JP6020093B2 - Alphabet reading estimation device - Google Patents

Description

  The present invention relates to an apparatus for estimating reading of an alphabetic character string.

  A conventional alphabet reading estimation apparatus uses a large amount of data in which an alphabet character string and its Japanese reading are associated with each other, and an gramm of a plurality of character strings obtained by decomposing an input alphabet character string and its Japanese reading. Multi-element linkage) frequency (for example, “tio, sho” and “n, n” chain frequency) is calculated and based on the transliteration model using this ngram frequency, the input alphabetic word is read in Japanese It was the device which estimates. (For example, Patent Document 1)

Japanese Patent No. 4084515

The alphabet reading estimation device disclosed in Patent Literature 1 is decomposed into partial alphabet character strings in units of predetermined character strings when an alphabet character string for estimating Japanese readings is input, and the decomposed partial alphabet characters The ngram frequency of the column is referenced from a large amount of data, and Japanese readings are estimated from the referenced data based on the transliteration model.

  In general, the alphabet reading estimation device can estimate reading more accurately if there is a longer character string corresponding to the alphabet character string input in the word dictionary. However, since it is necessary to store a large amount of data in the word dictionary in order to include a lot of longer character strings, an extra memory or the like is required, resulting in an increase in hardware cost. On the other hand, when the gram frequency table is configured with a small amount of data, there is a problem that the reliability of the data is lacking.

  An object of the present invention is to enable estimation of reading of a specific language such as Japanese of an alphabet with high reliability even if there is no large amount of data in a word dictionary.

  The present invention refers to a dictionary in which an input alphabet character string is decomposed into a plurality of partial alphabet character strings, and the partial alphabet character string and a plurality of partial readings that are candidates for reading are associated with each other and registered. In the alphabet reading estimation device for estimating partial reading for each partial alphabet character string, based on a dictionary partial alphabet character string from which a plurality of partial readings are estimated among a plurality of dictionary partial alphabet character strings constituting the alphabet character string The dictionary feature alphabet character string created in correspondence with the plurality of partial readings, and a partial score dictionary representing the frequency with which the reading of the dictionary feature alphabet character string is used as the partial reading, and the input Among the multiple partial alphabet strings that make up the alphabet string, multiple A characteristic alphabet character string creating unit that creates a plurality of characteristic alphabet character strings based on a partial alphabet character string on which partial reading is estimated, and a dictionary feature on which the plurality of partial readings are estimated with reference to the partial score dictionary A characteristic alphabet character string score calculation unit that calculates a score of the characteristic alphabet character string that matches the alphabet character string for each of the plurality of partial readings, and estimates a partial reading based on the score calculated for each of the plurality of partial readings And an alphabetic character string reading determination unit.

  The alphabet reading estimation device according to the present invention decomposes an input alphabet character string into partial alphabet character strings, further extracts characteristic alphabet character strings from the partial alphabet character strings, and scores them according to the frequency used as partial readings. Since the partial reading is estimated based on the score, the reading of a specific language with high reliability can be provided even with a small amount of data.

FIG. 3 is a diagram illustrating a device configuration of the alphabet reading estimation device according to the first embodiment. FIG. 3 is a diagram showing an example of a heading dictionary according to the first embodiment. FIG. 4 is a diagram showing an example of a partial score dictionary according to the first embodiment. The figure which shows the operation | movement flowchart of the alphabet reading estimation apparatus which concerns on Embodiment 1. FIG. FIG. 5 is a diagram showing an example of a headline series according to the first embodiment. FIG. 4 is a diagram illustrating an example of a characteristic alphabet character string according to the first embodiment. The figure which shows the example of the total score of the partial alphabetic character string which concerns on Embodiment 1. FIG. The figure which shows the example of the partial score dictionary which concerns on Embodiment 2. FIG. The figure which shows the example of the total score of the partial alphabet character string which concerns on Embodiment 2. FIG.

Embodiment 1
Hereinafter, Embodiment 1 of the present invention will be described with reference to FIGS. FIG. 1 is a diagram showing a device configuration of the alphabet reading estimation device according to the first embodiment. FIG. 2 is a diagram showing an example of a heading dictionary according to the first embodiment. FIG. 3 is a diagram showing an example of the partial score dictionary according to the first embodiment. FIG. 4 is a diagram showing an operation flowchart of the alphabet reading estimation apparatus according to the first embodiment. FIG. 5 is a diagram illustrating an example of a heading sequence according to the first embodiment. FIG. 6 is a diagram illustrating an example of a characteristic alphabet character string according to the first embodiment. FIG. 7 is a diagram showing an example of the total score of the partial alphabetic character strings according to the first embodiment.

  Hereinafter, the configuration of the alphabet reading estimation apparatus 1 according to Embodiment 1 will be described with reference to FIG.

  The alphabet reading estimation apparatus 1 according to Embodiment 1 includes an input device 10, an information processing unit 20, a storage unit 30, and an output device 40. Alphabetic character strings for estimating Japanese readings (hereinafter referred to as “input alphabet character strings”) are input from the input device 10 to the information processing unit 20, and Japanese readings are estimated by the information processing unit 20 to output the output device 40. Is output from. For example, when an input alphabet character string “reversal” is input from the input device 10, the information processing unit 20 creates a Japanese reading “reversal” and outputs it from the output device 40.

  The input device 10 is a device that inputs an input alphabet character string to the information processing unit 20. For example, a keyboard or a touch panel corresponds to this.

  The memory | storage part 30 is an apparatus which memorize | stores the various data (Heading dictionary 31 and the partial score dictionary 32) used for the estimation of the Japanese reading of an input alphabet character string. This is the case for hard disk drives, for example.

  The heading dictionary 31 is data in which an alphabet character string is decomposed into a plurality of character strings (hereinafter referred to as “dictionary partial alphabet character string”) and its Japanese reading (hereinafter referred to as “partial reading”). It is. The heading dictionary 31 is created in advance by assigning Japanese readings to an alphabetic character string and associating them with each other, and further taking correspondence between the alphabetic character string and the Japanese reading sound. For example, FIG. 2 is an example of the heading dictionary 31. The heading dictionary 31 is created from the alphabetic character strings “conversion”, “disposal”, “anniversary”, and “recognize”. If these alphabetic character strings are given Japanese readings and separated by sounds, “co / n / ver / sa / tio / n” = “ka / n / ba / se / sho / on”, “di / s / Po / sa / l ”=“ Di / Su / Po / The / Le ”,“ a / nni / ver / sa / ry ”=“ A / Di / Bar / Sa / Lee ”,“ re / co / g / Ni / ze ”=“ re / co / g / nai / z ”. For example, a partial reading “a” is associated with the line number 1 in the dictionary partial alphabet character string “a”. This is a heading dictionary 31 created from the acronym “a / nni / ver / sa / ry” = “a / ni / bar / sa / ree”. The dictionary partial alphabet character string “sa” has a plurality of corresponding partial readings such as “se”, “za”, and “sa”, so a plurality of partial readings are associated with each other.

  The partial score dictionary 32 is score information composed of partial scores 33 representing the frequency with which partial reading corresponding to the dictionary characteristic alphabet character string extracted from the dictionary partial alphabet character string is used. Here, the dictionary characteristic alphabet character string is a character string extracted from the dictionary partial alphabet character string based on the alphabet character string used when the index dictionary 31 was created. “Character string”, “Character string extracted by decomposing dictionary partial alphabet character string” and “Character string extracted including characters before or after the dictionary partial alphabet character string”. Moreover, the frequency used is a frequency which attaches a specific partial reading to one partial alphabet character string. Here, FIG. 3 is an example of the partial score dictionary 32 according to the first embodiment. The partial scoring dictionary 32 is created from the alphabet strings “conversion”, “disposal”, “anniversary”, “recognize”, the partial alphabet string “sa” obtained from these alphabet strings, and its partial readings. It is a thing. In the dictionary partial alphabet character string “sa”, the “dictionary partial alphabet character string” is “sa”, and the “character string extracted by decomposing the dictionary partial alphabet character string” is “s”, “a”, “dictionary partial alphabet” “O-sa”, “r-sa”, “sa + l”, “sa + t”, “po-sa”, “er-sa” are extracted from the character string including the preceding or following characters. It becomes. In FIG. 3, when the frequency of “sa” in the dictionary characteristic alphabet character string “o-sa” is read as “sa” (partial reading), the dictionary characteristic alphabet character string “o-sa” Of these, the partial score 33 of the line of partial reading “sa” is set high. In addition, “−” and “+” of the dictionary characteristic alphabet character string are “dictionary partial alphabet character string” and “previous character string extracted from the dictionary partial alphabet character string including the preceding or following characters”. It is a symbol indicating a connection relationship with “or a subsequent character”. For example, “−” indicates that “dictionary part alphabet character string” and “previous character” are connected, and “+” indicates that “dictionary part alphabet character string” and “following character” are connected. Indicates that The partial score dictionary 32 is created by assigning Japanese readings to an alphabetic character string and analyzing the correspondence of sounds. This partial score 33 is a score set from the knowledge of the designer and the results of performance tests on a large number of data.

  The information processing unit 20 is a device that estimates Japanese readings of an input alphabet character string. The information processing unit 20 includes an input alphabet character string decomposition unit 21, a headline series creation unit 22, an estimation target determination unit 23, a feature alphabet character string creation unit 24, a partial alphabet character string score calculation unit 25, and a partial alphabet character string reading determination unit. 26.

  The input alphabet character string decomposition unit 21 is a device that receives an input alphabet character string from the input device 10 and decomposes it into a plurality of partial alphabet character strings.

  The headline series creation unit 22 is an apparatus that creates a headline series 220 by referring to the headline dictionary 31 and assigning partial readings to partial alphabetic character strings obtained by decomposing an input alphabetic character string.

  The estimation target determining unit 23 is a device that selects a partial alphabet character string for estimating partial reading from among a plurality of partial alphabet character strings.

  The characteristic alphabet character string creating unit 24 is a device that extracts a characteristic alphabet character string from a partial alphabet character string. Characteristic alphabet string is a character string extracted from a partial alphabet string. Specifically, "partial alphabet string", "character string extracted by decomposing partial alphabet string" and "partial alphabet" It is a "character string extracted including the preceding or succeeding characters centering on the character string". For example, in the partial alphabet character string “sa” of the input alphabet character string “reversal”, “partial alphabet character string” is “sa”, “character string extracted by decomposing the partial alphabet character string” is “s”, “A” and “a character string extracted including the preceding or following character centered on a partial alphabetic character string” are “r-sa”, “sa + l”, “er-sa”, and “sa + l $”. The “$” mark means that there are no characters. Note that the “character string extracted by decomposing a partial alphabetic character string” is not created because the partial alphabetic character string cannot be decomposed if it is one character.

  The partial alphabet character string score calculation unit 25 is a device that refers to the partial score dictionary 32 and assigns and adds the score of the characteristic alphabet character string for each partial reading to calculate the total score 201.

  The partial alphabet character string reading determination unit 26 is a device that estimates partial reading of the partial alphabet character string based on the score added by the partial alphabet character string score calculation unit 25.

  The output device 40 is a device that displays Japanese readings of the input alphabet character string estimated by the information processing unit 20. For example, the display of a PC monitor or a car navigation system corresponds to this.

  Next, operation | movement of the alphabet reading estimation apparatus 1 which concerns on Embodiment 1 is demonstrated using FIG. In the description of FIG. 4, an example in which an input alphabet character string “reversal” is input from the input device 10 to the information processing unit 20 and a Japanese reading “reversal” is output to the output device 40 will be described. In this example, Japanese readings or partial readings are shown with katakana characters, but hiragana may be used.

  In step 1, the input device 10 inputs the input alphabet character string “reversal” to the information processing unit 20.

  In step 2, the input alphabet character string decomposition unit 21 decomposes the input alphabet character string “reversal” into a plurality of partial alphabet character strings. Note that the input alphabet character string decomposing unit 21 decomposes the input alphabet character string according to the classification of dictionary partial alphabet character strings registered in the heading dictionary 31. That is, in this example, the dictionary partial alphabet character strings registered in the heading dictionary 31 are “re”, “ver”, “sa”, “l”, and therefore the input alphabet character string decomposition unit 21 converts the input alphabet character string into “ re, ver, sa, l ”. In addition, although “s” and “a” exist in addition to “sa” as a registered dictionary partial alphabet character string in the heading dictionary 31, in this example, a method of decomposing with a longer character string “sa” is adopted. To do.

  In step 3, the headline series creation unit 22 creates a headline series 220 with reference to the headline dictionary 31. The heading series creation unit 22 searches the heading dictionary 31 for a dictionary partial alphabet character string that matches a plurality of decomposed partial alphabet character strings. Further, partial reading of the dictionary partial alphabet character string is assigned to the corresponding partial alphabet character string. Here, FIG. 5 is a diagram showing an example of the header series according to the first embodiment. In this example, the headline series creation unit 22 performs partial reading “re-read” on partial alphabet character strings “re”, “ver”, “sa”, and “l” of the input alphabet character string “re, ver, sa, l”. ”,“ Bar / bar ”,“ se / the / sa ”,“ le ”.

  In step 4, the estimation target determining unit 23 selects a partial alphabet character string for which partial reading is estimated from the partial alphabet character strings of the header series 220 created in step 3. In this example, “sa” is selected.

  In step 5, if a plurality of partial readings are associated with the partial alphabet character string selected in step 4, the process proceeds to step 6. On the other hand, if the selected partial alphabetic character string has one partial reading, the process proceeds to step 8.

  In step 6, the characteristic alphabet character string creating unit 24 creates a characteristic alphabet character string for the partial alphabet character string selected by the estimation target determining unit 23. Here, FIG. 6 is a diagram illustrating an example of the characteristic alphabet character string according to the first embodiment. Since the input alphabet character string input in step 1 is “reversal” and the partial alphabet character string selected in step 4 is “sa”, “partial alphabet character string” (in FIG. 6, “all center” is indicated). ) Becomes “sa”. “Character strings extracted by decomposing partial alphabetic character strings” (indicated as “center 1” and “center 2” in FIG. 6) are “s” and “a”, respectively. Further, “a character string extracted including the preceding or following character centering on the partial alphabetic character string” (in FIG. 6, the character string including the previous character is included up to the “before” and the previous two characters. The character string including the preceding character is indicated as “rear”, and the character string extracted including the character after the second character is indicated as “after”. , “Er−sa”, “sa + l”, and “sa + l $”.

  In step 7, the partial alphabet character string score calculation unit 25 refers to the partial score dictionary 32 and calculates the total score 34 of the partial alphabet character string. FIG. 7A shows an example of the total score 34 of the partial alphabet character string “sa” according to the first embodiment. The partial alphabet character string score calculation unit 25 refers to the partial score dictionary 32 and determines the score of the dictionary characteristic alphabet character string that matches the characteristic alphabet character string of the partial alphabet character string selected in step 4 for each corresponding partial reading. Set. In this example, since the selected partial alphabet character string is “sa” and the partial reading is “se / the / sa” in the heading series 220, the partial alphabet character string score calculation unit 25 has the feature of “sa”. The score of “s, a, sa, r-sa, er-sa, sa + l, sa + l $” of the alphabet string is assigned to each of the partial readings “se / the / sa”. Furthermore, the total score 34 is obtained by adding the partial scores 33 of all characteristic alphabet character strings for each partial reading. Since the characteristic alphabet character string “sa + l $” is not described in the partial score dictionary 501, the total score 34 of the partial alphabet character string is not affected.

  In step 8, the partial alphabet character string reading determination unit 26 refers to the total score 34 of the partial alphabet character strings, and estimates the highest one of the total partial readings 34 as the correct partial reading. In this example, the total score 34 of the partial scores 33 of the characteristic alphabet string “s, a, r-sa, er-sa, sa + 1” of the partial alphabet string “sa” is 1. Three points are the highest. Therefore, the partial alphabetic character string reading determining unit 26 estimates that the partial reading of “sa” is “sa”. If it is determined in step 5 that there is only one partial reading of the selected partial alphabet character string, step 8 determines that the partial reading is a correct reading.

  If it is determined in step 9 that Japanese readings have been estimated for all partial alphabetic character strings in the headline series 220, the process proceeds to step 10. On the other hand, if there is a partial alphabet character string for which Japanese reading is not yet estimated, the process returns to step 4 to repeat the same operation. That is, Japanese readings are also estimated for “re”, “ver”, and “l” that are partial alphabet character strings other than the partial alphabet character string “sa”. For example, FIG. 7B is an example of the total score 34 of the partial alphabetic character string “ver” according to the first embodiment. Similar to the estimation of partial reading of the partial alphabet character string “sa”, a partial reading “bar” having a high total score 34 is estimated as a correct partial reading.

  In step 10, when all partial readings are estimated in step 8, the output device 40 reads the input alphabetic character string reading “re, ver, sal” and the Japanese reading “re, bar, sa, le”. Is output to the output device 40.

  As described above, the operation of the alphabet reading estimation device 1 according to Embodiment 1 ends.

  In the description of the alphabet reading estimation apparatus 1 according to the first embodiment, the input alphabet character string is “reversal”, but this is only an example, and the input alphabet character string is not limited to this as long as it is an unknown input alphabet character string. is not.

  In addition, in the estimation target determination unit 23 according to the first embodiment, an example in which the partial alphabet character string “sa” to be estimated from the heading series 220 is first selected is shown, but the order in which the partial alphabet character strings are selected. Is not limited to this, and “re”, “ver”, “sa”, and “l” may be selected in order from the front.

  Furthermore, although the input alphabet character string showed the example which is English, it should just be a language using alphabets, such as French and German, and is not restricted to English. Moreover, although the partial reading to be output is Japanese reading, it is not limited to this, and another language such as Chinese may be used.

  Note that “feature obtained by decomposing and extracting a partial alphabet character string of the characteristic alphabet character string created by the characteristic alphabet character string creating unit 24” is extracted up to two characters before and after the partial alphabet character string, but is limited to this. Instead, it may be one or more characters before and after.

  As described above, the alphabet reading estimation device 1 according to Embodiment 1 uses the characteristic alphabet character string creating unit 24 to divide a partial alphabet character string into partial alphabet character strings for which reading is estimated, a partial alphabet character string, In addition, a partial alphabet character string and a character string including the character string before and after the partial alphabet character string are created as a characteristic alphabet character string, and the reading is estimated based on the plurality of characteristic alphabet character strings. Even when the data registered in the dictionary 31 is small, the reading can be estimated more accurately.

Embodiment 2
The second embodiment will be described below with reference to FIGS. FIG. 8 is a diagram showing an example of the partial score dictionary according to the second embodiment. FIG. 9 is a diagram showing an example of the total score of the partial alphabetic character strings according to the second embodiment. In addition, since the structure of the alphabet reading estimation apparatus 1 which concerns on Embodiment 2 is the same as that of the alphabet reading estimation apparatus 1 which concerns on Embodiment 1 demonstrated using FIG. 1, description is abbreviate | omitted. Note that the estimation target determining unit 23 according to Embodiment 1 has an arbitrary order of estimating partial readings. On the other hand, the estimation target determining unit 23 according to Embodiment 2 selects partial readings sequentially from the front of the input alphabet character string.

  The partial score dictionary 32 according to the first embodiment is score information including a dictionary characteristic alphabet character string, a partial reading, and a partial score 33. On the other hand, the partial score dictionary 32 according to the second embodiment employs a preceding partial reading instead of a partial reading. The preceding partial reading includes a partial partial reading (estimated partial reading) in addition to the partial reading of the dictionary partial alphabet character string for which the reading is estimated. FIG. 8 is a diagram showing an example of the partial score dictionary according to the second embodiment. The partial score 33 in the case of reading the preceding partial reading “bar / sa” for the dictionary characteristic alphabet character string “r-sa” is “0.6”. On the other hand, the partial score 33 in the case of reading the preceding partial reading “Bar Sa” for the dictionary characteristic alphabet character string “r-sa” is “0.4”. This is because, when the reading corresponding to the dictionary partial alphabet string before the dictionary partial alphabet string “sa” is “bar”, the partial reading of “sa” in the dictionary characteristic alphabet string “r-sa” is This means that there are more cases of reading “sa” than reading “se”. As with the partial score dictionary 32 according to the first embodiment, the partial score dictionary 32 according to the second embodiment is set based on the knowledge of the designer and the results of performance tests on a large number of data.

  Further, the partial alphabet character string score calculation unit 25 according to the first embodiment reads the characteristic alphabet character string of the partial alphabet character string selected by the estimation target determination unit 23 in the heading series 220 and the partial reading corresponding thereto. The extracted score was set according to the partial score dictionary 32. On the other hand, the partial alphabet character string score calculation unit 25 according to Embodiment 2 extracts the preceding partial reading instead of the partial reading. For example, FIG. 9 is an example of the total score of partial alphabetic character strings according to the second embodiment. In this example, it is assumed that the partial alphabet character string selected by the estimation target determining unit 23 is “sa”, and partial reading has already been estimated for the partial alphabet character strings “re” and “ver”. To do. The partial alphabet character string score calculation unit 25 includes the partial reading “bar” of the previous partial alphabet character string already estimated by the partial alphabet character string reading determination unit 26 and the partial alphabet character string selected by the estimation target determination unit 23. The “sa / the / sa” reading of “sa” is extracted and used as the preceding partial reading.

  In this example, since the partial alphabet character string “re” has no preceding partial reading in the estimation of partial reading of the first partial alphabet character string “re” obtained by decomposing the input alphabet character string “reversal”, It is assumed that partial reading is estimated by the same method as in the first mode.

  As described above, the alphabet reading estimation device 1 according to the second embodiment is configured to use the preceding partial reading by the heading dictionary 31 and the partial alphabet character string score calculation unit 25, so that one partial alphabet character string With respect to the partial reading, it is possible to calculate the total score 34 of the partial alphabet character string in consideration of the partial reading of the preceding partial alphabet character string, and it is possible to estimate a more accurate Japanese reading.

Embodiment 3
Hereinafter, the third embodiment will be described. In addition, in the alphabet reading estimation apparatus 1 which concerns on Embodiment 3, description is abbreviate | omitted about what is the same structure and operation | movement as the alphabet reading estimation apparatus 1 which concerns on Embodiment 1 thru | or Embodiment 2. FIG.

  In the alphabet reading estimation apparatus 1 according to the first or second embodiment, the partial reading of the heading dictionary 31 is a character string made up of Japanese (Katakana characters). On the other hand, the alphabet reading estimation device 1 according to Embodiment 3 is characterized in that the partial reading of the heading dictionary 31 is composed of a speech synthesis symbol and a katakana character.

  The symbol for speech synthesis is a symbol composed of Katakana characters, characters that indicate unvoiced syllables (sounds that do not involve vocal cord vibration), and characters that indicate the presence or absence of accents (sound pitch). For example, the symbol for speech synthesis of the alphabet character string “restaurant” is indicated as “less H% L L L L L L”. Here, H indicates that the voice is high, L indicates that the voice is low, and% indicates an unvoiced syllable. For example, “su% L” indicates that the syllable represented by the katakana character S is an unvoiced syllable and the voice is low. The correspondence between the alphabetic character string and the speech synthesis symbol is expressed as “re / s / tau / ra / nt” = “les H / s% L / t L / la L / n L”.

  As described above, the alphabet reading estimation apparatus 1 according to the third embodiment is configured such that the partial reading of the heading dictionary 31 is composed of the speech synthesizing symbols and the katakana characters. Can be estimated.

Embodiment 4
Hereinafter, the fourth embodiment will be described. In addition, in description of the alphabet reading estimation apparatus 1 which concerns on Embodiment 4, description is abbreviate | omitted about what is the same structure and operation | movement as the alphabet reading estimation apparatus 1 which concerns on Embodiment 1 thru | or Embodiment 3. FIG.

  In the alphabet reading estimation apparatus 1 according to the first to third embodiments, the partial score 33 of the partial score dictionary 32 is set based on the knowledge of the designer and the result of the performance test of a large number of data. On the other hand, in the alphabet reading estimation device 1 according to the fourth embodiment, the partial score 33 of the partial score dictionary 32 is set as a partial score 33 by a model parameter of a conditional random field model (CRF). Features.

The conditional random field model is expressed by the following equations (1) to (3). In Expressions (1) to (3), w is a model parameter represented by a vector, x ⁽ⁱ⁾ is the i-th input alphabet character string, y ⁽ⁱ⁾ is the i-th partial reading, and P (y ^(I) | x ⁽ⁱ⁾ ) is the probability (conditional probability ⁾ that y ⁽ⁱ⁾ will occur when x ⁽ⁱ⁾ is given, and C is a constant determined experimentally. | · | Represents the magnitude of the vector.

Note that φ (x ⁽ⁱ⁾ , y ⁽ⁱ⁾ ) is a function that returns a vector value. For φ (x ⁽ⁱ⁾ , y ⁽ⁱ⁾ ), an example of the k-th element of the vector is shown in Equation (4) as in Equation (4).

  Using the steepest gradient method or the like, a model parameter w that maximizes L (w) is obtained. The obtained model parameter w is set as a partial score 33.

  Thus, since the model parameter of the conditional random field model is set to the partial score 33, there is an effect that the partial score 33 can be appropriately and automatically set in the creation of a plurality of characteristic alphabet character strings. The construction time can be reduced.

  DESCRIPTION OF SYMBOLS 1 Alphabet reading estimation device, 10 input device, 20 Information processing unit, 30 Storage unit, 40 Output device, 21 Input alphabet character string decomposition unit, 22 Headline series creation unit, 23 Estimation target determination unit, 24 Feature alphabet character string creation Part, 25 partial alphabet character string score calculation part, 26 partial alphabet character string reading determination part, 31 heading dictionary, 32 partial score dictionary

Claims (10)

The input alphabet character string is decomposed into a plurality of partial alphabet character strings, and the partial alphabet is referred to by referring to a dictionary in which the partial alphabet character string and a plurality of partial readings that are candidates for reading are registered in association with each other. In the alphabet reading estimation device that estimates partial reading for each character string,
Corresponds to the dictionary partial alphabet character string created based on the dictionary partial alphabet character string that is estimated to be a plurality of partial readings among the plurality of dictionary partial alphabet character strings constituting the alphabet character string, and the plurality of partial readings A partial score dictionary representing the frequency with which the reading of the dictionary feature alphabetic character string is used as a partial reading,
A characteristic alphabet character string creating unit that creates a plurality of characteristic alphabet character strings based on a partial alphabet character string in which a plurality of partial readings are estimated among a plurality of partial alphabet character strings constituting the input alphabet character string; ,
A characteristic alphabet character string score calculation unit that calculates a score of the characteristic alphabet character string that matches a dictionary characteristic alphabet character string from which the plurality of partial readings is estimated with reference to the partial score dictionary. When,
A characteristic alphabet character string reading determination unit that estimates partial reading based on the score calculated for each of the plurality of partial readings;
An alphabet reading estimation device comprising: 2. The alphabet reading estimation device according to claim 1, wherein the characteristic alphabet character string creating unit generates a plurality of characteristic alphabet character strings by dividing the partial alphabet character string from which the plurality of partial readings are estimated. . The characteristic alphabet character string creating unit creates a plurality of characteristic alphabet character strings based on the partial alphabet character string from which the plurality of partial readings are estimated and a partial alphabet character string adjacent to the partial alphabet character string. The alphabet reading estimation apparatus according to claim 1 or 2, wherein the alphabet reading estimation apparatus is characterized. The plurality of characteristic alphabet character strings are created by adding a character string one or two characters before the partial alphabet character string to the partial alphabet character string from which the plurality of partial readings are estimated. The alphabet reading estimation apparatus according to claim 3. The plurality of characteristic alphabet character strings are created by adding a character string after one or two characters of the partial alphabet character string to the partial alphabet character string from which the plurality of partial readings are estimated. The alphabet reading estimation apparatus according to claim 3. The alphabet reading estimation apparatus according to claim 3, wherein the plurality of characteristic alphabet character strings are partial alphabet character strings from which the plurality of partial readings are estimated. The plurality of characteristic alphabet character strings are a character string obtained by adding a character string one or two characters before the partial alphabet character string to the partial alphabet character string from which the plurality of partial readings are estimated;
A character string obtained by adding a character string after one or two characters of the partial alphabet character string to the partial alphabet character string from which the plurality of partial readings are estimated;
4. The alphabet reading estimation apparatus according to claim 3, comprising a partial alphabet character string from which the plurality of partial readings are estimated. The partial scoring dictionary includes a partial reading corresponding to a dictionary partial alphabet character string immediately preceding a dictionary partial alphabet character string from which the plurality of partial readings in the alphabet character string are estimated. It consists of partial reading of the alphabet string,
The plurality of partial readings of the characteristic alphabet character string creating unit include partial readings of a partial alphabet character string immediately preceding a partial alphabet character string from which the plurality of partial readings are estimated. The alphabet reading estimation device according to claim 7. 9. The alphabet reading estimation apparatus according to claim 1, wherein the partial score dictionary includes the plurality of partial readings as readings and symbols for speech synthesis. The alphabet reading estimation apparatus according to claim 1, wherein the partial score dictionary determines a score using a conditional random field model.

Images