stored in the correction rule storage device 20. The rule combining device 60 includes a correction rule selecting unit 61 and a correction rule combining unit 62. In the correction rule combining process, the correction rule selection unit 61 applies a correction rule to be further applied in order to make the unknown word of the post-application sentence a known word with respect to the word given “△” by the label assigning device 30. The correction rule is detected from the correction rule storage device 20, and the correction rule combining unit 62 combines the correction rules and stores them in the correction rule storage device 20 as new correction rules.

For example, the correction rule selection unit 61 stores the correction rule stored in the correction rule storage device 20 for the unknown word (w ₁ , w ₂ ... W _n ) to which “Δ” is assigned by the label assigning device 30. Are applied to the label determination unit 35 in order, and the correction rule combination unit 62 newly corrects the correction rule when the determination result by the label determination unit 35 becomes “◯”. It is stored in the correction rule storage device 20 as a rule.

  Specifically, the unknown word “hidhi” included in the original text “... Say hid ...” of the text information read from the text corpus storage device 10 When the correction rule “i → i” is applied by the correction rule application unit 32, it is corrected to “hid”, but when “hid” is an unknown word, the label given by the label determination unit 35 Becomes “△”. The correction rule selection unit 61 detects the correction rule having the highest score among the correction rules stored in the correction rule storage device 20 for the unknown word to which the label “Δ” is assigned by the label addition device 30 in order. A correction rule is applied to 30 to give a label. For example, in the case of the above-described example, the correction rule selection unit 61 detects a correction rule that includes the characters “hi” or “do” in the first word among the correction rules stored in the correction rule storage device 20. Here, the rule combining device 60 determines that the correction rules “hi → hi” (score 2), “hid → hid” (score 40), “do → do” (score 1), “hidden → bad” (score 100). ), When “hidoku → bad” (score 100) is detected, the correction rule “hid →→ bad” with a high score and the correction rule “hidoku → bad” are detected as the correction rules to be applied.

  The correction rule selection unit 61 detects the correction rule detected as the correction rule to be applied by applying the correction rule “i → i” to the unknown word “hid” by the label assignment device 30. ”Is further applied to the label applying device 30. As a result, the correction rule selection unit 61, based on the word newly assigned with the label to the label attaching device 30, creates a new correction rule “... Is generated, and is stored in the correction rule storage device 20. Here, the correction rule generated by the rule combining device 60 is a very specific sentence, but is stored in the correction rule storage device 20 as a correction rule having a shorter appropriate length by the processing of the rule generalization device 70 described later. Is done. Further, when the unknown word is not resolved even by the correction rule detected by the rule combining device 60, another correction rule may be detected recursively and the process performed by the label applying device 30 may be performed.

  The rule generalization device 70 performs a generalization process of the correction rules stored in the correction rule storage device 20. In the modification rule generalization process, the rule generalization apparatus 70 has a prefix or suffix of one or more characters in which the first word and the second word are common among the modification rules stored in the modification rule storage device 20. The correction rule is detected, a correction rule is generated by combining the first word and the second word from which the common prefix or suffix word is deleted, and is stored in the correction rule storage device 20.

  For example, in the rule generalization device 70, the correction rule “X → Y” stored in the correction rule storage device 20 is “wxu → wyu” (where w, x, y, u are each one or more characters ( Column)), a correction rule “x → y” is newly generated and stored in the correction rule storage device 20. Specifically, the rule generalization device 70 uses the prefix “correct” and the suffix “I don't know” for the correction rule “I don't know if it is correct → I don't know if it is correct” stored in the correction rule storage device 20. ”Is deleted, and a correction rule“ Inokam → Inokaha ”is generated and stored in the correction rule storage device 20. In addition, for the correction rule “Troublesome ぉ ~ → I'm in trouble” stored in the correction rule storage device 20, the suffix “Troublesome” is deleted and the correction rule “Chayo ぅ ~ →" It is generated and stored in the correction rule storage device 20.

  Here, the correction rule to be generalized by the rule generalization device 70 is applied only to the correction rule generated by the rule combining device 60 or the active learning device 80 and stored in the correction rule storage device 20. Also good. In addition, the correct answer determination system 1 further performs specialization processing on the modified rule obtained in this way by the rule specializing device 50 or combining processing by the rule combining device 60, thereby creating a new one. A rule may be generated.

  The active learning device 80 cannot generate a new correction rule by the rule specializing device 50, the rule combining device 60, and the rule generalizing device 70, and can detect an unknown word whose label given by the label assigning device 30 is “Δ”. An input of a word to be replaced is accepted, a correction rule having the unknown word as the first word and the input word as the second word is generated and stored in the correction rule storage device 20. The active learning device 80 includes a priority determination unit 81 and is connected to the input terminal 89.

  The priority determination unit 81 counts the number of appearances in the original sentence with respect to the original word to which the label “Δ” is given by the label determination unit 35, and descends the appearance frequency of the original word to which the label “Δ” is given. Sort into. The active learning device 80 outputs, to the input terminal 89, the predetermined top N words (N is an arbitrary number) among the original words sorted in descending order by the priority determination unit 81.

  The input terminal 89 is a computer terminal including a control unit, a calculation unit, a display unit (display), an input unit (keyboard), and the like, and causes the display unit to display a word input from the active learning device 80 as a first word. . In addition, the input terminal 89 receives an input of a second word that replaces the first word displayed on the display unit, and transmits the input second word to the active learning device 80. The active learning device 80 generates a correction rule in which the first word transmitted to the input terminal 89 is associated with the second word transmitted from the input terminal 89 corresponding thereto, and the correction rule storage device 20 Remember me.

Next, with reference to FIGS. 3 to 8, an example of the operation of each unit included in the correct answer determination system 1 will be described.
FIG. 3 is a diagram illustrating an operation example of the label applying device 30. The applied sentence extraction unit 31 reads the sentence information to be analyzed from the text corpus storage device 10 (step S1), and reads the correction rule from the correction rule storage device 20 (step S2). The applied sentence extraction unit 31 extracts a pre-application sentence to which the correction rule can be applied from the sentence information to be analyzed, and inputs it to the morpheme analysis unit 34 (steps S3 and S4). Further, the application sentence extraction unit 31 inputs the pre-application sentence and the correction rule to the correction rule application unit 32 (step S5).

  The correction rule application unit 32 applies the correction rule to the pre-application sentence input from the application sentence extraction unit 31, generates a post-application sentence, and inputs it to the morpheme analysis unit 34 (steps S6 and S7). The morpheme analysis unit 34 performs morpheme analysis on each of the pre-application sentence input from the application sentence extraction unit 31 and the post-application sentence input from the correction rule application unit 32. The morpheme analyzer 34 inputs the result of the morpheme analysis for the pre-application sentence to the label determination unit 35 (steps S8 and S9). Further, the morpheme analysis unit 34 inputs the result of the morpheme analysis on the post-application sentence to the label determination unit 35 (steps S10 and S11). The label determination unit 35 performs label determination by comparing each of the words included in the analysis result of the pre-application sentence input from the morpheme analysis unit 34 and the analysis result of the post-application sentence. A label-added sentence that is a combination of the word and the sentence after application is output (step S12). Here, the pre-application sentence and the post-application sentence input to the morphological analysis unit 34 in step S4 and step S5 are the sentence information to be analyzed and any one of the words included in the sentence information to be analyzed is a user. The sentence information may be replaced with another word.

  FIG. 4 is a diagram illustrating an operation example of the scoring device 40. The scoring unit 41 receives as input the label-added sentence to which the label is given by the label determination unit 35 and the correction rule for which the score stored in the correction rule storage device 20 is not calculated (steps S22 and S23). A score is calculated for each rule and input to the correction rule registration unit 42 (steps S23 and S24). Based on the score calculated by the scoring unit 41, the correction rule registration unit 42 stores the combination of the pre-application sentence and the post-application sentence in the correction rule storage device 20 (step S25).

  FIG. 5 is a diagram illustrating an operation example of the rule specializing device 50. The rule specializing apparatus 50 receives the label-added sentence to which the label is given by the label determination unit 35 as an input (step S31), and generates a correction rule that generates a word concatenated words adjacent to the word in the label-added sentence. Are input to the scoring device 40 (steps S32 and S33). The scoring device 40 calculates a score for the input correction rule, and stores the correction rule for which the score exceeding the predetermined threshold is calculated in the correction rule storage device 20 (step S34).

  FIG. 6 is a diagram illustrating an operation example of the rule combining device 60. The correction rule selection unit 61 of the rule combining device 60 receives the label-added sentence to which the label is attached by the label attaching apparatus 30 as an input (step S41), extracts a combination of words whose label is “Δ”, and corrects the sentence to be corrected To the label applying device 30 (steps S44 and S45). In addition, the correction rule selection unit 61 reads from the correction rule storage device 20 a correction rule to be applied to the combination of words whose label is “Δ” from the correction rule storage device 20 (step S42). (Step S43). The labeling apparatus 30 generates a labeling sentence by applying the correction rule input from the correction rule selection unit 61 to the correction target sentence input from the correction rule selection unit 61 and inputs the sentence to the rule combining device 60 ( Steps S46 and S47). The correction rule combining unit 62 of the rule combining device 60 combines the correction rule based on the input label-added sentence, associates it with the corresponding portion of the original sentence stored in the correction rule storage device 20 (step S48), and corrects it. The data is stored in the rule storage device 20 (step S49).

  FIG. 7 is a diagram illustrating an operation example of the rule generalization device 70. The rule generalization device 70 reads the correction rule stored in the correction rule storage device 20 (step S51), and stores the correction rule generated by performing the rule generalization processing in the correction rule storage device 20 (step S52). ).

  FIG. 8 is a diagram illustrating an operation example of the active learning device 80. The priority determination unit 81 of the active learning device 80 receives a set of words determined as unknown words by the label determination unit 35 (step S61), sorts the unknown words in order of priority, and determines the number of unknown words. Are sent to the input terminal 89 (steps S62 and S62). When the input terminal 89 displays the unknown word transmitted from the priority determination unit 81 on the display unit, the user inputs a word corresponding to the unknown word to the input terminal 89. The input terminal 89 transmits the word input from the user to the active learning device 80 (step S64). The active learning device 80 generates a correction rule that associates the unknown word transmitted to the input terminal 89 with the word transmitted from the input terminal 89 in response to this, and stores the correction rule in the correction rule storage device 20 (step S65). .

  Next, an operation example of the correct answer determination system 1 will be described with reference to FIG. In the initial state, a general-purpose correction rule is stored in advance in the correction rule storage device 20 (step S71). Here, since the correct answer determination system 1 of the present embodiment includes the active learning device 80, the active learning device 80 may input a rule that the correction rule of the initial state is not stored. The rule specializing device 50 reads one correction rule stored in the correction rule storage device 20, inputs it to the label assigning device 30, obtains a label assignment sentence, and performs rule specialization processing (step S72).

  The rule combining device 60 applies the other correction rule stored in the correction rule storage device 20 to the sentence with the label “Δ” among the label addition sentences acquired in step 73, and the label addition apparatus When the label is “◯”, the correction rule is stored in the correction rule storage device 20 as a new correction rule (step S73). The rule specializing device 50 determines whether the specialization processing and the combination processing have been completed for all the correction rules that need specialization among the correction rules stored in the correction rule storage device 20 ( Step S74). If the rule specializing apparatus 50 determines that processing has not been performed for all the correction rules (step S74: NO), the process returns to step S72. If the rule specializing apparatus 50 determines that processing has been performed for all the correction rules (step S74: YES), the process proceeds to step S75.

  The rule generalization device 70 refers to the correction rule registration type of the correction rules stored in the correction rule storage device 20, and the correction rule generated and stored by the rule combining device 60 or the priority determination unit 81. A generalization process is performed (step S75). At this time, the generalization process is not performed for the correction rules stored in advance in the correction rule storage device 20 from the initial state. The rule generalization device 70 determines whether or not generalization processing has been performed for all of the correction rules that need to be generalized among the correction rules stored in the correction rule storage device 20 (step S76). The rule generalization device 70 returns to step S75 if the generalization processing is not performed for all the correction rules (step S76: NO). The rule generalization device 70 proceeds to step S77 if the generalization processing is performed for all the correction rules (step S76: YES).

  If a new correction rule is stored in the correction rule storage device 20 from step S71 to step S76, the rule specializing apparatus 50 performs the process of step S72 again (step S77: YES). If a new correction rule is not stored in the correction rule storage device 20 (step S77: NO), the active learning device 80 asks the input terminal 89 to perform active learning processing, and information input by the user Based on the above, a correction rule is generated and stored in the correction rule storage device 20 (step S78). If a new correction rule is stored in the correction rule storage device 20 in step S78 (step S79: YES), the process returns to step S75, and generalization processing by the rule generalization device 70 is performed. In step S78, if a new correction rule is not stored in the correction rule storage device 20 (step S79: NO), the correct answer determination system 1 ends the process. As described above, by repeatedly executing Step S72 to Step S74 and Step S75 to Step S79, it is possible to increase the accuracy of the correction rule and generate various correction rules.

  As described above, according to the present invention, the correct answer determination system 1 generates various correction rules by performing correction rule specialization, combination, generalization, and active learning processing, and the correction rule storage device. 20 can be stored.

  It should be noted that a program for realizing the function of the processing unit in the present invention is recorded on a computer-readable recording medium, and the program recorded on the recording medium is read into a computer system and executed, whereby label determination and correction are performed. A rule generation process may be performed. Here, the “computer system” includes an OS and hardware such as peripheral devices. The “computer system” includes a WWW system provided with a homepage providing environment (or display environment). The “computer-readable recording medium” refers to a portable medium such as a flexible disk, a magneto-optical disk, a ROM, and a CD-ROM, and a storage device such as a hard disk built in the computer system. Further, the “computer-readable recording medium” refers to a volatile memory (RAM) in a computer system that becomes a server or a client when a program is transmitted via a network such as the Internet or a communication line such as a telephone line. In addition, those holding programs for a certain period of time are also included.

  The program may be transmitted from a computer system storing the program in a storage device or the like to another computer system via a transmission medium or by a transmission wave in the transmission medium. Here, the “transmission medium” for transmitting the program refers to a medium having a function of transmitting information, such as a network (communication network) such as the Internet or a communication line (communication line) such as a telephone line. The program may be for realizing a part of the functions described above. Furthermore, what can implement | achieve the function mentioned above in combination with the program already recorded on the computer system, and what is called a difference file (difference program) may be sufficient.

DESCRIPTION OF SYMBOLS 1 Correct answer determination system 10 Text corpus storage device 20 Correction rule storage device 30 Label giving device 31 Applicable sentence extraction unit 32 Correction rule application unit 33 Known word storage unit 34 Morphological analysis unit 35 Label determination unit 40 Scoring device 41 Scoring unit 42 Correction rule registration unit 50 Rule specialization device 60 Rule combination device 61 Correction rule selection unit 62 Correction rule combination unit 70 Rule generalization device 80 Active learning device 81 Priority determination unit 89 Input terminal

Claims (10)

A known word storage unit that stores known words that are predetermined as known words;
Input of first sentence information composed of a plurality of words and second document information in which any one of the plurality of words included in the first sentence information is replaced with another word. An analysis unit that divides the received first sentence information and the second document information into a plurality of words;
The word divided by the analysis unit is compared with the known word stored in the known word storage unit, and if it is stored in the known word storage unit, it is determined as a known word, and the known word storage A correct answer determination unit that determines an unknown word if not stored in the unit;
Based on the determination result of the correct answer determination unit, the word based on the first sentence information is an unknown word, and the word based on the second sentence information corresponding to the word based on the first sentence information is a known word. A detection unit for detecting a combination of words,
A correct answer determination apparatus comprising: An analysis object information storage unit for storing the first sentence information;
A correction rule in which a correction rule in which a first word to be corrected among a plurality of words included in the first sentence information is associated with a second word that replaces the first word is stored. A storage unit;
The first word included in the first sentence information is extracted, a correction process is performed to generate second sentence information in which the extracted first word is replaced with the second word, and the first word A correction rule application unit that inputs sentence information and the second sentence information to the analysis unit;
The correct answer determination apparatus according to claim 1, further comprising: Of the combinations of words detected by the detection unit, the word based on the first sentence information is set as the first word, and the word based on the detected second sentence information is associated as the second word. A correction rule registration unit that generates the correction rule and causes the correction rule storage unit to generate the correction rule;
The correct answer determination apparatus according to claim 2, further comprising: The correction rule registration unit is a correction in which a word obtained by dividing a word based on the first sentence information detected by the detection unit and a word based on the second sentence information into words having a shorter number of characters is associated. The correct answer determination apparatus according to claim 3, wherein a rule is generated and is used in the correction rule storage unit. The correction rule registration unit is configured to connect a word based on the first sentence information detected by the detection unit and a word based on the second sentence information to another word adjacent to each other in each sentence. The correct answer determination apparatus according to claim 3 or 4, wherein the associated correction rule is generated and stored in the correction rule storage unit. The correction rule application unit, when the correct answer determination unit determines that both the word based on the first sentence information and the word based on the second sentence information are unknown words, the correction rule Perform another correction process based on the other correction rules stored in
The correction rule registration unit detects a word based on the first sentence information detected by the detection unit based on a processing result of a second correction process performed by the correction rule application unit as the first word, and detects the first word. 6. The correction rule storing the word based on the second sentence information as the second word and generating the correction rule to cause the correction rule storage unit to generate the correction rule. The correct answer determination apparatus according to item 1. When the correct answer determination unit determines that both the word based on the first sentence information and the word based on the second sentence information are unknown words, the first word corresponding to the first word An input unit that accepts input of two words;
The correction rule registration unit associates a word based on the first sentence information detected by the detection unit as the first word, and associates a word input to the post-correction input unit as the second word. The correct answer determination apparatus according to any one of claims 3 to 6, wherein the correction rule is generated and stored in the correction rule storage unit. An analysis target information storage device in which first sentence information composed of a plurality of words is stored;
A correction rule storage device that stores a correction rule in which a first word to be corrected and a second word that replaces the first word are associated;
A known word storage unit for storing known words that are predetermined as known words;
A correction rule applying unit that performs a correction process for generating the second sentence information by extracting the first word included in the first sentence information and replacing the extracted first word with the second word; ,
Input of first sentence information composed of a plurality of words and second document information in which any one of the plurality of words included in the first sentence information is replaced with another word. An analysis unit that divides the received first sentence information and the second document information into a plurality of words;
The word divided by the analysis unit is compared with the known word stored in the known word storage unit, and if it is stored in the known word storage unit, it is determined as a known word, and the known word storage A correct answer determination unit that determines an unknown word if not stored in the unit;
Based on the determination result of the correct answer determination unit, the word based on the first sentence information is an unknown word, and the word based on the second sentence information corresponding to the word based on the first sentence information is a known word. A correctness determination device comprising: a detection unit that detects a combination of words that is:
A correction rule registration device that generates the correction rule based on a detection result by the detection unit of the correct answer determination device and stores the correction rule in the correction rule storage device;
A correct answer determination system comprising: A correct answer determination method of a correct answer determination apparatus including a known word storage unit that stores a known word that is predetermined as a known word,
Input of first sentence information composed of a plurality of words and second document information in which any one of the plurality of words included in the first sentence information is replaced with another word. Dividing the received first sentence information and the second document information into a plurality of words;
The divided word is compared with the known word stored in the known word storage unit, and if it is stored in the known word storage unit, it is determined as a known word and stored in the known word storage unit. A step of determining that it is an unknown word if not,
Based on the determination result, the word based on the first sentence information is an unknown word, and the word combination based on the second sentence information corresponding to the word based on the first sentence information is a combination of words Detecting steps,
A correct answer determination method comprising: In a computer of a correct answer determination apparatus including a known word storage unit that stores a known word that is predetermined as a known word,
Input of first sentence information composed of a plurality of words and second document information in which any one of the plurality of words included in the first sentence information is replaced with another word. Dividing the received first sentence information and the second document information into a plurality of words;
The divided word is compared with the known word stored in the known word storage unit, and if it is stored in the known word storage unit, it is determined as a known word and stored in the known word storage unit. A step of determining that it is an unknown word if not,
Based on the determination result, the word based on the first sentence information is an unknown word, and the word combination based on the second sentence information corresponding to the word based on the first sentence information is a combination of words Detecting steps,
The correct answer judgment program that executes

Images