JP5541124B2 - Language processing device, speech synthesis device, language processing method, and language processing program - Google Patents

Description

  The present invention relates to a language processing device, a speech synthesizer, a language processing method, and a language processing program.

  Natural language processing is incorporated into various man machine interfaces. As one aspect, when a text is output to a computer, natural language processing is used to read out the text. When text is read out, a morpheme analysis is performed on a sentence included in the text to divide the sentence into morphemes, and then reading is given to each morpheme. By executing the morphological analysis in this way, a phonetic character string is generated from a character string input as text.

  Here, as an example, a case where a phonetic character string is generated from a kanji-kana mixed sentence is assumed. Such a “kanji-kana mixed sentence” may include a word having a plurality of readings. As a technique for appropriately reading such words, a language processing apparatus that reads words using a co-occurrence relationship is disclosed. For example, when using the co-occurrence relationship to distinguish the word “rice”, the language processing device assigns “rice” if the word “rice” appears together with an agriculture-related word. When it appears together with the word, “be” is given. In addition to such a language processing device, a personal name kanji processing system that generates a kana-kanji conversion dictionary or constructs a personal name search system using a Chinese kanji table or a Korean kanji table is also disclosed.

JP 2001-184345 A

  However, the above-described conventional technique has a problem in that when a foreign language of Kanji is included in the text, the foreign language cannot be read out accurately.

  For example, the above language processing apparatus merely selects an appropriate reading as Japanese using a co-occurrence relationship. For this reason, even if the Japanese text contains Chinese, Korean, and Taiwanese kanji characters, it is impossible to determine which character string is in another language. Therefore, when text is read out using the language processing device described above, an erroneous Japanese reading may be given to a character string in another language.

  Furthermore, even if the kana-kanji conversion dictionary constructed by the above-described personal name kanji processing system is used for morphological analysis, it is still impossible to determine which character string is a foreign language among sentences included in the text. In addition, it is unknown what words in other languages are included in the text as character strings, and there is no limit to the number of words that exist in each country. Therefore, it is practically difficult to register all the foreign language words necessary for the kana-kanji conversion dictionary.

  The disclosed technology has been made in view of the above, and provides a language processing device, a speech synthesizer, a language processing method, and a language processing program capable of accurately reading out other languages of kanji notation included in text. For the purpose.

  The language processing apparatus disclosed in the present application includes an analysis unit that divides the sentence into morphemes and gives a reading to each morpheme by executing morpheme analysis on the sentence including kanji using a predetermined dictionary. Furthermore, the language processing apparatus includes an extraction unit that extracts an unknown word expressed in kanji from the result of morphological analysis by the analysis unit. Furthermore, the language processing device has a calculation unit that calculates a foreign language score representing the probability that the unknown word in Chinese characters extracted by the extraction unit is a foreign language different from the national language listed in the dictionary. Furthermore, the language processing apparatus includes a determination unit that determines which country the unknown word of the kanji notation is based on the foreign language score calculated by the calculation unit. Furthermore, the language processing apparatus includes a reading generation unit that generates a reading of an unknown word in the Chinese character notation according to a determination result by the determination unit.

  According to one aspect of the language processing device disclosed in the present application, there is an effect that it is possible to accurately read out another language of the kanji notation included in the text.

FIG. 1 is a diagram illustrating the configuration of the speech synthesizer according to the first embodiment. FIG. 2 is a diagram illustrating a configuration example of information stored in the morpheme dictionary storage unit. FIG. 3A is a diagram illustrating a configuration example of a Chinese co-occurrence dictionary stored in the co-occurrence dictionary storage unit. FIG. 3B is a diagram illustrating a configuration example of a Korean co-occurrence dictionary stored in the co-occurrence dictionary storage unit. FIG. 4 is a diagram illustrating a configuration example of information stored in the Japanese dictionary storage unit. FIG. 5 is a diagram illustrating a configuration example of a Korean kanji reading dictionary stored in the foreign language dictionary storage unit. FIG. 6 is a diagram for explaining a method for creating a co-occurrence dictionary. FIG. 7 is a diagram illustrating an example of accent assignment. FIG. 8 is a flowchart illustrating the procedure of the reading generation process according to the first embodiment. FIG. 9 is a flowchart illustrating the procedure of the Chinese score calculation process according to the first embodiment. FIG. 10 is a block diagram illustrating the configuration of the speech synthesizer according to the second embodiment. FIG. 11 is a diagram illustrating a configuration example of information stored in the single kanji dictionary storage unit. FIG. 12 is a diagram for explaining a method of creating a single kanji dictionary. FIG. 13 is a flowchart illustrating a procedure of Chinese score calculation processing according to the second embodiment. FIG. 14 is a schematic diagram illustrating an example of a computer that executes a language processing program according to the third embodiment.

  Hereinafter, embodiments of a language processing device, a speech synthesis device, a language processing method, and a language processing program disclosed in the present application will be described in detail with reference to the drawings. Note that this embodiment does not limit the disclosed technology. Each embodiment can be appropriately combined within a range in which processing contents are not contradictory.

[Configuration of speech synthesizer]
FIG. 1 is a diagram illustrating the configuration of the speech synthesizer according to the first embodiment. A speech synthesizer 10 shown in FIG. 1 performs processing for outputting input text (text) as a voice, that is, so-called text-to-speech reading. Is.

  That is, the speech synthesizer 10 according to the present embodiment reads the unknown word after determining which foreign language other than Japanese is the unknown word in the kanji notation included in the text by the foreign language score. Generate. Therefore, in the speech synthesizer 10 according to the present embodiment, the probabilities that an unknown word in Kanji is a foreign language is evaluated as the foreign language score, and then the reading of the foreign language having a high foreign language score is used as the unknown word reading. Can be generated. For this reason, in the speech synthesizer 10 according to the present embodiment, even if the Japanese text includes other languages such as Chinese, Korean, Taiwanese, etc., the character strings of the other languages are included. It can prevent giving wrong Japanese readings. Therefore, according to the speech synthesizer 10 according to the present embodiment, it is possible to accurately read out the other language of the Chinese character notation included in the text.

  In the example of FIG. 1, the following description is given assuming that the text-to-speech function is implemented in a personal computer (PC). However, the disclosed apparatus is not limited to this, and any information processing apparatus can be used. Applicable. Examples of the information processing apparatus include a mobile phone, a PHS (Personal Handyphone System), a PDA (Personal Digital Assistant), a car navigation system, and the like.

  The speech synthesizer 10 shown in FIG. 1 includes a reception unit 11, a morpheme dictionary storage unit 12, a co-occurrence dictionary storage unit 13, a Japanese dictionary storage unit 14a, a foreign language dictionary storage unit 14b, and a corpus storage unit 15a. And a creation unit 15. Furthermore, the speech synthesizer 10 includes a language processing unit 16, a prosody generation unit 17, a synthesis unit 18, and an output unit 19. Note that the speech synthesizer 10 includes various functional units included in a known computer in addition to the functional units illustrated in FIG. An example is an input device such as a keyboard or a mouse. Other examples include display devices such as a monitor, a display, and a touch panel. A further example is an interface for communicating with an external device.

  The reception unit 11 is a processing unit that receives text input. As an example, the reception unit 11 receives text data as text input from a web page acquired by a browser (not shown). As another example, the accepting unit 11 accepts a text file designated via an input device (not shown) as input text. As a further example, the receiving unit 11 converts data hooked from an application program (not shown) into text data and receives it as input text.

  In addition, the reception unit 11 divides the received input text into units in which processing is performed by the language processing unit 16 described later. As an example, the reception unit 11 outputs the input text to the subsequent language processing unit 16 one sentence at a time by separating a character string included in the input text each time a delimiter such as a punctuation mark, a question mark, or an exclamation mark is detected. To do. Here, the following explanation is given assuming that one sentence is output to the language processing unit 16 to be described later, but the input text is output to the subsequent language processing unit 16 in arbitrary units such as the number of characters and the data size. it can.

  The morpheme dictionary storage unit 12 is a storage unit that stores a dictionary used for morpheme analysis by an analysis unit 16a described later. As one aspect, the morpheme dictionary storage unit 12 stores morphemes and readings in association with each other. The “morpheme” here refers to the smallest meaningful unit among the elements of the sentence. Note that the “reading” stored in the morpheme dictionary storage unit 12 is determined that the morpheme that has not been extracted as an unknown word by the extraction unit 16b described later or the unknown word in kanji notation is Japanese by the determination unit 16d described later. It is adopted when

  FIG. 2 is a diagram illustrating a configuration example of information stored in the morpheme dictionary storage unit. In the example of FIG. 2, the morpheme “China” is read as “Chugok”, the morpheme “no” is read as “no”, the morpheme “is” is read as “o”, and the morpheme “visit” is read as “homon”. Indicates reading. Furthermore, in the example of FIG. 2, the morpheme “s” is read as “sul”, the morpheme “sei” is read as “sei”, the morpheme “city” is read as “miyako”, and the morpheme “kama” is called “kama”. Indicates that the morpheme “mountain” is read as “Yama”. Among these, “sei”, “tou”, “kama”, and “yama” are extracted as described later when the analysis unit 16a described later is a single kanji character that does not form a predicate with other kanji characters. It is extracted as an unknown word by the part 16b. The morpheme dictionary shown in FIG. 2 is merely an example, and any number of parts of speech can be registered.

  The co-occurrence dictionary storage unit 13 is a storage unit that stores a co-occurrence dictionary in which co-occurrence relationships between other languages and Japanese are defined. Such a co-occurrence dictionary is used to calculate a foreign language score, and the log likelihood of a co-occurrence word associated with a word included in a sentence from which an unknown word is extracted by an extraction unit 16b described later is described later. It is used for calculation of a foreign language score by the calculation unit 16c.

  As one aspect, the co-occurrence dictionary storage unit 13 stores a notation and a log likelihood in association with each type of co-occurrence words. The term “type of co-occurrence words” as used herein means that a word included in a sentence in which a Chinese noun appears when a co-occurrence dictionary is created is registered in the co-occurrence dictionary as a co-occurrence word. Refers to the type of position occupied by the noun. For example, if a word appears at a position other than immediately after a Chinese noun, the type of co-occurrence word is classified as “general co-occurrence word”. When a word appears immediately after a Chinese noun, the type of co-occurrence word is classified as “immediate word”. In addition, “log likelihood” represents the probability that an unknown word appearing in a sentence together with a co-occurrence word is a foreign language.

  Here, a Chinese co-occurrence dictionary is illustrated as an example of the co-occurrence dictionary. FIG. 3A is a diagram illustrating a configuration example of a Chinese co-occurrence dictionary stored in the co-occurrence dictionary storage unit. In the example of FIG. 3A, “China”, “Principal”, and “Visit Japan” are registered as common co-occurrence words, and the log likelihood of each is “2.322”, “1.874”, “1.246”. Show. Further, in the example of FIG. 3A, “middle president” is registered as the immediately following word, and the log likelihood is “4.822”. Although details will be described later with reference to FIG. 6, when the co-occurrence dictionary is created, the word immediately after is more related to the Chinese noun than the general co-occurrence word. The degree is weighted larger than the log likelihood of the general co-occurrence word and registered in the co-occurrence dictionary.

  Furthermore, a Korean co-occurrence dictionary is illustrated as an example of a co-occurrence dictionary. FIG. 3B is a diagram illustrating a configuration example of a Korean co-occurrence dictionary stored in the co-occurrence dictionary storage unit. In the example of FIG. 3B, “Korea”, “North Korea”, and “Visit Japan” are registered as common co-occurrence words, and their log likelihoods are “2.401”, “2.521”, and “1.312”, respectively. Indicates. Further, in the example of FIG. 3B, “Korea” is registered as the immediately following word, and the log likelihood is “4.972”. Here, the co-occurrence dictionaries for Chinese and Korean are illustrated here, but any other co-occurrence dictionaries can be prepared as long as the language includes Chinese characters such as Taiwanese.

  The Japanese dictionary storage unit 14a is a storage unit that stores a Japanese phonetic dictionary. Such a Japanese phonetic dictionary is used to generate a phonetic reading described later in order to give a Japanese phonetic reading as an unknown word reading when the determining unit 16d described later determines that an unknown word in Chinese characters is Chinese. Used by part 16e. Here, whether to generate Japanese phonetic reading as an unknown word or to generate Chinese reading by the reading generation unit 16e described later determines whether to read Chinese characters in Chinese via an input device (not shown). The mode can be selected by specifying ON or OFF of “Mode”.

  As an aspect, the Japanese dictionary storage unit 14a stores kanji and reading aloud in association with each other. FIG. 4 is a diagram illustrating a configuration example of information stored in the Japanese dictionary storage unit. In the example of FIG. 4, the reading of the kanji “sei” is “se”, the reading of the kanji “to” is “t”, the reading of the kanji “li” is “li”, and the kanji “liu” "" Indicates that "Ryu" is read. Note that the Japanese phonetic reading dictionary shown in FIG. 4 is merely an example, and any number of kanji readings can be registered.

  The foreign language dictionary storage unit 14b is a storage unit that stores a dictionary related to reading kanji in other languages. Such a foreign language dictionary is used by a later-described reading generation unit 16e when the later-described determining unit 16d determines that an unknown word in Chinese characters is a foreign language other than Japanese. If it is determined that the unknown word in Chinese characters is Chinese, the foreign language dictionary storage unit 14b will read the later-described reading generation unit 16e on the condition that “Chinese reading mode” is set to ON. Used by.

  Here, as an example of the foreign language dictionary, a kanji reading dictionary for Korean is illustrated. FIG. 5 is a diagram illustrating a configuration example of a Korean kanji reading dictionary stored in the foreign language dictionary storage unit. In the example of Fig. 5, the reading of the Chinese character "Kin" in Korean is "Kim", the reading of the Chinese character "Kama" in Korean is "Pu", and the reading of the Chinese character "Yama" in Korean. Indicates “San” and the reading of the Chinese character “Li” in Korean is “I”. Note that the Korean kanji reading dictionary shown in FIG. 5 is merely an example, and any number of kanji readings can be registered. Also, here, an example has been described in which kanji and readings are stored in association with each other, but reading accents may be stored in association with each other. In addition, here, the Kanji reading dictionary for Korean has been exemplified, but the Chinese character reading dictionary for Chinese can also be realized by the same configuration.

  In this embodiment, when an unknown word in Chinese characters is Chinese, a dictionary used to generate an unknown word reading is determined depending on whether the “Chinese reading mode” is ON or OFF. However, the present invention is not limited to this example. For example, either a Japanese phonetic dictionary or a foreign language dictionary may be used exclusively, and in this case, it is not necessary to maintain a dictionary that is not used.

  The corpus storage unit 15a is a storage unit that stores a Japanese corpus, that is, a large-scale language material. Such a corpus is used when a co-occurrence dictionary is created by the creation unit 15 described later. For this reason, the accuracy of the foreign language score calculated by the calculation part 16c mentioned later can be improved, so that there are many sentences prepared as a corpus. In order to create a co-occurrence dictionary, sentences having a co-occurrence relationship between other languages and Japanese are effective samples. Therefore, it is preferable that the corpus includes more nouns in other languages such as Chinese nouns and Korean.

  The creation unit 15 is a processing unit that creates a co-occurrence dictionary using the corpus storage unit 15a. A method for creating the co-occurrence dictionary will be described with reference to FIG. FIG. 6 is a diagram for explaining a method for creating a co-occurrence dictionary. The example in FIG. 6 illustrates the case where a Chinese co-occurrence dictionary and a Korean co-occurrence dictionary are created. However, a national language co-occurrence dictionary including Chinese characters such as Taiwanese is created. The same applies to the above.

  As shown in FIG. 6, the creation unit 15 performs morphological analysis on the Japanese corpus read from the corpus storage unit 15a (step S51). As an example, in the corpus storage unit 15a "Hello. ... You go to China in Beijing. ... North Korea in Pyongyang 17 days, ..." corpus that is assumed a case that has been held. In this case, the creation unit 15 obtains the following result as a result of morphological analysis of each sentence included in the corpus. In other words, the creation unit 15, "Hello (interjection). ... China (noun) • (particle), Beijing (Chinese noun) to - (particle) go (verb) mass (auxiliary verb). ... North Korea (Noun), No (Participant), Pyongyang (Korean noun), Deta (Participant), 17 (Numeric), Sun (Numeric Particle), ... ". In addition, the middle point “•” in the parentheses indicates a break between a morpheme and a morpheme.

  Subsequently, the creation unit 15 creates a database by extracting only sentences including Chinese nouns from the results obtained by the previous morphological analysis (step SA52). For example, the preparation unit 15 includes a sentence “China (noun), no (particle), Beijing (Chinese noun), he (particle), go (verb), and mas (auxiliary verb)” including the Chinese noun “Beijing”. Extract. In order to detect Chinese nouns in the extraction of such sentences, it may be detected by matching with nouns in the Chinese language dictionary or included in the corpus by an instruction operation by the co-occurrence dictionary designer. It is also possible to mark in advance a Chinese noun. Here, although one sentence is illustrated as a sentence including Chinese nouns, it is assumed that a large number of sentences are actually extracted.

  Then, the creation unit 15 generates a set of words immediately after the Chinese noun, that is, a set of words immediately after the Chinese noun, from the database of sentences including the Chinese noun (step SA53-1). For example, as the set of words immediately after, in the example of the above sentence, “he (particle)” is generated, and “ni (particle)…. Is generated.

  Further, the creating unit 15 generates a set of words excluding the Chinese noun and the word immediately after it, that is, a set of general co-occurrence words, from the database of sentences including the Chinese noun (step SA53-2). For example, as a set of general co-occurrence words, "... China (noun), no (particle), go (verb), mas (auxiliary verb) ..." and the like are generated.

  Subsequently, the creation unit 15 calculates the frequency at which each word appears in the set to which the word belongs, for each of the two sets generated earlier, hereinafter referred to as “appearance frequency” (step SA54-1 and step SA54-2). ). In the example of FIG. 6, “2012 (times)” is calculated as the frequency of occurrence of the word “he (particle)” immediately after, and “1893 (times)” as the frequency of occurrence of the word “ni (particle)” immediately after. Is calculated. Furthermore, “203 (times)” is calculated as the appearance frequency of the word “middle principal (noun)” immediately after, and “159 (times)” is calculated as the appearance frequency of the word “middle prime minister (noun)” immediately after. . On the other hand, “4183 (times)” is calculated as the appearance frequency of the general co-occurrence word “ga (particle)”, and “4024 (times)” is calculated as the appearance frequency of the general co-occurrence word “ha (particle)”. Is done. Furthermore, “176 (times)” is calculated as the appearance frequency of the general co-occurrence word “China (noun)”, “165 (times)” is calculated as the appearance frequency of the general co-occurrence word “lead (noun)” “162 (times)” is calculated as the appearance frequency of the general co-occurrence word “visit to Japan (noun)”. In addition, although the case where the frequency | count that a word appears as an appearance frequency was illustrated here was illustrated, it is good also as calculating the ratio that each word appears in the set to which it belongs as an appearance frequency.

  Here, the words appearing at the top of the appearance frequency of the immediately following word and general co-occurrence words are adjuncts such as particles and auxiliary verbs, and these adjuncts have low correlation with Chinese nouns. Therefore, in the subsequent processing, a word whose part of speech is a noun from a set of words immediately after and a set of general co-occurrence words is set as an object to be placed in the co-occurrence dictionary. As a result, it is possible to preferentially register words having high correlation with Chinese nouns in the co-occurrence dictionary, and it is possible to improve the reliability of the foreign language score calculated using the words.

  Then, the creation unit 15 calculates the log likelihood of each word for each set of immediately following words and general co-occurrence words (steps SA55-1 and SA55-2). As an example of the log likelihood, a calculation formula “log likelihood = log (appearance frequency * 100% / number of sentences)” is used. Any value can be applied to the “base of logarithm” of the above calculation formula, but as an example, it is preferable to use a natural logarithm. The “number of sentences” in the above calculation formula indicates the number of sentences stored in a database of sentences including Chinese nouns. In the example of FIG. 6, “2.411” is calculated as the log likelihood of the word “middle principal (noun)” immediately after, and “1.156” is calculated as the log likelihood of the word “middle prime minister (noun)” immediately after. Is done. On the other hand, “2.322” is calculated as the log likelihood of the general co-occurrence word “China (noun)”, and “1.874” is calculated as the log likelihood of the general co-occurrence word “lead (noun)”. “1.246” is calculated as the log likelihood of “visit to Japan (noun)”.

  Thereafter, the creating unit 15 assigns a weight greater than the log likelihood of the general co-occurrence word to the log likelihood of the immediately subsequent word among the log likelihood of the immediately subsequent word and the log likelihood of the general co-occurrence word. Is registered in the co-occurrence dictionary storage unit 13 as a Chinese co-occurrence dictionary (step SA56). As an example of such weighting, the creation unit 15 sets the log likelihood of the immediately following word to twice the log likelihood of the general co-occurrence word, and multiplies the log likelihood of the immediately following word by “2”. Such weighting increases the proportion of the log likelihood of the immediately following word that has a higher correlation with Chinese nouns than the general co-occurrence word, which is reflected in the calculation of the foreign language score, thereby improving the reliability of the foreign language score. Can do. Note that the log likelihood weighting is not limited to the above example. That is, the weight given to the log likelihood of the immediately following word only needs to be higher than the weight given to the log likelihood of the general co-occurrence word, and an arbitrary value can be given to the weight to each log likelihood.

  As described above, the Chinese co-occurrence dictionary shown in FIG. 3A is created by the processing from step S51 to step SA56. This makes it possible to define an evaluation criterion for evaluating the probability that an unknown word in Chinese characters is Chinese. In addition, the process of said step SA53-1-SA55-1 and the process of said step SA53-2-SA55-2 can also be performed in parallel, and either is processed before or after. It does not matter as executing.

  On the other hand, the creation unit 15 creates a Korean co-occurrence dictionary in the same manner as the Chinese co-occurrence dictionary. Explaining this, the creation unit 15 creates a database by extracting only sentences containing Korean nouns from the results obtained by the previous morphological analysis (step SB52). For example, the creating unit 15 may include a sentence containing the Korean noun “Pyongyang” “... North Korea (noun), no (particle), Pyongyang (Korean noun), de (particle), 17 (numerical), and date (numerical particle). ), ... "is extracted. In order to detect Korean nouns when extracting such sentences, it is possible to detect Korean nouns by matching them with nouns in the Korean language dictionary, or include them in the corpus by an instruction operation by the co-occurrence dictionary designer. It is also possible to mark in advance a Korean noun. In addition, although one sentence was illustrated here as a sentence containing a Korean noun, in fact, many sentences shall be extracted.

  Then, the creation unit 15 generates a set of words immediately after the Korean noun, that is, a set of words immediately after the Korean noun, from a database of sentences including Korean nouns (step SB53-1). For example, as a set of words immediately after, in the example of the above sentence, “de (particle)” is generated, and “he (particle)… ・ Korea (noun), general secretary (noun)…”, etc. Shall be generated.

  Further, the creating unit 15 generates a set of words excluding the Korean noun and the word immediately after it, that is, a set of general co-occurrence words, from the database of sentences including Korean nouns (step SB53-2). For example, as a set of general co-occurrence words, “... North Korea (noun), (particle), 17 (numerical), date (numerical particle),.

  Subsequently, the creation unit 15 calculates the appearance frequency of each word for each of the two sets generated previously (Steps SB54-1 and SB54-2). In the example of FIG. 6, “1671 (times)” is calculated as the appearance frequency of the word “de (particle)” immediately after, and “1422 (times)” as the frequency of occurrence of the word “he (particle)” immediately after. Is calculated. Further, “160 (times)” is calculated as the appearance frequency of the word “Korea (noun)” immediately after, and “133 (times)” is calculated as the appearance frequency of the word “general secretary (noun)” immediately after. On the other hand, “2977 (times)” is calculated as the appearance frequency of the general co-occurrence word “no (particle)”, and “2889 (times)” is calculated as the appearance frequency of the general co-occurrence word “ha (particle)”. Is done. Furthermore, “156 (times)” is calculated as the appearance frequency of the general co-occurrence word “Korea (noun)”, and “161 (times)” is calculated as the appearance frequency of the general co-occurrence word “North Korea (noun)”. Then, “128 (times)” is calculated as the appearance frequency of the general co-occurrence word “visit to Japan (noun)”. In this case as well, in the subsequent processing, a word whose part of speech is a noun in a set of words immediately after and a set of general co-occurrence words is a target to be placed in the co-occurrence dictionary.

  Then, the creation unit 15 calculates the log likelihood of each word for each set of immediately following words and general co-occurrence words (step SB55-1 and step SB55-2). In the example of FIG. 6, “2.486” is calculated as the log likelihood of the immediately following word “Korea (noun)”, and “1.475” is calculated as the log likelihood of the immediately following word “general secretary (noun)”. The On the other hand, “2.401” is calculated as the log likelihood of the general co-occurrence word “Korea (noun)” and “2.521” is calculated as the log likelihood of the general co-occurrence word “North Korea (noun)”. “1.312” is calculated as the log likelihood of the word “visit to Japan (noun)”.

  Thereafter, the creating unit 15 assigns a weight greater than the log likelihood of the general co-occurrence word to the log likelihood of the immediately subsequent word among the log likelihood of the immediately subsequent word and the log likelihood of the general co-occurrence word. Is registered in the co-occurrence dictionary storage unit 13 as a Korean co-occurrence dictionary (step SB56). As an example of such weighting, the creation unit 15 sets the log likelihood weight of the immediately following word to twice the log likelihood of the general co-occurrence word, as in the case of the Chinese co-occurrence dictionary, The log likelihood of the immediately following word is multiplied by “2”.

  Thus, the Korean co-occurrence dictionary shown in FIG. 3B is created by the processing from step S51 to step SB56. This makes it possible to define an evaluation criterion for evaluating the probability that an unknown word in Kanji is Korean. In addition, the process of said step SB53-1-SB55-1 and the process of said step SB53-2-SB55-2 can also be performed in parallel, and either is processed before or after. It does not matter as executing. Further, the processes of steps SA52 to SA56 for creating a Chinese co-occurrence dictionary and the processes of steps SB52 to SB56 for creating a Korean co-occurrence dictionary can be executed in parallel. The process may be executed either before or after.

  Returning to the description of FIG. 1, the language processing unit 16 is a processing unit that executes natural language processing on the input text received by the receiving unit 11. As shown in FIG. 1, the language processing unit 16 includes an analysis unit 16a, an extraction unit 16b, a calculation unit 16c, a determination unit 16d, a reading generation unit 16e, a provision unit 16f, and a phonetic generation unit 16g. And further.

  Among these, the analysis unit 16a is a processing unit that uses the morpheme dictionary storage unit 12 to perform a morpheme analysis on a sentence included in the input text, thereby dividing the sentence into morphemes and giving a reading to each morpheme. is there. As an example, it is assumed that an input text “Visit Chengdu, China” is input from the reception unit 11. In this case, the analysis unit 16a searches the morpheme dictionary shown in FIG. 2 for a morpheme that matches the character string included in the input text, and sets the input text to “China”, “no”, “sei”, “city” ”,“ ”,“ Visit ”,“ do ”. At this time, the analysis unit 16a, in the character string included in the input text, “Cheng” and “City” are actually two-character words “Chengdu”, but are not registered in the morpheme dictionary. Recognize that these morphemes are single kanji characters. After that, the analysis unit 16a gives a reading to each morpheme, and reads “Chinese [Chugok], [No], Cheng (single kanji) [Se], Miyako (single kanji) [Miyako]], [o].・ The result of the morphological analysis is “Visit [Houmon]. Note that any known morpheme analysis method can be applied to the morpheme analysis.

  The extraction unit 16b is a processing unit that extracts unknown words from the result of morphological analysis by the analysis unit 16a. As an example, the extraction unit 16b extracts a character analyzed as a single kanji character by the analysis unit 16a from the morphemes included in the input text as an unknown word expressed in kanji. In the example of the input text above, the extraction unit 16b extracts “Chengdu” and “Chengdu” as unknown words in Chinese characters because “Chengdu” and “City” are recognized as single Chinese characters. It is assumed that words that are established as foreign words, such as “Beijing” and “Shanghai”, are registered in the morpheme dictionary and are not extracted as unknown words.

  The calculation unit 16c is a processing unit that uses the co-occurrence dictionary storage unit 13 to calculate a foreign language score representing the probability that the unknown word in Chinese characters extracted by the extraction unit 16b is a foreign language. Explaining this, the calculation unit 16c determines whether or not there is a co-occurrence word registered in the co-occurrence dictionary in the morpheme included in the input text, that is, the unknown word in kanji notation coexists with a noun in another language. It is determined whether or not a relationship can be established. At this time, if there is a co-occurrence word registered in the co-occurrence dictionary, the calculation unit 16c uses the log likelihood associated with the co-occurrence word to translate another language of the unknown word in kanji notation. Calculate the score. Then, the calculation unit 16c repeats the calculation of the other language score until there are no co-occurrence words registered in the co-occurrence dictionary in the morphemes included in the input text, and the newly calculated other language score is calculated until the previous time. Cumulatively add to the other language score that was cumulatively added. The calculation of the foreign language score is executed for each Chinese co-occurrence dictionary and Korean co-occurrence dictionary, that is, for each foreign language. Hereinafter, the foreign language score related to Chinese is referred to as “Chinese score”, and the foreign language score related to Korean is referred to as “Korean score”.

  As an example, such a foreign language score is calculated using a calculation formula “foreign language score = log likelihood of co-occurrence word / distance from unknown word to morpheme”. Here, in the above formula, the log likelihood of the co-occurrence word is divided by the distance from the unknown word to the morpheme. The closer the distance between the unknown word and the morpheme, the stronger the correlation between the two. This is because there is a higher possibility that is a foreign language. If no co-occurrence word registered in the co-occurrence dictionary exists, the foreign language score is calculated as zero. Also, if there are co-occurrence words of both the general co-occurrence word and the immediately following word in the co-occurrence dictionary, for example, if the log likelihood of the immediately following word is used for the other language score Good.

  In the example of the above text, “China” registered as a general co-occurrence word in the Chinese co-occurrence dictionary exists in the morphemes included in the input text. For this reason, the calculation unit 16c uses the log likelihood “2.322” associated with the general co-occurrence word “China” and the unknown word “Chengdu” included in the input text in the above calculation formula for the foreign language score. Substitute the distance “2” to the morpheme “China”. Thereby, the calculation unit 16c calculates the Chinese score “1.161” of the unknown word “Chengdu”. On the other hand, there is no general co-occurrence word or immediately following word registered in the Korean co-occurrence dictionary in the morphemes included in the input text. Therefore, the calculation unit 16c calculates the Korean score “0.000” of the unknown word “Chengdu”.

  The distance from the unknown word to the morpheme can be calculated by giving an arbitrary morpheme among the morphemes included in the input text as an origin and giving coordinates to other morphemes. As an example, the coordinate position can be given to each morpheme by setting the morpheme at the beginning of the sentence as the origin and incrementing the coordinate value by one as the morpheme advances from the origin by one. When the coordinates are given to the morpheme in this way, the distance value becomes a negative value depending on the positional relationship between the unknown word and the morpheme, so it is preferable to take the absolute value. In the example of the above text, since the coordinate of the morpheme “China” is “0” and the coordinate of the unknown word “Chengdu” is “2”, the calculation unit 16c calculates the morpheme “China” from the unknown word “Chengdu”. Is calculated as “2” by calculating | 2-0 |. The method for calculating the distance is not limited to the above method, and the number of morphemes from the unknown word to the target morpheme may be measured.

  The discriminating unit 16d is a processing unit that discriminates which country the unknown word of kanji is based on, based on the foreign language score calculated by the calculating unit 16c. Explaining this, the determination unit 16d compares the Chinese score with the Korean score. At this time, if the Chinese score is higher than the Korean score, the determination unit 16d further compares the Chinese score with a predetermined threshold, for example, “1.000”. If the Chinese score is higher than the threshold, it can be determined that the unknown word in Chinese characters is “Chinese”. On the other hand, if the Chinese score is less than or equal to the threshold value, it can be determined that the unknown word in Chinese characters is “Japanese”. If the Korean score is greater than or equal to the Chinese score, the determination unit 16d compares the Korean score with a predetermined threshold, for example, “1.000”. If the Korean score is higher than the threshold, it can be determined that the unknown word in Kanji is “Korean”. On the other hand, if the Korean score is equal to or less than the threshold, it can be determined that the unknown word in Kanji notation is “Japanese”.

  In the example of the input text above, the Chinese score is “1.161”, the Korean score is “0.000”, and the threshold is “1.000”. Therefore, since Chinese score> threshold> Korean score, it can be determined that the unknown word “Chengdu” is in Chinese.

  In addition, the threshold value compared with a Chinese score and a Korean score is not limited to said value, Arbitrary values are employable. As an example, if the title given to the input text is registered in the Chinese or Korean co-occurrence dictionary, there is a high possibility that sentences related to Chinese or Korean will be entered. The threshold of the corresponding country can be lowered as compared with the case where it is not registered. It goes without saying that different threshold values can be adopted between the Chinese score and the Korean score.

  The reading generation unit 16e is a processing unit that generates a reading of an unknown word in Chinese characters according to the determination result by the determination unit 16d. To explain this, the reading generation unit 16e determines that “Chinese score> Korean score” and “Chinese score> threshold”, that is, if it is determined that the unknown word is Chinese, It is further determined whether or not “reading mode” is OFF. At this time, when the “Chinese reading mode” is OFF, the reading generation unit 16e generates a Japanese phonetic reading for each single Chinese character corresponding to the unknown word, using the Japanese dictionary storage unit 14a. On the other hand, when the “Chinese reading mode” is ON, the reading generation unit 16e uses the Chinese kanji reading dictionary stored in the foreign language dictionary storage unit 14b to read Chinese readings into unknown words. Generate.

  When “Korean score ≧ Chinese score” and “Korean score> threshold”, that is, when it is determined that the unknown word is Korean, the reading generation unit 16e displays the foreign language dictionary storage unit 14b. A Korean reading is generated for an unknown word using the Korean kanji reading dictionary stored in. When “Chinese score ≦ threshold” or “Korean score ≦ threshold”, that is, when it is determined that the unknown word is Japanese, the reading generation unit 16e generates a new reading for the unknown word. Instead, the result of the morphological analysis by the analyzing unit 16a is output to the adding unit 16f as it is. Even when an unknown word is not extracted by the extraction unit 16b, the reading generation unit 16e outputs the result of the morphological analysis by the analysis unit 16a to the adding unit 16f as it is.

  In the example of the text described above, the reading generation unit 16e reads the unknown word “Chengdu” in kanji notation according to “sei” and “to” stored in the Japanese dictionary storage unit 14a. Generate a “sate”. For this reason, it is possible to prevent the incorrect word “Chengdu” from being assigned to the unknown word “Chengdu” as a result of an erroneous morphological analysis.

  The assigning unit 16f is a processing unit that assigns an accent to the reading of an unknown word in kanji notation generated by the reading generating unit 16e. As an example, the assigning unit 16f assigns an accent to the first mora, that is, one short syllable of the second character from the back of the unknown word in Chinese characters.

  FIG. 7 is a diagram illustrating an example of accent assignment. As shown in FIG. 7, when the unknown word is “Li Yi”, the assigning unit 16 f gives an accent to “Li [Li]” as the second character from the back to be [Li ′ Ho]. If the unknown word is “Liu Bei”, the assigning unit 16f gives an accent to the first mora “Ryu” of the second character “Liu” from the back, and [Ryu Ubi] To do. When the unknown word is “Masawa”, the assigning unit 16f gives an accent to the first mora “ta” of the second character “sawa [taku]” from the back, and [Mouta Kaku] To do. Also, when the unknown word is “Hu Jintao”, the adding unit 16f adds an accent to the first mora “ki” of the second character “Kin” from the back, and [Kokitoto] To do. In addition, when the unknown word is “blue crab meat”, the adding unit 16f adds an accent to the first mora “ro” of the second character “meat” from the back, ].

  The phonetic generation unit 16g is a processing unit that generates a phonetic character string of the input text from the result of the morphological analysis in which an accent is added to the reading of an unknown word written in Chinese characters by the adding unit 16f. As an example, the phonetic generation unit 16g generates a katakana character string, which is a phonetic character, from a kanji-kana mixed sentence in which an accent is added to an unknown word reading in kanji notation. If an unknown word is not extracted by the extraction unit 16b, or if the unknown word is determined to be Japanese by the determination unit 16d, a phonetic character string is generated from the result of the morphological analysis by the analysis unit 16a. Is done.

  The prosody generation unit 17 is a processing unit that generates a prosody corresponding to the input text based on the phonetic character string generated by the phonetic generation unit 16g. The “prosody” mentioned here is a general term for the characteristics of how to speak, such as pause, phoneme length and intonation. As one aspect, the prosody generation unit 17 synthesizes a voice to be synthesized by the synthesis unit 18 to be described later, that is, a pitch pattern (pitch) which is a phoneme time length that is the length of each phoneme of the synthesized speech or a pitch change pattern. pattern) and the like.

  The synthesis unit 18 is a processing unit that generates a speech waveform from the prosody generated by the prosody generation unit 17 and synthesizes speech. As one aspect, the synthesizer 18 artificially synthesizes speech by generating a speech waveform according to the prosody generated by the prosody generator 17, for example, phoneme time length or pitch pattern.

  The output unit 19 is an output unit that outputs sound. As an example, the output unit 19 outputs the synthesized speech input from the synthesis unit 18. An example of the output unit 19 includes a speaker.

  Note that various types of integrated circuits and electronic circuits can be employed for the reception unit 11, the creation unit 15, the language processing unit 16, the prosody generation unit 17, and the synthesis unit 18 illustrated in FIG. Further, a part of the functional unit included in the language processing unit 16 may be another integrated circuit or an electronic circuit. For example, examples of the integrated circuit include ASIC (Application Specific Integrated Circuit) and FPGA (Field Programmable Gate Array). Examples of the electronic circuit include a central processing unit (CPU) and a micro processing unit (MPU).

  The hardware of the morpheme dictionary storage unit 12, the co-occurrence dictionary storage unit 13, the Japanese dictionary storage unit 14a, the foreign language dictionary storage unit 14b, and the corpus storage unit 15a shown in FIG. Applicable. As an example, a semiconductor memory element such as a random access memory (RAM), a read only memory (ROM), or a flash memory can be employed. It should be noted that a storage device such as a hard disk or an optical disk may be adopted for the above five storage units.

[Process flow]
Next, the flow of processing of the speech synthesizer according to the present embodiment will be described. Here, after explaining the (1) reading generation process for generating a reading for an unknown word in Kanji notation, (2) calculating the other language score will be described.

(1) Reading Generation Processing FIG. 8 is a flowchart illustrating the procedure of reading generation processing according to the first embodiment. This reading generation process is started when a single input text is input to the language processing unit 16 by the receiving unit 11.

  As shown in FIG. 8, the analysis unit 16a uses the morpheme dictionary storage unit 12 to perform morpheme analysis on a sentence included in the input text (step S101). Subsequently, the extraction unit 16b extracts a character analyzed by the analysis unit 16a as a single kanji character from the morphemes included in the input text as an unknown word in kanji notation (step S102).

  At this time, when an unknown word is not extracted (No at Step S102), the reading generation unit 16e outputs the result of the morphological analysis by the analysis unit 16a to the adding unit 16f as it is (Step S103), and ends the processing. .

  Here, when an unknown word is extracted (Yes in step S102), the calculation unit 16c uses the co-occurrence dictionary storage unit 13 to calculate other language scores such as a Chinese score and a Korean score. "National language score calculation process" is executed (step S104).

  Thereafter, the determination unit 16d compares the Chinese score with the Korean score (step S105). At this time, if the Chinese score is higher than the Korean score (Yes at Step S105), the determination unit 16d further compares the Chinese score with a predetermined threshold, for example, “1.000” (Step S106). .

  If the Chinese score is higher than the threshold (Yes at Step S106), it can be determined that the unknown word in Chinese characters is “Chinese”. In this case, the reading generation unit 16e further determines whether or not “Chinese reading mode” is OFF (step S107).

  At this time, if the “Chinese reading mode” is OFF (Yes at Step S107), the reading generation unit 16e uses the Japanese dictionary storage unit 14a to store Japanese for each single Chinese character corresponding to the unknown word. Sound reading is generated (step S109), and the process is terminated.

  On the other hand, when the “Chinese reading mode” is ON (No at Step S107), the reading generation unit 16e uses the Chinese kanji reading dictionary stored in the foreign language dictionary storage unit 14b to identify an unknown word. A Chinese reading is generated (step S108).

  If the Chinese score is less than or equal to the threshold (No at step S106), it can be determined that the unknown word in Kanji is “Japanese”. In this case, the reading generation unit 16e outputs the result of the morphological analysis by the analysis unit 16a to the adding unit 16f without generating a new reading for the unknown word (step S103), and ends the process.

  If the Korean score is greater than or equal to the Chinese score (No at Step S105), the determination unit 16d compares the Korean score with a predetermined threshold, for example, “1.000” (Step S110).

  At this time, if the Korean score is higher than the threshold (Yes at Step S110), it can be determined that the unknown word in Kanji notation is “Korean”. In this case, the reading generation unit 16e generates a Korean reading for the unknown word using the Korean kanji reading dictionary stored in the foreign language dictionary storage unit 14b (step S111), and ends the process. .

  On the other hand, if the Korean score is equal to or less than the threshold (No at Step S110), it can be determined that the unknown word in Kanji is “Japanese”. In this case, the reading generation unit 16e outputs the result of the morphological analysis by the analysis unit 16a to the adding unit 16f without generating a new reading for the unknown word (step S103), and ends the process.

(2) Other Language Score Calculation Processing FIG. 9 is a flowchart illustrating a procedure of Chinese score calculation processing according to the first embodiment. This Chinese score calculation process is a process corresponding to step S104 shown in FIG. 8, and starts when an unknown word in Chinese characters is extracted by the extraction unit 16b. Here, the case where the Chinese score is calculated among the other language scores is illustrated, but the same applies to the case where the Korean score is calculated.

  As illustrated in FIG. 9, the calculation unit 16c sets various parameters (step S301). As an example, the calculation unit 16c initializes the Chinese score and the loop counter I to zero, and sets values for the position J of the unknown word and the number N of morphemes in the sentence.

  Then, the calculating unit 16c repeatedly executes the following steps S303 to S307 until the loop counter I becomes equal to the number N of words in the sentence (Yes at Step S302).

  That is, the calculation unit 16c determines whether or not the I-th morpheme is registered in the Chinese co-occurrence dictionary, that is, whether or not an unknown word in Chinese characters can have a co-occurrence relationship with a Chinese noun. (Step S303).

  At this time, when the I-th morpheme is registered in the Chinese co-occurrence dictionary (Yes in step S303), the calculating unit 16c calculates the logarithm corresponding to the I-th morpheme from the Chinese co-occurrence dictionary. The likelihood is extracted (step S304). The calculating unit 16c calculates the distance L from the unknown word to the morpheme by calculating the absolute value of the loop counter I-the position J of the unknown word, that is, | I-J | (step S305).

  Subsequently, the calculation unit 16c calculates the Chinese score of the I-th morpheme by the calculation formula “Chinese score = log likelihood corresponding to the I-th morpheme / distance L from unknown word to morpheme” to calculate “I Cumulative addition is added to the Chinese score that has been cumulatively added up to the “−1” th (step S306).

  Then, the calculation unit 16c increments the loop counter I (step S307), and proceeds to the above step S302. If the I-th morpheme is not registered in the Chinese co-occurrence dictionary (No at Step S303), the processing from Step S304 to Step S306 is not executed and the loop counter I is incremented ( Step S307).

  After that, when the loop counter I becomes equal to the number N of words in the sentence (No at Step S302), the Chinese score has been calculated for all the morphemes of the input text, and thus the process ends.

[Effect of Example 1]
As described above, the speech synthesizer 10 according to the present embodiment performs a morpheme analysis on a sentence including kanji using a morpheme dictionary, thereby dividing the sentence into morphemes and giving a reading to each morpheme. To do. Furthermore, the speech synthesizer 10 extracts an unknown word in Chinese characters from the result of morphological analysis. Furthermore, the speech synthesizer 10 calculates a foreign language score representing the probability that the unknown word in Kanji is a foreign language. Furthermore, the speech synthesizer 10 determines which country the unknown word in Kanji is based on the other language score. Furthermore, the speech synthesizer 10 generates a reading of an unknown word in Chinese characters according to the determination result.

  Therefore, in the speech synthesizer 10 according to the present embodiment, the probabilities that an unknown word in Kanji is a foreign language is evaluated as the foreign language score, and then the reading of the foreign language having a high foreign language score is used as the unknown word reading. Can be generated. For this reason, in the speech synthesizer 10 according to the present embodiment, even if the Japanese text includes other languages such as Chinese, Korean, Taiwanese, etc., the character strings of the other languages are included. It can prevent giving wrong Japanese readings. Therefore, according to the speech synthesizer 10 according to the present embodiment, it is possible to accurately read out the other language of the Chinese character notation included in the text.

  In addition, the speech synthesizer 10 according to the present embodiment calculates a foreign language score using a co-occurrence relationship between a morpheme other than an unknown word included in a sentence and a noun in another language. For this reason, according to the speech synthesizer 10 according to the present embodiment, it is possible to evaluate whether or not an unknown word expressed in Kanji can have a co-occurrence relationship with a noun in another language by using the other language score.

  Furthermore, the speech synthesizer 10 according to the present embodiment adds an accent to the reading of an unknown word in kanji notation, and the phonetic character of the input text from the result of the morphological analysis in which the accent is added to the reading of the unknown word in kanji notation. Generate a column. For this reason, the speech synthesizer 10 according to the present embodiment can generate a phonetic character string after appropriately adding an accent in addition to reading an unknown word. Therefore, according to the speech synthesizer 10 according to the present embodiment, it is possible to read out the other language of the kanji notation included in the text more accurately.

  In the first embodiment, the case where the foreign language score is calculated using the co-occurrence dictionary is illustrated. However, the disclosed apparatus is not limited to this, and the other language score can be calculated by another method. . Therefore, in the second embodiment, a case will be described in which a foreign language score is calculated using a single Chinese character dictionary in which the probability that a single Chinese character is a foreign language is defined.

  FIG. 10 is a block diagram illustrating the configuration of the speech synthesizer according to the second embodiment. In the present embodiment, the same reference numerals are given to functional units having the same functions as those of the speech synthesizer 10 according to the first embodiment, and the description thereof is omitted.

  Compared to the speech synthesizer 10 shown in FIG. 1, the speech synthesizer 30 shown in FIG. 10 replaces the co-occurrence dictionary storage unit 13, the corpus storage unit 15 a, the creation unit 15, and the language processing unit 16 with a single kanji dictionary. The point which has the memory | storage part 31, the foreign language noun memory | storage part 32a, the preparation part 15, and the language processing part 33 differs. Furthermore, the calculation part 33a shown in FIG. 10 differs in the processing content compared with the calculation part 16c shown in FIG.

  Among these, the single Chinese character dictionary storage unit 31 is a storage unit that stores a single Chinese character dictionary in which the uniqueness of a single Chinese character is defined. As an aspect, the single Chinese character dictionary storage unit 31 stores the Chinese character and Korean character of the single Chinese character in association with each single Chinese character. The “Chinese character” here refers to the probability that a single Chinese character is Chinese, and is calculated by calculating the log likelihood from the frequency at which the single Chinese character appears with a Chinese noun. Similarly, “Koreanness” refers to the probability that a single Chinese character is Korean, and is calculated by calculating the log likelihood from the frequency with which a single Chinese character appears together with a Korean noun. In the single kanji dictionary, the uniqueness of another language associated with the single kanji that forms the unknown word extracted by the extraction unit 16b is used for the calculation of the other language score by the calculation unit 33a described later.

  FIG. 11 is a diagram illustrating a configuration example of information stored in the single kanji dictionary storage unit. In the example of FIG. 11, the Chinese character of the single Chinese character “Kin” is “1.534” and the Korean character is “4.126”. In addition, the Chinese character of the single kanji character “Kama” is “1.478”, and the Korean character is “3.632”. In addition, the Chinese character of the single Chinese character “Yama” is “1.759”, and the Korean character is “1.758”. Also, the Chinese character of the single Chinese character “Li” is “3.335” and the Korean character is “3.411”. In the same manner, the Chinese character and Korean character of the single kanji characters “Cheng”, “Miyako”, “Liu” and “Kyo” are defined.

  The foreign language noun storage unit 32a is a storage unit that stores a noun list of other languages. As one aspect, the foreign language noun storage unit 32a stores a Chinese noun list and a Korean noun list. Such a noun list of other languages is used when a single kanji dictionary is created by the creation unit 32 described later. For this reason, the accuracy of the foreign language score calculated by the calculation part 33a mentioned later can be improved, so that there are many sentences prepared as a noun list.

  The creation unit 32 is a processing unit that creates a single kanji dictionary using the foreign language noun storage unit 32a. A method for creating this single kanji dictionary will be described with reference to FIG. FIG. 12 is a diagram for explaining a method of creating a single kanji dictionary. In the example of FIG. 12, the case of defining the Chinese character and Korean character of a single kanji character is illustrated. However, if it is a national language including a Chinese character, for example, Taiwanese, the character of the Taiwanese character can be defined similarly.

  As shown in FIG. 12, the creation unit 32 reads a Chinese noun list from the noun list stored in the foreign language noun storage unit 32a, and sets the Chinese noun list loaded in the Chinese noun list to 1 Two kanji strings are connected (step SA71). In the example of FIG. 12, the creating unit 32 concatenates Chinese nouns “Beijing”, “Nanjing”, “Liu Bei”, and so on to form a Chinese character string “Beijing Nanjing Liu Bei ...”.

  Then, the creating unit 32 calculates the frequency at which each Chinese character included in the Chinese character string of the Chinese noun appears in the Chinese character string, hereinafter referred to as “appearance frequency” (step SA72). In the example of FIG. 12, “37 (times)” is calculated as the appearance frequency of the Chinese character “Liu”, and “35 (times)” is calculated as the appearance frequency of the Chinese character “Kyo”. Furthermore, “32 (times)” is calculated as the appearance frequency of the Chinese character “Li”, and “26 (times)” is calculated as the appearance frequency of the Chinese character “Miyako”. Further, “25 (times)” is calculated as the appearance frequency of the Chinese character “Sei”.

  Then, the creation unit 32 calculates the log likelihood of each Chinese character (step SA73). As an example of the log likelihood, a calculation formula “log likelihood = log (appearance frequency * 100% / length of kanji string)” is used. Any value can be applied to the “base of logarithm” of the above calculation formula, but as an example, it is preferable to use a natural logarithm. In addition, “the length of the Chinese character string” in the above calculation formula indicates the total number of Chinese characters in the Chinese character string of the Chinese noun. In the example of FIG. 12, “3.502” is calculated as the log likelihood of the Chinese character “Liu”, and “3.466” is calculated as the log likelihood of the Chinese character “Kyo”. Furthermore, “3.335” is calculated as the log likelihood of the Chinese character “Li”, and “2.213” is calculated as the log likelihood of the Chinese character “Me”. Further, “2.086” is calculated as the log likelihood of the Chinese character “Sei”.

  On the other hand, the creation unit 32 executes the same processing for the Korean noun list as for the Chinese noun list. Explaining this, the creation unit 32 reads a Korean noun list from the noun list stored in the foreign language noun storage unit 32a, and converts the Korean nouns included in the Korean noun list into one kanji. Connect to the column (step SB71). In the example of FIG. 12, the creating unit 32 concatenates Korean nouns “Pyongyang”, “Kim Dae Jung”, etc. into a single Korean noun kanji string “Pyongyang Jin Dae Je ...”.

  Then, the creating unit 32 calculates the frequency at which each kanji character included in the kanji string of the Korean noun appears in the kanji string (step SB72). In the example of FIG. 12, “39 (times)” is calculated as the appearance frequency of the Chinese character “Kin”, and “37 (times)” is calculated as the appearance frequency of the Chinese character “Kama”. Further, “33 (times)” is calculated as the appearance frequency of the Chinese character “Li”, and “22 (times)” is calculated as the appearance frequency of the Chinese character “Mountain”.

  Then, the creation unit 32 calculates the log likelihood of each Chinese character (step SB73). In the example of FIG. 12, “4.126” is calculated as the log likelihood of the Chinese character “Gold”, and “3.632” is calculated as the log likelihood of the Chinese character “Kama”. Further, “3.411” is calculated as the log likelihood of the Chinese character “Li”, and “1.758” is calculated as the log likelihood of the Chinese character “Mountain”.

  Thereafter, the creation unit 32 sets the log likelihood of each Chinese character calculated from the Chinese character string of the Chinese noun as Chinese, and sets the log likelihood of each Chinese character calculated from the Chinese character string of the Korean noun as Korean. After that, the creation unit 32 collects the Chinese and Korean qualities of the same kanji and registers them in the single kanji dictionary storage unit 31 (step S74). Note that for a Chinese character that exists only in one of a Chinese noun kanji string or a Korean noun kanji string, the creation unit 32 sets a minimum value including zero, for example, “0.500” as Chinese character or Korean. After being given as narrative, it is registered in the single kanji dictionary storage unit 31.

  In this way, the single kanji dictionary shown in FIG. 11 is created by the processing of steps SA71 to SA73, steps SB71 to SB73, and step S74. As a result, as in the co-occurrence dictionary in the first embodiment, an evaluation criterion for evaluating the probability that an unknown word in Chinese characters is a foreign language can be defined. Note that the processes in steps SA71 to SA73 and the processes in steps SB71 to SB73 can be executed in parallel, or any of them can be executed first or later. .

  Here, as an example, it is assumed that an input text “Visit Busan” is input from the reception unit 11. In this case, the analysis unit 16a searches the morpheme dictionary shown in FIG. 2 for a morpheme that matches the character string included in the input text, and sets the input text to “Kama”, “Mountain”, “O”, “Visit” ”And“ Yes ”. At this time, the analysis unit 16a indicates that “Kama” and “Yama” in the character string included in the input text are actually two-letter words “Busan [Busan]”, but are registered in the morpheme dictionary. Because there are no, these morphemes are recognized as single kanji characters. After that, the analysis unit 16a gives a reading to each morpheme and reads “Kama (single kanji) [Kama], Yama (single kanji) [Yama], [o], visit [Houmon], and [sul]. The result of morphological analysis is obtained. Subsequently, since “Kama” and “Yama” are recognized as single kanji characters, the extraction unit 16 b extracts these “Busan” as unknown words in Kanji notation.

  The calculation unit 33a uses the single kanji dictionary storage unit 31 to calculate a foreign language score representing the probability that the unknown word in the kanji notation extracted by the extraction unit 16b is a foreign language. Here, the method for calculating the Chinese score is exemplified, but the method for calculating the Korean score is also the same. Explaining this, the calculation unit 33a extracts the Chinese character corresponding to the single Chinese characters constituting the unknown word in the Chinese character notation from the single Chinese character dictionary. And the calculation part 33a calculates a Chinese score by totaling the Chinese character of each single Chinese character. Then, the calculation unit 33a calculates a normalized Chinese score by dividing the Chinese score, which is the total value, by the number of characters in the unknown word. The reason for such normalization is to allow the determination unit 16d to determine which country the word is based on one threshold even if the number of characters of the unknown word is different each time the input text is received.

  In the example of the above text, the calculation unit 33a has the Chinese character “1.478” corresponding to the single kanji character “Kama” constituting the unknown word in the Chinese character notation and the Chinese character “ 1.759 "is extracted from the single kanji dictionary. Then, the calculation unit 33a adds up the Chinese character “1.478” of the single Chinese character “Ku” and the Chinese character “1.759” of the single Chinese character “Yama”, and then divides by the number of characters “2” of the unknown word “Busan”. To do. Thereby, the calculation unit 33a calculates the normalized Chinese score “1.618”. On the other hand, the calculation unit 33a calculates the Korean character “3.632” corresponding to the single kanji character “kama” that constitutes the unknown word of the kanji notation and the Korean character “1.758” corresponding to the single character “yama” to the single kanji dictionary. Extract from Then, the calculation unit 33a adds up the Chinese character “3.632” of the single Chinese character “Ku” and the Chinese character “1.758” of the single Chinese character “Yama”, and then divides by the number of characters “2” of the unknown word “Busan”. To do. Accordingly, the calculation unit 33a calculates a normalized Korean score “2.695”.

  Thus, the calculation unit 33a calculates a Chinese score “1.618” and a Korean score “2.695”. When the threshold used by the determination unit 16d is “2.000”, the Korean score> threshold> Chinese score, so the determination unit 16d determines that the unknown word “Busan” is Korean. For this reason, the reading generation unit 16e generates the reading “Pusan” for the unknown word “Busan” written in Chinese characters according to “Pug” and “San” stored in the foreign language dictionary storage unit 14b. . For this reason, it is possible to prevent “Kamayama” as a result of erroneous morphological analysis from being given to the unknown word “Busan”.

[Process flow]
FIG. 13 is a flowchart illustrating a procedure of Chinese score calculation processing according to the second embodiment. This Chinese score calculation process is a process corresponding to step S104 shown in FIG. 8, and starts when an unknown word in Chinese characters is extracted by the extraction unit 16b. Here, the case where the Chinese score is calculated among the other language scores is illustrated, but the same applies to the case where the Korean score is calculated.

  As illustrated in FIG. 13, the calculation unit 33a sets various parameters (step S501). As an example, the calculation unit 33a initializes the Chinese score and the loop counter I to zero and sets a value for the number N of unknown words.

  Then, the calculation unit 33a repeatedly executes the processes from step S503 to step S505 below until the loop counter I becomes equal to the number N of characters of the unknown word (Yes at step S502).

  That is, the calculation unit 33a extracts the Chinese character corresponding to the I-th kanji of the unknown word from the single kanji dictionary (step S503). Then, the calculation unit 33a cumulatively adds the Chinese score of the I-th Chinese character to the Chinese score that has been cumulatively added up to the “I-1” -th (step S504). Thereafter, the calculation unit 16c increments the loop counter I (step S505), and proceeds to the above step S502.

  When the loop counter I becomes equal to the number N of unknown word characters (No at Step S502), the calculation unit 33a divides the Chinese score by the number N of unknown word characters (Step S506). After normalizing the Chinese score in this way, the process ends.

[Effect of Example 2]
As described above, the speech synthesizer 30 according to the present embodiment determines whether the unknown word in kanji notation included in the text is any other language other than Japanese, based on the foreign language score. Generate word readings. Therefore, in the speech synthesizer 30 according to the present embodiment, the probabilities that an unknown word in Kanji is a foreign language is evaluated as a foreign language score, and a foreign language reading with a high foreign language score is regarded as an unknown word reading. Can be generated. For this reason, in the speech synthesizer 30 according to the present embodiment, even if the Japanese text includes other languages such as Chinese, Korean, Taiwanese, etc., the character strings in the other languages are included. It can prevent giving wrong Japanese readings. Therefore, according to the speech synthesizer 30 according to the present embodiment, as in the first embodiment, it is possible to accurately read out the other language of the Chinese character notation included in the text.

  Furthermore, the speech synthesizer 30 according to the present embodiment calculates a foreign language score using the frequency, likelihood, or probability of occurrence of a single Chinese character constituting an unknown word in Chinese characters as a foreign language word. For this reason, in the speech synthesizer 30 according to the present embodiment, it is possible to evaluate the probability that the character configuration of the unknown word is a foreign language by using the foreign language score.

  Although the embodiments related to the disclosed apparatus have been described above, the present invention may be implemented in various different forms other than the above-described embodiments. Therefore, another embodiment included in the present invention will be described below.

[Scope of application]
For example, in the first embodiment and the second embodiment described above, the case where the disclosed apparatus is implemented as the speech synthesis apparatus 10 or 30 is illustrated, but the implementation form of the disclosed apparatus is not limited to this. For example, only the function of the language processing unit 16 or 33 included in the speech synthesizer 10 or 30 can be applied as a language processing device.

  Moreover, in said Example 1 and Example 2, although the case where each other Example was implemented independently was illustrated, these Examples can be implemented combining suitably. As an example, a multilingual score calculated using a co-occurrence dictionary and a multilingual score calculated using a single kanji dictionary are subjected to statistical processing, for example, arithmetic mean or weighted average, and the other language score. It is also possible to further improve the reliability.

  In addition, each component of each illustrated apparatus does not necessarily need to be physically configured as illustrated. In other words, the specific form of distribution / integration of each device is not limited to that shown in the figure, and all or a part thereof may be functionally or physically distributed or arbitrarily distributed in arbitrary units according to various loads or usage conditions. Can be integrated and configured. For example, the reception unit 11, the creation unit 15, the language processing unit 16, the language processing unit 33, the prosody generation unit 17, or the synthesis unit 18 may be connected as an external device of the speech synthesizer 10 or 30 via a network. In addition, the reception unit 11, the creation unit 15, the language processing unit 16, the language processing unit 33, the prosody generation unit 17, or the synthesis unit 18 each have a separate device, and are connected to the network to cooperate with each other. You may make it implement | achieve the function of the synthesizing | combining apparatus 10 or 30. FIG.

[Language processing program]
The various processes described in the above embodiments can be realized by executing a prepared program on a computer such as a personal computer or a workstation. In the following, an example of a computer that executes a language processing program having the same function as that of the above embodiment will be described with reference to FIG.

  FIG. 14 is a schematic diagram illustrating an example of a computer that executes a language processing program according to the third embodiment. As illustrated in FIG. 14, the computer 100 according to the second embodiment includes an operation unit 110 a, a microphone 110 b, a speaker 110 c, a display 120, and a communication unit 130. The computer 100 further includes a CPU 150, a ROM 160, an HDD (Hard Disk Drive) 170, and a RAM (Random Access Memory) 180. These units 110 to 180 are connected via a bus 140.

  As shown in FIG. 14, the HDD 170 stores in advance a language processing program 170 that exhibits the same function as the language processing unit 16 or 33 shown in the first or second embodiment. The language processing program 170 may be integrated or separated as appropriate, similarly to each component of the language processing unit 16 or 33 shown in FIG. In other words, all data stored in the HDD 170 need not always be stored in the HDD 170, and only data necessary for processing may be stored in the HDD 170.

  Then, the CPU 150 reads the language processing program 170 a from the HDD 170 and expands it in the RAM 180. Thereby, as shown in FIG. 14, the language processing program 170a functions as a language processing process 180a. The language processing process 180a expands various data read from the HDD 170 in an area allocated to itself on the RAM 180, and executes various processes based on the expanded various data. The language processing process 180 includes, for example, processing executed by the language processing unit 16 or 33 shown in FIG. 1 or FIG. 10, for example, processing shown in FIG. 8, FIG. 9, or FIG. It should be noted that all the processing units virtually realized on the CPU 150 do not always have to operate on the CPU 150, and only the processing units necessary for the processing need only be virtually realized.

  The language processing program is not necessarily stored in the HDD 170 or the ROM 160 from the beginning. For example, each program is stored in a “portable physical medium” such as a flexible disk inserted into the computer 100, so-called FD, CD-ROM, DVD disk, magneto-optical disk, or IC card. Then, the computer 100 may acquire and execute each program from these portable physical media. In addition, each program is stored in another computer or server device connected to the computer 100 via a public line, the Internet, a LAN, a WAN, etc., and the computer 100 acquires and executes each program from these. It may be.

  The following supplementary notes are further disclosed with respect to the embodiments including the above examples.

(Supplementary Note 1) An analysis unit that performs reading of each morpheme after dividing the sentence into morphemes by performing morpheme analysis on the sentence including kanji using a predetermined dictionary;
An extraction unit for extracting an unknown word in kanji from the result of morphological analysis by the analysis unit;
A calculation unit that calculates a foreign language score representing the probability that the unknown word of the kanji notation extracted by the extraction unit is a different foreign language from the national language listed in the dictionary;
Based on the other language score calculated by the calculation unit, a determination unit that determines which country the unknown word of the Kanji notation is,
A language processing apparatus comprising: a reading generation unit configured to generate a reading of an unknown word in the Chinese character notation according to a determination result by the determination unit.

(Additional remark 2) The said calculation part calculates the said foreign language score using the co-occurrence relationship between the morpheme other than the said unknown word contained in the said sentence, and the word of a foreign language. The language processing apparatus according to 1.

(Additional remark 3) The said calculation part calculates the said foreign language score using the frequency, likelihood, or probability that the single Chinese character which comprises the unknown word of the said Chinese character notation appears as a word of another language. The language processing apparatus according to attachment 1.

(Additional remark 4) The provision part which provides an accent to the reading of the unknown word of the Chinese character notation produced | generated by the said reading production | generation part,
Additional notes 1, 2, or further comprising: a phonetic generation unit that generates a phonetic character string of the sentence from a result of a morphological analysis in which an accent is added to reading of an unknown word written in Chinese characters by the adding unit 4. The language processing device according to 3.

(Additional remark 5) The analysis part which gives a reading to each morpheme after dividing the above-mentioned sentence into a morpheme by performing a morpheme analysis to a sentence containing Kanji using a predetermined dictionary,
An extraction unit for extracting an unknown word in kanji from the result of morphological analysis by the analysis unit;
A calculation unit that calculates a foreign language score representing the probability that the unknown word of the kanji notation extracted by the extraction unit is a different foreign language from the national language listed in the dictionary;
Based on the other language score calculated by the calculation unit, a determination unit that determines which country the unknown word of the Kanji notation is,
According to the determination result by the determination unit, a reading generation unit that generates a reading of an unknown word in the kanji notation,
An assigning unit that gives an accent to the reading of an unknown word in kanji notation generated by the reading generation unit;
A phonetic generation unit that generates a phonetic character string of the sentence from a result of morpheme analysis in which an accent is added to the reading of an unknown word written in Chinese characters by the adding unit;
A prosody generation unit that generates a prosody corresponding to the sentence based on the phonetic character string generated by the phonetic generation unit;
A speech synthesis apparatus comprising: a synthesis unit that generates a speech waveform from the prosody generated by the prosody generation unit and synthesizes speech.

(Appendix 6)
By performing morphological analysis on sentences containing kanji using a predetermined dictionary, the sentence is divided into morphemes and then given readings to each morpheme,
Extract unknown words in kanji from the result of the morphological analysis,
Calculating a foreign language score representing the certainty that the unknown word in the kanji notation is different from the national language listed in the dictionary;
Based on the other language score, determine which country the unknown word of the kanji notation is,
The language processing method characterized by performing the process which produces | generates the reading of the unknown word of the said Chinese character notation according to the said discrimination | determination result.

(Appendix 7)
By performing morphological analysis on sentences containing kanji using a predetermined dictionary, the sentence is divided into morphemes and then given readings to each morpheme,
Extract unknown words in kanji from the result of the morphological analysis,
Calculating a foreign language score representing the certainty that the unknown word in the kanji notation is different from the national language listed in the dictionary;
Based on the other language score, determine which country the unknown word of the kanji notation is,
A language processing program that executes a process of generating an unknown word reading of the kanji notation according to the discrimination result.

(Appendix 8) The computer
Using a predetermined dictionary, each sentence is divided into morphemes by performing morphological analysis on each of the sentences containing kanji,
Of the sentences divided into morphemes, extract sentences containing nouns in other languages that are different from the national languages listed in the dictionary and are composed of kanji,
The frequency of the first morpheme of the sentence including the foreign language noun and the morpheme immediately after the foreign language noun appears in the plurality of sentences, the foreign language noun and the foreign language noun Calculating the frequency with which the second morpheme, excluding the word immediately after, appears in the plurality of sentences,
Using a frequency at which the first morpheme appears, when a word that has a co-occurrence relationship with the first morpheme appears, a first log likelihood representing the probability that the word is the other language is calculated And
Using the frequency of appearance of the second morpheme, when a word that has a co-occurrence relationship with the second morpheme appears, a second log likelihood representing the probability that the word is the other language is calculated And
A co-occurrence dictionary of the first morpheme and the second morpheme is created by associating the first morpheme and the first log likelihood and associating the second morpheme and the second log likelihood. A method for creating a co-occurrence dictionary characterized by executing processing.

(Appendix 9) The computer
As the process of calculating the frequency, a process of calculating a frequency at which the morpheme appears in the plurality of sentences is performed on a morpheme that is a noun of the first morpheme or the second morpheme. The method for creating the co-occurrence dictionary according to appendix 8.

(Appendix 10) The computer
The weighting process which gives a weight larger than said 1st log likelihood to said 2nd log likelihood among said 1st log likelihood and said 2nd log likelihood is performed, It is characterized by the above-mentioned. 8. A method for creating a co-occurrence dictionary according to 8 or appendix 9.

(Appendix 11) The computer
Acquire multiple nouns in other languages that are different from the country that uses kanji-kana mixed sentences as the national language,
Concatenate multiple foreign language nouns into one kanji string,
For each single kanji character included in the concatenated kanji character string, calculate the frequency at which the single kanji character appears in the kanji character string,
Using the frequency at which the single kanji appears, calculate a log likelihood representing the probability that the single kanji is the other language,
A method for creating a single Chinese character dictionary, comprising: executing a process of creating a dictionary for the single Chinese character by associating the single Chinese character and the log likelihood.

DESCRIPTION OF SYMBOLS 10 Speech synthesizer 11 Reception part 12 Morphological dictionary storage part 13 Cooccurrence dictionary storage part 14a Japanese dictionary storage part 14b Other language dictionary storage part 15a Corpus storage part 15 Creation part 16 Language processing part 16a Analysis part 16b Extraction part 16c Calculation part 16d Discriminator 16e Reading generator 16f Applier 16g Phonetic generator 17 Prosody generator 18 Synthesizer 19 Output unit

Claims (7)

An analysis unit that gives a reading to each morpheme after dividing the sentence into morphemes by performing morphological analysis on sentences containing kanji using a predetermined dictionary;
An extraction unit for extracting an unknown word in kanji from the result of morphological analysis by the analysis unit;
A calculation unit that calculates a foreign language score representing the probability that the unknown word of the kanji notation extracted by the extraction unit is a different foreign language from the national language listed in the dictionary;
Based on the other language score calculated by the calculation unit, a determination unit that determines which country the unknown word of the Kanji notation is,
A language processing apparatus comprising: a reading generation unit configured to generate a reading of an unknown word in the Chinese character notation according to a determination result by the determination unit.   The said calculation part calculates the said foreign language score using the co-occurrence relationship between the morpheme other than the said unknown word contained in the said sentence, and the word of another language. Language processor.   The said calculation part calculates the said foreign language score using the frequency, likelihood, or probability that the single Chinese character which comprises the unknown word of the said Chinese character notation appears as a word of another language. The language processing device described. An assigning unit that gives an accent to the reading of an unknown word in kanji notation generated by the reading generation unit;
A phonetic generation unit that generates a phonetic character string of the sentence from a result of a morphological analysis in which an accent is added to reading of an unknown word written in Chinese characters by the adding unit. Or the language processing apparatus of 3. An analysis unit that gives a reading to each morpheme after dividing the sentence into morphemes by performing morphological analysis on sentences containing kanji using a predetermined dictionary;
An extraction unit for extracting an unknown word in kanji from the result of morphological analysis by the analysis unit;
A calculation unit that calculates a foreign language score representing the probability that the unknown word of the kanji notation extracted by the extraction unit is a different foreign language from the national language listed in the dictionary;
Based on the other language score calculated by the calculation unit, a determination unit that determines which country the unknown word of the Kanji notation is,
According to the determination result by the determination unit, a reading generation unit that generates a reading of an unknown word in the kanji notation,
An assigning unit that gives an accent to the reading of an unknown word in kanji notation generated by the reading generation unit;
A phonetic generation unit that generates a phonetic character string of the sentence from a result of morpheme analysis in which an accent is added to the reading of an unknown word written in Chinese characters by the adding unit;
A prosody generation unit that generates a prosody corresponding to the sentence based on the phonetic character string generated by the phonetic generation unit;
A speech synthesis apparatus comprising: a synthesis unit that generates a speech waveform from the prosody generated by the prosody generation unit and synthesizes speech. Computer
By performing morphological analysis on sentences containing kanji using a predetermined dictionary, the sentence is divided into morphemes and then given readings to each morpheme,
Extract unknown words in kanji from the result of the morphological analysis,
Calculating a foreign language score representing the certainty that the unknown word in the kanji notation is different from the national language listed in the dictionary;
Based on the other language score, determine which country the unknown word of the kanji notation is,
The language processing method characterized by performing the process which produces | generates the reading of the unknown word of the said Chinese character notation according to the said discrimination | determination result. On the computer,
By performing morphological analysis on sentences containing kanji using a predetermined dictionary, the sentence is divided into morphemes and then given readings to each morpheme,
Extract unknown words in kanji from the result of the morphological analysis,
Calculating a foreign language score representing the certainty that the unknown word in the kanji notation is different from the national language listed in the dictionary;
Based on the other language score, determine which country the unknown word of the kanji notation is,
A language processing program that executes a process of generating an unknown word reading of the kanji notation according to the discrimination result.

Images