JP2007156545A - Symbol string conversion method, word translation method, its device, its program and recording medium - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a device for converting a first word belonging to a certain language system into a second word corresponding to the other language system. <P>SOLUTION: A word translation device 5B is provided with: a word output part 5 for preparing a database by referring to a transliteration probability model 7, and using approximation under the consideration of a word set maximizing conditioned probability with the character history of the word set as conditions, and for retrieving a second word corresponding to an input first word; a conversion candidate retrieving part 40 for outputting a third word extracted from document data acquired by electronic equipment 50 connected to a communication network NW based on the first word as a conversion candidate to the second word; and a conversion possibility calculating part 30 for calculating status transition weight based on the third word acquired by the conversion candidate retrieval part 40 by referring to a composite database acquired by compounding the history relating to characters comprising the third word with the database prepared by the word output part 5. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a symbol string conversion method, a word translation method, an apparatus thereof, a program thereof, and a recording medium used in, for example, an information search system, a question answering system, and a machine translation system.

  Conventionally, for example, in an information retrieval system, a question answering system, and a machine translation system, cross-language conversion, that is, conversion (translation) from a source language word or compound word (hereinafter simply referred to as a word) to a target language word. May be necessary. When such cross-language conversion is necessary, a database describing word conversion rules such as a dictionary is generally used.

  Further, a technique called “transliteration” is known in which word translation (symbol string conversion) is regarded as character-by-character conversion by focusing on characters (symbols) constituting a word (for example, Non-Patent Document 1). Non-Patent Document 2). According to this transliteration, for example, even when there are many types of languages to be converted, it is expected that it is not always necessary to prepare various dictionaries corresponding thereto.

  In the transliteration technique disclosed in Non-Patent Document 1, a probability model indicating the certainty of transliteration (the certainty of conversion of a word) is created in advance using a symbol corresponding to the pronunciation of a word. From the words (characters) in the language, the words (characters) in the target language that have the greatest probability are obtained. Specifically, when creating a probabilistic model from learning data that includes multiple sets of corresponding word pairs in both languages, the probability that the source language characters are converted to the source language pronunciation and the source language pronunciation The probability of changing to the pronunciation of the target language and the probability that the pronunciation of the target language is converted into characters of the target language are statistically obtained, and the probability of transliteration is calculated by the product of the probabilities.

In addition, the transliteration technique disclosed in Non-Patent Document 2 realizes transliteration from words expressed in katakana (Japanese) to words expressed in alphabet (English). Specifically, in this transliteration technique, a probability model indicating the conversion probability from each letter in which katakana notation is replaced with Roman letter notation to each letter of English words is created in advance, and transliteration is performed using this probability model. Do the letter. In this case, not only conversion in units of characters but also conversion probabilities between a plurality of characters that can be associated with each other before and after the character of interest (a plurality of characters) are used.
K. Knight et al, “Machine Transliteration”, Computational Linguistics, 1998, vol.24, No.4, p.599-612 E. Brill et al. “Automatically Harvesting Katakana-English Term Pairs from Search Engine Query Logs” in Proceedings of the 6th Natural Language Processing Pacific Rim Symposium, 2001, p.393-399

However, the transliteration technique (symbol conversion technique) described above has the following problems.
In other words, the technique disclosed in Non-Patent Document 1 has a problem in that a probability model cannot be created unless both the source language word and the target language word in the learning data are known. . Moreover, it is difficult to take correspondence between pronunciations in language pairs with different pronunciation systems.

  On the other hand, with the technology disclosed in Non-Patent Document 2, it is possible to make the correspondence with the alphabet easier by converting Katakana to Roman letters, but it is compatible with various languages other than Japanese-English translation. In order to realize the character, there is a problem that another information having the same effect as the pronunciation information is further required.

  Accordingly, the present invention has been made in view of the above-described problems, and provides a technique capable of converting a symbol string belonging to a predetermined symbol system into a corresponding symbol string belonging to an arbitrary symbol system. The purpose is to do.

  In order to solve the above-mentioned problem, the symbol string conversion method according to claim 1 is a symbol string conversion using a symbol co-occurrence frequency in a symbol string set which is a combination of symbol strings having the same meaning belonging to different symbol systems. A symbol string conversion method for a device, wherein the symbol string conversion device includes a step of inputting a first symbol string belonging to a first symbol system, and a symbol set in the symbol string set in the co-occurrence frequency and the symbol string set. Estimating a second symbol string belonging to a second symbol system corresponding to the input first symbol string using the frequency of appearance order, and outputting the estimated second symbol string And the step of executing.

  According to this procedure, the symbol string conversion apparatus estimates a second symbol string composed of symbols that are likely to appear in an order corresponding to the appearance order of the symbols constituting the input first symbol string. can do. Therefore, even if the input first symbol string is an unlearned (unknown) symbol string that is not registered in advance in the learning database used for calculating the co-occurrence frequency, the second symbol string is output. It becomes possible to do. Here, when using the frequency of the appearance order of the symbol set, it is also possible to use the state transition weight generated by performing processing such as taking the logarithm of the probability value of the appearance order and reversing the sign. When using this state transition weight, the symbol string converter searches for the second symbol string that minimizes the state transition weight.

  Further, a symbol conversion probability model creating apparatus for creating a symbol conversion probability model storing a co-occurrence frequency used by the symbol string converting apparatus according to claim 1 uses a symbol conversion probability model creating method as described below. be able to. That is, the symbol conversion probability model creation device configures the first symbol string and the second symbol string, respectively, based on the first symbol string and the second symbol string stored as data in the learning database. A degree of association (co-occurrence frequency, chi-square value, etc.) between the first symbol and the second symbol, and a symbol corresponding to one of the first symbol and the second symbol Calculating a virtual relevance using a virtual empty symbol when there is no symbol, and creating an inter-symbol relevance database that stores the calculated relevance and virtual relevance as data, and the learning database The degree of association between symbols associated between the first symbol string and the second symbol string based on the data stored in the data and the data stored in the inter-symbol association degree database And virtual Create a symbol string set consisting of two symbol string pairs that are correlated so that the sum or product of each relevance is maximized, and create a symbol string set database that stores the generated symbol string pairs as data And referring to data stored in the symbol string set database, calculating the co-occurrence frequency as a frequency of appearance order of the symbol string set, and creating the symbol conversion probability model May be executed.

  Further, the symbol string conversion method according to claim 2 is the symbol string conversion method according to claim 1, wherein the step of estimating the second symbol string is based on the co-occurrence frequency. The frequency of the appearance order is calculated for each symbol string set composed of a set in which the symbol string and the second symbol string are associated with each other in symbol units, and the frequency of the appearance order is calculated based on the result of the calculation. The second symbol string is estimated using the frequency of the appearance order with respect to the searched symbol string set.

  According to such a procedure, the symbol string conversion device has a frequency of appearance order among the frequencies of appearance order in a symbol string set composed of a set in which the first symbol string and the second symbol string are associated with each other in symbol units. Since the symbol string set is estimated by using an approximation that considers only the symbol set with which the correspondence is maximized, the solution search space can be reduced by pruning the search.

  The symbol string conversion method according to claim 3 is the symbol string conversion method according to claim 1 or 2, wherein the step of estimating the second symbol string is based on the co-occurrence frequency. The frequency of the appearance order is calculated for each symbol string set composed of a set in which the first symbol string and the second symbol string are associated in symbol units, and based on the result of the calculation, Searching for a symbol string set having the highest frequency of appearance order, creating a database for storing the frequency of appearance order relating to the searched symbol string set as data, and referring to the database, the second And searching for a symbol string.

  According to such a procedure, the symbol string conversion device stores, as data, the frequency of the appearance order related to the symbol string set searched using approximation that considers only the symbol set associated with the highest frequency of the appearance order. A database is created, and the second symbol string is searched with reference to the created database. Therefore, when the input first symbol string is a learned (known) symbol string registered in advance in the learning database used for calculating the co-occurrence frequency, the first symbol string is stored in the learning database. The second symbol string registered in pairs with the string can be output as a conversion result.

  The word translation method according to claim 4 is the word translation method according to any one of claims 1 to 3, wherein the symbol string is a word composed of characters. The symbol string conversion device acquires document data from an electronic device connected to a communication network based on an input word, and a predetermined number of words from the acquired document data. And a step of extracting as a conversion candidate from the inputted word.

  According to such a procedure, the symbol string conversion device functions as a word translation device, and converts a first word belonging to an arbitrary language system into a corresponding second word in another language system. The characters used in the language system are preferably phonetic characters. Then, the symbol string conversion device extracts words from the document data acquired from the communication network as conversion candidates from the first word input for translation. Therefore, even if the input first word is an unlearned (unknown) word that is not registered in advance in the learning database used for calculating the co-occurrence frequency, the existing first word extracted from the communication network exists. A word can be adopted and output as a translation result.

  The word translation device according to claim 5 is a word translation device using the simultaneous occurrence frequency of characters in a word set which is a combination of words having the same meaning belonging to different language systems, wherein the first language A second word corresponding to the input first word by using the input means for inputting the first word belonging to the system, and the frequency of the simultaneous occurrence and the appearance order of the character set in the word set; And a word search means for estimating a second word belonging to the language system and an output means for outputting the estimated second word.

  According to such a configuration, the word translation device can estimate a second word composed of characters that are likely to appear in an order corresponding to the appearance order of the characters constituting the input first word. it can. Here, for example, the first word can be written in katakana and the second word can be written in alphabet. According to this word translation apparatus, even if the input first word is an unlearned (unknown) word that is not registered in advance in the learning database used to calculate the co-occurrence frequency, the second word A word can be output. Here, when the frequency of the appearance order of character sets is used, it is also possible to use the state transition weight generated by performing processing such as taking the logarithm of the probability value of the appearance order and reversing the sign. When this state transition weight is used, the word translation device searches for a second word that minimizes the state transition weight.

  Further, the word translation device according to claim 6 is the word translation device according to claim 5, wherein the word search means is configured to determine the first word and the second word based on the co-occurrence frequency. Is calculated for each word set consisting of a set of characters associated with each other, and based on the result of the calculation, a word set having the maximum frequency of appearance is searched for. The second word is estimated using the frequency of the appearance order related to the searched word set.

  According to such a configuration, the word translation device has the highest appearance order frequency among the appearance order frequencies in the word set composed of a set in which the first word and the second word are associated in character units. Since the word is estimated using an approximation that considers only the character sets that are associated, the solution search space can be reduced by pruning the search.

  Further, the word translation device according to claim 7 is the word translation device according to claim 5 or 6, wherein the first word and the second word are converted into characters based on the co-occurrence frequency. The frequency of the appearance order is calculated in an arbitrary word set composed of pairs associated in units, and based on the result of the calculation, the word set that maximizes the frequency of the appearance order is searched, and this search is performed. Database creation means for creating a database for storing the frequency of appearance order related to the word set as data, wherein the word search means searches for the second word with reference to the database. To do.

  According to such a configuration, the word translation device stores a database that stores, as data, the frequency of appearance order related to a word set searched using approximation that considers only the character set associated with the highest frequency of appearance order. The second word is searched with reference to the created database. Therefore, when the input first word is a learned (known) word registered in advance in the learning database used for calculating the co-occurrence frequency, the first word is paired with the first word in the learning database. It becomes possible to output the 2nd word registered by (2) as a translation result.

  Further, the word translation device according to claim 8 is the word translation device according to any one of claims 5 to 7, wherein communication is performed based on the first word input to the input means. Document data acquisition means for acquiring document data from an electronic device connected to the network, and a conversion candidate for extracting a predetermined number of words as conversion candidates from the first word from the acquired document data And an extraction means.

  According to such a configuration, the word translation device extracts words from the document data acquired from the communication network as conversion candidates from the first word input for translation. Therefore, even if the input first word is an unlearned (unknown) word that is not registered in advance in the learning database used for calculating the co-occurrence frequency, the existing first word extracted from the communication network exists. A word can be adopted and output as a translation result.

  A symbol string conversion program according to claim 9 causes a computer to execute the symbol string conversion method according to any one of claims 1 to 3. By being configured in this way, a computer in which this program is installed can realize each function based on this program.

  A word translation program according to claim 10 causes a computer to execute the word translation method according to claim 4. By being configured in this way, a computer in which this program is installed can realize each function based on this program.

  A recording medium according to an eleventh aspect is characterized in that the symbol string conversion program according to the ninth aspect or the word translation program according to the tenth aspect is recorded. By being configured in this way, a computer equipped with this recording medium can realize each function based on a program recorded on this recording medium.

  According to the present invention, a symbol string belonging to a predetermined symbol system can be converted into a corresponding symbol string belonging to an arbitrary symbol system. In particular, it is possible to perform conversion in consideration of the appearance order of known symbol conversion results without using information such as pronunciation and Romanization rules.

Hereinafter, embodiments of the present invention will be described with reference to the drawings as appropriate.
[Configuration of word translation system]
(First embodiment)
FIG. 1 is a diagram illustrating a configuration example of a word translation system including a word translation apparatus according to the first embodiment of the present invention. The word translation system (symbol string conversion system) 1 includes a first word (symbol string) that is a conversion source character string and a second word that is a conversion destination character string corresponding to the first word. Using the transliteration probability model in which the co-occurrence probability (co-occurrence frequency) of the character to be stored is stored as data, the input first word is converted into the second word and output. Here, the first word is composed of a plurality of first characters belonging to the first language system. Similarly, the second word is composed of a plurality of second characters belonging to the second language system. The co-occurrence probability is a probability that the appearance of the first character and the appearance of the second character as a conversion result of the first character occur at the same time. Hereinafter, the first word may be referred to as a source word, the first character as a source character, the second word as a target word, and the second character as a target character.

As shown in FIG. 1, the word translation system 1 includes a storage device 2, a storage device 3, a transliteration probability model creation device (symbol conversion probability model creation device) 4, and a word translation device (symbol string conversion device). And 5.
The storage device 2 stores a learning database 6 and is a storage means such as a general hard disk.
The learning database 6 is a set of source words and target words.

The storage device 3 stores a transliteration probability model (symbol conversion probability model) 7 and is a storage means such as a general hard disk.
The transliteration probability model 7 stores the transliteration probability from the source character to the target character, and the co-occurrence probability of the source character and the target character as data.

  The transliteration probability model creation device (symbol conversion probability model creation device) 4 and the word translation device (symbol string conversion device) 5 are general computers (computers) such as a CPU (Central Processing Unit) and a RAM. (Random Access Memory), ROM (Read Only Memory), HDD (Hard Disk Drive), KB / CRT (Key Board / Cathode Ray Tube), and an input / output interface.

The transliteration probability model creation device (symbol conversion probability model creation device) 4 obtains the correspondence between the source character and the target character based on the learning database 6, and calculates the transliteration probability between the source character and the target character. A transliteration probability model 7 is created by modeling as an N-gram model determined in consideration of the previous (N-1) transliteration results.
The word translation device (symbol string conversion device) 5 outputs a target word corresponding to a source word using a transliteration probability model 7 with one source word as an input.

[Configuration of transliteration probability model creation device]
FIG. 2 is a functional block diagram illustrating a configuration example of the transliteration probability model creation device illustrated in FIG. 1.
As shown in FIG. 2, the transliteration probability model creation device 4 includes an input means 10, a storage means (RAM, etc.) 11, a character-to-character relationship database creation means (a symbol-to-symbol relationship degree database creation means) 12, a word A set database creation means (symbol string set database creation means) 13, an occurrence probability calculation means 14, and a writing means 15 are provided.

The input means 10 is an input interface for inputting source characters and target characters from the learning database 6 and outputting them to the inter-character relevance degree database creating means 12 and the word set database creating means 13. The input means 10 inputs a database creation instruction or the like from the input device M. The input device M is a pointing device such as a mouse or a keyboard, for example.
The storage means 11 includes a RAM, a ROM, and an HDD, and stores an inter-character relevance database (inter-symbol relevance database) 16 and a word group database (symbol string group database) 17 in the HDD. Is.

The inter-character relevance database 16 stores the statistical relevance between the source character and the target character as data. Here, the association level Assoc (s, t) is a scale indicating that the target character t is likely to appear as a transliteration candidate for the source character s. For example, the target word T ₀ corresponding to the source word S ₀ including the source character s includes many target characters t, or the target word T ₁ not corresponding to the source word S ₀ includes too many target characters t. If not, the relevance level Assoc (s, t) is high. Specifically, the degree of association can be the co-occurrence frequency, the chi-square value used for statistical tests, φ ² that is a value obtained by normalizing the chi-square value to a range of 0 to 1, and the like. .
The word set database 17 stores a word set made up of two sets of words associated with each other so that the product of the relevance between the characters associated with the source word and the target word is maximized. Is stored as

The character-to-character relevance database creating means 12 calculates a statistical relevance between the source character and the target character based on the data stored in the learning database 6, and of the source character and the target character. When there is no character corresponding to any of the characters, a virtual relevance degree using the virtual empty character φ is calculated, and the inter-character relevance degree database 16 using the calculated relevance degree and the virtual relevance degree as data is created. Is. In the present embodiment, as the virtual relevance, the virtual relevance degree Assoc (s, φ _t ) when the source character s does not correspond to any character of the target character t, and the target character t which of the source character s Two types are used: a virtual relevance level Assoc (φ _s , t) when it does not correspond to a character.

  The word set database creation means 13 uses the data stored in the learning database 6 and the data stored in the inter-character relevance database 16 to inter-character spacing associated between the source word and the target word. Generate a word set consisting of two word sets that are associated with each other so that the product of the relevance level and the virtual relevance level is maximized (optimal), and the generated word set is data The word set database 17 is created.

  Here, a specific example until the word set database 17 is created will be described with reference to FIG. FIG. 3 is an explanatory diagram showing a specific example until the word set database shown in FIG. 2 is created. Here, the first language system is Japanese (Katakana) and the second language system is English (alphabet) 5.

  As shown in (a) of FIG. 3, “ice cream” is assumed as the first word (source word) S, and “ice cream” is assumed as the second word (target word) T. As assumed here, the number of words is not limited to one, but may be a compound word (for example, ice cream) including a plurality of words.

Here, the first word S is assumed to be composed of m first characters (source characters s ₁ , s ₂ ,..., S _m ). Therefore, in the case of “ice cream”, since the first word S is m = 7 as shown in FIG. 3B, seven character IDs (s _{1 to} s ₇ ) are attached. It will be.

Similarly, it is assumed that the second word T is composed of n second characters (target characters t ₁ , t ₂ ,..., T _n ). Therefore, in the case of “ice cream”, since the second word T is n = 8 as shown in FIG. 3C, eight character IDs (t _{1 to} t ₈ ) are attached. It will be. Here, it is assumed that the character string is combined while ignoring the white space, but the white space may be treated as a character. Of course, other symbols such as underbars may be handled in the same manner, not limited to spaces.

When FIG. 3B and FIG. 3C are compared, the number of characters is different (m <n). In the present embodiment, the correspondence between characters is one-to-one correspondence between one character of the first character and one character of the second character, and an empty character is required when there is no corresponding character. That is, when the first word S and the second word T are associated with each other so as to be optimal in character units, the first word has an empty character φ (character ID “φ” as shown in FIG. Two _s ") are inserted. Similarly, in the second word, one empty character φ (character ID “φ _t ”) is inserted as shown in FIG. When the correspondence is optimized in this way, the number of characters becomes equal. If the number at this time is 1 (el), it can be generally expressed as l ≧ m and l ≧ n. In this case, l = 9.

As shown in (f) of FIG. 3, a word set is generated from each of the words with blank characters shown in (d) of FIG. 3 and (e) of FIG. Then, the correspondence A between the characters of both languages constituting this word set is set as A = a ₁ , a _{2 ,.} The element of association A, that is, the character set ID is represented by a _i = (s _j , t _k ). Here, s _j is s _1, ..., one of s _m, or a phi _s, t _k is t _1, ..., is that one or phi _t of t _n.

Further, in the present embodiment, when the correspondence is optimized, I s ₁ before placing an _{empty, ..., s m, t 1} , ..., and those associating without changing their order each character of t _n To do. In other words, there is a relationship of J> j and K> k for a _I = (s _J , t _K ) where I> i. Specifically, as shown in (f) of FIG. 6, each element of the character set (I, i) and character set (S, c) is compared in the character set IDs “a ₂ ” and “a ₃ ”. Then, the order of “i” and “su” on the first word side (source side) is the same as the order of the original “ice cream”, and “order” on the second word side (target side) The order of “i” and “c” is the same as that of the original “ice cream”. That is, the order is not changed by the association.

On the other hand, for example, as shown in FIG. 6G, each element of the character set (I, e) and the character set (S, c) in the character set IDs “a ₂ ” and “a ₃ ” In comparison, the order of “i” and “su” on the first word side (source side) is the same as the order of the original “ice cream”, but on the second word side (target side) “ The order of “e” and “c” is reversed from the original “ice cream” order. That is, the order is changed by the association. In short, in the present embodiment, the association shown in (g) of FIG. 6 is eliminated by the optimum association, and the association is made as shown in (f) of FIG.

Based on the formula (1), the word set database creating means 13 creates a correspondence (optimum correspondence) A ^ (A hat) that maximizes the product of the degree of association between characters and the degree of virtual association. ) In Expression (1), Assoc (a _i ) is the degree of association between the source character and target character of the character set a _i for which a predetermined association “A” has been made, and “argmax _A (y), “y = f (A)” means to obtain “A” when y is maximum.
Further, the word set database creating means 13 may obtain an optimum association A ^ based on the formula (2). In this case, the association is calculated so that the sum of the relevance between characters and the virtual relevance is maximized.

Returning to FIG. 2, the description of the configuration example of the transliteration probability model creation device 4 will be continued.
The occurrence probability calculation means 14 refers to the data stored in the word set database 17 to determine the co-occurrence probability of the appearance order of the character set of the source character and the target character in the source word and the target word constituting the word set. The probability (frequency of appearance order) is calculated, and the transliteration probability model 7 is created. Here, the probability of the appearance order is a conditional probability on the condition that the state transition of each character until the focused source character or target character appears is a condition. That is, the occurrence probability calculation means 14 determines the probability that the appearance of a certain source character and the appearance of the target character, which is the transliteration result of the source character, occur at the same time (N−1) sources immediately before the certain source character. The transliteration probability model 7 is created using the character history and the history of (N-1) target characters immediately before the target character. For example, when the source character (character ID “s _j ”) and the target character (character ID “t _k ”) described with reference to FIG. 3 are used, the source character and the target character are represented in the associated word group. The probability (co-occurrence probability) P (a _i ) of appearing character sets (character set ID “a _i ”) is the immediately preceding (N−1) character sets (a _i−1 ,..., A _{i−N +1} ) with a conditional probability. N is a numerical value indicating “N” in the N-gram language model. Further, hereinafter, the simple probability means the co-occurrence probability.

Therefore, the occurrence probability calculation means 14 calculates a conditional probability P (a _i | a _i−1 ,..., A _{i−N + 1} ) using the word set database 17. Here, when a large value is set for N, the majority of conditional probabilities become “0”, and as a result, the versatility of the probability model deteriorates. Smoothing is performed using the probability value when the value is small (for example, 1, 2, 3). Thereby, since the conditional probability of the immediately preceding (N−1) character does not become “0”, a probability value that is not “0” can be given to any transliteration result. As the smoothing process, a known smoothing technique applied to an N-gram language model used for natural language processing or speech recognition can be used (for example, “stochastic language model” Kenji Kita, University of Tokyo (See Publishing, 1999, Chapter 3, Languages and Calculations-4).

  The writing means 15 writes the probability value calculated by the occurrence probability calculating means 14 into the storage device 3 (see FIG. 1) as the transliteration probability model 7.

  The inter-character relevance database creating means 12, the word set database creating means 13, and the occurrence probability calculating means 14 are such that the CPU expands a predetermined program stored in the ROM or the like of the storage means 11 to the RAM. It is realized by executing.

[Configuration of word translation device]
4 is a functional block diagram illustrating a configuration example of the word translation apparatus illustrated in FIG.
Based on the transliteration probability model 7 created by the transliteration probability model creation device 4, the word translation device 5 translates the source characters constituting the source word input from the input device M into the target characters. Translation (conversion) into words is realized, and the translated target words are output to the output device D.

(Translation principle)
Here, the principle of translation (symbol string conversion) in the word translation apparatus 5 will be described based on mathematical expressions. In the description of the translation principle, the term “target word T” includes an accurately translated word corresponding to the source word S on a one-to-one basis, and a word similar thereto, so to speak, It means what can be called target word candidates.

  The co-occurrence probabilities that the input source word S and the target word T including the exact translation result appear as a word set in the transliteration probability model 7 are the source character of the input source word S and the target word T Are different depending on the correspondence A between the characters and the target character. At this time, the co-occurrence probability P (S, T, A) of the input source word S and the target word T can be expressed as a product of conditional probabilities as shown in Expression (3).

  Since there are many possibilities as the correspondence A between characters, the source word S and the target word T appear as a set of words that are transliterated based on the transliteration probability model 7 in consideration of all of them. The typical probability P (S, T) is expressed by the equation (4).

  According to the above equation (4), in order to accurately determine the probability P (S, T), the sum of the probability values for each association A must be calculated. However, considering all the correspondences A, the calculation becomes enormous, which is not practical. Therefore, in order to reduce the amount of calculation, the following approximation is introduced in this embodiment. That is, as shown in Expression (5), the association A when the probability value for the association A is maximized is adopted as the optimum association A ′.

  Then, an approximation is performed in consideration of only the optimum association A ′ expressed by the above-described equation (5). A known Viterbi algorithm can be used for such calculation. By this approximation, the probability P (S, T) shown in the above equation (4) is approximated as in equation (6). In the specific calculation of Expression (6), Expression (3) described above is used.

  The word translation device 5 considers the correspondence between arbitrary characters with respect to an arbitrary target word T as a search for the optimal target word T ′ with respect to the source word S, and further satisfies the above-described equation (6). Since the search is performed, the solution search space can be reduced by pruning the search.

  As a search for the above-described equation (6), that is, a method for searching for the optimum target word T ′, in this embodiment, a known finite state machine called WFST (Weighted Finite State Transducer) is used. Efficient search (see Non-Patent Document 1). In this WFST, weights for state transitions are defined in advance, and a sequence of source characters can be input and a sequence of target characters can be output.

  It is possible to integrate the functions of a plurality of finite state machines by combining a plurality of WFSTs (see Non-Patent Document 1). That is, from the source word S, a first WFST that translates and generates an intermediate language word I indicating a language that is neither the language of the source word S nor the target word T, and the target word T is determined from the intermediate language word I. You may make it implement | achieve translation from the source word S to the target word T by synthesize | combining with 2nd WFST which carries out translation production | generation. With this configuration, for example, even if a learning database for realizing transliteration between the language of the source word S and the language of the target word T cannot be used, the language of the source word S and the intermediate language If the learning database for realizing transliteration between and the learning database for realizing transliteration between the intermediate language and the language of the target word T are respectively used, the first WFST and the second WFST can be created. Here, the number of intermediate languages used for conversion in the translation from the source word S to the target word T is not limited to one, and if there is a learning database for realizing transliteration, a plurality of types are available. It is also possible to intervene intermediate languages.

  Specifically, in the present embodiment, the word translation device 5 is configured by combining one type of WFST as a set of one type of WFST database and a WFST search program as shown below, but synthesizing it via an intermediate language. A plurality of types of WFST that can be calculated may be used. In this case, a plurality of types of transliteration probability models 7 are used.

(Specific example of configuration)
In order to realize the translation (symbol string conversion) principle described above, the word translation device 5 has an input means (first input means) 21, a storage means 22, and a state transition information database creation as shown in FIG. Means (database creation means) 23, word search means 24, output means (first output means) 25, and state transition information database 26 are provided.

The input means (first input means) 21 is an input interface, which inputs a source word (first word) from the input device M and outputs it to the state transition information database creation means 23 and the word search means 24. Is. Further, the input means 21 inputs the source character string and the target character string from the transliteration probability model 7 and outputs them to the state transition information database creation means 23.
The storage means 22 includes a RAM, a ROM, and an HDD, and stores a state transition information database 26 in the HDD.

The state transition information database 26 corresponds to the WFST database described above. This state transition information database 26 corresponds to the probability of the appearance order of character sets in a word set consisting of a set in which a source word and a target word associated with the source word are associated with each other in character units. The weight (state transition weight) is stored as data together with the transition source state and the transition destination state. Note that the appearance order probability itself may be stored instead of the weight.
In addition, the state transition information database 26 specifically includes a source character sequence of a word set stored in the transliteration probability model 7 as input correspondence data. Further, the state transition information database 26 has, as output correspondence data, the target character series of the word set stored in the transliteration probability model 7 and the weight of the probability value of the above-described equation (6) as the state transition weight. .

  The state transition information database creation means (database creation means) 23 refers to the data stored in the transliteration probability model 7 and creates the state transition information database 26. The state transition information database creation means 23 selects a word set having the highest probability of the appearance order of the character set in the word set consisting of the source word and the target word associated with the source word. A weight corresponding to the probability obtained by using an approximation to be considered (corresponding to the above-described equation (6)) is calculated as a state transition weight. Each probability value for calculating the probability of the appearance order is obtained in advance.

Here, the state transition weight calculated by the state transition information database creation unit 23 will be described. Conditional probability P of the formula (3) described above _{(a i | a i-1} , ..., a i-N + 1) Conditions a _{i-1 in, ..., a i-N +} 1 of the history. This history is a series of state transitions corresponding to each character set a _i . Specifically, attention is focused on the i-th character set a _i (s _j , t _k ). This character set _{a i (s j, t k} ) is input with a source character s _j, it corresponds to the state transition for outputting the target character t _k. This character set of _{a i (s j, t k} ) until the advent of, just before the (N-1) number of character sets a _i-1, ..., via the state transition sequence of a _{i-N + 1} Yes. Therefore, the weight corresponding to the conditional probability P (a _i | a _i−1 ,..., A _{i−N + 1} ) for the state transition corresponding to the character set a _i (s _j , t _k ). Is assigned as a state transition weight. Here, the state transition weight is assumed to be the logarithm of the logarithm of the conditional probability, that is, −logP (a _i | a _i−1 ,..., A _{i−N + 1} ). Here, the base of the logarithm is 2, for example. When the source character s _j is an empty character φ (character ID “φ _s ”), it is realized as a state transition performed irrespective of the input source character (this is called an ε transition). When the target character t _k is the empty character φ (character ID “φ _t ”), this is realized as a state transition without output.

The word search means 24 estimates a second word corresponding to the input first word, and corresponds to the above-described WFST search program. The word search means 24 searches (estimates) the optimum target word T that satisfies the above-described equation (6) corresponding to the input source word S with reference to the data stored in the state transition information database 26. And output to the output means 25. Specifically, the word search means 24, the source character s ₁ constituting the source word S input, ..., when the input corresponding data sequentially state transition s _m information database 26, epsilon transition is also taken into account Thus, the output correspondence data in the state transition information database 26 is searched, and the target character string having the smallest state transition weight is selected from the character series (target character string) corresponding to the searched output correspondence data. In the present embodiment, the word search unit 24 selects the target character string having the minimum state transition weight, but the present invention is not limited to this, and a plurality of target character strings may be selected. In this case, the top several target words are output as conversion candidates. The word search means 24 may output the value of the state transition weight together with the target word (target character string). In this case, the value of this state transition weight can be used as a conversion possibility indicating how likely the translation (symbol string conversion) is between the target word and the input source word.

The state transition information database creation means 23 and the word search means 24 described above are realized by the CPU developing and executing a predetermined program stored in the ROM or the like of the storage means 22 on the RAM. is there.
The output means (first output means) 25 is an output interface to the output device D, and outputs the target word searched by the word search means 24 to the output device D. The output device D is a display device such as a liquid crystal display, for example.

[Operation of transliteration probability model creation device]
The operation of the transliteration probability model creation device 4 will be described with reference to FIG. 5 (refer to FIG. 2 as appropriate).
FIG. 5 is a flowchart showing the operation of the transliteration probability model creation apparatus shown in FIG.
The transliteration probability model creation device 4 calculates the relevance between characters (symbols) between the source character and the target character based on the data stored in the learning database 6 by the inter-character relevance database creating means 12. The interrelationship database 16 is created (step S1).
Subsequently, the transliteration probability model creation device 4 uses the word set database creation unit 13 to calculate the product of the relevance based on the data stored in the learning database 6 and the data stored in the inter-character relevance database 16. A word (symbol string) set that maximizes is generated, and a word set database 17 is created (step S2).
Subsequently, the transliteration probability model creation device 4 uses the occurrence probability calculation unit 14 to simultaneously generate characters in each word (source word and target word) of the word set based on the data stored in the word set database 17. The probability is calculated as a conditional probability with a history as a condition, and a transliteration probability model (symbol conversion probability model) 7 is created (step S3).

[Operation of word translation device]
The operation of the word translation device 5 will be described with reference to FIG. 6 (see FIG. 4 as appropriate).
FIG. 6 is a flowchart showing the operation of the word translation apparatus shown in FIG.
Based on the transliteration probability model (symbol conversion probability model) 7 by the state transition information database creation means 23, the word translation device 5 relates to the source characters and target characters that constitute the word set and the target word, respectively. The state transition weight corresponding to the conditional probability of the character (symbol) is calculated, and the state transition information database 26 is created in advance (step S11).
And the word translation apparatus 5 inputs the 1st word (symbol string) which is a translation object from the input device M as a source word by the input means 21 in the state by which the state transition information database 26 was created beforehand ( Step S12).
Subsequently, based on the state transition information database 26 created in advance in step S11, the word translation device 5 uses the word search unit 24 to select the second word (the target word corresponding to the first word (source word)) The symbol string is searched (step S13).
Subsequently, the word translation device 5 outputs the searched second word (target word) as a translation result (step S14). Thereby, the output device D displays the target word. The word translation device 5 may output the value of the state transition weight together with the target word.

  According to the first embodiment, a word (first word) belonging to a predetermined language system can be converted into a corresponding word (second word) belonging to an arbitrary language system. Moreover, in the word translation system 1, the transliteration probability model creation device 4 does not use information such as pronunciation and romaji rules, and sets the first and second word pairs registered in the learning database 6. The transliteration probability model 7 is created using only the set. For this reason, the word translation device 5 using the transliteration probability model 7 has the problem of processing a word whose pronunciation is unknown, the problem of association between pronunciations, knowledge for notation conversion represented by Romanization, and the like. It is possible to perform translation in consideration of the history of known symbol conversion results without the necessity. As a result, for example, the translation dictionary in English (alphabet, phonogram) document search systems using Japanese katakana (phonetic characters), similar question answering systems, and machine translation systems cannot be covered by translation dictionaries. You will be able to handle words. In addition, although the word translation apparatus 5 demonstrated in the best mode provided with the state transition information database preparation means 23 and the state transition information database 26, these are not essential structures.

(Second Embodiment)
FIG. 7 is a diagram illustrating a configuration example of a word translation system including a word translation apparatus according to the second embodiment of the present invention.
The word translation system (symbol string conversion system) 1A includes a first word (source word) and one or more third words that are candidates for conversion from the source word to the second word (target word). Input. This word translation system 1A is the same as the word translation system 1 shown in FIG. 1 except that it includes a word translation device (symbol string conversion device) 5A. For convenience of explanation, this word translation system 1A has the same configuration. Are denoted by the same reference numerals, and description and drawings are omitted as appropriate.

As shown in FIG. 7, the word translation device (symbol string conversion device) 5 </ b> A includes a word output unit 5 and a conversion possibility calculation unit 30.
The word output unit 5 points to the word translation device 5 (first embodiment) shown in FIG. 4 and is given the same reference numerals.
The conversion possibility calculation unit 30 indicates the likelihood of translation (symbol string conversion) between the third word input from the outside of the word translation device 5A and the input source word. Are output as probability values.

  When the conversion possibility calculation unit 30 receives both the source word (first word) S and the third word as input, and performs the association such that the product of the probabilities expressed by the above equation (6) is maximized. Have the function of calculating and outputting the weights of these as probabilities for them. That is, the conversion possibility calculation unit 30 calculates the conversion possibility (likelihood) of the third word from the source word S based on the transliteration probability model 7. At that time, the finite state machine created by the state transition information database creating means 23 (see FIG. 4) is used.

[Configuration of convertibility calculator]
FIG. 8 is a functional block diagram illustrating a configuration example of the convertibility calculation unit illustrated in FIG.
In the present embodiment, as a method for the convertibility calculation unit 30 to search for an optimal state transition sequence, a finite state automaton that accepts the state transition information database 26 (WFST database) and the target character string constituting the target word T is used. A weighted finite state automaton (WFSA) obtained by combining with (FSA: Finite State Automaton) is used. In the present embodiment, the WFSA specifically includes a WFSA database and a WFSA search program.

  As shown in FIG. 7, the convertibility calculation unit 30 includes an input unit (second input unit) 31, a storage unit 32, a combined state transition information database creation unit 33, as shown in FIG. , State transition weight calculating means 34 and output means (second output means) 35.

  The input means (second input means) 31 is an input interface, and one or more third words that are conversion candidate words to the target word (second word) of the source word (first word) Is input from the input device M and output to the state transition weight calculation means 34. In addition, the input unit 31 inputs the state transition information database 26 from the word output unit 5 and outputs it to the combined state transition information database creation unit 33.

The memory | storage means 32 contains RAM, ROM, and HDD, and memorize | stores the synthetic | combination state transition information database 36 in HDD.
The combined state transition information database 36 corresponds to the WFSA database described above, and combines the history of the third character constituting the third word to be input and the data stored in the state transition information database 26. It is stored as data. In FIG. 8, only one combined state transition information database 36 is shown. However, in the second embodiment, each third word scheduled to be input and the state transition information database 26 are combined to be input. The composite state transition information database is created in advance for the number of planned third words.

  The synthesized state transition information database creation unit 33 synthesizes the history related to the third character constituting the third word to be input and the data stored in the state transition information database 26 of the word output unit 5 and synthesizes them. The composite state transition information database 36 using the result as data is created. The synthesized state transition information database creating unit 33 creates an FSA necessary for synthesis from the third word input from the input unit 31. Note that the combined state transition information database creating unit 33 may input the FSA created in advance from the input unit 31 and synthesize the database.

The state transition weight calculation means 34 corresponds to the above-described WFSA search program. The state transition weight calculation means 34 refers to the data stored in the combined state transition information database 36, and inputs the source characters (first characters) constituting the first word input to the input means 21, and the input The state transition weight described above is calculated as the probability of the appearance order of the character set composed of the third character constituting the third word input to the means 31. The conditional probability itself may be calculated instead of the state transition weight.
Specifically, the state transition weight calculating unit 34 sequentially combines the third character constituting the third word, and the combined state transition in which the history of the third character constituting the third word is synthesized as FSA. In the case of the input correspondence data in the information database 36, the total value of the state transition weights from the source character (first character) to the third character is calculated in consideration of the ε transition. Then, this calculation process is executed for each of the combined state transition information databases 36 corresponding to the input third word, and the total value is the minimum value among the plurality of input third words. 3 is searched, and the minimum value at that time is output to the output means 35 as the possibility of conversion.

  The composite state transition information database creation means 33 and the state transition weight calculation means 34 described above are realized by the CPU developing and executing a predetermined program stored in the ROM or the like of the storage means 32 on the RAM. Is.

  The output means (second output means) 35 is an output interface to the output device D, and outputs the state transition weight (or probability) of the third word selected by the state transition weight calculation means 34 to the output device D. To do. Note that the state transition weights to be output may be each value or a total value.

[Operation of the convertibility calculator]
The operation of the convertibility calculation unit 30 will be described with reference to FIG. 9 (refer to FIG. 8 as appropriate).
FIG. 9 is a flowchart showing the operation of the convertibility calculation unit shown in FIG.
The conversion possibility calculation unit 30 uses the combined state transition information database creation unit 33 to convert the history of the third characters (symbols) constituting one or more third words (symbol strings) that are known to be input into the state transition. Combining with the information database 26, a combined state transition information database 36 is created in advance (step S21).
And the conversion possibility calculation part 30 is the conversion candidate of the 1st word (symbol string) as a source word from the input device M by the input means 31 in the state which produced the synthetic | combination state transition information database 36 beforehand. A third word (symbol string) is input (step S22).
Subsequently, the convertibility calculating unit 30 causes the state transition weight calculating unit 34 to change from the first character (symbol) constituting the first word to the third character (symbol) constituting the third word. The third word having the smallest total value of the state transition weights is selected (step S23).
Subsequently, the convertibility calculation unit 30 outputs the state transition weight of the third word selected by the state transition weight calculation unit 34 by the output unit 35 (step S24). Thereby, the output device D displays the state transition weight as the conversion possibility.

  In the above description of the second embodiment, the state transition weight calculation unit 34 calculates the total value of the state transition weights, and this total value is the minimum value among the plurality of input third words. In the above description, the third word is searched for. However, only the total value or each value of the state transition weight may be output. In this case, the user may visually check the state transition weight displayed on the output device D and select the third word that is the minimum value at that time.

  According to the second embodiment, when a plurality of words (third word) are input as conversion candidates of the source word (first word), the third word is converted from the source word (first word) to the third word. The certainty of conversion into words can be obtained, and the accuracy of translation can be improved. Even if the source word (first word) is an unlearned (unknown) word that is not registered in the learning database 6 in advance, the third word is translated from the source word (first word). It can also be adopted as a result (conversion candidate). In this case, a third word whose conversion possibility is higher than a predetermined value may be output (displayed) as a translation result.

(Third embodiment)
FIG. 10 is a diagram illustrating a configuration example of a word translation system including a word translation apparatus according to the third embodiment of the present invention.
The word translation system (symbol string conversion system) 1B is a word that is a candidate for conversion to the second word (target word) of the source word that is input to the word translation device 5B together with the first word (source word). 3 words are acquired from outside the word translation device 5B.
This word translation system 1B is the same as the word translation system 1A shown in FIG. 7 except that it includes a word translation device (symbol string conversion device) 5B. For convenience of explanation, this word translation system 1B has the same configuration. Are denoted by the same reference numerals, and description and drawings are omitted as appropriate.

As shown in FIG. 10, the word translation device (symbol string conversion device) 5 </ b> B includes a word output unit 5, a conversion possibility calculation unit 30, and a conversion candidate search unit 40.
The conversion candidate search part 40 inputs the word group extracted based on the document data acquired from the electronic device 50 connected to the communication network NW to the conversion possibility calculation part 30 as a third word.
The communication network NW is composed of, for example, the Internet.
The electronic device 50 is a storage device such as a computer (information processing device) such as a Web server or a hard disk device including a database, for example.

[Configuration of conversion candidate search unit]
FIG. 11 is a functional block diagram illustrating a configuration example of the conversion candidate search unit illustrated in FIG.
As shown in FIG. 11, the conversion candidate search unit 40 includes input means 41, storage means 42, document data acquisition means 43, conversion candidate extraction means 44, and output means 45.

The input means 41 is an input interface, which inputs a source word (first word) from the input device M and outputs it to the document data acquisition means 43. The input unit 41 inputs document data from the communication network NW and outputs it to the document data acquisition unit 43.
The storage means 42 includes a RAM, a ROM, and an HDD, and stores data such as document data input from the input means 41, various operation programs, and the like.

  The document data acquisition unit 43 acquires document data from the electronic device 50 connected to the communication network NW based on the first word (source word) input to the input unit 41. The document data obtaining unit 43 retrieves a document including the input source word by using a known document retrieval method on the Internet or a document retrieval method for a document database. The number of documents to be acquired may be designated from the input device M, or the number of documents designated in advance may be stored in the storage means 42.

  The conversion candidate extraction unit 44 extracts a predetermined number of third words from the document data acquired by the document data acquisition unit 43 and outputs the third word to the output unit 45. The extraction method by the conversion candidate extraction unit 44 is arbitrary, and for example, matching by a regular expression using a character code used in the target language may be used. Note that the number of words to be extracted may be designated from the input device M, or the number of words designated in advance may be stored in the storage means 42.

The document data acquisition unit 43 and the conversion candidate extraction unit 44 are realized by the CPU developing and executing a predetermined program stored in the ROM or the like of the storage unit 42 on the RAM. .
The output unit 45 is an output interface to the output device D, and outputs the third word extracted by the conversion candidate extraction unit 44 to the output device D.

[Operation of conversion candidate search unit]
The operation of the conversion candidate search unit 40 will be described with reference to FIG. 12 (see FIG. 11 as appropriate).
FIG. 12 is a flowchart showing the operation of the conversion candidate search unit shown in FIG.
The conversion candidate search part 40 inputs the 1st word (symbol string) which is a translation object from the input device 41 as a source word by the input means 41 (step S31).
Subsequently, the conversion candidate search unit 40 acquires document data from the communication network NW based on the input first word (source word) by the document data acquisition unit 43 (step S32).
Subsequently, the conversion candidate search unit 40 uses the conversion candidate extraction unit 44 to extract a third word (symbol string) that is a conversion candidate from the acquired document data (step S33).
And the conversion candidate search part 40 outputs the extracted 3rd word to the conversion possibility calculation part 30 by the output means 45 (step S34). Thereby, in the conversion possibility calculation part 30, the 3rd word will be input into the synthetic | combination state transition information database preparation means 33 (refer FIG. 8) by the input means 31 (refer FIG. 8).

  According to the third embodiment, the word extracted from the document data acquired from the communication network can be input as the third word, and the convertibility of the third word can be calculated. Therefore, even when an appropriate second word is not searched, if the conversion possibility of the third word acquired from the communication network is an appropriate result, the third word is adopted as a conversion candidate. Is possible.

  As mentioned above, although each embodiment of this invention was described, this invention is not limited to these, In the range which does not change the meaning, it can implement variously. For example, in each embodiment, a word composed of characters belonging to a certain language system is a conversion target (translation target). However, the “language” in this case is not limited to a natural language, It may be a symbol system based on rules. In this case, the present invention can be realized as a symbol string conversion method, a symbol string conversion device, and a symbol string conversion program for converting a symbol string composed of symbols belonging to this symbol system.

  In the word translation device 5A of the second embodiment and the word translation device 5B of the third embodiment, the input means 21, storage means 22, and output means described in the word translation device 5 of the first embodiment, respectively. 25, input means, storage means, and output means are provided separately, but these may have a common configuration, or the input device M and the output device D may be shared.

  Further, the word translation device 5B of the third embodiment has been described on the assumption that the word conversion possibility acquired from the communication network NW is calculated, but the conversion possibility calculation unit 30 is not an essential component. When the word translation device 5B does not include the convertibility calculation unit 30, for example, when the word output unit 5 determines that the source word (first word) is an unknown word, it is acquired from the communication network NW. The one or more words may be used as conversion candidates from the source word (first word) and output (displayed) to the output device D.

  Next, a plurality of examples in which the effect of the present invention has been confirmed will be described. In each example, word conversion was performed when the source language was Japanese (indicated in katakana) and the target language was in English (indicated in alphabet).

[Example 1]
In the word translation system 1 (see FIG. 1), a transliteration probability model 7 is created in advance by the transliteration probability model creation device 4, and the source word “Donald” is created using the word translation device 5 of the first embodiment. The target word “donald” was acquired.
In this case, the transliteration probability model creation device 4 created the transliteration probability model 7 as follows.
First, in the learning database 6, as illustrated in FIG. 13A, pairs of katakana notation 1301 words and alphabet notation 1302 words are stored.

Further, the inter-character relevance database creating means 12 (see FIG. 2) of the transliteration probability model creating apparatus 4 uses φ ² (s, t) represented by the equation (7) as the relevance. This φ ² (s, t) is a value obtained by normalizing the chi-square value to a range of 0 to 1 (for details, see WA Gale and KW Church, “Identifying word correspondances in parallel texts Proceedings of the 4th DARPA workshop on Speech and Natural Language, 1991).

  Here, freq (*) indicates the number of word groups in which the symbol * appears in the learning database 6. That is, freq (s) indicates the number of word pairs in which the source character s appears, freq (t) indicates the number of word pairs in which the target character t appears, and freq (s, t) indicates both words that appear. Indicates the number of pairs. L is the total number of all word groups stored in the learning database 6.

This inter-character relevance database creating means 12 deletes the blank on the target character side that appears as a break between English words and treats it as a series of words.
In the created inter-character relevance database 16, as illustrated in FIG. 13B, the relevance 1312 with the source character is stored for each target character 1311 in the learning database 6. For example, when the target character 1311 is “a”, the source character “A” has a relevance level of “0.312370273233768”, and the source character “La”, the source character “NA”, etc. have a predetermined relevance level. ing. Similarly, when the target character 1311 is “b”, it is indicated that the source character “B” has an association degree of “0.247172957562107”.

Further, the inter-character relevance database creating means 12 calculates the geometric mean of the relevance between the source character s and other target characters as the relevance Assoc (s, φ _t ) with the empty character φ _t with the source character s. As the degree of association Assoc (φ _s , t) between the empty character φ _s and the target character t, the geometric average of the degree of association between the target character t and another source character was used.

  Next, the word set database creation means 13 (see FIG. 2) of the transliteration probability model creation device 4 performs the learning database 6 as shown in FIG. 13A and the character spacing as shown in FIG. Using the relevance database 16, for each word set in the learning database 6, an association between characters that satisfies the above-described equation (1) is obtained, and a word set database 17 as illustrated in FIG. 14 is obtained. create.

  As shown in FIG. 14, the association 1401 describes the associated target character and source character connected by a colon. Here, <eps> is a symbol representing the empty character φ, <s>: <s> at the left end of each line indicates the starting point of the word, and </ s>: </ s> at the right end of each line is A symbol representing the end point of a word.

  Next, the occurrence probability calculation means 14 (see FIG. 2) of the transliteration probability model creation apparatus 4 created a trigram model of N = 3 from the word set database 17 shown in FIG. Here, the occurrence probability calculation unit 14 actually created a unigram model with N = 1, a bigram model with N = 2, and a trigram model with N = 3. As a result, the transliteration probability model 7 stores the co-occurrence probability of each character set in three types as shown in FIG.

As shown in FIG. 15A, the unigram model (N = 1) includes occurrence probability 1501, first notation 1502, and smoothing coefficient 1503.
The occurrence probability 1501 indicates a logarithm of the probability that a transliterated character set will occur regardless of the immediately preceding word.
A first notation 1502 is a character set notation. A smoothing coefficient 1503 is a coefficient for smoothing, and is used to settle N-gram probabilities of N> 1.

As shown in FIG. 15B, the bigram model (N = 2) includes data of an occurrence probability 1511, a second notation 1512, a third notation 1513, and a smoothing coefficient 1514.
The occurrence probability 1511 indicates a logarithm of the probability that a transliterated character set occurs depending on the immediately preceding word.
The second notation 1512 is a notation of the immediately preceding character set.
A third notation 1513 is a character set used for obtaining the occurrence probability 1511.
The smoothing coefficient 1514 is a coefficient for smoothing, and is used to estimate the probability of N gram of N> 2.

The trigram model (N = 3) includes data of occurrence probabilities 1521, fourth notation 1522, fifth notation 1523, and sixth notation 1524, as shown in FIG.
The occurrence probability 1521 indicates a logarithm of the probability that a transliterated character set occurs depending on the immediately preceding two words.
The fourth notation 1522 is the notation of the previous character set.
The fifth notation 1523 is the notation of the immediately preceding character set.
The sixth notation 1524 is a character set used to determine the occurrence probability.

  In creating these models, the occurrence probability calculation means 14 used a known language model creation tool “CMU-Cambridge SLM Toolkit” (P. Clarkson et a1, “Statistical language modeling using the CMU-Cambridge toolkit”, in Proceedings of EUROSPEECH97, 1997, p.2707-2710). At this time, each character set connected by a colon shown in FIG. 14 or FIG. 15 is treated as one symbol.

  As a smoothing method for these models, the occurrence probability calculation means 14 uses a well-known Witten-Bell smoothing method (IH Witten and TC Bell, “The zero-frequency problem: Estimating the prob-abilities of novel events in adaptive text compression ”, IEEE Transaction of Information Theory, 1991 vol. 37, no. 4, p. 1085-1094).

  Next, the word translation device 5 according to the first embodiment uses the state transition information database creation unit 23 (see FIG. 4) to change the state transition as illustrated in FIG. 16 from the transliteration probability model 7 illustrated in FIG. An information database 26 is created. At this time, each character set connected by the colon shown in FIG. 15 is disassembled, the source character is used as input correspondence data of the state transition information database 26, and the target character is output correspondence of the state transition information database 26. Used as data.

As shown in FIG. 16, the state transition information database 26 stores a state identification 1601, a first state number 1602, a second state number 1603, a source character 1604, a target character 1605, and a state transition weight 1606. is doing.
The state identification 1601 indicates an initial state “I”, a transition state “T”, and an end (acceptance) state “F”.
The first state number 1602 indicates a transition source state number. However, in the initial state “I” and the end (acceptance) state “F”, the state number of the initial state and the state number of the end state are shown. The second state number 1603 indicates a transition destination state number.
The source character 1604 is input correspondence data corresponding to an input symbol, and is generated by a source character obtained by decomposing each character set connected by a colon shown in FIG.
A target character 1605 is output correspondence data corresponding to an output symbol, and is generated by a target character obtained by decomposing each character set connected by a colon shown in FIG.
The state transition weight 1606 is a weight (state transition weight) given to the transition. However, in the initial state “I” and the end (acceptance) state “F”, the initial state weight and the end state weight are shown. In FIG. 16, the weight in the initial state and the weight in the end state are substantially “0”.

  This state transition information database 26 reflects the context information (conditional probability) of the transliteration probability model 7. Specifically, in the transliteration probability model 7 illustrated in FIG. 15, the simultaneous occurrence probability of the transliterated character set is determined by the last two transliterated character sets (in the case of (c) in FIG. 15). Yes. The co-occurrence probability (context information or conditional probability) at this time is held in the first state number 1602 and the second state number 1603 (each state) in FIG.

  With each database maintained as described above, the source word “Donald” is input as the input word 1701 to the word translation apparatus 5 of the first embodiment as an example as shown in FIG. Input and translated. That is, the source words are converted into target words using WFST in which the state transition information database 26 illustrated in FIG. 16 is a WFST database and the word search means 24 (see FIG. 4) is a WFST search program. An output example of the word search means 24 at this time is shown in FIG.

In the output example shown in FIG. 17B, a corresponding state number 1711, an input symbol 1712, an output symbol 1713, and a state transition weight 1714 are stored.
The correspondence state number 1711 is a state number indicating the state of a character set that constitutes a word set including a source word and a corresponding target word. Here, the state of the character set reflects the conditional probability.
An input symbol 1712 indicates a sequence of source characters when an optimum association is executed for the source word. When there is no character corresponding to the target word, <eps> is described instead of the empty character φ. <S> indicates the start point of the word, and </ s> is a symbol indicating the end point of the word.

The output symbol 1713 is similar to the input symbol 1712 regarding the target word. When the series of output symbols 1713 are connected, “donald” is generated as the output word 1721 as shown in FIG. The generated “donald” is displayed on the output device D (see FIG. 4).
The state transition weight 1714 is a value obtained by reversing the log of the conditional probability.
In this example, the conversion is correctly performed according to the spelling, but even if the conversion result is not spelled correctly, if the search space can be enlarged and more conversion candidates can be obtained, for example, in the information retrieval system, It can be used by including it in the query.

  Moreover, you may make it the word translation apparatus 5 output the value of the state transition weight 1714 shown to (b) of FIG. 17 with a target word. In this case, the value of this state transition weight can be used as the possibility of conversion indicating how likely the translation (character conversion) is between the target character and the input source character. If the total value of the state transition weights 1714 shown in FIG. 17B is output, it is possible to indicate how likely the translation is between the target word and the input source word.

[Example 2]
The second embodiment is obtained by adding the following contents to the first embodiment. That is, in the word translation system 1A (see FIG. 7), using the word translation device 5A of the second embodiment, three third candidates are converted as conversion candidates for the source word “leopard” that is the first word. The possibility of conversion to each word when “leopard”, “lion”, and “leopon” are input was calculated. All alphabets are replaced with lowercase letters.

  Therefore, in the convertibility calculation unit 30 (see FIG. 8) of the word translation device 5A, the synthesized state transition information database creation unit 33 acquires a third word, for example, “leopard” via the input unit 31, As shown in FIG. 18, an FSA (finite state automaton) that is a database corresponding to the “leopard” is created. This FSA includes a transition source state number 1801, a transition destination state number 1802, an input character notation 1803, an output character notation 1804, and a state transition weight 1805. For example, if the current state is “1 (one)” and the symbol “l (el)” is received as the input character notation 1803 in the row where the transition source state number 1801 is “1 (one)”, the state transition The weight 1805 is set to “0”, the symbol “<eps>” is output as the output character notation 1804 (in this case, nothing is done), and the transition destination state number 1802 is set to the state “2”. Transition. Note that the notation is as described above.

  Then, the combined state transition information database creation unit 33 (see FIG. 8) combines the FSA illustrated in FIG. 18 and the state transition information database 26 illustrated in FIG. As a result, a combined state transition information database 36 (WFSA database) in which “leopard” is combined is created. Similarly, although not shown, an FSA is created for the other third words “lion” and “leopon” to be input, and is synthesized with the state transition information database 26 illustrated in FIG. Each state transition information database 36 is created. Then, the word translation device 5A refers to the created state transition weight information database 36 by the state transition weight calculation means 34 (see FIG. 8), for example, when “leopaard” is input, and the state transition weight Calculate Similarly, state transition weights are calculated for “lion” and “leopon”, respectively. The calculation results at that time are shown in FIG.

  As shown in FIGS. 19A to 19C, the convertibility calculation unit 30 outputs an output word 1901 corresponding to each third word input as a conversion candidate, and a cumulative weight 1902. Here, the cumulative weight 1902 is the total value of the state transition weights from the source character to the third character. For example, with respect to the input “leopard”, “leopard” is output as the output word 1901 and “12.799” is output as the cumulative weight 1902 as shown in FIG. Similarly, for “lion”, “23.4622” is output as the cumulative weight 1902 as shown in FIG. 19B, and for “leapon”, as shown in FIG. 19C, “17.98” is output as the cumulative weight 1902. Therefore, in this example, the cumulative weight 1902 of “leopard” searched for “leopard” is the smallest. In other words, among the three input third words, “leopaard” has the highest conversion possibility.

  Note that the conversion possibility calculation unit 30 may output the calculation result in the form of FSA (finite state automaton) that is a database, as shown in (d) to (f) of FIG. These FSAs include a corresponding state number 1911, an input character notation 1912, an output character notation 1913, and a state transition weight 1914. Each item is the same as the FSA shown in FIG.

  Further, instead of the third embodiment, the third words “leopaard”, “lion” and “leopon” and the state transition information database 26 are combined at a time to create a unique combined state transition information database 36. You may make it create. In this case, each input third word has a probability calculated based on this unique combined state transition information database 36 and has the highest conversion possibility (in this case, “leopard”). Are output as the first candidate, and are output in descending order of possibility of conversion. As a result, the conversion possibilities can be compared at a time.

[Example 3]
Example 3 is obtained by adding the following contents to Example 2. That is, in the word translation system 1B (see FIG. 10), using the word translation device 5B of the third embodiment, the source word “Super-Kamiokande”, which is the first word, is a third conversion candidate. The word was acquired using the Internet, and the possibility of conversion to the correct “Super-Kamiokande” was calculated.

  The conversion candidate search unit 40 of the word translation device 5B searches the document containing the source word “Super-Kamiokande” by the document data acquisition unit 43 (see FIG. 11) and stores it in the storage unit 42 as a database. An example of the database stored in the storage means 42 is shown in FIG. As shown in FIG. 20A, this database includes a document title 2001 and a URL 2002 indicating the address of a home page on which the document is posted, and stores ten document data (No. 501-510). is doing.

  The conversion candidate search unit 40 accesses the URL 2002 of each document data (No. 501-510) by the conversion candidate extraction unit 44 (see FIG. 11), and selects a word suitable as a conversion candidate for the source word “Super-Kamiokande”. Extraction was performed as shown in FIG. Here, as shown in FIG. 20B, the conversion candidate extracting unit 44 extracts a total of ten alphabetic words extracted one by one from each document data (No, 501-510) as extracted words 2011. Is output to the conversion possibility calculation unit 30 as the third word.

  Thereby, the convertibility calculation unit 30 performs the operation described in the second embodiment on the ten third words input from the conversion candidate search unit 40, and calculates the convertibility (weight). ,Output. An example of the output data is shown in FIG. As shown in (c) of FIG. 20, the output data has extracted words 2021 and convertibility (weight) 2022 as items. According to this output data, “Super-Kamiokande” No. 611 has a conversion possibility (weight) 2022 of “19.9722”, and this weight is the smallest value among the ten extracted words 2021. Yes. In other words, No. 611 “Super-Kamiokande” has the highest conversion possibility. As a result, the word translation device 5B obtains a plurality of conversion candidate words using the Internet for the source word “Super-Kamiokande”, and among them, conversion to the correct “Super-Kamiokande” is possible. We were able to ask for the greatest sex.

  According to the third embodiment, when the input source word (similarly the target word) is not registered in the learning database 6 when the transliteration probability model 7 used by the word translation device 5B is created. However, the source word can be converted into the most likely target word. That is, in the first stage, the third word acquired on the Internet is synthesized as FSA in the state transition information database 26, and in the second stage, the first word and the acquired third word are input. The conversion possibility is calculated based on the composite state transition information database 36 created in advance, and the third word having the highest conversion possibility is selected. Therefore, the word translation apparatus 5B does not depend on a fixed dictionary, but uses actual words acquired via the communication network NW as conversion candidates, so that the practicality becomes high. For this reason, it is suitable for the translation of proper nouns, such as a person name and a place name, and the translation of articles, news, etc. in which new words are used one after another.

It is a figure which shows the structural example of the word translation system containing the word translation apparatus which concerns on the 1st Embodiment of this invention. It is a functional block diagram which shows the structural example of the transliteration probability model creation apparatus shown in FIG. It is explanatory drawing which shows the specific example until the word set database shown in FIG. 2 is created. It is a functional block diagram which shows the structural example of the word translation apparatus shown in FIG. It is a flowchart which shows operation | movement of the transliteration probability model creation apparatus shown in FIG. It is a flowchart which shows operation | movement of the word translation apparatus shown in FIG. It is a figure which shows the structural example of the word translation system containing the word translation apparatus which concerns on the 2nd Embodiment of this invention. It is a functional block diagram which shows the structural example of the conversion possibility calculation part shown in FIG. It is a flowchart which shows operation | movement of the conversion possibility calculation part shown in FIG. It is a figure which shows the structural example of the word translation system containing the word translation apparatus which concerns on the 3rd Embodiment of this invention. It is a functional block diagram which shows the structural example of the conversion candidate search part shown in FIG. It is a flowchart which shows operation | movement of the conversion candidate search part shown in FIG. It is explanatory drawing of the Example which converts a katakana into an alphabet, (a) is an example of a learning database, (b) is an example of an inter-character relationship degree database. It is explanatory drawing which shows an example of the word set database with a correspondence between characters produced from the learning database shown to (a) of FIG. It is explanatory drawing which shows an example of the transliteration probability model database shown in FIG. 1, (a) is a unigram, (b) is a bigram, (c) has each shown the data of a trigram. It is explanatory drawing which shows an example of the state transition information database shown in FIG. It is explanatory drawing which shows an example of word translation, (a) is an input word, (b) is a search result, (c) has shown the output word. It is explanatory drawing which shows an example of FSA regarding a 3rd word. It is explanatory drawing which shows an example of the search result at the time of inputting the 3rd word to the word translation apparatus which concerns on 2nd Embodiment. It is explanatory drawing which shows an example of the search of the 3rd word by the conversion candidate search part of the word translation apparatus concerning 3rd Embodiment, (a) is a search document, (b) is an extraction word, (c) is output Results are shown.

Explanation of symbols

1 (1A, 1B) Word translation system (symbol string conversion system)
2 storage device 3 storage device 4 transliteration probability model creation device (symbol conversion probability model creation device)
5 Word translation device (word output unit)
5A, 5B Word translation device (symbol string conversion device)
6 Learning database 7 Transliteration probability model (symbol conversion probability model)
10 input means 11 storage means 12 character-to-character relevance database creation means (symbol relevance degree database creation means)
13 Word set database creation means (symbol string set database creation means)
14 occurrence probability calculation means 15 writing means 16 inter-character relevance database (inter-symbol relevance database)
17 Word set database (symbol string set database)
M input device 21 input means (first input means)
22 storage means 23 state transition information database creation means (database creation means)
24 word search means 25 output means (first output means)
26 state transition information database D output device 30 convertibility calculation unit 31 input means (second input means)
32 storage means 33 composite state transition information database creation means 34 state transition weight calculation means 35 output means (second output means)
36 Compound State Transition Information Database 40 Conversion Candidate Search Unit 50 Electronic Device N Communication Network 41 Input Means 42 Storage Means 43 Document Data Acquisition Means 44 Conversion Candidate Extraction Means 45 Output Means

Claims (11)

A symbol string conversion method of a symbol string conversion device that uses a symbol co-occurrence frequency in a symbol string set that is a combination of symbol strings of the same meaning belonging to different symbol systems,
The symbol string converter is
Inputting a first symbol string belonging to the first symbol system;
A second symbol string belonging to a second symbol system corresponding to the input first symbol string is estimated using the co-occurrence frequency and the frequency of the appearance order of the symbol groups in the symbol string set And steps to
Outputting the estimated second symbol string;
A symbol string conversion method comprising: The step of estimating the second symbol string includes:
Based on the co-occurrence frequency, the frequency of the appearance order is calculated for each symbol string set composed of a set in which the first symbol string and the second symbol string are associated with each other in symbol units, Based on the calculation result, a symbol string set that maximizes the frequency of the appearance order is searched, and the second symbol string is estimated using the frequency of the appearance order related to the searched symbol string set. The symbol string conversion method according to claim 1. The step of estimating the second symbol string includes:
Based on the co-occurrence frequency, the frequency of the appearance order is calculated for each symbol string set composed of a set in which the first symbol string and the second symbol string are associated with each other in symbol units, Based on the calculation result, searching for a symbol string set that maximizes the frequency of the appearance order, and creating a database that stores the frequency of the appearance order related to the searched symbol string set as data;
Searching the second symbol string with reference to the database;
The symbol string conversion method according to claim 1 or 2, characterized by comprising: The symbol string conversion method according to any one of claims 1 to 3,
A word translation method in which the symbol string is a word composed of letters,
The symbol string converter is
Obtaining document data from an electronic device connected to a communication network based on an input word;
Extracting a predetermined number of words from the acquired document data as conversion candidates from the input words;
A word translation method, further comprising: A word translation device using the co-occurrence frequency of characters in a word set, which is a combination of words of the same meaning belonging to different language systems,
An input means for inputting a first word belonging to the first language system;
A word search for estimating a second word belonging to a second language system corresponding to the input first word, using the co-occurrence frequency and the appearance order frequency of the word set Means,
Output means for outputting the estimated second word;
A word translation device comprising: The word search means includes
Based on the co-occurrence frequency, the frequency of the appearance order is calculated for each arbitrary word set composed of a set in which the first word and the second word are associated with each other in units of characters. The second word is estimated using the frequency of the appearance order relating to the searched word set based on the search for the word set having the highest appearance order frequency. Item 6. The word translation device according to Item 5. Based on the co-occurrence frequency, the frequency of the appearance order is calculated for each arbitrary word set composed of a set in which the first word and the second word are associated with each other in units of characters. A database creation means for searching for a word set having the maximum frequency of appearance order and creating a database for storing the frequency of appearance order related to the searched word set as data;
The word translation device according to claim 5 or 6, wherein the word search means searches for the second word with reference to the database. Document data acquisition means for acquiring document data from an electronic device connected to a communication network based on the first word input to the input means;
Conversion candidate extraction means for extracting a predetermined number of words as conversion candidates from the first word from the acquired document data;
The word translation device according to any one of claims 5 to 7, further comprising:   A symbol string conversion program for causing a computer to execute the symbol string conversion method according to any one of claims 1 to 3.   A word translation program for causing a computer to execute the word translation method according to claim 4.   A recording medium in which the symbol string conversion program according to claim 9 or the word translation program according to claim 10 is recorded.

Images