JP2015069359A - Translation device and translation program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To achieve highly accurate sign language translation.SOLUTION: A translation device which performs sign language translation on input data includes: input character division means for dividing the input data in each predetermined character; translating means for translating the character divided by the input character division means using a translation model in which a combination of learning data of a translation target language and sign language is learned in advance by a predetermined phrase; and translation result output means for outputting a translation result by the translating means.

Description

  The present application relates to a translation apparatus and a translation program, and more particularly, to a translation apparatus and a translation program for realizing highly accurate sign language translation of proper nouns.

  There are techniques for translating from the original language into various target languages corresponding to the intended use. For example, in the case of translation of a proper noun, the proper noun reading is used. For example, when “Fukushima” is translated into English, the reading is “Fukushima” written in Roman letters. Conventionally, as described above, there is a translation technique based on reading of proper nouns (see, for example, Patent Document 1 and Patent Document 2).

  In recent years, translation into sign language has attracted attention as one of the target languages. Sign language is an important means of communication for the hearing impaired. Sign language is the first language, especially for people who have lost their hearing in their innate or early childhood, and are easier to understand than Japanese. For this reason, it is preferable to present sign language information rather than Japanese characters. The sign language can be presented using a video such as CG (Computer Graphics). Therefore, when translating from the original language into sign language, first the sign language word string is converted from the original language, the CG corresponding to each word is extracted from the converted sign language word string, and the extracted CG is connected to sign language video. Is generated.

JP 2005-92682 A JP-T-2005-520251

  By the way, many proper nouns and the like in sign language are translated for each character. For example, in the case of “Matsue”, the character “Matsu” is translated into the sign language word {pine}, and the character “E” is translated into the finger character {e} (enclosed in {} above). The part represents one word of sign language, and the same applies to the following description).

  Therefore, in the translation of proper nouns using sign language, since the reading of proper nouns is rarely used, the conventional translation method cannot be applied as it is.

  In one aspect, an object of the present invention is to provide a translation device and a translation program for realizing highly accurate sign language translation.

  In order to solve the above problems, the present invention employs means for solving the problems having the following characteristics.

  In one aspect, a translation device in a translation device that performs sign language translation of input data, an input character dividing unit that divides the input data into predetermined characters, and a character divided by the input character dividing unit in advance Translation means for performing translation using a translation model in which a combination of learning data of a translation target language and sign language is learned in a predetermined phrase unit, and translation result output means for outputting a translation result by the translation means.

  The translation program in one aspect is a translation program for causing a computer to function as each unit included in the translation apparatus described above.

  In sign language translation, it is possible to realize particularly accurate translation of proper nouns.

It is a figure which shows an example of a function structure of a translation apparatus. It is a figure which shows the example of a function expansion for finger character conversion. It is a flowchart which shows an example of a translation model learning process. It is a flowchart which shows an example of a translation process. It is a figure which shows an example of a replacement model. It is a figure which shows an example of a language model. It is a figure which shows the comparative example with the conventional method.

<About this embodiment>
In the present embodiment, for example, a translation technique between a translation target language (for example, Japanese) and sign language is provided. In the present embodiment, for example, translation from Japanese (original language) to sign language (target language) may be performed, or translation from sign language (original language) to Japanese (target language) may be performed. In this case, the Japanese language that is the translation target language can be either the original language or the target language. In the following, as an example, a method of translating Japanese as an original language and sign language (Japanese sign language) corresponding to the Japanese as a target language will be described.

  Here, generally, four methods of (1) Kanji sign language, (2) finger characters, (3) Kanji sign language + finger characters, and (4) fixed translation are used for proper noun expressions in sign language.

(1) Kanji sign language Kanji sign language means, for example, translation when a proper noun is divided into characters and replaced with sign language corresponding to each character. For example, the kanji sign language corresponding to the Japanese kanji “Fuku” is {happiness}, which is a semantically similar sign language word. Using this, if it is “Fukushima”, it will be divided into “Fuku” and “Island”, and it will be in two words: Kanji sign language corresponding to “Fuku” {Happy} and Kanji sign language representing “Island” {Island}. Expressed. In many cases, these words can be expressed more concisely than “(2) finger characters” to be described later. For example, these words are often used when there is no fixed translation.

(2) Finger characters Finger characters mean, for example, translation when the proper noun reading kana is represented by finger characters. In sign language, all 50 Japanese kana characters are defined as finger characters. Although finger characters are highly expressive, there is a problem that it takes time to express one word, so it is not often used in Japanese translation. However, finger characters are often used when it is difficult to translate kanji sign language, foreign place names, and katakana.

(3) Kanji sign language + finger character Kanji sign language + finger character means, for example, a translation technique in which “(1) Kanji sign language” and “(2) finger character” described above are combined. For example, in the case of “Nagano”, “Long” is represented by Kanji sign language using the {Long} sign language word, and “No” is represented by {N} finger characters. For example, finger characters are often used for kanji with short readings such as “no”.

(4) Fixed translation A fixed translation represents a case where a sign language word corresponding to a certain proper noun is already determined. For example, in the case of “Hiroshima”, in sign language, the shape of the torii gate of Itsukushima Shrine is expressed by finger movement. Fixed translations are often very distinctive expressions, meanings can be conveyed reliably and can be expressed succinctly, so if there is a fixed translation, it is often used preferentially.

  In this way, except for special cases where there is a fixed translation in sign language, proper noun translation is realized by dividing proper nouns into characters and replacing them with corresponding sign language words or finger characters for each character. .

  In the translation of proper nouns from Japanese to Japanese sign language, Kanji characters contained in Japanese proper nouns are replaced with sign language words or finger characters. Therefore, it is preferable to divide the proper noun for each character and determine for each character which sign language word or finger character the character corresponds to.

  Therefore, in this embodiment, a sign language word or a finger character corresponding to each character is learned using a machine translation technique. Note that the machine learning of the correspondence between Japanese characters and sign language words is not very accurate. The reason is that the difference between the number of Japanese characters appearing per sentence and the number of words in sign language is large, and the difference between the number of differences in Japanese characters and the number of differences in sign language words is large.

  Therefore, in this embodiment, appropriate correspondence is easy by dividing a sentence pair used for learning (for example, a sentence spare corresponding to Japanese and sign language) into phrases and shortening the learned sentence pair. To be.

  Hereinafter, embodiments in which a translation apparatus and a translation program are suitably implemented will be described in detail with reference to the drawings.

<Functional configuration example of translation device>
FIG. 1 is a diagram illustrating an example of a functional configuration of the translation apparatus. The translation apparatus 10 in FIG. 1 includes a translation model learning unit 11, a word input unit 12, a fixed translation translation unit 13, an input character dividing unit 14, a translation unit 15, a finger character conversion unit 16, and a translation result output. And means 17.

  The translation model learning unit 11 learns the translation model in advance and stores the translation model used by the translation unit 15. For example, the translation model learning unit 11 learns a translation model by dividing a sentence pair used for learning (for example, a sentence spare corresponding to Japanese and sign language) into phrases. Thereby, the translation model can be learned by appropriately associating the original language (Japanese) with the target language (Japanese sign language).

  The word input unit 12 accepts input of a character example to be translated (a character string (sentence) consisting of one word or a plurality of words) as an example of input data. The input word in the word input unit 12 is output to the fixed translation translating unit 13 and the input character dividing unit 14. The word input unit 12 can input a word character string by various input units such as character input using an operation button on a keyboard or a touch panel, or voice input using a microphone, but is not limited thereto. Absent. For example, the word input means 12 may input news manuscript data, script data such as movies and dramas, and the like.

  If there is a word included in the fixed translation dictionary 21 among the word strings input by the word input means 12 using the preset fixed sign dictionary 21 for sign language translation, Translate into a fixed sign language corresponding to the word. The fixed translation dictionary 21 includes words of proper nouns such as place names and personal names such as “Kobe → {Kobe}”, “Tokyo → {Tokyo}”, “Yokohama → {Yokohama}”, and the like. Is stored in association with Japanese sign language. The result translated by the fixed translation translating means 13 is output to the input character dividing means 14.

  The input character dividing unit 14 divides words other than the characters translated by the fixed translation translating unit 13 among the word character strings input by the word input unit 12 in character units. The input character dividing unit 14 outputs the divided characters to the translating unit 15.

  The translation unit 15 translates the divided character string using the translation model learned by the translation model learning unit 11. The translation unit 15 outputs the translation result to the finger character conversion unit 16. For the translation in the translation means 15, a known statistical translation tool kit such as “moses” can be used, but is not limited thereto.

  Based on the result of the translation means 15, the finger character conversion means 16 uses the remaining characters that have not been translated by the fixed translation translation means 13 and the translation means 15 among the word character strings input by the word input means 12 as finger characters. Convert. The finger character conversion means 16 outputs the finger character conversion result to the translation result output means 17.

  The translation result output means 17 integrates the translation result of the fixed translation translation means 13, the translation result of the translation means 15, and the output result of the finger character conversion means 16 into the word character string input by the word input means 12. Output the corresponding translation result.

  For example, the translation result output means 17 can convert the translation result into a sign language CG with a preset CG character and output it to a screen or the like, but is not limited to this, for example, a character string of the translation result (For example, “Kamaishi → {Kama (finger character)} {stone}”) or the like may be output.

<Functional Configuration Example of Translation Model Learning Unit 11>
Here, a functional configuration example in the translation model learning unit 11 described above will be specifically described. As shown in the example of FIG. 1, the translation model learning unit 11 includes, for example, a learning data storage unit (sentence spare) 31, a sentence division unit 32, a learning data storage unit (phrase pair) 33, and a character unit division unit 34. And a replacement model learning means 35, a language model learning means 36, and a translation model storage means 37. Note that the learning data storage means (sentence spare) 31 and the learning data storage means (phrase pair) 33 may be configured as one learning data storage means.

  The learning data storage means (sentence spare) 31 stores, for example, sentence pairs of the original language and the target language divided in predetermined word units. In this embodiment, the learning data storage means (sentence spare) 31 stores learning data of pairs for each sentence (sentence) of Japanese and Japanese sign language, for example.

  The sentence division unit 32 inputs the sentence pair (sentence spare) stored in the learning data storage unit (sentence spare) 31 and outputs the result of word alignment (for example, word alignment, word association). Use to split into shorter phrase pairs. Moreover, the sentence division means 32 stores in the learning data storage means (phrase pair) 33 for each sentence phrase. For obtaining word alignment, a known tool kit such as “GIZA ++” can be used, but is not limited thereto. For example, “GIZA ++” is a tool that calculates the probability value of a word for use in statistical translation, and can calculate the probability value of the correspondence relationship between words. Therefore, the sentence dividing means 32 can calculate “GIZA ++” and use the result to divide into phrase pairs and store them in the learning data storage means (phrase pair) 33.

  The character unit dividing unit 34 divides Japanese (original language) into character units among the phrase pairs stored in the learning data storage unit (phrase pair) 33. A phrase pair composed of the Japanese divided into character units and a sign language word string corresponding to the Japanese is output to the replacement model learning means 35.

  The replacement model learning unit 35 learns a replacement model for the phrase pair divided by the character unit by the character unit dividing unit 34. The replacement model is a compilation of translation probabilities between multiple Japanese characters and multiple words in sign language. A specific example of the replacement model will be described later.

  The language model learning means 36 learns the contents as a language model based on the contents of the learning data storage means (sentence spare) 31. The language model is generated from the data stored in the learning data storage means (sentence spare) 31. For the generation of the language model in the language model learning means 36, for example, a known statistical language model creation tool kit such as “SRILM” can be used, but it is not limited to this.

  The translation model storage unit 37 stores the content of the replacement model obtained from the replacement model learning unit 35 and the language model obtained from the language model learning unit 36 as a translation model.

  As described above, the translation model learning unit 11 divides a sentence pair used for learning into phrase units in advance, thereby shortening the learned sentence pair and appropriately and easily associating with the word character string to be translated. And enables high precision translation.

<Function expansion example for finger character conversion (word-reading association)>
Here, the finger character conversion means 16 in this embodiment described above, for example, when there is an input character that is not in the translation model at the time of translation, or when “en (finger character)” or the like is obtained as a result of translation For example, a finger character using a reading corresponding to the character is obtained. This finger character conversion is performed using, for example, a preset finger character conversion dictionary or the like, but it is necessary to associate kanji and reading. Therefore, the translation apparatus 10 may extend the function by providing a function for associating the reading with the word character string input to the finger character conversion means 16.

  FIG. 2 is a diagram illustrating an example of function expansion for finger character conversion. In the example of FIG. 2, only the part related to input / output to the finger character conversion means 16 is shown in the configuration of the translation apparatus 10 described above, which is another embodiment of the translation apparatus.

  In the example of FIG. 2, it includes a word input unit 12, a word reading input unit 41, a word-reading association unit 42, and a finger character conversion unit 16.

  The word reading input means 41 inputs a reading corresponding to the word (for example, kanji notation of a proper noun) input by the word input means 12 described above. The word-reading association means 42 associates kanji and readings using, for example, the input of the proper noun kanji notation from the word input means 12 and the proper noun reading from the word reading input means 41.

  Note that the word-reading association unit 42 is not limited to the association of proper nouns, and may associate readings with other words. Moreover, the word reading input means 41 may input a reading manually and may acquire it automatically from a Chinese character notation using a dictionary. For example, when “Sonoda” is input, the word-reading association unit 42 outputs which reading corresponds to which character, such as “Sonoda”.

  The finger character conversion means 16 uses the processing result described above, for example, when “en (finger character)” is obtained as the translation result of “garden” from the translation means 15, for example, Utilizing that the reading is “Sono”, it can be converted to “Sono (finger character)”.

<Example of translation model learning process>
Next, an example of the translation model learning process in the present embodiment will be described using a flowchart. FIG. 3 is a flowchart illustrating an example of the translation model learning process. In the example of FIG. 3, the translation model learning means 11 divides a sentence spare of an original language (Japanese language) and a target language (Japanese sign language) included in learning data stored in advance, and generates a phrase pair. Store (S01).

  Next, the translation model learning unit 11 divides the phrase pair Japanese described above in character units (S02). Next, the translation model learning unit 11 learns a replacement model from the phrase pair using the divided characters (S03).

  Next, the translation model learning unit 11 learns a language model from a sentence spare included in the learning data (S04). Note that the timing of the process of S04 is not limited to this, and may be performed before the process of S01 described above, for example.

  Next, the translation model learning unit 11 stores the replacement model obtained by the process of S03 and the language model obtained by the process of S04 (S05). The translation model learning process described above is performed before the translation process described later.

<Example of translation processing>
Next, an example of translation processing in the present embodiment will be described using a flowchart. FIG. 4 is a flowchart illustrating an example of translation processing. In the example of FIG. 4, the translation apparatus 10 receives an input of a word to be translated (including a character string or the like) by the word input unit 12 or the like (S11).

  Next, the translation apparatus 10 translates the word input in the process of S11 into a fixed sign language using the preset fixed sign dictionary 21 or the like for sign language translation (S12). Next, the translation apparatus 10 divides a part (for example, a new expression word or a new proper noun) that cannot be translated using the fixed translation dictionary 21 in units of characters with respect to the word input in the process of S11 ( S13). Next, the translation apparatus 10 translates it into sign language in units of characters using the translation model stored in the translation model storage means 37 (S14).

  Next, if there is a character that could not be translated even after the processing up to S14, the translation device 10 converts the character into a finger character (S15), and the translation result obtained by the above-described processing (input in the processing of S11). The final translation result for the completed word is output (S16).

  Here, translation apparatus 10 determines whether or not to continue translation of other words or the like (S17). When translation is continued (YES in S17), the process returns to S11. Moreover, the translation apparatus 10 complete | finishes a translation process, when not continuing a translation (in S17 NO). With the processing described above, appropriate translation from Japanese into sign language can be realized.

<Examples of various data>
Next, various data examples used in the present embodiment will be described with reference to the drawings.

<Example of replacement model>
FIG. 5 is a diagram illustrating an example of a replacement model. The replacement model is data learned by the replacement model learning means 35. In the example of FIG. 5, the replacement model items include, for example, “Japanese (original language) notation”, “sign language (target language) notation”, “various probabilities”, and “(within the replacement model) the same sign language expression. There are “the number of lines”, “the number of lines having the same Japanese expression (within the replacement model)”, etc., but it is not limited thereto.

  In the replacement model, using a pair of learning data, the combination of Japanese (original language) and Japanese sign language (target language), and various probabilities, “sign language word → Japanese translation probability (likelihood) "Product of co-occurrence probability for each Japanese word when sign language word → Japanese", "Translation probability (likelihood) of Japanese word → sign language word", "Japanese word → Sign language word Learning of “product of co-occurrence probability for each word in sign language” and “a numerical value given uniformly (e)”.

  In addition, {pt} of “sign language notation” shown in FIG. 5 indicates a pointing action of a sign language speaker such as a CG character, and {pt3} indicates a pointing action to a person other than himself or the other party or a person. . In addition, {N} of “sign language notation” represents a nod of a sign language speaker such as a CG character performing sign language. In this embodiment, nodding is not used in proper nouns and can be ignored for translation.

  In this embodiment, in order to generate a translation model used for translation from a Japanese word to a sign language word, it is sufficient if there is data of “translation probability (likelihood) of Japanese word → sign language word” in the example of FIG. Other probabilities may not be included in the replacement model. These parameters such as various probability values are obtained by machine learning from various data such as “the number of lines having the same sign language expression” and “the number of lines having the same Japanese expression”. Note that “a numerical value (e) given uniformly” is a value used for adjustment of the numerical value at the time of translation, and is not limited to the example of FIG. 5 and may not be included in the replacement model. .

  Here, the translation using the translation model in the present embodiment can be performed, for example, for each character, but is not limited to this, and can also be performed by translating a plurality of characters and phrases. For example, since “Fukushima” is included in the learning data, “Fukushima” also appears in the replacement model shown in FIG. 5, and this information is also referred to during translation. Therefore, in the translation processing according to the present embodiment, when a character string that does not appear in the order in the learning data, such as “Shimafuku”, is translated one character at a time.

<Example of language model>
FIG. 6 is a diagram illustrating an example of a language model. The language model is data learned by the language model learning means 36. The items of the language model shown in the example of FIG. 6 include, for example, “probability of word alignment (log likelihood)”, “arrangement of words”, “backoff probability”, and the like. Absent. Since the items used in the translation processing in the present embodiment are “probability of word alignment (log likelihood)” and “arrangement of words”, “backoff probability” is not included in the language model. May be.

  The “back-off probability” is used to calculate the probability of a word string that did not appear in the 3-word sequence when learning, for example, “3-gram” that considers up to 3-word sequence.

  In the example of FIG. 6, for example, the numerical value of the backoff probability in the case of {happy} is −0.18566667 in the learning data in the word string of 3 words following {happy} in the learning data. It represents the probability that something that did not appear will appear. For example, in the case of a 3-word sequence that does not appear in the learning data, the 3-word sequence model cannot be used, so the probability is determined using a shorter model, such as a 2-word sequence model, 1-word sequence model, etc. Represents a value. As described above, the backoff probability is used as an example of a coefficient used in an operation such as multiplication when a lower-order model is used.

<Examples of translation model learning and translation>
Next, a specific example of translation model learning and translation in this embodiment will be described. For example, as an example of a sentence spare stored in the learning data storage means (sentence spare) 31,
“Japanese: Nagano / Ha / Morning / From / Sunny”
"Sign language: {long} / {no [finger]} / {morning} / {from} / {clear} / {dream}"
Suppose there is. Note that “/” described above is a label indicating the division of the divided words.

From this data, for example, “GIZA ++” or the like obtains the correspondence between words, and creates a phrase pair based on it.
1. "Japanese: Nagano / ha", "Sign language: {long} / {no [finger]}"
2. “Japanese: Morning / From”, “Sign Language: {Morning} / {From}”
3. “Japanese: Sunny / De /”, “Sign Language: {Sunny} / {Dream}”
And three phrase pairs can be generated. In the present embodiment, the replacement model described above is learned after Japanese of the phrase pair is divided into character units.

  Examples of learned phrase pairs include “long → {long}”, “field → {no [finger]}”, “ha → (no corresponding sign language word)”, “morning → {morning}”, “ You can learn from “Kara / Ra → {From}”, “Sunny / Re / Ru → {Sunny}”, “De / Shi / yo / U → {Dream}”, etc. Here, “/” is a label indicating the division of the divided characters.

  In the present embodiment, by performing translation using the learning result (translation model) described above, it is possible to realize appropriate translation utilizing, for example, a sign language proper noun expression method.

  Next, examples of translation processing in the present embodiment will be described.

<Example 1>
For example, when “Kato” is input as an input word, the translation apparatus 10 performs translation using the fixed translation dictionary 21, and when “Kato → {Kato}” exists in the fixed translation dictionary 21, {Kato } Is output.

  In the case of the first embodiment, the translation device 10 includes translation sign language corresponding to all characters of the input word in the fixed translation dictionary 21, so that the translation by the translation unit 15 and the finger character by the digit conversion unit 16 No conversion is performed.

<Example 2>
For example, when “Fukushima” is input as an input word, the translation device 10 performs translation using the fixed translation dictionary 21 first, and since the input word is not included in the fixed translation dictionary 21, the translation model storage unit 37. Is translated using the translation rules of “Fuku → {Happy}” and “Island → {Island}” from the translation model stored in, and a translation result of {Happy} {Island} is output.

  In the case of the second embodiment, the translation device 10 includes the translation sign language corresponding to all the input words in the translation model stored in the translation model storage unit 37. Not performed.

<Example 3>
For example, when “Sonoda” is input as an input word, the translation device 10 performs translation using the fixed translation dictionary 21 first, and does not include the input word in the fixed translation dictionary 21. “Garden → {En (finger)}” “Ta → {Ta}” is obtained.

  Also, the translation device 10 processes the finger character by the finger character conversion means 16 because there is an instruction to convert the portion of “Garden → {En (finger character)}” into a finger character.

  Here, in the third embodiment, the above-mentioned word-reading association unit 42 obtains an association of “Sonoda” with “Sonoda”. Therefore, the translation apparatus 10 converts the translation of “zono” into {sono (finger character)} and finally obtains a translation result of {sono (finger character)} {field}.

<Example 4>
For example, when “Kamaishi” is input as an input word, the translation device 10 first translates using the fixed translation dictionary 21, and there is “stone → {stone}” in the fixed translation dictionary 21. Is not included in the dictionary, translation is performed using the translation model stored in the translation model storage unit 37. However, there is no translation rule because the character “Kama” is not learned.

  Therefore, the translation apparatus 10 uses the finger character conversion means 16 to convert “Kama” into a finger character. Therefore, “Kama → (Finger character)” and “Stone → {Stone}” are obtained as input to the finger character conversion means 16.

  As for “Kama”, since “Kamaishi” is read as “Kamaishi”, “Kama” can be obtained by using, for example, a known mono-ruby imparting technique. As for “Kama”, “Kama” may be associated with the word-reading association means 42 described above.

  Therefore, “Kama → (finger character)” becomes “Kama → Kama (finger character)”. As a result, the finger character conversion means 16 converts the “kama” into a finger character, and finally a translation result of {Kama (finger character)} {stone} is obtained.

  The translation results shown in the first to fourth embodiments can be output as CG sign language by a CG character, for example. Moreover, although Examples 1-4 mentioned above were only a word, the same process can be performed even if it is a character string (sentence) containing several words.

<Translation procedure>
That is, in the translation method in the present embodiment, when a word such as a proper noun is input, for example, the proper noun is first divided for each character, and a sign language word corresponding to each character is acquired from the replacement model. As a result, a plurality of translation candidates can be acquired for each character, and then they are combined in all patterns.

  For example, from the above-mentioned “score of the part to convert characters (translation probability)” obtained from the replacement model and “probability of word arrangement of translation results (probability of word arrangement)” obtained from the language model, “ P (S) · P (T | S) "is calculated, and the largest one is output as the translation result. Here, P (S) represents the likelihood obtained from the language model, and P (T | S) represents the likelihood obtained from the replacement model. P (S) means “likelihood of output sentence”, and P (T | S) is “probability that the original sentence is the sentence T when the sentence“ S ”is input. "Means.

For example, in the case of “Fukushima”, when the translation candidate and score (translation probability) for each character are first calculated from the replacement model, the relationship between Japanese → sign language and its likelihood is:
Lucky → {happy}: likelihood 0.6
Lucky → {Fuku [finger]}: Likelihood 0.5
Island → {Island}: likelihood 0.7
Island → {shore} {island}: likelihood 0.2
And so on.

Next, when the likelihood of combining each is obtained from the language model,
{Happy} {Island}: Likelihood 0.6
{Fuku [finger]} {island}: likelihood 0.3
{Happy} {shore} {island}: likelihood 0.2
{Fuku [finger]} {Kishi} {Island}: Likelihood 0.1
Etc. From these results,
P ({happy} {island}) = 0.6 * 0.7 * 0.6 = 0.252
P ({Hook [Finger]} {Island}) = 0.5 * 0.7 * 0.3 = 0.105
P ({happy} {shore} {island}) = 0.6 * 0.2 * 0.2 = 0.024
P ({Hook [Finger]} {Kishi} {Island}) = 0.5 * 0.2 * 0.1 = 0.01
Since {happiness} {island} has the maximum score (0.252), this is output as the final translation result.

<Comparative example>
Next, a comparative example of translation results between this embodiment and the conventional method will be described. FIG. 7 is a diagram showing a comparative example with the conventional method. In the comparative example, “GIZA ++” and “grow-diag-final-and” were used for word alignment and translation model generation. “Moses” was used as the decoder, and “SRILM” was used for learning the language model.

  Since Japanese proper nouns are character strings of several characters, 3-gram is adopted as a language model, and for example, 21995 sentence pairs of an existing sign language news corpus are used as learning data. In addition, personal names and place names were used as proper nouns.

  First, a total of 200 proper names consisting of 100 names randomly extracted from the Japanese surname database and 100 place names randomly extracted from Japanese city names were translated by 3 native sign language speakers. 96 person names and 82 place names that matched the expression of two or more speakers were used as comparison data.

  As a result, as shown in FIG. 7, in the conventional method (the translation model (Baseline method) directly learned from the sentence spare without being divided into phrases), 74 out of 96 words (accuracy rate 77. 1%), 52 of 82 place names were correct, with a correct rate of 63.4%. In this embodiment (this method), 79 out of 96 words (correct) Rate was 82.3%), and the place name was 56 out of 82 words (correct answer rate 68.3%) and correct.

  In other words, the translation accuracy of this method improved by 5.2 points for human names, 4.9 points for place names, and 5.0 points for all proper nouns (total). Therefore, according to the present embodiment as described above, sign language translation with high accuracy can be realized.

<Execution program>
Here, the translation device 10 described above includes, for example, a volatile storage device (storage device) such as a CPU (Central Processing Unit), a RAM (Random Access Memory), and a nonvolatile storage device (such as a ROM (Read Only Memory)). Storage device), an input device such as a mouse, a keyboard, and a pointing device, a display device for displaying images and data, and a computer having an interface device for communicating with the outside.

  Accordingly, the above-described functions of the translation apparatus 10 can be realized by causing the CPU to execute a program describing these functions. These programs can also be stored and distributed in a recording medium such as a magnetic disk (floppy (registered trademark) disk, hard disk, etc.), an optical disk (CD-ROM, DVD, etc.), or a semiconductor memory.

  That is, by generating an execution program (translation program) for causing a computer to execute the processing in each configuration described above and installing the program in, for example, a general-purpose personal computer or server, the above-described translation model learning processing or translation Processing can be realized. In addition, about the process by the execution program in this embodiment, it is not limited to this.

  As described above, according to the present embodiment, it is possible to realize highly accurate translation of proper nouns and the like in sign language translation. For example, a Japanese word as an example of the original language is associated with a sign language word as an example of the target language, and the sentence is divided into phrases based on the result, and the Japanese character is based on the divided result. By machine learning of the association between a sign language word and a sign language word, the number of characters and the number of sign language words in the learning data pair are reduced, and the learning accuracy of the association is improved. Therefore, more accurate sign language translation can be realized by using the translation model obtained by the learning method described above.

  Note that, for example, the association between Japanese words and sign language words is easy to associate because the number of appearances of words in the sentence having the same meaning is somewhat similar. In the present embodiment, when the learning data is divided, if there is a clear difference in the number of words between Japanese and Japanese sign language, the division result may be removed as incorrect. For example, this time, if there is a difference of more than twice the number of words and 5 or more, it is removed as an error, thereby enabling more appropriate translation.

  In this embodiment, not only sign language but also translation performance may be improved by learning a translation model after dividing a sentence. For example, this is effective when the learning data is not a complete “direct translation corpus”.

  Conventionally, for example, some programs have been broadcast in sign language, but it is difficult to secure a sign language interpreter, and information can be presented in sign language for sudden disasters such as at night. In some cases, sign language images with improved translation accuracy can be provided by applying the technique of the present embodiment described above.

  In the above-described embodiment, an example of translation from Japanese (original language) to Japanese sign language (target language) is shown. However, the present invention is not limited to this example. It can be applied to translations such as other language sign language, and can also be applied to a technique for translating Japanese sign language into Japanese.

  The preferred embodiments of the present invention have been described in detail above. However, the disclosed technology is not limited to the specific embodiments, and various modifications can be made within the scope of the gist of the present invention described in the claims. Modifications and changes are possible. It is also possible to combine all or some of the components of the above-described embodiments.

DESCRIPTION OF SYMBOLS 10 Translation apparatus 11 Translation model learning means 12 Word input means 13 Fixed translation translation means 14 Input character division means 15 Translation means 16 Finger character conversion means 17 Translation result output means 21 Fixed translation dictionary 31 Learning data storage means (sentence spare)
32 sentence division means 33 learning data storage means (phrase pair)
34 Character unit dividing means 35 Replacement model learning means 36 Language model learning means 37 Translation model storage means 41 Word reading input means 42 Word-reading association means

Claims (5)

In a translation device that performs sign language translation of input data,
Input character dividing means for dividing the input data into predetermined characters;
Translation means for translating a character divided by the input character dividing means using a translation model in which a combination of learning data of a translation target language and sign language is learned in a predetermined phrase unit,
A translation apparatus comprising translation result output means for outputting a translation result by the translation means. A translation model learning means for learning the translation model;
The translation model learning means includes sentence dividing means for dividing learning data for each sentence of the language to be translated and the sign language included in preset learning data into sentences in phrases.
Character unit dividing means for dividing the phrase pair learning data obtained by the sentence dividing means into character units;
The translation apparatus according to claim 1, wherein the translation model is learned using a phrase divided by the character unit dividing unit and a language model corresponding to learning data for each sentence. Before the input data is translated by the translation means, there is a fixed translation translation means for translating the input data using a preset fixed translation dictionary,
The translation apparatus according to claim 1, wherein the translation unit performs translation using the translation model for a word that cannot be translated by the fixed translation translation unit in the input data.   4. A finger character conversion means for converting the word into a sign language of a finger character when there is a word that could not be translated by the fixed translation means and the translation means in the input data. The translation device described in 1.   The translation program for functioning a computer as each means which the translation apparatus of any one of Claims 1 thru | or 4 has.

Images