JP2018067125A - Reading estimation device and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To highly accurately produce a Kanji string with mono-ruby (Furigana where Kana is associated with each character of a compound) without using manually created rules.SOLUTION: A Kanji duplication unit 11 of a reading estimation device 1-1 sequences and duplicates a single Kanji per character constituting an entry word to produce a duplicated single Kanji. A Japanese dictionary storage unit 10 and a proper noun dictionary storage unit 12 store duplicated single Kanji association data for a duplicated single Kanji constituting an entry word and reading about the entry word and reading data thereof. A Kanji and reading association unit 13 calculates an association probability value by machine learning per combination for associating a duplicated single Kanji constituting an entry word with reading per character, and produces duplicated single Kanji association data associating a duplicated single Kanji with reading on the basis of the association probability value. A Kanji restoration unit 14 reads duplicated single Kanji association data from the proper noun dictionary storage unit 12, restores the duplicated single Kanji to the original single Kanji, and produces an entry word with mono-ruby of a proper noun.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a reading estimation device and a program for estimating reading of a Chinese character.

  The furigana attached to kanji is called ruby. There are two types of ruby: mono-ruby with furigana for each kanji (single kanji) and group ruby for the whole kanji string. For example, if the Chinese character string “Chuo-ku” of proper noun is represented by mono-ruby, it becomes “chuo-ku”, and if it is represented by group-ruby, it becomes “chuo-ku”.

  Conventionally, a ruby automatic assigning method for automatically assigning ruby to a Chinese character is known. Many ruby automatic assignment methods are related to group ruby (see, for example, Patent Documents 1 and 2 and Non-Patent Documents 1 and 2), but there are also methods related to mono-ruby (for example, Patent Documents 3 and 3). .

  Non-Patent Document 3 automatically applies a mono-ruby method in which a kanji string and its reading are automatically associated for each single kanji by machine learning. Then, after the automatic association process, the result is corrected using a rule created manually.

  FIG. 8 is a flowchart for explaining a conventional automatic association process. The conventional automatic association processing device (reading estimation device) performs the reading shown for each single kanji character by associating the kanji string and its reading for each single kanji character by machine learning in the processing shown in FIG. Estimate (mono ruby).

  The reading estimation apparatus inputs a kanji string and its reading data (step S801), divides the kanji string into single kanji characters for each character, and divides the reading data into readings for each character (step S802).

  For example, the reading estimation device inputs a kanji string “center” and its reading “chuo”, divides the kanji string “center” into “middle” and “middle”, and converts the readings “chuo” into “chu” “u” “ Divide into “O” and “U”.

  The reading estimation device generates all combinations in which a single kanji for each character constituting the kanji string and a reading for each character constituting the reading data are associated, and the corresponding probability value for each combination is generated by machine learning. Is calculated (step S803).

  In the above-described example, the reading estimation apparatus has “middle: chu”, “middle: u”, “middle: o”, “middle: chu”, “middle: u”, and “middle: ”Are generated. “:” Indicates correspondence.

Then, the reading estimation apparatus applies an EM algorithm as machine learning, for example, to various kanji strings and readings, and calculates a corresponding probability value for each combination. For example, for the Chinese character strings “center”, “junior high school”, “medium” and their readings “chuo”, “chugaku”, “chuuki”, the corresponding probability values for each combination are as follows.
P (central | U) = 0.748288, P (central | o) = 1.00000, P (Study | U) = 0.040140,
P (Study | Ga) = 1.00000, P (Study | Ku) = 1.00000, P (Ki | U) = 0.503310,
P (ki | ki) = 1.00000, P (medium | chu) = 1.00000,
Incidentally, the corresponding probability value of other combinations is 0.

  The reading estimation apparatus associates a single kanji character with a reading based on the corresponding probability value for each combination (step S804). In the above example, the reading estimation device applies the kanji character string “center” and the reading “chuo” (reading “chu” “u1” “o” “u2” for each character in descending order of the corresponding probability values. Then, a single Chinese character and a reading are associated with each other, and “middle: chu”, “middle: u1”, “middle: o”, “middle: u2” are specified. “U1” indicates the first “U”, and “U2” indicates the second “U”.

When the kanji character string “center” and the reading “chuo” are applied in descending order of the corresponding probability values, they are as follows.
P (Medium | Ju) = 1.00000
P (center | o) = 1.00000
P (central | U 1) = 0.748288
P (Center | U2) = 0.748288

  The reading estimation device generates and outputs a kanji string with mono-ruby (step S805). In the above example, the reading estimation device generates and outputs a Chinese character string “Chuo” with mono-ruby for the Chinese character string “center” and the reading “chuo”.

  Thus, in the automatic matching process of the prior art shown in FIG. 8, for example, when the kanji string “center” and the reading “chuo” are associated with single kanji and reading by machine learning, the kanji with monorubi The column will be “Chu”.

  The reason for this incorrect correspondence is that “medium” readings are not only “chu” but also “naka”, and “center” readings are mostly “o”. This is because the corresponding probability value is P (middle | w) <P (center | w). This correspondence probability value is greater for the kanji character string “center” and the reading “chuo” than the probability that the reading “u” is associated with the single kanji character “middle” rather than the probability that the reading “u” is associated with the single kanji character “middle”. It is high. In other words, the first reading “c” of the reading “chuo” is associated with the single kanji character “middle” rather than the single kanji character “middle”.

After such an incorrect association, the result is corrected using a rule created manually. For example, the following rule a is used.
<Rule a>
The reading “Ju” is not associated with a kanji by only one character, and the next reading is also associated with the kanji.

  “Chu” associates the single kanji character “middle” with the reading of only one character. By using the rule a, the next reading “c” is also associated with the single kanji character “middle”. It is done. As a result, “chu” is corrected to “chu”.

  On the other hand, since the first reading “U” of the reading “Woo” is associated with the single Chinese character “Middle”, the remaining reading “O” is associated with the Chinese character “Center”. As a result, for the proper noun kanji character string “center” and the reading “chuo”, the kanji character strings “middle (chu)” and “middle (ou)” with mono-ruby correctly associated are obtained.

JP2015-138273A JP 2004-151847 A JP-A-6-274475

Toru Nagano, Shinsuke Mori, Masafumi Nishimura, "Simultaneous Reading and Accent Estimation for Speech Synthesis", Journal of Information Processing Society of Japan, No.47, Vol.6, pp.1793-1801, 2006 Jun Hatori, Kumi Suzuki, “Predicting Japanese Reading Based on Machine Translation”, The 17th Annual Conference of the Association for Natural Language Processing, pp.579-589, 2011 Taro Miyazaki, Naoto Kato, “Autonomous Monorubi Assignment Method”, The 19th Annual Conference of the Language Processing Society, pp.34-37, 2013

  As described above, the conventional mono-ruby automatic assignment method uses a machine learning to associate a single kanji and a reading for a kanji string and its reading, generates a kanji string with mono rubi, and manually creates a rule. Use to correct the estimated Kanji string with mono-ruby.

  However, in order to increase the accuracy of kanji strings with mono-ruby in the conventional mono-ruby automatic assignment method, it is necessary to use many rules. Since the rules are created manually in advance, there is a problem that it takes time.

  Therefore, the present invention has been made to solve the above-mentioned problems, and its object is to provide a reading estimation apparatus and program capable of generating a kanji string with monorubi with high accuracy without using a manually created rule. Is to provide.

  In order to solve the above-described problem, the reading estimation apparatus according to claim 1 estimates a reading for each single kanji character from a headword and its reading data, and generates a kanji string with monorubi. A dictionary storage unit for inputting word reading data and storing a dictionary including the headword, the reading data, and association data, and a single kanji for each character constituting the kanji string that is the headword. Correspondence between a kanji duplicating unit that generates a plurality of duplicate single kanji characters for each single kanji character, and a plurality of duplicate single kanji characters constituting the headword and a reading for each character constituting the reading data The associated kanji / reading association unit for generating the association data, and the association data are read from the dictionary storage unit, and the duplicate single kanji character included in the association data is restored to the original A kanji restoration unit that restores to a single kanji and generates a kanji string with monoruby of the headword, and the kanji and reading association unit uses the dictionary stored in the dictionary storage unit, Corresponding probability value indicating the degree of correspondence between the duplicate single kanji characters and the readings for each combination in which each of the plurality of duplicate single kanji characters constituting the headword and the readings for each character constituting the reading data correspond to each other The association data is generated based on the correspondence probability value, and the dictionary storage unit stores the association data generated by the kanji and reading association unit. .

  The reading estimation apparatus according to claim 2 estimates a reading for each single kanji from a kanji string and its reading data, and inputs the kanji string and the reading data in a reading estimation apparatus that generates a kanji string with monorubi, A single kanji character for each character constituting the kanji string is duplicated, a plurality of duplicate single kanji characters are generated for the single kanji character, and a plurality of duplicate single kanji characters constituting the kanji string and the reading data are stored. An initializing unit that generates initial association data composed of readings for each character that constitutes, and the duplicate single kanji characters included in the initial association data generated by the initializing unit are associated with the readings For each combination, a corresponding probability value indicating the degree to which the duplicate single kanji character corresponds to the reading is calculated, and based on the correspondence probability value, each of a plurality of duplicate single kanji characters constituting the kanji string and the reading Correspondence data generation means for generating associated correspondence data, and the duplicate single kanji character included in the association data generated by the association data generation means is restored to the original single kanji character, and the kanji string A kanji string generating unit with mono ruby that generates a kanji string with mono rubi.

  The reading estimation apparatus according to claim 3 is the reading estimation apparatus according to claim 1 or 2, wherein each single kanji character constituting the kanji string is duplicated, and 2 for each single kanji character. Two duplicate single kanji characters are generated.

  Furthermore, the Japanese sentence text generation apparatus with mono-ruby according to claim 4 is a Japanese sentence text generation apparatus with mono-ruby that generates a Japanese sentence text with mono-ruby from a Japanese sentence text, and extracts a kanji string from the Japanese sentence text. Then, the Chinese character string extraction unit that generates the reading data of the Chinese character string and the association data generated by the reading estimation device according to any one of claims 1 to 3 are stored corresponding to the Chinese character string. And the association data corresponding to the kanji string extracted by the kanji string extraction unit is read from the memory, the duplicate single kanji included in the association data is restored to the original single kanji, and the monorubi A kanji character sequence generated by the reading estimation unit; and a reading estimation unit that generates a kanji character sequence with the kanji character sequence included in the Japanese sentence text. Replaced kanji string attached, characterized in that it is provided with a, and a text generation unit which generates the Monorubi with Japanese sentence text.

  Furthermore, the sign language CG translation device of claim 5 translates Japanese text to generate sign language CG data, and synthesizes the sign language CG data representing the movement of the mouth with the sign language CG data. 5. A Chinese character string extraction unit that extracts a Chinese character string from the Japanese sentence text and generates reading data of the Chinese character string, and an association generated by the reading estimation device according to claim 1. A memory in which data is stored corresponding to the kanji string, and the association data corresponding to the kanji string extracted by the kanji string extraction unit are read from the memory, and a copy included in the association data A reading estimator that restores a single kanji to the original single kanji and generates a kanji string with monorubi, and a plurality of single kanji characters that compose the kanji string with monorubi generated by the reading estimator For this, the mouth CG data generating unit that generates the mouth CG data corresponding to the single kanji and the mouth CG data generated by the mouth CG data generating unit are combined with the sign language CG data. And a CG synthesis unit.

  Furthermore, a program according to a sixth aspect causes a computer to function as the reading estimation apparatus according to any one of the first to third aspects.

  As described above, according to the present invention, a kanji string with mono-ruby can be generated with high accuracy without using a manually created rule.

It is a block diagram which shows the structural example of the reading estimation apparatus of Example 1. FIG. 6 is a flowchart illustrating a processing example of the reading estimation apparatus according to the first embodiment. It is a figure explaining a proper noun dictionary. It is a block diagram which shows the structural example of the reading estimation apparatus of Example 2. FIG. 10 is a flowchart illustrating a processing example of the reading estimation apparatus according to the second embodiment. It is a block diagram which shows the structural example of the Japanese sentence text production | generation apparatus with a Monorbi. It is a block diagram which shows the structural example of a sign language CG translation apparatus. It is a flowchart explaining the conventional automatic matching process.

Hereinafter, embodiments for carrying out the present invention will be described in detail with reference to the drawings.
〔Overview〕
First, as an outline of the present invention, an example of generating a Chinese character string “Chu”, “Oo”, and “Ku” with monorubi from a Chinese character string “Chuo-ku” and its reading data “Chuu-Ok”. Will be described.

  The reading estimation apparatus according to the embodiment of the present invention first divides the kanji string “center” into single kanji characters “middle”, “center”, and “ku”, and reads the reading data “chuouku” as in the conventional technique. ”,“ U ”,“ O ”,“ U ”and“ K ”.

The reading estimation apparatus duplicates the single kanji characters “middle”, “center”, and “ku” into two characters, and adds the subscript symbols “B” and “E”. As a result, the single Chinese characters “middle”, “middle B”, “middle E” (“middle B / middle E”) are generated, and the single kanji characters “middle” to “middle B” “middle E” (“ “B” and “E” are generated, and “K” and “K” are generated from the single kanji character “K”.
“Medium” → “Medium B / Medium E”
“Center” → “Center B / Center E”
"City" → "City B / City E"

  Here, the two-character break after copying is /. “B” means beginning, and “E” means end. Further, single Chinese characters “middle B”, “middle E”,.

  The reading estimation apparatus is configured to copy each of a plurality of duplicate single kanji characters constituting the kanji string for the duplicate single kanji character string “middle B / middle E / middle B / middle E / ku B / ku E” and the reading data “chuouku”. All combinations corresponding to the reading of each character constituting the reading data are generated. Then, the reading estimation device uses various kanji strings and reading data to calculate a corresponding probability value indicating the degree to which the duplicate single kanji and reading correspond for each combination by machine learning.

  Accordingly, for example, “U” of “Chu” can be expressed at the end of the reading of the single Chinese character “Middle” and not at the beginning of the single Chinese character “Center”. That is, the corresponding probability value calculated by machine learning is P (medium E | c)> P (center B | c).

  This correspondence probability value is the probability that the reading “U” is associated with the duplicate single kanji character “Chu E” for the Chinese character string “Chuo-ku” and the reading “Chuu-Oku”. "Is higher than the probability of matching. That is, as will be described later, the first reading “U” of the reading “CHUOKU” is associated with the single Chinese character “middle” rather than the single Chinese character “middle”.

  The reading estimation apparatus associates the duplicate single kanji and the reading based on the correspondence probability value for each combination, and generates correspondence data. As a result, when the one having a large correspondence probability value is applied, the correspondence data indicating the combination of the correspondence between the duplicate single kanji and the reading is “middle B: chu”, “middle E: u”, “center B: o”, “ Central E: U, “B: Ku”, “E: φ”. Here, “φ” indicates that there is no association.

  The reading estimation device removes the symbols B and E from the correspondence data and restores the original single kanji character, thereby correctly matching the kanji strings “Chu”, “Oh”, “ A “ku” is generated.

  In this way, when automatically associating a single kanji character with a reading from the kanji character string and its reading data, a character string in which the order is clearly specified for the single kanji character is generated, and the character string and the Automatic alignment between readings. As a result, it is possible to generate a kanji string with mono-ruby with high accuracy without using manually created rules.

  The reading estimation device duplicates one single kanji character into two characters and generates two duplicate single kanji characters, but duplicates one kanji character into three or more characters and reproduces three or more duplicate single kanji characters. Kanji characters may be generated. However, in general, a single kanji is often read as two characters. Therefore, by duplicating one character of a single Chinese character into two characters and generating two duplicate single Chinese characters, the position of a reading in the single Chinese character can be expressed more correctly.

[Reading estimation device / Example 1]
Next, a reading estimation apparatus according to the first mode for embodying the present invention (Example 1) will be described. FIG. 1 is a block diagram illustrating a configuration example of the reading estimation apparatus according to the first embodiment. The reading estimation apparatus 1-1 includes a Japanese dictionary storage unit 10, a kanji replication unit 11, a proper noun dictionary storage unit 12, a kanji / reading association unit 13, and a kanji restoration unit 14.

  The Japanese dictionary storage unit 10 inputs headwords in the Japanese dictionary and their reading data and stores them. In addition, the Japanese dictionary storage unit 10 receives the duplicate single kanji characters constituting the headword from the kanji duplicating unit 11, and the initial correspondence consisting of the duplicate single kanji characters constituting the headword and the reading for each character constituting the reading data. Generate and store attached data. Further, the Japanese dictionary storage unit 10 receives and stores duplicate single kanji association data from the kanji and reading association unit 13.

  The kanji duplication unit 11 inputs single kanji for each character constituting the entry word from the Japanese dictionary storage unit 10, orders and duplicates the single kanji, and generates a duplicate single kanji. Then, the Chinese character replicating unit 11 outputs the replicated single Chinese character to the Japanese dictionary storage unit 10. The same applies to the proper noun dictionary storage unit 12.

  The proper noun dictionary storage unit 12 inputs headwords and their reading data of the proper noun dictionary and stores them. In addition, the proper noun dictionary storage unit 12 inputs the duplicate single kanji characters constituting the entry word from the kanji duplication unit 11, and initially corresponds to the duplicate single kanji character constituting the entry word and the reading for each character constituting the reading data. Generate and store attached data. Furthermore, the proper noun dictionary storage unit 12 inputs and stores the duplicate single kanji association data from the kanji and reading association unit 13.

  The kanji / reading association unit 13 reads the initial association data from the Japanese dictionary storage unit 10 and the proper noun dictionary storage unit 12. Then, the kanji / reading association unit 13 associates the duplicate single kanji and the reading included in the initial association data, and generates duplicate single kanji correspondence data. Then, the kanji / reading association unit 13 outputs the duplicate single kanji association data to the Japanese dictionary storage unit 10 or the proper noun dictionary storage unit 12 that has read out the corresponding initial association data.

  The kanji restoration unit 14 reads the duplicate single kanji character association data from the proper noun dictionary storage unit 12, restores the duplicate single kanji character to the original single kanji character, generates a proper noun mono-ruby entry word, and outputs it.

  FIG. 2 is a flowchart illustrating a processing example of the reading estimation apparatus 1-1 according to the first embodiment. The reading estimation apparatus 1-1 receives the headword of the Japanese dictionary and its reading data, performs initial setting of the Japanese dictionary, and generates initial association data (step S201). The reading estimation apparatus 1-1 performs initial setting for all input headwords and their reading data, and generates initial association data for each headword and its reading data.

  For example, the Japanese dictionary storage unit 10 inputs a Japanese dictionary entry word “Ken-Oo” and its reading data “Ken-o” and stores them in the memory. The Japanese dictionary storage unit 10 divides the headword “Ken-Oo” into single-kanji characters “Ken” and “Oo” in units of one character, and also reads the reading data “Ken-o” for each character. Divide into “O” and “U”. Then, the Japanese dictionary storage unit 10 outputs the single kanji characters “Zen” and “Zo” constituting the headword “Ken-Oo” to the Kanji duplication unit 11.

  The kanji duplicating unit 11 inputs the single kanji characters “Ken” and “middle” constituting the headword “Kenzoku” from the Japanese dictionary storage unit 10, orders the single kanji characters “Ken” and “middle”, and changes one character into two. Duplicate the characters and add the subscripts “B” and “E” to generate the duplicate single kanji characters “Category B / Category E” “Central B / Central E”. Then, the kanji duplicating unit 11 outputs the duplicate single kanji characters “Category B / Category E” and “Chuo B / Chuo E” that constitute the headword “Chuo Central” to the Japanese dictionary storage unit 10.

  The Japanese dictionary storage unit 10 inputs the duplicate single kanji characters “Zone B / Zone E” and “Zo B / Zo E” constituting the headword “Keno” from the Kanji duplicating unit 11. Then, the Japanese dictionary storage unit 10 reads the single character Kanji “Category B / Category E”, “Chuo B / Chuo E” that constitutes the headword “Keno” and the reading for each character that constitutes the reading data “Kenou”. Initial correspondence data consisting of “Ken / O / U” is generated and stored in the memory. The initial association data in this case is “Band B / Band E / Center B / Center E” and “Ken / N / O / U”.

  In step S201, the reading estimation apparatus 1-1 inputs various headwords in the Japanese dictionary and their reading data, generates initial association data for each input headword and its reading data, Store as part.

  The reading estimation apparatus 1-1 shifts from step S201 to input headwords and their reading data of the proper noun dictionary, performs initial setting of the proper noun dictionary, and generates initial association data (step S202). The reading estimation apparatus 1-1 performs initial setting for all input headwords and their reading data, and generates initial association data for each headword and its reading data.

  For example, the proper noun dictionary storage unit 12 inputs the headword “center” of the proper noun dictionary and its reading data “chuo” and stores them in the memory. Similar to the Japanese dictionary storage unit 10, the proper noun dictionary storage unit 12 divides the headword “middle” into single Chinese characters “middle” and “middle” and also reads the reading data “chuo” for each character. ”,“ U ”,“ O ”, and“ U ”. The kanji duplicating unit 11 duplicates the single kanji characters “middle” and “middle” in order to generate duplicate single kanji characters “middle B / middle E” and “middle B / middle E”.

  Here, for each character reading, for example, “chu” of “Chuuou”, “Kyu” of “Kyushu”, “ku”, “shu”, etc. ”,“ Sacchi ”and“ Sac ”are included. In addition, a roaring sound and a prompt sound such as “jack” of “jack” are included. That is, the readings “chu” “kyu” “shu” “ho” “satsu” “jack” are treated as a single character reading.

  As with the Japanese dictionary storage unit 10, the proper noun dictionary storage unit 12 stores the duplicate single kanji characters “middle B / middle E”, “middle B / middle E” and the reading data “chuo” that constitute the headword “middle”. Initial correspondence data composed of reading “chu / u / o / u” for each character to be formed is generated. Then, the proper noun dictionary storage unit 12 stores the initial association data in the memory. In this case, the initial association data is “middle B / middle E / middle B / middle E” and “chu / u / o / u”.

  In step S202, the reading estimation apparatus 1-1 inputs various headwords and their reading data of the proper noun dictionary, generates initial association data for each input headword and its reading data, and stores the proper noun dictionary. Store as part.

  FIG. 3 is a diagram for explaining the proper noun dictionary stored in the memory by the proper noun dictionary storage unit 12. The proper noun dictionary is composed of headwords, readings, and association data. In the above example, the proper noun dictionary includes the headword “middle”, the reading “chuo”, and the association data (initial association data in step S202) “middle B / middle E / middle B / middle E” “chu”. / U / o / u "is stored.

  The Japanese dictionary stored in the memory by the Japanese dictionary storage unit 10 also has the same data structure as the proper noun dictionary shown in FIG. In the Japanese dictionary, the headword “Ken-Oo”, reading “Ken-o”, and association data (initial association data in Step S201) “Ben-B / E-E / O-B / O-E” “Ken / N / "O / U" is stored.

  Returning to FIG. 2, the reading estimation apparatus 1-1 shifts from step S <b> 202 to use the Japanese dictionary stored in the Japanese dictionary storage unit 10 and the proper noun dictionary stored in the proper noun dictionary storage unit 12. Thus, the duplicate single kanji characters constituting the headword for the initial association data are associated with the readings. And the reading estimation apparatus 1-1 produces | generates duplicate single Chinese character matching data (step S203).

  Specifically, the reading estimation apparatus 1-1 uses all the initial association data included in the Japanese dictionary and the proper noun dictionary for each combination of duplicate single kanji characters and readings for all the initial association data. The corresponding probability value is calculated. Then, the reading estimation apparatus 1-1 associates the duplicate single kanji characters constituting the headword with the reading based on the correspondence probability value, and generates duplicate single kanji character association data for each headword and the reading data. To do.

  In the above example, the Japanese dictionary storage unit 10 reads out various initial association data such as initial association data “zone B / zone E / center B / center E” and “ke / n / o / u” from the memory. And output to the kanji and reading association unit 13. In addition, the proper noun dictionary storage unit 12 reads the initial association data from the memory and outputs it to the kanji and reading association unit 13.

  The kanji / reading association unit 13 inputs the initial association data “zone B / zone E / center B / center E”, “ke / n / o / u”, etc., from the Japanese dictionary storage unit 10. . Further, the kanji / reading association unit 13 obtains initial association data “middle B / middle E / middle B / middle E”, “chu / u / o / u”, etc., from the proper noun dictionary storage unit 12. input.

  The kanji / reading association unit 13 generates a combination that associates the duplicate single kanji included in the initial association data with the reading for each character. Then, the kanji / reading association unit 13 applies an EM algorithm as machine learning, for example, and calculates a correspondence probability value for each combination. For example, the combination of the initial association data “Category B / Category E / Central B / Central E” “Ken / N / O / U” is “Category B: Ke”, “Category B: N”, and “Category E”. : “E” and “E”: “Center B: Ke” “Center B: N” ... “Center E: Ke” “Center E: N”. In addition, the combinations of the initial association data “middle B / middle E / middle B / middle E” and “chu / u / o / u” are “middle B: chu”, “middle B: u”, and “middle E”. : Chu, “Medium E: U”, “Central B: Chu”, “Central B: U”, “Central E: Chu”, “Central E: U”, and so on.

  The kanji / reading association unit 13 outputs correspondence probability values for each combination to the Japanese dictionary storage unit 10 and the proper noun dictionary storage unit 12. The Japanese dictionary storage unit 10 and the proper noun dictionary storage unit 12 store corresponding probability values for each combination as machine learning data.

  Based on the correspondence probability value for each combination, the kanji / reading association unit 13 associates the duplicate single kanji and the reading for the initial association data, and generates association data. When the initial correspondence data “Bench B / Band E / Central B / Central E” and “Ken / N / O / U” in the above example are applied, the correspondence data between the duplicate single kanji and the reading The (replicated single kanji character association data) is “zone B (K) zone E (N) center B (O) center E (U)”.

  The kanji / reading association unit 13 outputs the duplicate single kanji association data “zone B (K) zone E (n) center B (e) center E (c)” to the Japanese dictionary storage unit 10. The Japanese dictionary storage unit 10 inputs the duplicate single kanji association data “Band B (K) B E (N) B B (O) B E (U)” from the Kanji and reading association unit 13. Store in memory.

  Similarly, the kanji / reading association unit 13 receives the initial association data “middle B / middle E / middle B / middle E” “chu / u / o / u” of the proper noun dictionary input from the proper noun dictionary storage unit 12. ”Is generated as duplicated single kanji character association data“ middle B (chu) middle E (c) middle B (g)) middle E (c) ”.

  The kanji / reading association unit 13 outputs the duplicate single kanji association data “middle B (chu), middle E (c), middle (B), middle (E), c)” to the proper noun dictionary storage unit 12. The proper noun dictionary storage unit 12 inputs the replicated single kanji association data “middle B (chu) middle E (c) B (o) middle E (c)” from the kanji and reading correspondence unit 13, Store in memory.

  Referring to FIG. 3, in the proper noun dictionary, as the correspondence data, the initial correspondence data “Middle B / Middle E / Middle B / Middle E” “Ju / U / O / U” and duplicate single kanji characters are supported. The attached data “middle B (chu) middle E (c) center B (e) center E (c)” is stored.

  Returning to FIG. 2, the reading estimation apparatus 1-1 shifts from step S203 to restore the headword of the proper noun dictionary (step S204), and generates and outputs a headword with mono-ruby of the proper noun (step S204). S205). The reading estimation apparatus 1-1 restores a headword for all the headwords and their reading data that have been input, and generates and outputs a headword with a proper noun for each headword and its reading data.

  In the above example, the proper noun dictionary storage unit 12 reads the duplicate single kanji association data “middle B (chu) middle E (c) B (o) middle E (c)” from the memory and sends it to the kanji restoration unit 14. Output.

  The kanji restoring unit 14 inputs the duplicate single kanji character association data “middle B (chu) middle E (c) central B (g) central E (c)” from the proper noun dictionary storage unit 12. Then, the kanji restoration unit 14 removes the symbols B and E from the duplicate single kanji association data “middle B (chu) middle E (c) middle B (g) middle E (c)”, as shown in FIG. , The association data “Chu, Chu, C, and C” is generated.

  The kanji restoring unit 14 restores the single kanji for each character constituting the original headword from the association data “middle (chu) middle (c) center (g)” (c), as shown in FIG. In addition, the headwords “Chu” and “Ou” with mono-ruby correctly associated are generated.

  The reading estimation device 1-1 calculates a corresponding probability value every time a new kanji string and reading data are input, and the kanji string, reading data, and initial association data included in the Japanese dictionary and proper noun dictionary. The duplicate single kanji character association data and the correspondence probability value may be updated.

  As described above, according to the reading estimation apparatus 1-1 of the first embodiment, the kanji duplicating unit 11 generates a duplicate single kanji by ordering and duplicating the single kanji for each character constituting the headword. The Japanese dictionary storage unit 10 generates and stores initial association data composed of duplicate single kanji characters constituting the entry word and readings for each character constituting the read data for the entry word and its reading data in the Japanese dictionary. To do. Similarly, the proper noun dictionary storage unit 12 generates and stores initial association data for headwords and their reading data in the proper noun dictionary.

The kanji / reading association unit 13 uses the initial correspondence data included in the Japanese dictionary and the proper noun dictionary of the Japanese dictionary storage unit 10 and the proper noun dictionary storage unit 12 to combine the combined single kanji and readings. A correspondence probability value is calculated for each, and based on the correspondence probability value, a single kanji character that constitutes a headword included in the initial association data and a reading are associated with each other. Generate kanji correspondence data. The Japanese dictionary storage unit 10 and the proper noun dictionary storage unit 12 store duplicate single kanji association data.

  The kanji restoring unit 14 reads the duplicate single kanji character association data from the proper noun dictionary storage unit 12, returns the duplicate single kanji character to the original single kanji character, and generates and outputs a monolingual headword with a proper noun.

  Thereby, it is not necessary to use a rule created manually when generating a headword with mono-ruby (a kanji string). Further, since the single kanji and the reading are associated with each other using the duplicate single kanji, the reading position of one character (for example, “u”) is read for the reading (“chu”) of the single kanji (for example, “middle”). ", The corresponding probability value considering the last position) can be obtained. That is, for a plurality of single Chinese characters constituting the Chinese character string, the single Chinese characters and the readings can be associated with each other so that the reading of one character is correctly positioned in the single Chinese character. For example, for the Chinese character string “center”, “Chu” of “Chu” can be positioned at the end of the reading of the single Chinese character “Middle” and not at the beginning of the single Chinese character “Center”. . Therefore, it becomes possible to generate a kanji string with mono-ruby with high accuracy.

[Reading estimation device / Example 2]
Next, a reading estimation apparatus according to the second mode for embodying the present invention (Example 2) will be described. The reading estimation device according to the second embodiment is a device that functionally represents the reading estimation device 1-1 according to the first embodiment, and the processing is substantially the same.

  FIG. 4 is a block diagram illustrating a configuration example of the reading estimation apparatus according to the second embodiment, and FIG. 5 is a flowchart illustrating a processing example of the reading estimation apparatus according to the second embodiment. This reading estimation apparatus 1-2 includes an initialization unit 20, a correspondence data generation unit 21, and a kanji string generation unit 22 with mono-ruby. The reading estimation device 1-2 inputs a kanji string and its reading data, generates and outputs a kanji string with mono-ruby.

  The initialization unit 20 includes a Chinese character string and reading division unit 23, a Chinese character duplication unit 24, and an initialization unit 25. The kanji string and reading division means 23 inputs the kanji string and its reading data (step S501), divides the kanji string into single kanji characters, and divides the reading data into readings of one character unit (step S501). S502). For example, the Chinese character string “center” is divided into single Chinese characters “middle” and “center”, and the reading data “chuo” is divided into readings “chu” “c” “o” “c”.

  The kanji duplicating means 24 orders the single kanji characters divided by the kanji string and the reading dividing means 23 to duplicate them into two characters, and adds the subscript symbols “B” and “E” to produce the duplicate single kanji characters. Generate (step S503). For example, duplicate single kanji characters “middle B” and “middle E” are generated from the single kanji character “middle”, and duplicate single kanji characters “middle B” and “middle E” are generated from the single kanji character “middle”.

  The initialization unit 25 generates initial association data composed of a single kanji character constituting the kanji string and a reading for each character constituting the reading data (step S504). For example, as shown in FIG. 3, initial association data “middle B / middle E / middle B / middle E” and “chu / u / o / u” are generated.

  The association data generation unit 21 includes a combination generation unit 26, a probability calculation unit 27, a dictionary 28, and a duplicate single kanji association data generation unit 29. The association data generation unit 21 inputs a kanji string, reading data, and initial association data from the initialization unit 20 and stores them in the dictionary 28.

  The combination generation unit 26 generates a combination in which the duplicate single Chinese character included in the initial association data is associated with the reading for each character (step S505). For example, for the initial association data “middle B / middle E / middle B / middle E” and “chu / u / o / u”, the combinations “middle B: chu”, “middle B: u”, and “middle E: “Chu”, “Medium E: U”, “Central B: Chu”, “Central B: U”, “Central E: Chu”, “Center E: U”,.

  The probability calculating means 27 uses, for example, an EM algorithm as machine learning using all the initial association data stored in the dictionary 28, and calculates a corresponding probability value for each combination of all duplicate single kanji characters and readings. (Step S506). Then, the probability calculation means 27 stores the corresponding probability value for each combination of the copied single kanji and the reading in the dictionary 28 as machine learning data.

For example, initial correspondence data “middle B / middle E / middle B / middle E” “chu / u / o / u”, kanji string “junior high school” and reading data “ Initial correspondence data “Chu B / Chu E / Study B / Study E” “Chu / U / G / K”, Chinese character string “Oki” and initial correspondence data “Oki” “Chu B” EM algorithm is applied using “/ center E / ki B / ki E” and “o / u / ki”, and a corresponding probability value for each combination is calculated as follows.
P (middle B | oh) = 0.999849, P (middle E | w) = 0.555835, P (middle E | ki) = 0.029560
P (Study B | Ga) = 0.406560, P (Study B | Ku) = 0.406560, P (Study E | Ga) = 0.406560
P (Study E | Ku) = 0.406560, P (Ki B | Ki) = 0.485220, P (Ki E | Ki) = 0.485220
P (Medium B | Ju) = 1.00000, P (Medium E | U) = 0.440979, P (Medium E | Ga) = 0.186880
P (Medium E) = 0.186880
Incidentally, the corresponding probability value of other combinations is 0.

  The duplicate single kanji character association data generation means 29 inputs the correspondence probability value for each combination from the probability calculation means 27 or reads the correspondence probability value for each combination from the dictionary 28. Then, the duplicate single kanji character association data generation means 29 is applied from the one having the highest correspondence probability value, and the plural single kanji character correspondence data which is the correspondence data between the duplicate single kanji character and the reading having the highest correspondence probability value for each reading. Is generated (step S507). Then, the duplicate single Chinese character association data generation means 29 stores the plural single Chinese character association data in the dictionary 28.

For example, when the kanji character string “center” and the reading data “chuo” are applied from the one with the highest corresponding probability value for each combination, it is as follows. As shown in FIG. (Chu) Middle E (U) Center B (O) Center E (C) "is generated.
P (Medium B | Ju) = 1.00000
P (Center B | O) = 0.998849
P (Center E | U) = 0.555835
P (Medium E | U) = 0.440979

  The dictionary 28 stores kanji strings, reading data, initial association data, and duplicate single kanji correspondence data, and also stores corresponding probability values for all combinations of duplicate single kanji characters and readings.

  The Monorbi-added kanji string generation unit 22 reads the duplicate single kanji character association data from the dictionary 28 of the association data generation unit 21, removes the symbols B and E from the duplicate single kanji character association data, and restores the original single kanji character ( Step S508). Then, the kanji string generating unit 22 with mono-ruby generates and outputs a kanji string with mono-ruby (step S509). For example, the symbols B and E are removed from the duplicate single kanji character association data “middle B (chu) middle E (c) middle B (g) middle E (c)”, and the kanji string “middle (chu)” with mono-ruby “ "Ou" is generated.

  Note that the reading estimation apparatus 1-2 may calculate the corresponding probability value and update the dictionary 28 every time a new Chinese character string and reading data are input.

  As described above, according to the reading estimation apparatus 1-2 of the second embodiment, the initialization unit 20 inputs a kanji character string and its reading data, converts the kanji character string into a single kanji character, and reads the reading data in units of one character. Are divided, and a single kanji character is ordered to generate a duplicate single kanji character, and initial correspondence data including a duplicate single kanji character constituting the kanji string and a reading for each character constituting the reading data is generated.

  The association data generation unit 21 calculates the correspondence probability value by the EM algorithm using all the initial association data for each combination of duplicate single kanji and reading included in the initial association data. Then, the association data generation unit 21 generates a plurality of single kanji character association data, which is association data between the copied single kanji character and the reading, based on the correspondence probability value.

  The Monorbi-attached kanji string generation unit 22 restores the duplicate single kanji included in the duplicate single-kanji association data to the original single kanji, and generates a monorubi-added kanji string.

  Thereby, like the reading estimation apparatus 1-1 of Example 1, it becomes possible to generate | occur | produce a Chinese character string with a mono rubi with high precision, without using the rule created manually.

[Japanese sentence text generator with mono-ruby]
Next, as an example using the functions of the reading estimation apparatus 1-1 shown in FIG. 1 and the reading estimation apparatus 1-2 shown in FIG. 4, a Japanese sentence text generation apparatus with mono-ruby will be described. FIG. 6 is a block diagram illustrating a configuration example of a Japanese sentence text generating apparatus with mono-ruby. This Japanese sentence text generation apparatus 2 with mono-ruby includes a kanji string extraction unit 30, a reading estimation unit 31, and a text generation unit 32.

  The Japanese sentence text generation device 2 with mono-ruby extracts a kanji string from the Japanese sentence text, generates a kanji string with mono-ruby by estimating a single kanji reading constituting the kanji string, Is generated. For example, the Japanese sentence text “We are in Ishikawa Prefecture”, while the Japanese sentence text with mono-ruby “I (Tachi) is the Ishikawa Prefecture (Ken) Will be generated. "

  The kanji string extraction unit 30 inputs the Japanese sentence text that is the object of mono-ruby assignment, extracts the kanji string from the character string that constitutes the Japanese sentence text, and uses the Japanese dictionary (not shown) to extract the extracted kanji string Generate reading data. Then, the Chinese character string extraction unit 30 outputs the Chinese character string and its reading data to the reading estimation unit 31. In the above example, the kanji strings “we” and “Ishikawa” are extracted from the Japanese sentence text “We are in Ishikawa” and their reading data “Watashitachi” and “Ishikawaken” are generated.

  The reading estimation unit 31 includes a memory (not shown in FIG. 6) in which the proper noun dictionary shown in FIG. 1 or the dictionary 28 shown in FIG. 4 is stored. The reading estimation unit 31 inputs the kanji string and its reading data from the kanji string extraction unit 30, and reads out the duplicate single kanji character association data corresponding to the kanji string and the reading data from the memory. Then, the reading estimation unit 31 performs the same processing as the kanji reconstruction unit 14 shown in FIG. 1 or the kanji string generation unit 22 with mono ruby shown in FIG. Restore to kanji and generate kanji strings with mono-ruby.

  The reading estimation unit 31 outputs the kanji string with mono-ruby to the text generation unit 32. In the above example, the kanji string “I (Tatashi)” with mono-ruby is generated from the duplicate single Kanji character association data corresponding to the Kanji string “us” and its reading data “Watashitachi”. In addition, a kanji string “Ishikawa (Kawa) Prefecture (Ken)” with mono-ruby is generated from the kanji string “Ishikawa Prefecture” and the duplicate single Kanji character association data corresponding to the reading data “Ishikawa Ken”.

  The proper noun dictionary stored in the memory (not shown) of the reading estimation unit 31 is generated by the configuration shown in FIG. 1 and the processing shown in FIG. The dictionary 28 stored in the memory (not shown) of the reading estimation unit 31 is generated by the configuration shown in FIG. 4 and the processing shown in FIG. The proper noun dictionary and the dictionary 28 may be downloaded from the outside via a network. Further, the proper noun dictionary may be generated by the reading estimation unit 31 including the configuration shown in FIG. 1, and the dictionary 28 is generated by the reading estimation unit 31 including the configuration shown in FIG. It may be.

  The text generation unit 32 inputs a kanji string with mono-ruby from the reading estimation unit 31 and also inputs Japanese sentence text to be mono-rubbed. Then, the text generation unit 32 replaces the kanji character string in the character string constituting the Japanese sentence text with the kanji character string with mono-ruby, and generates the Japanese sentence text with mono-ruby. The text generator 32 outputs a Japanese sentence text with mono-ruby.

  In the above example, the kanji strings "we" and "Ishikawa" in the Japanese text "We are in Ishikawa" are the kanji strings with mono-ruby "I (Tachi)" (Tachi) " Ishi River (Kawa) Prefecture (Ken) ". Then, a Japanese sentence text with mono-ruby “I am in Tachi, Ishikawa, Ken” is generated.

  As described above, according to the Japanese sentence text generating device 2 with mono-ruby, a Chinese character string is extracted from the Japanese sentence text, and the same processing as the reading estimation devices 1-1 and 1-2 of the first and second embodiments is performed. Then, by estimating the readings of duplicate single kanji characters constituting the kanji string, a kanji string with monoruby is generated. And the Japanese sentence text production | generation apparatus 2 with a mono rubi produces | generates the Japanese sentence text with a mono rubi containing the kanji string with a mono rubi.

  As a result, it is possible to generate kanji strings with mono-ruby with high accuracy without using manual rules, and as a result, it is possible to generate Japanese text text with mono-ruby with high-precision kanji strings with mono-ruby. It becomes.

[Sign Language CG Translation Device]
Next, a sign language CG translation device will be described as another example using the functions of the reading estimation device 1-1 shown in FIG. 1 and the reading estimation device 1-2 shown in FIG. FIG. 7 is a block diagram illustrating a configuration example of a sign language CG translation apparatus. The sign language CG translation device 3 includes a sign language CG translation unit 33, a kanji string extraction unit 34, a reading estimation unit 35, a mouth type CG data conversion unit (mouth type CG data generation unit) 36, and a CG synthesis unit 37.

  The sign language CG translation device 3 generates sign language CG data of the Japanese sentence text, extracts a kanji string from the Japanese sentence text, estimates a reading of a single kanji character constituting the kanji string, and generates a kanji string with mono-ruby. Generate. Then, the sign language CG translation device 3 generates mouth type CG data representing the movement of the mouth for each single kanji character of the kanji string with mono-rubi, and synchronizes the mouth language for each single kanji character synchronized with the sign language CG data of the Japanese sentence text. A type CG is synthesized.

  For example, the sign language CG data corresponding to the Japanese sentence text “We are in Ishikawa Prefecture” is generated, and the Chinese character string “I (Tachi)” with mono-ruby “Ishi River (Kawa) ) Prefecture (ken) "is generated, and mouth type CG data for each single Chinese character is generated. Then, the sign-type CG data is combined with mouth-type CG for each single kanji, and the single kanji “I (Watashi)”, “Tachi”, “Ishi (Ishi)”, “Kawa (Kawa)”, “Ken (Ken)” CG data in which a sign language word and its mouth shape are synchronized is generated.

  The sign language CG translating unit 33 inputs Japanese sentence text and translates the Japanese sentence text to generate sign language CG data. Specifically, the sign language CG translation unit 33 converts Japanese sentence text into a plurality of sign language words using a technique such as statistical translation, and a word motion corresponding to each of the plurality of sign language words is not illustrated. Read from motion DB (beta base). Then, the sign language CG translating unit 33 sequentially connects a plurality of word motions and generates sign language CG data representing a series of sign language sentences. The sign language CG translation unit 33 outputs the sign language CG data to the CG synthesis unit 37. For example, sign language CG data corresponding to the Japanese sentence text “We are in Ishikawa Prefecture” is generated.

  The Chinese character string extraction unit 34 and the reading estimation unit 35 are the same as the Chinese character string extraction unit 30 and the reading estimation unit 31 shown in FIG. In the above example, the Chinese character string “I (Tachi)” and “Ishi River (Ken)” with mono-ruby are generated.

  The mouth type CG data conversion unit 36 inputs a kanji string with monorubi from the reading estimation unit 35, and reads out mouth type CG data corresponding to a single kanji character constituting the kanji string with monoruby from a mouth DB (not shown). Convert single kanji into mouth type CG data. Then, the mouth type CG data conversion unit 36 outputs the mouth type CG data for each single Chinese character to the CG synthesis unit 37. In the above example, mouth type CG data for each single Chinese character “I”, “Tachi”, “Ishi”, “Kawa”, and “Ken” is output.

  The CG synthesizing unit 37 inputs sign language CG data from the sign language CG translation unit 33 and also inputs mouth type CG data for each single kanji from the mouth type CG data conversion unit 36, and further subtitle data of Japanese text Input audio data. Then, the CG synthesis unit 37 synthesizes the sign language CG data with the mouth type CG data, the caption data, and the audio data in synchronism, and generates and outputs the CG data. In the above example, the sign language words corresponding to the single kanji characters "I", "Tachi", "Ishi", "Kawa", "Ken" and their mouth type are synchronized CG Data is generated.

  Thereby, based on CG data, an image of a CG character in which a sign language word corresponding to a single Chinese character and its mouth shape are synchronized is generated, the image and caption data are displayed on the screen, and audio data is reproduced. .

  As described above, according to the sign language CG translating device 3, the sign language CG data of the Japanese sentence text is generated, the kanji string is extracted from the Japanese sentence text, and the reading estimation apparatuses 1-1 and 1-1 of the first and second embodiments. In the same process as in 1-2, the reading of duplicate single kanji characters constituting the kanji string is estimated to generate a kanji string with mono-ruby. Then, the sign language CG translation device 3 generates mouth type CG data for each single kanji character corresponding to the kanji string with monorubi, and synchronizes the mouth type CG for each single kanji character with the sign language CG data of the Japanese text. , CG data is generated.

  As a result, a kanji string with monorubi can be generated with high accuracy without using a manually created rule, and as a result, CG data in which the movement of the sign language and the movement of the mouth corresponding to the single kanji is reproduced. It becomes possible.

  Although the present invention has been described with reference to the first and second embodiments and application examples thereof, the present invention is not limited to the first and second embodiments and the like, and various modifications can be made without departing from the technical idea thereof. It is. For example, in the first and second embodiments, the reading estimation apparatuses 1-1, 1-2, and the like calculate a corresponding probability value for each combination in which a single kanji character constituting a kanji string is matched with a reading for each character. A plurality of single-kanji character correspondence data in which a duplicate single-kanji character and a reading are associated with each other is generated by applying a high correspondence probability value. On the other hand, the reading estimation devices 1-1, 1-2, etc. calculate the total value of the corresponding probability values of the combinations corresponding to the reading data of the Chinese character string (reading for each character constituting the character string), and the total value A plurality of single-kanji character association data in which the duplicate single-kanji character and the reading are associated with each other may be generated.

  In addition, as a hardware configuration of the reading estimation apparatuses 1-1 and 1-2, the Japanese sentence text generation apparatus 2 with mono-ruby, and the sign language CG translation apparatus 3 according to the embodiment of the present invention, a normal computer can be used. . The reading estimation devices 1-1 and 1-2, the Japanese sentence text generation device 2 with mono-ruby, and the sign language CG translation device 3 are a volatile storage medium such as a CPU and a RAM, a non-volatile storage medium such as a ROM, and an interface. It is comprised by the computer provided with etc.

  The functions of the Japanese dictionary storage unit 10, the kanji replication unit 11, the proper noun dictionary storage unit 12, the kanji and reading association unit 13, and the kanji restoration unit 14 provided in the reading estimation apparatus 1-1 describe these functions. Each program is realized by causing the CPU to execute the program. In addition, each function of the initialization unit 20, the association data generation unit 21, and the kanji string generation unit 22 with mono-ruby included in the reading estimation apparatus 1-2 is also executed by causing the CPU to execute a program describing these functions. Realized. The functions of the kanji string extraction unit 30, the reading estimation unit 31, and the text generation unit 32 included in the Japanese sentence text generation apparatus 2 with mono-ruby are also realized by causing the CPU to execute a program describing these functions. Is done. The functions of the sign language CG translation unit 33, the kanji string extraction unit 34, the reading estimation unit 35, the mouth type CG data conversion unit 36, and the CG synthesis unit 37 included in the sign language CG translation apparatus 3 also describe these functions. Each is realized by causing the CPU to execute the program.

  These programs can be stored and distributed on a storage medium such as a magnetic disk (floppy (registered trademark) disk, hard disk, etc.), optical disk (CD-ROM, DVD, etc.), semiconductor memory, etc., and sent and received via a network. You can also

DESCRIPTION OF SYMBOLS 1 Reading estimation apparatus 2 Japanese sentence text generation apparatus 3 with a Monorbi Sign language CG translation apparatus 10 Japanese dictionary storage part 11 Kanji replication part 12 Proper noun dictionary storage part 13 Kanji and reading matching part 14 Kanji restoration part 20 Initialization part 21 Association data generation unit 22 Monorbi-added kanji string generation unit 23 Kanji string and reading division unit 24 Kanji duplication unit 25 Initialization unit 26 Combination generation unit 27 Probability calculation unit 28 Dictionary 29 Duplicate single kanji association data generation unit 30 and 34 Kanji Column extraction unit 31, 35 Reading estimation unit 32 Text generation unit 33 Sign language CG translation unit 36 Mouth type CG data conversion unit 37 CG synthesis unit

Claims (6)

In the reading estimation device that estimates the reading for each single kanji from the headword and its reading data, and generates a kanji string with mono-ruby,
A dictionary storage unit for inputting a headword and reading data of the headword, and storing a dictionary including the headword, the reading data, and association data;
A kanji duplication unit that duplicates each single kanji character constituting the kanji string that is the headword and generates a plurality of duplicate single kanji characters for the single kanji character;
A kanji and reading association unit that generates the association data in which each of a plurality of duplicate single kanji characters constituting the headword is associated with a reading for each character constituting the reading data;
A kanji restoration unit that reads the association data from the dictionary storage unit, restores the duplicate single kanji characters included in the association data to the original single kanji characters, and generates a kanji string with monorubi for the entry word; Prepared,
The kanji and reading association unit is
Using the dictionary stored in the dictionary storage unit, for each combination in which each of a plurality of duplicate single kanji characters constituting the headword corresponds to reading for each character constituting the reading data, the duplicate Calculating a correspondence probability value indicating a degree of correspondence between a single kanji character and the reading; generating the correspondence data based on the correspondence probability value;
The dictionary storage unit
The reading estimation apparatus characterized by storing the association data generated by the kanji and reading association unit. In the reading estimation device that estimates the reading for each single kanji from the kanji string and its reading data, and generates a kanji string with mono-ruby,
The kanji string and the reading data are input, the single kanji for each character constituting the kanji string is duplicated, a plurality of duplicate single kanji characters are generated for the single kanji, and the kanji string is configured. An initializing unit for generating initial association data composed of a plurality of duplicate single kanji characters and readings for each character constituting the reading data;
For each combination of the duplicate single kanji and the reading included in the initial association data generated by the initialization unit, a corresponding probability value indicating the degree to which the duplicate single kanji and the reading correspond An association data generating means for generating association data in which each of a plurality of duplicate single kanji characters constituting the kanji string is associated with the reading based on the correspondence probability value;
A kanji string generation unit with mono-ruby that restores the duplicate single-kanji characters included in the association data generated by the association data generation means to the original single kanji and generates a kanji string with mono-ruby of the kanji string;
A reading estimation apparatus characterized by comprising: In the reading estimation apparatus according to claim 1 or 2,
A reading estimation apparatus, wherein each single kanji character constituting the kanji string is duplicated, and two duplicate single kanji characters are generated for each single kanji character. In a Japanese sentence text generator with mono ruby that generates Japanese sentence text with mono rubi from Japanese sentence text,
A kanji string extraction unit that extracts kanji strings from the Japanese sentence text and generates reading data of the kanji strings;
Memory in which the association data generated by the reading estimation device according to any one of claims 1 to 3 is stored corresponding to the Chinese character string;
The association data corresponding to the kanji string extracted by the kanji string extraction unit is read from the memory, the duplicate single kanji included in the association data is restored to the original single kanji, and the kanji string with mono-ruby is obtained. A reading estimator to generate,
A text generation unit that replaces the kanji string included in the Japanese sentence text with the kanji string with mono-ruby generated by the reading estimation unit, and generates the Japanese sentence text with mono-ruby;
A Japanese sentence text generator with mono-ruby, characterized by comprising: In a sign language CG translation device that translates Japanese text to generate sign language CG data, and synthesizes mouth type CG data representing mouth movements with the sign language CG data,
A kanji string extraction unit that extracts kanji strings from the Japanese sentence text and generates reading data of the kanji strings;
A memory in which the association data generated by the reading estimation device according to any one of claims 1 to 4 is stored corresponding to the Chinese character string;
The association data corresponding to the kanji string extracted by the kanji string extraction unit is read from the memory, the duplicate single kanji included in the association data is restored to the original single kanji, and a kanji string with mono-ruby is generated. A reading estimator to
A mouth type CG data generating unit that generates the mouth type CG data corresponding to the single kanji character for each of a plurality of single kanji characters constituting the kanji string with mono ruby generated by the reading estimation unit;
A CG synthesis unit that synthesizes the mouth CG data generated by the mouth type CG data generation unit with the sign language CG data;
A sign language CG translation apparatus comprising:   A program for causing a computer to function as the reading estimation apparatus according to any one of claims 1 to 3.

Images