JP2009048472A - Morpheme candidate generation device and computer program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To achieve an unlimited length of unknown word morpheme candidates, exclusion of extremely improper unknown word morpheme candidates, and avoidance of reduction in generation probability of the unknown word morpheme candidates. <P>SOLUTION: The device comprises: a morpheme division point probability calculation means which calculates a morpheme division point probability in which a space between words in an input sentence is a division point of morpheme; a morpheme probability calculation means which calculates a morpheme probability in which a character string in the input sentence is a morpheme; and a morpheme candidate determination means which determines all character strings which will be morpheme candidates in the input sentence. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a morpheme candidate generation device and a computer program.

  Japanese morpheme analysis is a process of dividing a Japanese sentence into morphemes and determining grammatical attributes (part of speech, inflection type, inflection form, etc.) of each morpheme. In general, a morpheme dictionary is used for the processing. Specifically, in the Japanese sentence to be analyzed (input sentence), all character strings that match the headings of the morpheme dictionary are set as morpheme candidates, and the candidates are expressed in a graph structure. Determine the sequence and sequence of grammatical attributes. However, since new words (particularly proper nouns and technical terms) appear, it is difficult to register all morphemes in the dictionary in advance unless the analysis target is limited. Therefore, in morpheme analysis, an unknown word process for analyzing an unknown word morpheme that is not registered in the morpheme dictionary is required. In the unknown word processing, a character string that is not registered in the morpheme dictionary and may be a morpheme is estimated from the input sentence, and this is added to the graph structure as a morpheme candidate. If the correct morpheme is not added to the graph structure as a morpheme candidate in this unknown word process, it is important to generate a correct unknown word morpheme and use it as a morpheme candidate.

The following are known as unknown word processing in the conventional morphological analysis.
(1) Processing by length heuristics (for example, see Non-Patent Document 1)
When all the arbitrary character strings are handled as morpheme candidates, the order of the number of candidates increases to n ² (n is the number of characters in the input sentence). Therefore, in order to reduce the number of candidates, the character string length of morpheme candidates is limited by heuristics. In the prior art described in Non-Patent Document 1, all character strings of 5 characters or less and consecutive katakana character strings are candidates for unknown words.
(2) Processing by character type heuristics (for example, see Non-Patent Document 2)
In the method of generating unknown word candidates by heuristics based on character types, all character strings that meet the rules are morpheme candidates. In the prior art described in Non-Patent Document 2, unknown word candidates are generated by heuristics using information such as character types, and morphological analysis is performed based on CRF as an identification model.
(3) Processing by n-gram generation model (for example, see Non-Patent Document 3)
In the prior art described in Non-Patent Document 3, the generation probability of an unknown word is calculated using a character n-gram generation model and used to generate an unknown word candidate.
(4) Process for determining only division into morphemes first (for example, see Non-Patent Document 4)
In the prior art described in Non-Patent Document 4, as an unknown word analysis method, only division into morphemes is determined in advance without considering the part of speech.
Kiyotaka Uchimoto, Satoshi Sekine, Hitoshi Isahara, “Morphological analysis based on the maximum entropy model -Solution of unknown word problem-, Natural language processing, Vol.8, No.1, pp.127-141, Jan.2001. Taku Kudo, Atsushi Yamamoto, Yuji Matsumoto, “Japanese Morphological Analysis Using Conditional Random Fields”, SIG-NL-161 (13), pp.89-96, 2004. Masaaki Nagata, “Japanese morphological analysis using statistical language model and N-best search”, Transactions of Information Processing Society of Japan, Vol.40, No.9, pp.3420-3431, Sep.1999. Tetsuji Nakagawa, Yuji Matsumoto, “Chinese / Japanese word segmentation using word-level and character-level information”, IPSJ Journal, Vol.46, N0.11, pp.2714-2727, Nov.2005.

However, the above-described conventional technology has the following problems.
(1) In processing by length heuristics, there is a problem that unknown words longer than that cannot be analyzed if the number of characters is limited.
(2) In the processing based on the character type heuristics, if a character string having a feature that does not appear in the learning data is used as a morpheme candidate, there is a problem that the identification model cannot be appropriately identified and causes an analysis error. Therefore, it is important to calculate the possibility of becoming a morpheme at the stage of generating a morpheme candidate from an input sentence so as not to generate an extremely inappropriate morpheme candidate. Even when the longest matching method or the minimum number of division method of the conventional method is used, inappropriate unknown word candidates cause a decrease in analysis accuracy.
(3) In the process using the n-gram generation model, the generation probability of a character string that does not appear at all or almost in the learning data is small. Since it is considered that the unknown word often does not appear at all or hardly in the learning data, the generation probability of the character string of the unknown word tends to be small, and it is difficult to appropriately generate the unknown word morpheme candidate.
(4) Compared to the process of determining only the division into morphemes, it is advantageous for us to consider grammatical attributes such as parts of speech at the same time, because more information can be taken into account. I believe.

  The present invention has been made in view of such circumstances, and its purpose is not to limit the length of unknown word morpheme candidates, to eliminate extremely inappropriate unknown word morpheme candidates, unknown It is an object to provide a morpheme candidate generation device and a computer program capable of realizing that the generation probability of word morpheme candidates is not reduced.

  In order to solve the above-mentioned problem, the morpheme candidate generation device according to the present invention includes a morpheme division point probability calculation unit that calculates a morpheme division point probability that a character has a division point between characters in an input sentence, and the morpheme division point. A morpheme probability calculation unit that calculates a morpheme probability that a character string in the input sentence is a morpheme using a probability, and a morpheme candidate that determines character strings that are all morpheme candidates in the input sentence based on the morpheme probability And a determination unit.

  The morpheme candidate generation device according to the present invention includes a feature template table that stores a feature template for creating a feature from the input sentence, and a feature vector definition table that defines a feature that is a component of a feature vector, The morpheme segmentation point probability calculation means creates a feature using the feature template, examines a match between the feature and the definition of the feature vector, and a feature vector having a predetermined value only for the matched component (for example, a match) The value of the calculated component is “1”, and the value of the component that did not match is “0”), and the morpheme segmentation point probability is calculated by the maximum entropy method probability model using the feature vector And

  In the morpheme candidate generation device according to the present invention, the morpheme candidate determination means, for each character in the input sentence, uses the character as the first character of the morpheme candidate and connects subsequent characters. The morpheme probability of the character string is calculated, the morpheme probability is determined by a threshold value, and it is determined whether the character string is a morpheme candidate from the threshold determination result. The product of the probability and the probability that the subsequent character is not a morpheme division point is held (step 207 in FIG. 10), and the product that holds the morpheme division point probability between characters immediately after the connected character string The morpheme probability of the connected character string is obtained by multiplying by (step 204 in FIG. 10).

  In the morpheme candidate generation device according to the present invention, the morpheme candidate determination unit determines, based on the threshold value, a product of a morpheme division point probability between certain characters and a probability that a subsequent character is not a morpheme division point. (Step 208 in FIG. 10), it is determined from this threshold determination result whether to omit the process of concatenating subsequent subsequent characters for the first character.

A computer program according to the present invention uses a morpheme division point probability calculation function for calculating a morpheme division point probability that a morpheme division point is between characters in an input sentence, and a character string in the input sentence using the morpheme division point probability. A morpheme probability calculation function that calculates a morpheme probability that is a morpheme and a morpheme candidate determination function that determines a character string that is a morpheme candidate based on the morpheme probability are realized by a computer.
As a result, the morpheme candidate generation device described above can be realized using a computer.

  According to the present invention, morpheme candidates can be automatically generated from an input sentence. The generated morpheme candidates have features such as unlimited length, exclusion of extremely inappropriate morpheme candidates, and the generation probability of morpheme candidates does not decrease because it does not depend on the n-gram generation model. . Thereby, it becomes possible to contribute to realization of appropriate unknown word processing.

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

  The present invention generates a morpheme candidate from an input sentence. Based on a probability that the character string is a morpheme for an arbitrary character string in the input sentence (hereinafter referred to as “morpheme probability”), the character string Is determined as a morpheme candidate. When calculating the morpheme probability, a probability that a space between characters in the input sentence is a morpheme division point (hereinafter referred to as “morpheme division point probability”) is used.

First, the morpheme probability and the morpheme division point probability according to the present embodiment will be described.
Add the end symbol "#" in the beginning of a sentence and the end of the sentence of the input sentence S ', the morphemes w _i, the processed sentence _{S {S = w 1 w 2} ··· w i ··· w n = w 1 n , W ₁ = w _n = “#”}. n is the number of morphemes. At this time, the division probability into the morpheme sequence w ₁ ⁿ is “P (w ₁ ⁿ | S)”. Furthermore, when each morpheme is approximated so as not to depend on other morphemes, equation (1) is obtained. In the formula (1), “P (w _i | S)” represents the probability that a certain character string in the processing target sentence S is a morpheme.

Here, “P (w _i | S)” is transformed into character unit processing. The character S is “c _j ”, and the processing target sentence S is {S = c ₁ c ₂ ... C _j ... C _m = c ₁ ^m }. m is the number of characters. The morpheme sequence w ₁ ⁿ is divided into whether the morpheme string w ₁ ⁿ is divided between all characters in the character string c ₁ ^m of the process target statement S (delimiter) It can be expressed by non-delimiter. Therefore, as shown in FIG. 1, the character c _j and the character c _{j + 1} are expressed as “s _j = 1” when they are division points and “s _j = 0” when they are not division points ( That is, s _j ∈ {0,1}). If {w _i = c _p ^q }, as shown in FIG. 2, the morpheme w _i can be expressed by {S = c ₁ ^m } and “s _p−1 ^q ”. Thereby, Formula (2) is obtained.

Here, “P (s _p−1 ^q | S)” is approximated so that “s _j ” does not depend on other “s”, and the condition {S = c ₁ ^m } is changed to “s _j Limit to the a characters immediately before and the a characters immediately following. Thereby, Formula (3) is obtained.

In the present embodiment, the value of equation (3) is used for the morpheme probability. A character string having a morpheme probability equal to or higher than a threshold is set as a morpheme candidate. In equation (3), if “s _j = 1”, “P (s _j | c _{j−a + 1} ^{j + a} )” is the morpheme division point probability. In this embodiment, the maximum entropy method is used as a probability model for calculating “P (s _j | c _{j−a + 1} ^{j + a} )”. In addition to characters, character types such as katakana, hiragana, and kanji can also be used as features used in the maximum entropy method. A probabilistic model of “P (s _j | c _{j−a + 1} ^{j + a} )” by the maximum entropy method is expressed by equation (4).

However, “F (c _{j−a + 1} ^{j + a} , s _j )” is a feature vector. The feature vector F (c _{j−a + 1} ^{j + a} , s _j ) characterizes the characters c _{j−a + 1} ^{j + a} and “s _j ”. The feature vector F (c _{j−a + 1} ^{j + a} , s _j ) is composed of components having 0 or 1 as values. “Λ” is a vector of weight parameters corresponding to each component of the feature vector F (c _{j−a + 1} ^{j + a} , s _j ) (hereinafter referred to as “weight parameter vector”). “·” Represents an inner product. “Z” is a normalization term of the probability model. Since the normalization term Z is “s _j ε {0, 1}”, it is expressed by equation (5).

  In the present embodiment, the morpheme division point probability is calculated by the equations (4) and (5).

Next, the morpheme candidate generation device according to the present embodiment will be described.
FIG. 3 is a block diagram showing a configuration of the morpheme candidate generation device 1 according to an embodiment of the present invention. In FIG. 3, the morpheme candidate generation device 1 includes a morpheme division point probability calculation unit 2, a morpheme probability calculation unit 3, a morpheme candidate determination unit 4, a feature template table 5, a probability model parameter table 6, and a feature vector definition table 7.

  The morpheme division point probability calculation unit 2 calculates a morpheme division point probability between each character in the input sentence. The morpheme probability calculation unit 3 calculates the morpheme probability of each character string in the input sentence. The morpheme candidate determination unit 4 determines whether each character string in the input sentence is a morpheme candidate based on the morpheme probability.

The feature template table 5 stores feature templates. FIG. 4 shows a configuration example of the feature template table 5. In the example of FIG. 4, 18 types of feature templates are stored in the feature template table 5. In FIG. 4, only the feature template related to the character is shown, but the feature template related to the character type can also be stored. In the feature template related to the character type, in FIG. 4, the character (for example, “c _j ^{j + 1} ”) in the feature template related to the character may be replaced with the expression of the character type.

The probability model parameter table 6 stores a weight parameter vector Λ. The feature vector definition table 7 stores information defining the feature vector F (c _{j−a + 1} ^{j + a} , s _j ). FIG. 5 shows a configuration example of the feature vector definition table 7. As shown in FIG. 5, the feature vector definition table 7 shows the correspondence between features and component vector component numbers. The feature template table 5, the probability model parameter table 6 and the feature vector definition table 7 are prepared in advance and stored in the memory of the morpheme candidate generation device 1.

  Note that the morpheme candidate generation device 1 according to the present embodiment may be realized by dedicated hardware, or is configured by a computer system such as a personal computer, and the morpheme candidate generation device 1 shown in FIG. The function may be realized by executing a program for realizing the function of each part of the above on the CPU.

  Next, operations of the morpheme candidate generation device 1 illustrated in FIG. 3 will be sequentially described.

First, the operation of the morpheme division point probability calculation unit 2 will be described with reference to FIG. FIG. 6 is a flowchart showing a procedure of morpheme division point probability calculation processing according to the present embodiment. In the morpheme division point probability calculation process, a morpheme division point probability d _{j in} which an inter-character j in the input sentence S ′ is a morpheme division point (s _j = 1) is calculated. Hereinafter, description will be made with specific examples as appropriate.

In FIG. 6, in step 101, a terminal symbol “#” is added to the head and end of the input sentence S ′, and the processed sentence S after the addition is held in the memory. Here, “I am Japanese” is used as a specific example of the input sentence S ′. The number of characters in this input sentence example is seven. In step 101, “#I am Japanese #” is stored in the memory as the processing target sentence S. This processing target sentence example has 9 characters. FIG. 7 shows an example of a data holding format on the memory. As shown in FIG. 7, the character variable c _j , the character in the processing target sentence S, and the variable j indicating the character spacing are paired and held in the memory. The variable j indicates the space between characters immediately after the character of the character variable c _j .

  In step 102, the value “1” between the first characters is assigned to the variable j indicating the space between characters to be calculated. However, {j = 1, 2,..., M}, and the number m between characters is ““ number of characters in input sentence S ′ ”+ 1”. In the case of “I am Japanese,” the number m between letters is “7 + 1 = 8”.

  In step 103, with respect to all character templates in the feature template table 5, a feature is created using the feature template for each character j. However, there may be no feature corresponding to the feature template.

In the example of the feature template table 5 in FIG. 4, there are 18 types of feature templates. Therefore, for each of the 18 types of feature templates, a feature is created using the feature template, and a total of 18 features are created. As a specific example of the created feature, FIG. 8 shows a case where “j = 3”. FIG. 8 shows 18 features when “j = 3” for “#I am a Japanese #” as the processing target sentence S. For example, when the feature template “<c _j ^{j + 1} , s _j > = <c ₃ ⁴ , s ₃ >” of template number 1 is used, the character “c” of the character variable c ₃ held in the memory is Two characters are concatenated up to the character “day” of the character variable c ₄ to create a feature <“has day”, s ₃ >. At this time, it s ₃ is used 1. This is because the morpheme division point probability d _{j in} which the inter-character j becomes the morpheme division point (s _j = 1) is obtained. As a result, the feature <"is a day", 1> is created. Similarly, for example, when the feature template “<c _j−1 ^{j + 3} , s _j > = <c ₂ ⁶ , s ₃ >” of the template number 9 is used, the character variable c ₂ stored in the memory Concatenate five characters from the character “I” to the character “people” in the character variable c ₆ to create a feature <“I am Japanese”, 1>. Since there are no corresponding features for the two types of feature templates of template numbers 7 and 14 in FIG. 8, 16 features are created in the example of FIG.

  In step 104, the feature vector definition table 7 is searched for each feature created in step 103 using the feature as a search key. In step 105, the component number of the feature found as a result of the search in step 104 is acquired from the feature vector definition table 7.

In step 106, a feature vector “F (c _{j−a + 1} ^{j + a} , s _j ) = {f ₁ , f ₂ , f ₃ ,..., F _r }” is created. Here, f _k is the value of the component number k. k is a natural number from 1 to r. r is the maximum value of the component number. In the feature vector component f _k , the value of the component with the component number acquired in step 105 is set to “1”, and the value of the component with the component number not acquired is set to “0”.

For example, as an example of the results of steps 104 and 105, it is assumed that only the features of template numbers 16 and 17 are found from the feature vector definition table 7 and only component numbers 3 and 5 are acquired at “j = 3”. At this time, in step 106, the feature vector “F (c _{j−a + 1} ^{j + a} , s _j ) = (f) where the component values of only the component numbers 3 and 5 are“ 1 ”at“ j = 3 ”. ₁ , f ₂ , f ₃ , f ₄ , f ₅ , f ₆ ,..., F _r ) = (0,0,1,0,1,0,..., 0) ”is created.

In step 107, the morpheme dividing point probability d _j for the inter-character _j is calculated by the above equations (4) and (5) and stored in the memory. The inter-character j morpheme dividing point probability d _j is obtained by using the inter-character j feature vector F (c _{j−a + 1} ^{j + a} , s _j ) and the weight parameter vector Λ in the probability model parameter table 6. calculate. Here, in equation (5), F (c _{j−a + 1} ^{j + a} , 1) in the case of “s ′ = 1” is the feature vector F (c _{j−a + 1} ^{j +} ) obtained in step S106. ^a , s _j ). F (c _{j−a + 1} ^{j + a} , 0) in the case of “s ′ = 0” performs the above steps S104, 105, and 106 with “s _j = 0”, and the resulting feature vector F (c _{j-a + 1} ^{j + a} , s _j ). Incidentally, as shown in FIG. 5, when the next component number of component number "s _j = 1" in the feature other than s _j is the same in feature vector definition table 7 is defined as "s _j = 0" is F (c _{j−a + 1} ^{j + a} , 0) in the case of “s ′ = 0” is the feature vector F (c _{j−a + 1} ^{j + a} , s _j ) obtained in step S106, Only the component value of the component number obtained by incrementing the component number whose value is “1” is set to “1” (for example, “(s _j = 1) and F (c _{j−a + 1} ^{j + a} , s _j ) = (0,0,1,0,1,0), then `` s _j = 0 '' and F (c _{j-a + 1} ^{j + a} , s _j ) = (0,0,0, (In this case, in order to obtain F (c _{j−a + 1} ^{j + a} , 0) in the case of “s ′ = 0”, steps S104 and 105). , 106 is not necessary).

In step 108, if “j = m”, since the morpheme division point probabilities d _j for all j between characters have been calculated, the processing of FIG. 6 ends. on the other hand. If “j ≠ m”, “1” is added to the variable j to return to step 103 in order to calculate the morpheme dividing point probability between the next characters.

By repeating the above steps 103 to 107 m times, morpheme division point probabilities (m) between characters are created and stored in the memory. FIG. 9 shows an example of a data holding format on the memory. As shown in FIG. 9, a character variable c _j , a character _j in the processing target sentence S, a variable j indicating a space between characters, and a morpheme division point probability d _j are stored in a memory as a set.

Next, operations of the morpheme probability calculation unit 3 and the morpheme candidate determination unit 4 will be described with reference to FIG. The morpheme probability calculation unit 3 and the morpheme candidate determination unit 4 operate in cooperation. FIG. 10 is a flowchart showing a procedure of morpheme candidate determination processing according to the present embodiment. The morpheme candidate determination process determines whether the character string is a morpheme candidate for an arbitrary character string in the input sentence S ′. Note that in the morpheme candidate determination process, the character variable c _j , the variable j indicating the character in the processing target sentence S, and the morpheme division stored in the memory as a set by the above morpheme division point probability calculation process Use point probabilities d _j .

In FIG. 10, in step 201, an initial value “2” is substituted for variable p. In step 202, the morpheme division point probability d _p-1 is substituted for the variable e. In step 203, the variable p is substituted for the variable q. In step 204, the product of the variable e and the morpheme division point probability _dq is substituted for the variable f. The variable f is the probability string c _p ^q linked each character from a character variable c _p to the character variable c _q is morpheme (morpheme probability).

In step 205, the value of the variable f and the threshold value Th are compared. As a result, if the variable f is greater than or equal to the threshold Th, the process proceeds to step 206, where the character string c _p ^q is output as a morpheme candidate. On the other hand, if the variable f is less than the threshold value Th, the process proceeds to step 207. In step 207, the product of the variable e and “1-morpheme division point probability d _q ” is substituted into the variable e.

  In step 208, the value of the variable e and the threshold value Th are compared. As a result, when the variable e is less than the threshold value Th, the process proceeds to step 211. On the other hand, if the variable e is greater than or equal to the threshold Th, the process proceeds to step 209.

  In step 209, “1” is added to the variable q. In step 210, the variable q is compared with the value of the number m between characters. As a result, if the variable q is greater than or equal to the number m between characters, the process proceeds to step 211. On the other hand, if the variable q is less than the number m between characters, the process returns to step 204.

  In step 211, “1” is added to the variable p. In step 212, the variable p is compared with the value of the number m between characters. As a result, when the variable p is greater than or equal to the number m between characters, the processing of FIG. On the other hand, if the variable p is less than the number m between characters, the process returns to step 202.

  In the morpheme candidate determination process in FIG. 10 described above, it is determined whether or not the character string is a morpheme candidate for all combinations of character strings in the input sentence S ′. Specifically, for each character in the input sentence S ′, the subsequent character is concatenated with that character as the first character of the morpheme candidate, and the morpheme probability (variable f) of each concatenated character string And the morpheme probability is determined by the threshold Th, and it is determined from the threshold determination result whether the character string is a morpheme candidate.

Further, “the product of the morpheme division point probability d _p−1 and (1− morpheme division point probability d _p ) (the variable e as a result of step 207) is determined based on the threshold value Th. It is determined whether or not to omit the process of concatenating subsequent characters after the step 207. When the variable e as a result of step 207 becomes less than the threshold value Th, each of the character variable c _p to the character variable c _q is determined. It is clear that the morpheme probability is less than the threshold Th for a character string obtained by concatenating characters after the character variable c _{q + 1} to a character string c _p ^q obtained by concatenating characters. This process omits the subsequent process of concatenating the subsequent characters, and by this omission, the morpheme probabilities are calculated for all the concatenated combinations of the subsequent characters for a certain first character, and the amount of calculation is larger than that in the case of threshold determination. Drastically reduce Door can be.
Also, when connecting subsequent characters, using the variable e (the variable e as a result of step 207) that has been calculated so far, compared to the case of calculating equation (3) without using the variable e. Thus, the calculation of the morpheme probability f can be made efficient.

  According to the above-described embodiment, morpheme candidates can be automatically generated from the input sentence. The generated morpheme candidates have characteristics such as unlimited length, exclusion of extremely inappropriate morpheme candidates, and the generation probability of morpheme candidates does not decrease because it does not depend on the n-gram generation model. . Thereby, it becomes possible to contribute to realization of appropriate unknown word processing.

Next, an example of generating the probability model parameter table 6 and the feature vector definition table 7 according to the present embodiment will be described.
FIG. 11 is a block diagram illustrating a configuration of the probability model parameter generation device 20 according to the present embodiment. In FIG. 11, the probability model parameter generation device 20 includes a feature vector definition table generation unit 21, a feature vector definition table recording unit 22, a probability model parameter estimation unit 23, and a feature template table 5. The feature template table 5 is the same as that included in the morpheme candidate generation device 1 of FIG.

The parameter of the probability model (weight parameter vector Λ) is estimated by performing MAP (Maximum A Posteriori) estimation using “Gaussian prior”. Specifically, the weight parameter vector Λ is calculated so as to maximize the log likelihood L _{Λ of the} entire learning data. The learning data includes a learning sentence and information indicating whether each character j in the learning sentence is a morpheme dividing point (s _j = 1) or not (s _j = 0). The weight parameter vector Λ is calculated by the equations (6) and (7). The optimal solution can be obtained using “L-BFGS” which is a quasi-Newton method. As for “L-BFGS”, for example, “DCLiu and J. Nocedal,“ On the limited memory BFGS method for large scale optimization. ”Math. Program, Vol. 45, Issue 3, pp. 503-528, December 1989. "It is described in.

In the equation (7), σ ² is a hyper parameter. The hyper parameter σ ² is prepared in advance. || ・ || represents the norm.

A feature vector is created using a feature template. For example, if a feature <c _j ^{j + 1} = “value is”, s _j = 1> is created from the feature template <c _j ^{j + 1} , s _j > in FIG. 4, the presence / absence of the feature is returned. The feature function f can be defined by equation (8). A feature vector is constructed by using such a feature function f as a component of the feature vector.

In FIG. 11, the feature vector definition table generation unit 21 uses the feature template table 5 to create a feature vector from learning data. Specifically, a character string is cut out with each feature template for a learning sentence of learning data, and a character string that appears in the learning sentence a predetermined number of times (for example, three times) or more is used for the feature. Then, a set of the character string and “s _j = 1” and a set of the character string and “s _j = 0” are created as features.

  The feature vector definition table recording unit 22 records a set of the feature created by the feature vector definition table generating unit 21 and the component vector component number. Thereby, the feature vector definition table 7 shown in FIG. 5 is created. The probability model parameter estimator 23 calculates the weight parameter vector Λ using the equations (6) and (7).

A program for realizing the functions of the morpheme candidate generation device 1 shown in FIG. 3 is recorded on a computer-readable recording medium, and the program recorded on the recording medium is read into a computer system and executed. The morpheme candidate generation process may be performed. Here, the “computer system” may include an OS and hardware such as peripheral devices.
Further, the “computer system” includes a homepage providing environment (or display environment) if a WWW system is used.
“Computer-readable recording medium” refers to a flexible disk, a magneto-optical disk, a ROM, a writable nonvolatile memory such as a flash memory, a portable medium such as a DVD (Digital Versatile Disk), and a built-in computer system. A storage device such as a hard disk.

Further, the “computer-readable recording medium” means a volatile memory (for example, DRAM (Dynamic Random Access Memory)), etc., which hold programs for a certain period of time.
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.

  As mentioned above, although embodiment of this invention was explained in full detail with reference to drawings, the specific structure is not restricted to this embodiment, The design change etc. of the range which does not deviate from the summary of this invention are included.

It is explanatory drawing for demonstrating the relationship between the character in the character string which concerns on this invention, and a morpheme division | segmentation point. It is explanatory drawing for demonstrating the relationship between the character in the character string which concerns on this invention, and a morpheme division | segmentation point. It is a block diagram which shows the structure of the morpheme candidate production | generation apparatus 1 which concerns on one Embodiment of this invention. It is a structural example of the feature template table 5 shown in FIG. It is a structural example of the feature vector definition table 7 shown in FIG. It is a flowchart which shows the procedure of the morpheme division | segmentation point probability calculation process which concerns on one Embodiment of this invention. It is an example of the data retention format on the memory of the morpheme candidate production | generation apparatus 1 shown in FIG. It is a specific example of the feature created by the morpheme division point probability calculation processing according to an embodiment of the present invention. It is an example of the data retention format on the memory of the morpheme candidate production | generation apparatus 1 shown in FIG. It is a flowchart which shows the procedure of the morpheme candidate determination process which concerns on one Embodiment of this invention. It is an Example which produces | generates the probability model parameter table 6 and the feature vector definition table 7 which are shown in FIG.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... Morphological candidate production | generation apparatus, 2 ... Morphological division point probability calculation part, 3 ... Morphological probability calculation part, 4 ... Morphological candidate determination part, 5 ... Feature template table, 6 ... Probability model parameter table, 7 ... Feature vector definition table

Claims (5)

A morpheme segmentation point probability calculation means for calculating a morpheme segmentation point probability that a character segment between input characters is a morpheme segmentation point;
A morpheme probability calculation means for calculating a morpheme probability that a character string in the input sentence is a morpheme using the morpheme dividing point probability;
Based on the morpheme probability, morpheme candidate determination means for determining character strings that are all morpheme candidates in the input sentence;
A morpheme candidate generation device characterized by comprising: A feature template table for storing a feature template for creating a feature from the input sentence;
A feature vector definition table that defines the feature that is a component of the feature vector, and
The morpheme segmentation point probability calculating means creates a feature using the feature template, examines a match between the feature and the definition of the feature vector, creates a feature vector having a predetermined value only for the matched component, A morpheme segmentation point probability is calculated by a maximum entropy method probability model using the feature vector.
The morpheme candidate generation device according to claim 1, wherein: The morpheme candidate determining means, for each character in the input sentence, using the character as the first character of the morpheme candidate, connecting subsequent characters, calculating the morpheme probability of each connected character string, The morpheme probability is determined by a threshold value, and it is determined whether the character string is a morpheme candidate from the threshold determination result, and the morpheme division point probability between the characters immediately before the first character and the subsequent characters are divided. Holding the product with the probability that is not a point, and obtaining the morpheme probability of the connected character string by multiplying the product that holds the morpheme division point probability between the characters immediately after the connected character string,
The morpheme candidate generation device according to claim 1, wherein: The morpheme candidate determination means determines a product of a morpheme division point probability between certain characters and a probability that a subsequent character is not a morpheme division point based on the threshold, and from the threshold determination result, Decide whether to omit the process of concatenating subsequent characters after
The morpheme candidate generation device according to claim 3. A morpheme division point probability calculation function for calculating a morpheme division point probability that a character in an input sentence is a morpheme division point;
Using the morpheme dividing point probability, a morpheme probability calculation function for calculating a morpheme probability that the character string in the input sentence is a morpheme;
A morpheme candidate determination function for determining a character string that is a morpheme candidate based on the morpheme probability;
A computer program for causing a computer to realize the above.

Images