JP3446342B2 - Natural language processing method and speech synthesizer - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a natural language processing method suitable for use in natural language processing of Japanese texts and the like, and speech synthesis based on the information obtained as a result, and speech synthesis. Regarding the device.

[0002]

2. Description of the Related Art In a conventional speech synthesizer, phonological information and prosody information are obtained by, for example, natural language processing of a sentence containing kanji and kana characters, and the input kanji and kana characters are mixed based on the phonological information and prosody information. It is designed to synthesize a voice corresponding to a sentence. In this case, in order to bring the synthetic speech closer to human speech, a pose to be inserted into a sentence containing kanji and kana is usually requested as one piece of prosody information.

In the conventional speech synthesizer, the pose is set at the position of a punctuation mark in a sentence containing kanji and kana, for example.

[0004]

Therefore, in the past, since the insertion position of the pose was set without sufficiently considering the grammatical rules, there is a problem that the obtained synthetic sound becomes unnatural. It was

Then, for example, "prosodic rules for synthesis of Japanese sentence speech" (Kawai, Hirose, Fujisaki: The Acoustical Society of Japan,
Vol.50, No.6, pp.433-442, 1994), a pose setting method based on a dependency relationship is also proposed. However, this method mainly requires semantic information, and since a general-purpose technology for processing such semantic information in natural language has not been established, only limited text is targeted for natural language processing. It was difficult to perform natural language processing on general-purpose texts that do not have any. That is, with this method, it was difficult to set the pose insertion position for all input sentences so that the synthesized sound could be heard naturally.

Further, in the past, when the number of beats (corresponding to the unit of kana characters) (the number of beats) (the number of mora) (the number of sounding units corresponding to the kana notation) between inserted poses is too large, Since the pose was set again at half the total number of beats without taking grammatical information into account, the position of the pose was sometimes unnatural.

Therefore, conventionally, there has been a problem that it is difficult to hear the synthesized sound or it is difficult to understand its contents.

The present invention has been made in view of such a situation, and makes it possible to set a pose at an appropriate position in a sentence, thereby making it possible to obtain, for example, a natural synthesized voice. Is.

[0009]

A natural language processing method according to the present invention morphologically analyzes an input sentence in Japanese, and from the morphological analysis result of the input sentence, finds a grammatical relationship between morphemes forming a compound word. A compound word is identified based on the defined compound word rule, and from the identification result of the compound word, the phrase is identified based on the phrase rule that defines the grammatical relationship between the morphemes forming the phrase, , A bunsetsu consisting of one or more morphemes, which forms an input sentence, is identified statistically based on the bunsetsu rule that defines the grammatical relationship between the morphemes forming the bunsetsu. In addition, it is a natural language processing method that applies a pose setting rule that defines the position of a pose to be inserted in a sentence and sets the position of a pose to be inserted between consecutive clauses, which is a compound word rule, a clause rule, or a consecutive clause rule. , The pose setting rule and the join condition include the join condition for joining the phonemes and the additional information to be added to the compound word, the clause, or the consecutive phrase obtained by joining the morphemes satisfying the join condition. Accordingly, it is characterized by being described using additional information.

In this natural language processing method, the conjunctive clause rule defines a relational modification rule that defines the relationship between the morphemes that form the continuation modified conjunctive clause, and a adnominal modification rule that defines the relationship between the morphemes that form the adnominal modified conjunctive clause. In the case of including the phrase, it is possible to identify the linked modified linked sentence or the linked modified linked phrase based on the linked modification rule or the linked modification rule from the bunsetsu identification result.

When the morpheme analysis result includes at least a morpheme character string and part-of-speech information, the compound word rule, bunsetsu rule, continuous bunsetsu rule, or pause setting rule can be described using the morpheme analysis result. Part of speech information includes
In addition to the part-of-speech, it is possible to include a higher-level classification that integrates the part-of-speech and a lower-level classification that subdivides the part-of-speech. Further, the sub-class can include at least the inflectional form.

[0012] The additional information can be a high-level classification in which the morpheme's part-of-speech is integrated or a sub-classification in which the morpheme's part-of-speech is subdivided.

If the compound word rule, the clause rule, or the continuous clause rule further describes a separation condition for separating two combined morphemes, the combined morphemes can be separated according to the separation condition.

When a compound word rule is applied between a compound word composed of a plurality of morphemes and an adjacent morpheme which is an adjacent morpheme, it is adjacent to an adjacent morpheme among a plurality of morphemes forming the compound word. Compound word rules can be applied between morphemes and adjacent morphemes. Further, when a bunsetsu rule is applied between a bunsetsu consisting of a plurality of morphemes and an adjacent morpheme which is a morpheme adjacent to the bunsetsu, a morpheme adjacent to the adjacent morpheme among the plurality of morphemes forming the bunsetsu, Phrase rules can be applied between morphemes. Furthermore, when a consecutive phrase clause or a pause setting rule is applied between a conjunctive phrase composed of a plurality of morphemes and an adjacent morpheme which is an adjacent morpheme, it is adjacent to an adjacent morpheme among a plurality of morphemes forming the consecutive phrase. A consecutive clause rule or a pause setting rule can be applied between a morpheme and an adjacent morpheme.

In the pause setting rule, the pause is set between the last morpheme of the morphemes forming the preceding consecutive bunsetsu and the head morpheme of the morphemes forming the subsequent consecutive bunsetsu in two adjacent consecutive bunsetsus. The insertion frequency can be statistically obtained from the reading result of the actual sentence and can be created. Furthermore, the pose setting rule is a priority rule in which a setting condition for setting a pose between the last morpheme and the first morpheme is described with a priority representing a priority order for setting a pause in between, When the narrowing conditions for narrowing down the pose setting position are included, and the narrowing rule that describes the position is determined, a candidate for the position to set the pose is determined based on the priority rule, and based on the narrowing rule based on the candidates. , You can determine the final position to set the pose.

The speech synthesizer of the present invention performs a natural language processing on an input sentence in Japanese to generate phonological information and prosody information corresponding to the input sentence (for example, the language processing unit shown in FIG. 1). 1 and prosody processing unit 8)
And a synthesizing unit (for example, the acoustic processing unit 11 shown in FIG. 1) that synthesizes a voice corresponding to the input sentence based on the phoneme information and the prosody information, and the prosody information is a pause inserted in the input sentence. The position of the pose is obtained by the natural language processing method according to any one of claims 1 to 12.

[0017]

In the natural language processing method of the present invention, the input sentence is morphologically analyzed, and based on the morphological analysis result, based on the compound word rule, the bunsetsu rule, and the continuous bunsetsu rule that define the grammatical relationship between the morphemes, Compound words, clauses, and consecutive clauses are identified. Then, the position of the pose to be inserted between the consecutive clauses is set by sequentially applying the pose setting rules statistically obtained to the identification result. A compound word rule, a bunsetsu rule, or a continuous bunsetsu rule includes a join condition for joining morphemes, and additional information added to the compound word, bunsetsu, or consecutive bunsetsu obtained by joining morphemes satisfying the join condition. , Pose setting rules and connection conditions are described using additional information as necessary. Therefore, by combining morphemes, additional information can be given to the entire combined morpheme, and the pose setting rule and the combining condition are described using the additional information. It is possible to set the pose according to.
That is, it is possible to insert a pause at an appropriate position in the input sentence without particularly limiting the input sentence.

In the speech synthesizer of the present invention, the Japanese input sentence is subjected to natural language processing to obtain phonological information and prosodic information corresponding to the input sentence, and based on the phonological information and prosodic information. , The voice corresponding to the input sentence is synthesized. In this case, the position of the pause inserted in the input sentence, which is one of the prosody information, is obtained by the natural language processing method according to any one of claims 1 to 12. Therefore, it is possible to obtain a synthetic sound that is natural and easy to understand without particularly limiting the input sentence.

[0019]

1 is a block diagram showing the configuration of an embodiment of a speech synthesizer to which the present invention is applied. In this speech synthesizer, a synthesized voice corresponding to a Japanese sentence, for example, a mixture of kanji and kana (for example, composed of text data) or the like (hereinafter, simply referred to as an input sentence) is generated.

This speech synthesizer is roughly divided into a computing unit 1 and a memory unit 2. The arithmetic unit 1 is composed of a language processing unit 3, a prosody processing unit 8, and a sound processing unit 11, which are three basic processing units, and the memory unit 2 is a dictionary (a Kanji dictionary or the like) used in the arithmetic unit 1. , Rules (morphological analysis rules, connection rules, pause setting rules, parameter generation rules) and data (prosodic control model, phoneme piece data) are stored.

The language processing unit 3 includes a morphological analysis unit 4, a combination processing unit 5, a pose setting processing unit 6, and a phonetic symbol generation unit 7.
The input sentence is processed by natural language, and the phoneme information (reading (pronunciation) of the input sentence) and the prosodic information (for example, prosodic phrase, accent phrase (accent type), and pose) of a voice (synthetic sound) to be synthesized are synthesized. Position, its length, and other information) are generated (extracted).

The prosody processing unit 8 comprises a prosody control model parameter generation unit 9 and a prosody data generation unit 10. The prosody control model parameter generation unit (hereinafter referred to as parameter generation unit) 9 is the language processing unit 3. Parameters (model control parameters) for driving a model for controlling the prosodic features of the synthetic sound are generated based on the generated phonological information and prosody information according to the parameter generation rules stored in the memory device 2. It is done like this.

The prosody data generation unit 10 uses various parameters (prosody control models) stored in the memory device 2 for controlling the prosody characteristics of the synthesized voice, using the parameters generated by the parameter generation unit 9. It is adapted to drive and generate specific numerical values (numerical values calculated from a model) representing the prosodic characteristics of the synthetic sound, that is, prosody data.

The acoustic processing unit 11 performs voice synthesis based on the phoneme information or the prosody data generated by the language processing unit 3 or the prosody data generation unit 10, respectively, and synthesizes the resultant synthesized sound (not shown). It is designed to be sent to an audio section and output from a built-in speaker, for example.

The information generated by the language processing unit 3 (phonological prosody information described later) is not only the parameter generation unit 9 in the subsequent stage but also the prosody data generation unit 10 and the acoustic processing unit 11 via the parameter generation unit 9. Is also being supplied to. This is because the phonological information included in the phonological prosody information is used in the processes of the prosody data generation unit 10 and the acoustic processing unit 11.

Next, the operation will be described. When the input sentence is input to the language processing unit 3, the input sentence is subjected to natural language processing.

That is, first, the morpheme analysis unit 4 morphologically analyzes the input sentence using a dictionary and morpheme analysis rules, decomposes it into morphemes, and further determines its pronunciation (reading), part-of-speech information, and beat number. It

Further, the morpheme analysis unit 4 adds an accent (an accent when a word is uttered by only one (hereinafter, appropriately referred to as a basic accent)) to each morpheme analyzed. And, if necessary,
Processing such as vowel lengthening and vowel devoicing is performed. Then, the processing result is supplied to the combination processing unit 5.

In the combination processing unit 5, the morphemes forming the input sentence are combined based on the combination rule, and the combination result is separated as necessary, whereby the continuous sentence clause (phrase) (one piece) of the input sentence is obtained. At the same time that the unit consisting of the above clauses (and therefore also the unit consisting of one or more morphemes) is identified, the position at which a pose is set between consecutive clauses is identified.

Then, the pose setting processing section 6 sets a pose between the identified consecutive clauses based on the pose setting rule. Note that the pose setting processing unit 6 sets a pose and a priority (to be described later) for the pose, and then takes into consideration the beat count between the poses to determine the final pose position (hereinafter, referred to as appropriate). , Pose information) is set.

After setting the pose information, from the setting result,
The phonetic symbol generator 7 obtains the structure of the input sentence,
Based on the structure of this sentence, prosodic phrases and other information are obtained. Furthermore, in the phonetic symbol generation unit 7, from the processing result of the pose setting processing unit 6 and the morphological analysis result,
Information in which phonological information and prosodic information are expressed by symbols or characters (since this information includes phonological information and prosodic information, hereinafter, referred to as phonological prosodic information) is generated.

When morphemes (words) are connected to form a sentence, the position of the morpheme's accent may change depending on whether the morpheme is uttered in the sentence or the morpheme is uttered independently. Therefore, the phonetic symbol generation unit 7 determines the position of the accent in the phrase, that is, the accent phrase, based on, for example, a predetermined accent movement rule when the phonological prosody information is generated.

As described above, the language processing unit 3 reads the input sentence (phonological information) and the prosodic phrase, accent phrase, which is necessary for controlling the prosodic features of the synthesized speech corresponding to the input sentence. Phonological prosody information, which is prosody information including other prosody information, is generated and output to the parameter generation unit 9.

In the parameter generation unit 9, based on the phonological prosody information from the language processing unit 3 and the parameter generation rule stored in the memory device 2, for example, the fundamental frequency (pitch frequency) of the synthesized voice, the phoneme duration, And a parameter for driving a prosody control model stored in the memory device 2, which controls prosody features such as power (for example, power in units of phonemes).

That is, a so-called Fujisaki model, for example, is stored in the memory device 2 as a prosody control model for obtaining the fundamental frequency, and a so-called quantification is used as a prosody control model for obtaining the duration and power. When the model according to the first class is stored, the parameter generation unit 9 sets, for example, the size of the phrase command or the accent command and the position thereof (at the time of issuing the phrase command or the accent command) as the parameters for driving the Fujisaki model. Then, as a parameter for driving the model according to the quantification type 1, a category of factors (for example, a phoneme to be speech-synthesized, phonemes before and after it) is set. Here, the pose information set by the pose setting processor 6 is used to determine the position of the phrase command in the parameter generator 9.

As the prosody control model, other than the above-mentioned ones, it is possible to use other conventionally known ones. However, in the parameter generation unit 9, the prosody stored in the memory device 2 is used. It is necessary to generate the parameters corresponding to the control model.

In the prosody data generator 10, the memory device 2
As the prosody control model stored in, for example, a Fujisaki model or a model according to the quantification type 1 is driven by the parameters generated (set) by the parameter generation unit 9, and thereby prosody data is output. That is, when the Fujisaki model is driven, the fundamental frequency or the like as the prosody data is output, and when the model according to the quantification type 1 is driven, the power and the duration time or the like as the prosody data are output. .

Concretely, as the prosody data, the fundamental frequency value for each phrase of the synthetic sound, the number of frames indicating the duration of each phoneme, the number of frames indicating the power value of each syllable, and the like are included in the acoustic processing unit 11. Is output to.

The sound processing section 11 is composed of, for example, a synthesizer for performing speech synthesis by a method similar to the conventional one, that is, for example, a speech synthesizer by waveform segment connection, a so-called formant synthesizer, or the like. The sound processing unit 11 is, for example, a voice synthesizer by waveform segment connection, and is stored in the memory device 2, for example, C
When speech unit data (eg, digitized waveform data) necessary for regular speech synthesis is stored in units of V, CVC / VCV, etc., the sound processing unit 11 is generated by the language processing unit 3. Phonological information of phonological prosody information,
Based on the prosody data (specific numerical values of the fundamental frequency, duration, and power time, etc.) output from the prosody data generation unit 10, necessary speech segment data (speech segment data) is converted into a continuous speech waveform. To be connected. Further, the acoustic processing unit 11 inserts a pause into the voice waveform based on the pause information in the phonological prosody information generated by the language processing unit 3. Then, in the synthesized voice unit, the voice waveform is subjected to necessary processing such as D / A conversion processing, and output from the built-in speaker.

As described above, the synthesized voice corresponding to the input sentence is output.

Next, the details of the processing in the morphological analysis unit 4 constituting the language processing unit 3 will be described.

In the morpheme analysis unit 4, as described above, the input sentence is decomposed into morphemes corresponding to words based on the dictionary and the morpheme analysis rule, and further, its pronunciation (reading), accent, part-of-speech information, and beat. The number etc. is determined. The part-of-speech information of these is given to each morpheme by referring to the part-of-speech information (including the inflectional form) included in the morpheme analysis rule, for example, as shown in FIGS.

The part-of-speech information (including inflectional forms) shown in FIGS. 2 to 4 is also included in the combination rule and the pose setting rule. Also, the part-of-speech information mentioned here (including inflectional forms)
In addition to so-called part-of-speech, for example, "number", "alphabet", and other symbols are also included (hereinafter, these are also treated as part-of-speech). Further, the part-of-speech information (including the inflectional forms) includes upper-class classifications in which part-of-speech is integrated and sub-classifications in which part-of-speech is subdivided, and also includes inflectional forms. Here, in the following, in order to distinguish the part-of-speech information (inflectional form) shown in FIGS. 2 to 4 from the part-of-speech information given to the morpheme, it is referred to as a part-of-speech information dictionary.

In the part-of-speech information dictionary, part-of-speech information is classified and registered in a hierarchical structure with "all parts-of-speech" as the highest hierarchy (highest classification).

That is, for example, the subordinate classification "Ka line five-stage verb" in FIG. 2 is included in the higher-level part of speech "verb", and the part-of-speech "verb" is included in the higher classification "defective". Further, the "defective" is included in a higher-level classification called "independent word", and the "independent word" is included in "all parts of speech" in the highest classification.

Further, for example, the "proper noun" in FIG. 2 is included in a higher part-of-speech called "noun" (in this case, "noun" is a superclass of "proper noun"), and "noun" is It is included in the higher-level category called "independent word". The "independent word" is included in the "total part of speech" of the highest classification as described above.

Further, the part-of-speech information output from the morphological analysis unit 4 as the morphological analysis result includes the inflectional form as necessary. That is, the morpheme analysis result is output as, for example, "auxiliary verb / conjunctive form". As shown in FIG. 4, the inflectional form is defined separately from the hierarchical structure at the end of the part-of-speech information dictionary.

Since the inflectional form represents the utilization of a part of speech (however, there is a part of speech in which a verb or auxiliary verb is utilized), the part of speech can be subdivided according to the inflectional form. Therefore, the inflectional form is treated as a subclass.

Although not shown in FIGS. 2 to 4, the morpheme itself can be described as the lowest classification of each part of speech if necessary. That is, for example, it is possible to describe a specific numerical value as a further subclass of "numerical" which is a subclass of "noun".

By setting the part-of-speech information in such a hierarchical structure, an efficient description can be made when describing various rules. That is, for example, the rules for all of the numbers, alphabets, hiragana, katakana, Greek letters, Russian letters, and single kanji shown in FIG. 3 are described. If you do, you don't have to describe all of them, but just the higher part of speech, "single kanji".

As described above, by describing the morpheme itself, for example, a concrete numerical value as the lowest classification, there is a certain principle for "numerical words" excluding the concrete numerical value. It is easy to apply such rules and apply exceptional rules only to the specific numerical values. In this case, the rules can be described in detail, so that the accuracy of the processing can be improved (a pose corresponding to the actual utterance can be set).

Next, FIG. 5 shows the morphological analysis result of the morphological analysis unit 4 when, for example, the input sentence "A man was pressed down by a clerk last night and arrested by Marunouchi police for suspicion of theft." ing. In the morphological analysis unit 4,
The input sentence is analyzed into morphemes, and with reference to the part-of-speech information dictionary and the like shown in FIGS. 2 to 4 (for example, morpheme analysis rules), part-of-speech information and the like are added to each morpheme, but in FIG. Only the items related to the present invention are shown.

That is, in FIG. 5, the morpheme, its pronunciation, and the type of accent included in the input sentence "A man was held down by a clerk last night and arrested by the Marunouchi police for suspicion of theft."
The part-of-speech information and the number of beats are sequentially shown from the leftmost column to the right.

Here, the pronunciation (reading) of the morpheme is shown in hiragana as a general rule, but the long sound is shown by "-" and the nasal voice is shown by "o". The accent nucleus is the accent nucleus (in Japanese, the utterance level drops from high level to low level immediately after the accented beat, but this falling part is the accent nucleus). It is indicated by an accent type indicating whether or not However, those with no accent nucleus are of type 0. Further, no accent is attached to the attached words (for example, particles and punctuation marks), and a "*" mark is shown in the column of the accent.

Further, the part-of-speech information in FIG. 5 is the part-of-speech information described in the part-of-speech information dictionary shown in FIGS. 2 to 4. In addition, in the part-of-speech information in FIG.
Subsequent classifications are shown below /, and below. That is, for example, "noun (adverb usage)" in the first line of the part-of-speech information indicates that the part-of-speech of the morpheme is "noun" and the usage is adverbial. However,
The part-of-speech information of the morpheme "Yesterday" is actually "noun / common noun (adverb usage)" from FIG. 2, but in FIG. 5, the description becomes complicated, so it is referred to as "noun (adverb usage)". is there.

Further, for example, "auxiliary verb / continuous form" in the 18th line of the part-of-speech information indicates that the form and the part-of-speech is "auxiliary verb" and its utilization is "continuous form".

The number of beats of each morpheme is shown as a number in principle, but the number of beats of punctuation marks is not given, and is shown as *.

Next, the details of the combination processing unit 5 that processes the above morphological analysis results and the details of the pose setting processing unit 6 in the subsequent stage will be described.

FIG. 6 and FIG. 7 show examples of combining rules used in the combining processing section 5.

This combining rule is, for example, (1) to (9)
9 items are described as one unit. The pose setting rule referred to by the pose setting processing unit 6 is (1)
The condition consisting of 8 items except (9) of (9) to (9) is described as one unit (the pose setting rule will be described later).

That is, in the combination rule, from left end to right, (1) condition number, (2) previous morpheme heading,
(3) Part of speech of previous morpheme, (4) Subclassification of previous morpheme, (5) Utilization of previous morpheme, (6) Heading of morpheme after, (7) Part of speech of morpheme after, (8) After The 9 categories of the morpheme subclass and (9) flag information are described as one unit. Note that it is not always necessary to describe all nine items, and only the necessary items among them may be described. In FIG. 6 and FIG. 7 (the same applies to the pose setting rule of FIG. 8 described later), the item “Don't Care” is NULL.

Among the above 9 items, (2) to (8)
Indicates that two adjacent morphemes in the input sentence are compound words, bunsetsu, continuous bunsetsu, and
Is considered to be a condition (joining condition) that must be met when it is grammatically recognized (or two adjacent morphemes in the input sentence are compound words, bunsetsus, consecutive bunsetsus, and (Conditions that should be met when the grammatical is not allowed to configure (division conditions) has been).

The combination rule (the same applies to the pose setting rule), which is a set of conditions as described above, is the result of morphological analysis described in FIG. 5 (character string representing a morpheme (morpheme character string), accent, part-of-speech information ( (Including inflectional forms), beats, etc.). Note that in the combination rules shown in FIGS. 6 and 7 (the same applies to the pose setting rule (pause priority rule) shown in FIG. 8), the morpheme character string and part-of-speech information (including inflectional forms) in the morpheme analysis result are It is used. In addition, (2) to (8) (the same applies to the pause setting rule) among the nine items that form one unit of the combining rule,
It is also described using the flag information of (9).

In the condition number of (1), for example, unique (unique) numbers are described in ascending order.
The number of this condition must be described. Also, the numbers do not have to be continuous and may be discontinuous.

In the morpheme heading before (2), a morpheme character string representing the previous (preceding) morpheme of two adjacent morphemes is described, and in the morpheme heading after (6),
A character string representing a later morpheme is described. That is, in (2) and (6), the morpheme shown in the leftmost column of the morpheme analysis result shown in FIG. 5 is described. So, for example, (2) is "Marunouchi," (6) is "police,"
Each is described.

In the part-of-speech of the previous morpheme of (3), the part-of-speech of the previous morpheme of two adjacent morphemes is described.
The part of speech of the morpheme after (7) describes the part of speech of the subsequent morpheme. The part-of-speech of a morpheme here is a subclass of the part-of-speech information column in the morpheme analysis result shown in FIG.
In other words, the usages such as "(adverb usage)" enclosed in parentheses and the parts (parts of speech and subclasses below) excluding inflectional shapes (inflections) such as "/" It Therefore, for example, (3) is "noun / sahen",
(7) is described as "one-stage verb".

The sub-classification of the previous morpheme in (4) is the usage of the sub-classification of the previous morpheme of the two adjacent morphemes (the part enclosed in parentheses in the part-of-speech information in FIG. 5). ) Is described, and in the subclass of morpheme after (8),
The usages of the latter subclasses of morphemes are described. Note that nouns have many subclasses (usages) such as "(independent usage)" and "(formal noun usage)" in addition to "(adverb usage)" shown in FIG. The illustration is omitted. Therefore, for example, (4) is described as “(adverb usage)” and (8) is described as “(independent usage)”.

In (5) the utilization of the previous morpheme, the utilization (inflectional form) of the subclass of the previous morpheme of the two adjacent morphemes is described. The use of morphemes here means "... form" such as, for example, "preformed form" after / in the part-of-speech information in the morpheme analysis result shown in FIG.
Means the part that is said to be. Therefore, for example, (5)
Is described as "preformed".

In the flag information of (9), information (additional information) to be added to a compound word, phrase, or continuous phrase obtained by combining morphemes satisfying the conditions of (2) to (8) is described. (Details will be described later).

The associative rules shown in FIGS. 6 and 7 are compound word rules that define the grammatical relationship between the morphemes that make up a compound word (conditions that two morphemes that make up a compound word must satisfy grammatically. ), A bunsetsu rule that defines the grammatical relationship between the morphemes forming the bunsetsu (conditions that the two morphemes forming the bunsetsu must be grammatically satisfied), and a grammatical relationship between the morphemes forming the consecutive modified consecutive bunsetsu. Adjacent modified adjunct clause rules (conditions that the two morphemes that compose adjunct modified conjunctive clauses must be grammatically satisfied), and adnominal modified adjunct clause rules that define the grammatical relationship between the morphemes that constitute adjunct modified adjunct clauses (adjective modifier) The two morphemes that make up the conjunctive clause include conditions that must be satisfied grammatically.

In the combination processing unit 5, it is determined whether or not the two adjacent morphemes match the compound word rule, the bunsetsu rule, the consecutive modified consecutive bunsetsu rule, or the adnominal modified consecutive bunsetsu rule, and it is determined that they match the respective rules. 2 of them if
The two morphemes are combined to form a compound word, bunsetsu, consecutive modified bunsetsu, or adnominal modified bunsetsu.

Here, as described above, there are cases where a combination of two morphemes based on a combination rule does not form a compound word, a clause, a consecutive modified consecutive clause, or an adnominal modified consecutive clause. Therefore, in order to deal with such an exceptional case, a compound word rule, a bunsetsu rule, a continuous modified bunsetsu rule,
Alternatively, the adnominal modified continuous clause rule includes a division rule (division condition) for dividing (separating) two joined morphemes.

Here, hereinafter, the rules for combining morphemes as a compound word, a bunsetsu, a consecutive modified consecutive bunsetsu, or an adjunct modified consecutive bunsetsu will be referred to as a compound word joining rule, a bunsetsu joining rule,
These are called the modified binding rules for continuous use or the modified binding rules for adnominal bonds, respectively. Further, a rule for dividing a morpheme that is combined as a compound word, a bunsetsu, a linked modified linked bunsetsu, or an adornment modified linked bunsetsu is referred to as a compound word division rule, a bunsetsu division rule, a linked modification division rule, or an adjunct modification division rule.

Therefore, the compound word rule, the bunsetsu rule, the continuous modified bunsetsu rule, or the union modified continuous bunsetsu rule is the compound word combination rule and the compound word division rule, the bunsetsu combination rule and the bunsetsu division rule, the continuous modification tie rule and the continuous modification split. It can be said that each of them is composed of a rule, or a adnominal modification combining rule and an adnominal modification division rule.

Of the conditions (rules) shown in FIGS. 6 and 7, the condition number of (1) is 1000, 2000, 3000, 4000, 500.
The 0, 6000, 7000, and 8000 series are compound word combination rules, compound word division rules, bunsetsu combination rules, bunsetsu division rules, tandem qualification join rules, tandem qualification split rules, tandem qualification join rules, or tandem modifier splits, respectively. Equivalent to the rules. 6 and 7, conditions (rules) in the 4000 and 6000 series, that is, the bunsetsu division rule and the continuous modification division rule are not shown.

Although the combination processing section 5 is actually a program, by combining the rules as described above,
It is possible to easily specify the rule to be used for each module of the program that performs processing according to each rule. That is, for each module, the rule to be used can be specified by the number of the condition (1) (such as 1000 series or 2000 series).

Next, FIG. 8 shows an example of a pose setting rule used in the pose setting processing section 6. The pause setting rules include statistics (frequency statistics) when poses are inserted between consecutive clauses (and thus also between morphemes) included in many sample sentences (such as sentences written in newspapers). Take (this is done by having the actual announcer read the sentence, for example), and the grammatical condition that the morpheme satisfies when the pose is inserted between the morphemes, which is obtained from the statistical results ( Pose setting conditions) are described.

The pose setting rule includes a pose priority rule in which a setting condition for setting a pose between morphemes is added with a priority indicating a priority order for setting a pause, and a setting position of the pose. And a pause narrowing rule in which narrowing conditions for narrowing down are described. According to the pose priority rule, candidates for positions at which poses are set are determined, and from among the candidates, final positions at which poses are set are determined based on the pose narrowing rules.

The one shown in FIG. 8 is a pause priority rule which constitutes a pause setting rule, which constitutes the combining rule (FIGS. 6 and 7) as described above (1) to (9). Of the items, the items except (9) are described as one unit.

In this pause priority rule, when a pose is frequently inserted between morphemes satisfying a certain setting condition, the setting condition is given a high priority, and the pose is inserted less frequently. For example, it is given a low priority.

As described above, by assigning the priority to the setting condition, it is possible to preferentially set the pause for the morphemes that satisfy the setting condition of the high priority. Therefore, the priority can be said to be the priority for setting the pose.

Here, for example, 1 is the highest priority,
Hereinafter, it is assumed that as the number increases, it decreases.

In FIG. 8, the setting conditions of the priorities 1 to 6 have the condition numbers (1) of 11000, 12000, 13000, 140.
They are collectively described in the 00, 15000, 16000 series.

Therefore, which priority setting condition is used in the pose setting processing unit 6 can be easily designated by the condition number of (1), as in the case of the above-described combination processing unit 5.

The pose narrowing-down condition which constitutes the pose setting rule will be described later.

Next, in FIG. 9, the morpheme analysis result (FIG. 5) is combined by the combination processing unit 5 based on the combination rule, and the processing result is converted into a pause priority rule by the pose setting processing unit 6. The result of the pause priority processing based on FIG. In addition to the pause priority process, the pose setting processing unit 6 also performs a pose narrowing process based on a pose narrowing rule as described later. Further, FIG. 10 shows a flowchart for explaining the operation of the combination processing unit 5 and the pause setting processing unit 6. The combination processing unit 5 performs the processes of steps S1 to S9, and the pose setting processing unit 6 performs the processes of steps S10 and S11.

That is, in the combination processing section 5, the reference beat rate is set in step S1. Here, the reference number of beats means the number of beats between two poses used as a reference in the process of the pose setting processing unit 6 (however, punctuation marks are not included in the number of beats). Here, for example, the standard beats are A and B.
Two types are set. Also, the standard beat A is
For example, the range is set to 11 beats (mora) or more and less than 26 beats (mora) (in this case, the reference beat rate A represents the range, and henceforth referred to as the reference beat rate range A), and the reference beat rate B is, for example, 4
It shall be 5 beats (mora).

Then, in step S2, a compound word is identified from the morphological analysis result (FIG. 5) based on the compound word combination rule (compound word combination processing is performed).

Here, of the morphemes identified as the compound word, the leading morpheme or the trailing morpheme is expressed as "head" or "tail", respectively. In addition, a morpheme that remains as a compound word without being combined with another morpheme is expressed as “arc”. The punctuation marks are left as they are.

That is, in step S2, as shown in FIGS.
The conditions (rules) in the 1000s from the condition number 1001 of the combination rules of are applied between two adjacent morphemes in the morphological analysis result of FIG. 5, and it is determined whether any of the conditions is satisfied. To be done. When the two morphemes satisfy either condition, the front morpheme is the "head" and the rear morpheme is the "tail".

Specifically, in FIG. 5, for example, the morphemes "last night" and "male" and "Marunouchi" and "police" are all before the number 1001 of the condition number (1) shown in FIG. The morpheme's part of speech is "noun", and the latter morpheme's part of speech is "noun". Therefore, in this case, "last night" =
Head, "male" = tail (hereinafter appropriately referred to as "(last night""male")),"Marunouchi" = head, "police" = tail (("Marunouchi""police")) .

If three or more morphemes continuously satisfy the condition, the leading morpheme is the "head" and the trailing morpheme is the "tail", and all the morphemes sandwiched between them are the morphemes. It is referred to as the "body". For example, the compound word "Marunouchi Central Police" consisting of the morphemes "Marunouchi", "center" and "police" whose part of speech is "noun" is the morpheme "Marunouchi" and "center",
And "center" and "police" satisfy the condition of 1001, so "Marunouchi" = head, "center" = trunk, "police" respectively.
= Tailed (In this case, "Marunouchi" and "center" are first combined, then ("Marunouchi" and "center"), and then "center" and "police" are combined , (("Marunouchi""center")"police")).

After the above compound word combination processing, step S3
Then, the compound word dividing process is performed based on the compound word dividing rule. That is, in step S3, FIG.
The conditions in the 2000s from the condition number 2001 in the associative rule of 2 are adjacent to each other in the result of the compound word combining process described above.
It is applied between two morphemes (however, they are combined to form a compound word), and it is determined whether any of the conditions is satisfied. If the two morphemes satisfy any of the conditions, the previous morpheme that has been designated as the "head" is the "arc".
In addition, the morpheme after being made a "tail" is changed to "fox".
This compound word division processing is performed in order to exclude a combination of morphemes combined as a compound word that is exceptionally unfavorable to be identified as a compound word. In addition, each processing of steps S5, S7, and S9 described later also includes steps S4 and S9.
This is performed in order to exceptionally exclude (divide) the morphemes combined in the respective processes of 6 and S8.

Specifically, in FIG. 5, for example, the morphemes “last night” and “male” satisfy the condition of 1001. Therefore, as described above, “last night” = head and “male” = tail. It However, the sub-classification of the previous morpheme "Yoruyo" is "(adverb usage)" (Fig. 5), and the part of speech of the latter morpheme "male" is "noun". The morpheme meets the conditions of 2001 in Fig. 6, and therefore "last night"
= Fox, "male" = fox The morphemes before and after the division may be “tail” and “head” in addition to “arc” and “arc”. That is, for example, if two morphemes that are "head-to-tail" are divided into "arc" and "arc", for example, six morphemes that are "head-body-body-body-tail" are divided at the center. If it is done, it is considered to be "caudal torso" and "caudal torso", and the divided front and back morphemes are "tail" and "head".

The processing results of the compound word combination processing and compound word division processing described above are shown in the "word" column in FIG.

Here, a basic processing method adopted in the combination processing section 5 and the pose setting processing section 6 will be described. That is, in the combination processing unit 5 and the pose setting processing unit 6, when two or more morphemes are combined, the leading or trailing morpheme is made to be a morpheme representing the entire combined morpheme. Therefore, for example, focusing on the part-of-speech, for the "Marunouchi police" combined by the compound word combination process, the part-of-speech is "noun" for both "Marunouchi" and the morpheme "police" at the end. Even if one of them becomes the representative, the whole part of speech is "noun", but for "Marunouchi police to" obtained as a result of the phrase combination processing (step S4) described later, the leading morpheme "Marunouchi"
The part-of-speech of the combined morpheme is different because the part-of-speech of is a noun and the part-of-speech of the ending morpheme "ni" is a case particle. become.

That is, in the case where the combining rule or the pose setting rule is applied between, for example, "to Marunouchi police" and the (adjacent) morpheme in front of it shown in FIG. Of the morpheme "re" and "to Marunouchi police",
A coupling rule or a pose setting rule is applied between the morpheme "re" and the leading morpheme "Marunouchi".

In addition, for example, in the case of "to Marunouchi police", if the combining rule or the pose setting rule is applied between "to Marunouchi police" and the (adjacent) morpheme "theft" that follows it, Between the last morpheme "ni" adjacent to the morpheme "theft" and the morpheme "theft", the combining rule or the pose setting rule is applied.

Furthermore, for two or more morphemes that are combined, a combining rule or a pause setting rule is applied between adjacent morphemes at the end and the beginning. Specifically, the following processing is performed. (As described above, one morpheme itself is enclosed in square brackets "", and a combination of two or more morphemes is surrounded by round brackets ().
Surrounded by).

That is, for example, when (“clerk” “ni”) and (“hold” “re”) are adjacent to each other, (“clerk”)
The trailing morpheme "ni" representing "ni") and the leading morpheme "holding" representing ("holding""are")
If the conditions for combining morphemes are met, (“clerk” “ni”)
And ("press""re") are combined into one, resulting in (("clerk""ni")("press""re")).

Therefore, the morphemes to which the combining rule or the pose setting rule is applied are only those "head", "tail" and "arc", not the ones "body".

Next, after the processing of step S3, step S
4, the bunsetsu joining process is performed based on the bunsetsu joining rule, whereby the bunsetsu is identified from the compound word identification result. That is, the result of the compound word division processing is shown in FIG.
The clause linkage rules in the 3000s starting from 3001 are applied.
As a result, for example, "(Marunouchi""police") and the following "ni" shown in the "word" column of FIG. 9 are regarded as "tails" morpheme "police" and "fox". The morpheme "ni"
From the morphological analysis result shown in FIG. 5, the condition of 3004 shown in FIG. 6 is satisfied, so that they are combined and expressed as ((“Marunouchi” “police”) “ni”). That is, the morpheme "Marunouchi",
"Police" and "ni" are "head", "body", and "tail", respectively.
It is said that

After the phrase combination process described above, the process proceeds to step S5, and the phrase division process is performed based on the phrase division rule. That is, the bunsetsu division rule is applied between the morphemes that are combined to form a bunsetsu, and the bunsetsu that is not recognized as a bunsetsu is released.

Specifically, although not shown in FIGS. 5 and 6, as the segment division rule, for example, 4001, "police", noun, NULL, NULL, NULL, case particle, NULL, NULL are described. In this case, the morpheme “police” and “ni” that compose the ((“Marunouchi” “police”) “ni”) identified as a phrase satisfy the condition of this 4001,
("Marunouchi""police") and "ni" will be divided.

The results obtained by the phrase combination process and the phrase segmentation process are shown in the column of "bunsetsu" in FIG. Note that FIG.
Shows a processing result when the above condition No. 4001 is not described in the combining rule.

After that, the process proceeds to steps S6 to S9 in sequence.
Based on the continuous modification combination rule, the continuous modification division rule, the continuous modification combination rule, or the continuous modification division rule, the continuous modification combination process, the continuous modification division process, the continuous modification combination process, or the continuous modification split process is performed, respectively.

That is, from the result of identifying the bunsetsu, the continuous modified linked bunsetsu and the adnominal modified linked bunsetsu are identified. Note that step S
The processing of 6 and S7 may be performed after the processing of steps S8 and S9 instead of before the processing. However,
Simulations have shown that a more natural synthesized sound can be obtained by performing the processing of steps S6 and S7 before the processing of steps S8 and S9.

Specifically, in the continuous modification and binding process, as shown in FIG.
Processed as shown in the column of "bunsetsu", for example, "(clerk""ni") and ("holding""are")
From the morphological analysis result of FIG. 5, the last "ni" representing "ni") and the "holding" representing ("holding""are") satisfy the condition of No. 5004 in FIG. ("Clerk"
"Ni") ("holding""rear"),
As shown in the column of clause 1 in FIG. 9, “clerk”, “ni”, “holding”, and “re” are “head”, “trunk”, “trunk”, and
It is called a "tail".

Next, in the continuous modification division processing, although not shown in FIG. 5 and FIG.
For example, 6001, NULL, case particle, NULL, NULL, “press”, general verb, NUL
If L, NULL was described, it was identified as a continuous modified clause ((“clerk” “ni”) (“hold” “re”))
Since the morpheme "ni" and "holding down" that compose are satisfying the condition of No. 6001, these are ("clerk""ni") and ("holding down").
It will be divided into two parts.

The results obtained by the continuous modification combination processing and the continuous modification division processing are shown in the column "Phrase 1" of FIG. Note that FIG. 9 shows a processing result when the above-mentioned condition No. 6001 is not described in the combining rule.

In the adjoining modified binding process or the adjoining modified dividing process, the same processes as the adjoining modified joining process or the adjoining modified dividing process are performed with reference to the adjoining modified joining rule or the adjoining modified dividing rule. The processing result as shown in the column "Phrase 2" of 9 is obtained.

Here, the part-of-speech information output as a result of the morphological analysis is only the one described in the part-of-speech information dictionary shown in FIGS. 2 to 4 described above. That is, for example, if a subclass that subdivided the part of speech is described, the subclass can be obtained as the part of speech information. Specifically, as shown in FIG. 2, for example, as a subclass of “ordinary noun”,
If "(adverb usage)" (an ordinary noun of adverb usage) is described, a subclass "(adverb usage)" can be obtained as part-of-speech information when a morpheme falls under that. The same applies to the upper classification, as shown in FIG. 2, for example, by describing "independent word" as a higher classification of "noun", the morpheme is a lower classification thereof, for example, "(adverb usage ) ”, The upper classification“ independent word ”is also obtained as the part-of-speech information.

As described above, the method of classifying morphemes by directly describing the part-of-speech information in the part-of-speech information dictionary is as follows.
Hereinafter, it is referred to as a static classification method.

By the way, for example, the noun "arrest" is
It is also used as a noun, like "arrest", and as a change verb, like "arrest". Therefore, if you focus only on “arrest”, the part of speech will be a “noun”, but in order to insert a pose at a more natural position, after “arrest”, a particle such as “wo” is added. It is necessary to distinguish and classify the morpheme of "arrest" depending on whether it continues or is followed by an inflectional ending (for example, "Suru").

Therefore, in the case of making the distinction by the static classification method, as shown in FIG. 2, the above-mentioned "arrest" means that the "arrest" is a noun. In addition to this, the part-of-speech information "noun / sa-hen [sa-change]" that means that it is a verb should be described in the part-of-speech information dictionary ([sa-change] means "arrest" Means that it was certified as a verb).

In this case, the morphological analysis unit 4
In addition to the morpheme "arrest", when the morpheme following it is followed, when the morpheme of "sa variant inflection" continues, as part of speech information, "noun" or The part-of-speech information "noun / sa-hen [sa-hen]", which means that it is a changed "verb", is output. As a result, in subsequent processing, "arrest"
Will be treated as a morpheme of "noun / sa-hen [sa-hen]".

Next, in order to classify the morpheme of "arrest" as described above, it is necessary to pay attention to the morpheme following the morpheme. Therefore, it is possible to carry out as follows. That is, in the morphological analysis, "noun / sahen" is output as the part of speech information of "arrest" (therefore,
In the part-of-speech information dictionary, "noun / sahen" is described,
"Do not describe" noun / sa-hen "[sa-chang]", and after "arrest" of "noun" alone,
When the morpheme of "sa-inflectional ending" is continued, the combination rule is described so as to combine them.

Then, when the morpheme "arrest" of the part-of-speech information "noun / sa-hen" and the morpheme of the following "sa-inflectional ending" are combined based on the combination rule (FIGS. 6 and 7), " Additional information indicating that the morpheme "arrest" is a verb that changes (additional information called [change])
Is added to the combined result (in this case, "arrest" constituting the combined result).

After that, taking into consideration the part-of-speech information "noun / sa-hen" obtained as a result of the morpheme analysis and the additional information [sa-chang] added as described above, the part-of-speech information of "arrest" is added. Is treated as a "noun / sa-hen [sa-hen]".

On the other hand, if the morpheme following the morpheme “arrest” of the part-of-speech information “noun / sa-hen” is other than “sa-hen-inflectional ending”, even if they are combined, additional information is particularly added. Otherwise, "arrest" will be treated as it is as a morpheme of "noun / sahen".

In this case, "arrest" means "noun / sa change [sa change]" by adding or not adding additional information corresponding to the morpheme following it.
Or it is treated as "noun / sahen". That is, the classification of the combination result dynamically changes depending on the combination of the morphemes to be combined. Therefore, the method of classifying morphemes by the above method can be called a dynamic classifying method in contrast to the above static classifying method.

In the dynamic classification method, at first, there is a classification of "noun / sahen" which means a noun, but there is no classification of "noun / sahen [sachange]" which means a verb. Only when the additional information “sa change” is added in the processing, the classification “noun / sa change [sa change]” appears.

The flag information of (9) which constitutes one unit of the associative rule shown in FIGS. 6 and 7 corresponds to the above-mentioned additional information, and for example, one of the flags shown in the column of “clause” in FIG. The morpheme "arrest", "sa", which are combined to form a clause,
The "arrest" of "re" and "ta" is classified into "noun / sa strange [sa change]" by the dynamic classification method.

That is, first, from the morphological analysis result of FIG.
The part of speech of "arrest" or "sa" is "noun / sa-hen" or "sa-inflectional ending", respectively, and therefore 3001 in FIG.
It is said that the clause is composed ("arrest""sa") because it satisfies the conditions of the clause clause combination rule. That is, "arrest" =
Head, "sa" = tail.

In this case, the flag information No. 3001 (9) is described as "head [change]". This "head [sa change]" means that the first morpheme (the morpheme called "head") of the morpheme obtained by combining morphemes that satisfy the condition (condition 3001) is This means that the information “change” is added. If the flag information is described as “tail [sa change]”, the term “tail [sa change]” means that the tail morpheme (of the tail morphemes among the morphemes combined with each other that satisfies the condition). It means that the information "sa change" is added to the "morpheme" which is called "tail".

Therefore, in this case, the part-of-speech information of the first morpheme "arrest" of the combined result ("arrest""sa") is:
Changed from "noun / sa change" to "noun / sa change [sa change]".

In the following, "re" adjacent to the last "sa" representing ("arrest""sa") is 3006 in FIG.
(("Arrest""sa")"re")
Furthermore, the last "re" representing (("arrest""sa")"re") and the adjacent "ta" are 3010 in FIG.
Because the condition of the number is satisfied (((“arrest” “sa”) “re”)
"Ta"). That is, "arrest", "sa",
"Re" and "ta" are "head", "body", "body",
It is called a "tail".

The pause setting rule and the combining rule are described using flag information as necessary, and the flag information added to the morpheme "arrest" is then "arrest".
However, it is taken into consideration when applying the pose setting rule and the combining rule. That is, when the morpheme adjacent to the morpheme "arrest" identified as constituting the phrase is a "particle" such as "de", the part of speech of the morpheme "de" or "arrest" is "particle", respectively. Or "noun-sa-hen [sa-hen]", the condition 5001 shown in FIG. 7 is satisfied, and these conditions are combined to form a continuous modified consecutive clause.

If the morpheme "arrest" is not added with the flag information "sa change", the part of speech of the morpheme "arrest" is "noun / sa-hen", and therefore 5001 shown in FIG. Turn modified binding conditions are not met, so they will not be bound.

The flag information is [sa change] for morphemes.
It can be used not only when adding part-of-speech information, but also when adding other part-of-speech information (or information other than part-of-speech information).

That is, from the input sentence "never arrested" as shown in FIG. 11A, for example, FIG.
A morphological analysis result as shown in is obtained. Note that FIG.
Although not shown in FIGS. 2 to 4, the part of parentheses in (b) (the same applies to FIG. 12B described later) is shown in FIG.

First, the “adverb (declarative adverb)” “never” and the “noun / sahen” “arrest” do not meet any of the join conditions shown in FIGS. 6 and 7 and are not joined. . Next, the morphemes “arrest” and “sa” are combined with (“arrest” “sa”) as described above, and the flag information “sa change” is added to “arrest”. That is, at this stage, the "arrest" is regarded as "noun / sa strange [sa change]".
Then, (“arrest” “sa”) and “re” are also combined as described above to be ((“arrest” “sa”) “re”).

Next, "re" representing (("arrest""sa")"re") and "no" following it are shown in FIG. 11 (b).
From the morphological analysis result of, satisfy the condition of No. 3009 in FIG.
Therefore ((("arrest""sa")"re")"not"). Further, in this case, the flag information of the number 909 (9) describes "head [absent]", so the combined result ((("arrest""sa")"re")"absent") The flag information "not" is added to the morpheme "arrest" at the beginning of the.

The flag information may be sequentially added at each stage of combining the morphemes after the morpheme at the head (or the end) of the combination result is determined, or
For example, it is possible to add as follows. That is, after the head (or end) morpheme of the combined result is determined, the flag information is stored in, for example, a stack, and the stored combined flag information is obtained after the final combined result is obtained. You may add it collectively.

Therefore, ((("arrest""sa")"re")
The part-of-speech information for "arrest" that composes "no" is
Sa change [sa change] [never].

Next, ((("arrest""sa")"re")
From the result of the morphological analysis of FIG.
Meet condition 10 and therefore (((("arrest""sa")
"Re") "not") "like").

As described above, the input sentence shown in FIG. 11 (a) has the clauses "never" and (((("arrest""sa")).
"Re") "not") "like"). And this phrase (morpheme) "never" and ((("arrest""sa")"re")"not")"like") (morpheme) "arrest" mean "arrest" Since the part-of-speech information of "is noun / sa strange [sa change] [not]",
No. 5003 of the consecutive modified qualification combining rule is satisfied, and thus, at this stage, it is combined as a constituent of the consecutive modified qualification clause.
As shown in (c), ("never"(((("arrested"
"Sa") "re") "not") "like")).

In the input sentence "never arrested", the relationship that has a dependency with "never" is "not" which represents negation, and the grammatical property of this "not" is the flag "not". By being reflected in “arrest” by information, as described above, “never”,
((("Arrest""sa")"re")"not")"like") will be combined to form a continuous modified clause.

Next, for example, from the input sentence "Hakozaki Inter 3 km away" as shown in FIG. 12A, the morphological analysis result as shown in FIG. 12B is obtained. From this morphological analysis result, "3" and "km-mail" are shown in FIG.
It satisfies the 1002 compound word combination rule, and is defined as (“3” “km”). Then, "kilometer" representing ("3""kilometer") and its adjacent "destination" satisfy the compound word combination rule of 1003 in FIG.
((“3” “kilometer”) “destination”). Further, in this case, since the flag information of the compound word combination rule of 1003 is “tail [distance direction]”, the compound word ((“3” “km”) “forward” obtained as a result of the combination. )
Flag information "distance direction" is added to the morpheme "first" at the end of the.

As a result, (("3""km")
The "preceding" representing "preceding") and the subsequent "no" satisfy the clause coupling rule 3003 in FIG. 6 and are thus combined as a clause, and thus ((("3""km")
"Destination") "No"). Furthermore, in this case, since the flag information of the clause connection rule of No. 3003 is "tail [distance direction]", the clause ((("3"
Flag information "distance direction" is added to the morpheme "no" at the end of "kilometer") "forward") "no").
That is, in this case, the flag information added to the morpheme "preceding" is, so to speak, added to the morpheme "no", and the part-of-speech information of "no" becomes "case particle [distance direction]".

On the other hand, "Hakozaki interchange" and "o" are shown in FIG.
From the morphological analysis result of 2 (b), the bunsetsu combination rule of No. 3004 in FIG. 6 is satisfied, and thus the bunsetsu is combined to form a bunsetsu, which is (“Hakozaki Inter” “wa”).

Phrase ((("3""km")
If you consider the relationship between "(destination))" no ") and (" Hakozaki interchange "" o "), (((" 3 "" km "))
The part-of-speech information of the morpheme "no" or "Hakozaki Inter" representing """""""" or "(Hakozaki Inter""wo") is "case particle [distance direction]" or "proper noun (place name)", respectively. ) ”, They satisfy the adnominal modified binding rule numbered 7003 in FIG. 7. As a result, ((("3""km")"destination")"no") and ("Hakozaki interchange""wo") are combined to form an adnominal-modifying continuous clause, and ((((" 3 "" kilometer ")" destination ")
"No") ("Hakozaki interchange""wa")).

However, since the morphemes “no” and “Hakozaki Inter” also satisfy the adnominal modification division rule of No. 8001 in FIG. 7, they are combined to form an adnominal modifier continuous clause ((((“3” “kilometer” )"Previous")
(“Hakozaki interchange” “o”)) is the original phrase (((“3” “km”) “destination”) “no”),
("Hakozaki Inter""wo") and.

Therefore, the result of the combining process is shown in FIG.
As shown in (c).

That is, in this case, since the morpheme "no" is added with the flag information "distance direction", unlike the usual relationship between a case particle and a noun, ((("3""km"). ) "Ahead") "No") and ("Hakozaki Inter"
"W") is not combined as it constitutes an adnominal-modified conjunctive clause. In this way, by providing the flag information, it is possible to perform so-called exceptional processing in which other specific morphemes are not combined before or after the specific morpheme.

As will be described later, ((("3""kilometer")"destination")"no") and ("Hakozaki interchange") that were not combined to form an adnominal-modified continuous clause.
Between "" and "," that is, between adnominal-modified conjunctive clauses, are candidates for setting a pose. That is, "3 kilometers ahead"
There is a high possibility that a pose will be inserted between "Hakozaki Inter".

Therefore, when the present invention is applied to a voice synthesis process for guiding a destination in a car navigation system, for example, a phrase representing a distance such as "3 km ahead" and "Hakozaki Inter A pose is likely to be inserted between a phrase representing a place name corresponding to a destination such as "." According to the experiment conducted by the inventor of the present application,
It has been known that when a pose is inserted between a phrase representing a distance and a phrase representing a place name corresponding to a destination, the synthesized voice becomes easy to hear.

After the processing in the merging processing unit 5, that is, the processing in steps S1 to S9 in FIG. 10 is completed and the consecutive clauses of the input sentence are identified, the process proceeds to step S10, in which the pose setting processing unit 6 first Pause priority processing is performed.

That is, in the pose setting processing unit 6, the morpheme in which the column of "phrase 2" in FIG. 9 is "arc" or "tail",
Between the following (the next) morpheme (between adnominal-modified conjunctive clauses), the morpheme is “no”, and the column of “phrase 1” in FIG. 9 is “tail”, and the morpheme immediately after that. During this period, the pause priority rule shown in FIG. 8 is applied.

Note that the pause priority rule is applied between the morpheme in which the column of "phrase 2" in FIG. 9 is "arc" or "tail" and the morpheme that follows it (union). It may be only between the modified consecutive clauses), but especially when the number of beats of one sentence in the input sentence is large, the morpheme is “no” and the column of “phrase 1” in FIG. 9 is “tail”, Simulations have shown that applying a pause priority rule to a morpheme immediately after that also gives a more natural and easy-to-hear synthetic sound.

As a result of applying the pause priority rule, for example,
From the morphological analysis result of FIG. 5, the morpheme “last night” that is regarded as “lone” and the morpheme “male” that is the next “head” are
Since the condition of No. 12001 of priority 2 is satisfied, the flag of priority 2 indicating that the pose is a candidate for the position to be inserted is set to “last night”.

Also, for example, the morpheme "ha" which is "tail"
Then, the morpheme “clerk”, which is the next “head”, satisfies the condition 13001 No. 3 of priority 3 in FIG. 8 from the morpheme analysis result of FIG. 5, so the flag of priority 3 is “ha”. Be built up. In the same manner, flags (priority flags) are set as shown in the column of "interval" in FIG. In the column of "between" in FIG. 9, a number indicating the priority is shown at the position where the flag is set.

Further, in the pause priority processing, a subtotal of the number of beats between the positions where the priority flag is set is obtained.
However, for the first and last priority flags in the input sentence, the subtotal of the number of beats from the beginning of the input sentence to the first priority flag, and the subtotal of the number of beats from the last priority flag to the end of the input sentence. Are required respectively.

The numbers in the "beat" column at the right end of FIG. 9 represent this subtotal value. For example, for “last night” (3 beats), the number of beats is 3 beats. For example, “ha” (1 beat)
Is a total of 4 beats with "male" (3 beats), and for "re" (2 beats), "clerk" (4 beats)
The total number of beats is 10 beats.

After the pause priority processing, the process proceeds to step S11, where the pause narrowing processing is performed according to the pause narrowing rules, and the position for inserting the pose with the priority flag set (the morpheme with the flag set). , And the next morpheme), the final setting position of the pose is determined.

Here, the pose narrowing rule is actually
It is a model of a method in which an announcer puts pauses at meaning breaks in a sentence when the announcer reads the newspaper. Note that this pause narrowing rule is also described in a table format, so to speak, like the pause priority rule shown in FIG.

13 and 14 are flowcharts for explaining the details of the pause narrowing down process. Note that this flowchart reflects the pose narrowing-down rule. That is, in this flowchart,
Includes pose refinement rules (however,
As described above, the conditions described in the table format are prepared, and whether or not the conditions are satisfied is determined in the pause narrowing process (step S11)).

In the pose narrowing down process, first, in step S21 of FIG. 13, a pose is set immediately after a morpheme to which priority flags of priority 1 and 2 are set, for example. A morpheme (immediately after) to which a priority flag of priority 1 or 2 is set corresponds to a clause boundary or a phrase boundary, respectively. Also, the length of the pose set at the clause boundary is made relatively long, and the length of the pose set at the phrase boundary is made slightly shorter than that at the clause boundary.

Then, the process proceeds to step S22, and a predetermined determination process is performed. That is, in step S22, it is determined whether or not there is a morpheme that is not a target of the determination processing (undetermined), for example, a priority flag of priority 3 or 4 is set. If it is determined in step S22 that there is an undetermined morpheme to which the priority flag of priority 3 or 4 is set, the process proceeds to step S23, and the position of the morpheme (hereinafter, appropriately referred to as a morpheme of interest) (that As a result of setting the pause immediately after the morpheme), it is determined whether or not the number of beats between the poses falls within the reference beat rate range A. In step S23, the number of beats between the poses is
When it is determined that the number does not fall within the reference beat rate range A, step S24 is skipped and the process returns to step S22. That is, in this case, the pose setting (additional setting) is not performed.

On the other hand, if it is determined in step S23 that the number of beats between the poses is within the reference number of beats range A, the process proceeds to step S24, the pose is set at the position of the morpheme of interest, and the process returns to step S22.

The processing of steps S22 to S24 is first performed from the beginning of the input sentence toward the end of the sentence for the morpheme to which the priority flag of priority 3 is set, and then the beginning of the input sentence again. From the end of the sentence toward the end of the sentence, the morpheme with the priority flag of priority 4 set.

Here, in the case shown in FIG. 9, first, a pose is set immediately after "last night", "rear", and "ta" (step S21). Next, since the priority of the priority flag set to "ha" is 3, whether or not the beats between the pauses fall within the reference beat range A when a pause is set there. Is determined. When a pose is set at the position of "ha", the number of beats from immediately after "last night" to "ha" in which the pose is already set is 4 beats in the "beat" column of FIG. The number of beats from immediately after "ha" to "reset" where the pause has already been set is 10 beats, so all of them fall within the reference beat range A (11 beats to less than 26 beats as described above). No, so no pose is set at the "ha" position.

If, as a result of setting a pause at the position of "ha", the beats before and after that pose fall within the reference beat range A, the pose is set at that position ( Step S24). When the pose is additionally set in this way, whether or not to set the next pose (the processing of steps S22 to S24) is performed in consideration of the additionally set pose.

Furthermore, the number of beats between pauses is the reference number of beats range A.
The reason why the poses are set depending on whether or not they fit within the range corresponds to the fact that the number of beats between the poses when the announcer reads aloud is not extremely small or large. Further, in the present embodiment, since the model (reading model) obtained by making the announcer read a relatively long sentence such as a newspaper is regularized (pause narrowing regularization), the reference beat range A is set as described above. Although the number of beats is 11 or more and less than 26, when the input sentence is a short sentence, the reference beat rate range A (same as the reference beat number B) is set to a small range such as 3 mora or more and less than 11 mora. It is possible to set. Further, it is possible to change (dynamically change) the reference beat number range A (the same applies to the reference beat number B) in accordance with the beat number of one sentence in the input sentence.

On the other hand, if it is determined in step S22 that there is no morpheme to which an undetermined priority flag of priority 3 or 4 is set, step S31 in FIG.
And the number of beats between the poses that have been set at the moment,
It is determined whether or not there is a portion having a reference beat count B (46 beats as described above) or more.

Here, the reference number of beats B corresponds to the number of beats which is assumed to be the limit at which the announcer can read aloud in one breath (in one breath) and is not difficult for the listener to hear. In step S31 and subsequent steps, if there is a pause between beats that exceeds the reference beat B, based on the result (pause narrowing rule) obtained by statistically processing the announcer's reading result, in principle, more Pause is set (additional setting) from the position where the priority flag of high priority is set, and when there is no part where the beat count between pauses is the reference beat B or higher, the pose is set. (Additional settings) (Reset) ends.

Both the reference beat rate range A and the reference beat rate B are common in that the announcer does not make the listener feel unnatural reading. However, the reference beat range A is a high-priority range in which the announcer prefers to pose if possible (unless semantically unnatural). The reference beat rate B differs in that it is a beat rate determined because it is difficult for the announcer to read aloud in one breath (constrained by the physical constraints of the announcer), and the listener has difficulty in hearing. Therefore, the part where the beats between the poses is the reference beat B or more is a part where it is preferable not to put the poses if possible (unless the announcer can read aloud in one breath and the listener is hard to hear). That is, it is a low part as the priority order of the pose.

If it is determined in step S31 that there is no portion where the number of beats between the poses set at the present time is equal to or greater than the reference number of beats B, the process ends.

Here, in the case shown in FIG. 9, the position to which the pose is set before the processing of step S31 is immediately after "last night", "gone", and "ta".
From the beginning of the sentence to “last night”, from “last night” to “rear”,
Since the number of beats from “re” to “ta” or from “ta” to the end of the sentence is 3, 14, 25, or 0 beats respectively, there is no portion having a reference beat B or more, and the processing is not performed. finish. Therefore, in the case shown in FIG. 9, the position immediately after “last night”, “rear”, and “ta” is the final pose setting position.

On the other hand, if it is determined in step S31 that there is a portion in which the number of beats between the preset poses is equal to or greater than the reference beat B, the process proceeds to step S32, and the pauses (hereinafter referred to as appropriate). It is a particle "ha" that is flagged with a priority of 3 in (between pauses), and the part-of-speech information of the morpheme following it is "independent word" (non-verbal independent word) It is determined whether something exists.

In step S32, it is determined that there is a particle "ha" flagged with the priority 3 and the morpheme part of speech information that follows it is a non-verbal independent word between the target poses. If so, the process proceeds to step S33, where a pose is set (immediately after) the particle "ha",
It returns to step S31.

Here, in the case shown in FIG. 9, if the number of beats from immediately after "Last night" to immediately after "Play" is not less than the reference beat B, a flag of priority 3 is set. The particle "ha" (this "ha" and the "clerk" following it "satisfy the condition 13001 in FIG. 8) satisfy the condition of step S32, so in step S33,
The pose will be set at that position.

On the other hand, it is determined that there is no particle "ha" flagged with the priority 3 and the part of speech information of the morpheme that follows it is a non-verbal independent word between the attention poses. In this case, the process proceeds to step S34, and it is determined whether or not there is a morpheme set with a priority flag of priority 5 between the focus poses. If it is determined in step S34 that a morpheme with a priority flag of priority 5 exists between the target poses, the process proceeds to step S35.
A pose is set immediately after the morpheme, and step S31 is performed.
Return to. When there are a plurality of morphemes flagged with a priority level of 5 between the target poses, the pose is set to the position of the morpheme closest to the center (middle) between the target poses. .

If it is determined in step S35 that there is no morpheme set with the priority flag of priority 5 between the noticed poses, the process proceeds to step S36, and priority 6 is given between the noticed poses. It is determined whether there is a morpheme flagged. If it is determined in step S36 that a morpheme with a priority flag of priority 6 exists between the target poses, step S3
7, the pose is set immediately after the morpheme, and the process returns to step S31. In addition, when there are a plurality of morphemes flagged with the priority flag of priority 6 between the noticed poses, as in the case of step S35 described above, the center (middle) between the noticed poses is the most. A pose is set at the position of a nearby morpheme.

On the other hand, if it is determined in step S36 that no morpheme with the priority flag of priority 6 exists between the target poses, the process ends.

In step S36, it is determined that there is no morpheme set with the priority flag of priority 6 between the target poses, and the processing is terminated when the position is not unnatural as a break in meaning. It corresponds to the case where there is no, and it should be read in one breath.

Further, in FIG. 14, the determination process for the morpheme to which the priority flag of priority 4 is set is omitted, but this is performed by using the pause priority rule shown in FIG. The reason is that, even if the morpheme determination process for which the priority flag of priority 4 is set is performed, there is almost no effect on the final pose setting position. Therefore, it does not mean that the determination process for a morpheme with a priority flag of priority 4 set should not be performed.

The pose set in this way (pose information)
Based on the above, phonological prosody information is generated in the phonetic symbol generation unit 7 of FIG. 1, and further, through the prosody processing unit 8 and the acoustic processing unit 11, synthetic voices with inserted pauses are output.

As described above, the compound word rule, the bunsetsu rule, or the conjunctive clause rule is a compound word, a bunsetsu, or a conjunctive clause obtained by combining a morpheme that joins morphemes and morphemes that satisfy the join condition. Since the pose setting rule is described using the flag information as necessary, including the flag information to be added to, the flag information is given to the entire combined morphemes by combining the morphemes. Further, since the pose setting rule and the combining condition are described by using the flag information, it is possible to set the pose according to the combination of the morphemes to be combined. That is, it is possible to insert a pause at an appropriate position in the input sentence without particularly limiting the input sentence.

Further, since the input sentence is morphologically analyzed and the compound word, the bunsetsu, the consecutive modified consecutive bunsetsu, and the adnominal modified consecutive bunsetsu are identified from the morphological analysis result, the unit (identification result) formed at each stage of the processing is identified. By using this, various other processes can be performed in addition to the pose setting process. That is, for example, it is possible to perform the process of adding an accent in the unit of the identified compound word, or the process of adding the accent or the pitch in the unit of the identified phrase.

Further, the part-of-speech information output as a result of the morphological analysis includes not only so-called part-of-speech, but also a higher-level classification or a subdivided lower-level classification, so that it is easy to describe a combination rule and a pose setting rule. As
For example, compared to the case of setting a pose based only on the part-of-speech of a morpheme, it is possible to set a highly accurate pose that can handle exceptional cases (set a pose at a more natural position). It will be possible.

Further, since the combination rule and the pose setting rule are described by using the morpheme character string obtained as a result of the morpheme analysis, the part-of-speech information, etc., it is not necessary to add or input other information. Modules that make up the processing system can be easily constructed, and general-purpose Japanese text can be processed.

Also, based on the pose priority rule created by statistically obtaining the frequency of appearance of poses between consecutive phrases, and the pose narrowing down rule created from actual human reading results, the final pose Since the setting position is determined, the pose can be set at a natural position.

Further, the combination rule and the pose setting rule are
Instead of being written as a program in the module that performs each process, it is provided in a table format, so to speak, separately from the program, so that the order of application and expansion / revision can be easily performed.

Further, as described above, the reference beat number range A and the reference beat number B can be changed, so that a natural and easy-to-understand synthesized voice corresponding to the input sentence to be processed can be obtained. Becomes

The case where the present invention is applied to a voice synthesizing apparatus has been described above.
For example, it can be applied to a voice translation device that translates a Japanese input sentence into another language, or conversely translates another language into Japanese and outputs it by voice. Since the position where the Japanese pose is inserted matches the semantic break in the Japanese syntax, it is possible to improve the parsing accuracy and improve the conversion accuracy to other languages. It becomes possible to output a voice (synthesized sound) that is natural and easy to hear when translated and output as voice.

Further, the present invention is, for example, a voice recognition device,
It can be applied to a word processor capable of voice input, a document creation device, and the like. When applied to a voice recognition device, utterances before and after a pose is inserted can be narrowed down as a recognition target, and a combination pattern of morphemes before and after a pause is regularized (described in the pause setting rule). Therefore, the recognition accuracy can be improved.

Further, when applied to a word processor or a document creation device, when performing "kana-kanji conversion" of a sentence or the like input by voice, it is possible not to perform the kanji conversion processing of the portion that spans the pauses. Therefore, the conversion efficiency can be improved.

Further, in the present embodiment, the combining rule includes a rule (division rule) for dividing the combined morphemes (composite word division rule, phrase division rule, continuous modification division rule, adnominal modification division rule). However, this division rule does not necessarily have to be described in the combining rule.

Further, in the present embodiment, the combination rule and the pause setting rule are described by using the morpheme character string and the part-of-speech information in the morpheme analysis result. It is also possible to describe by using the result of morphological analysis such as.

Further, in this embodiment, only the grammatical information is described in the combining rule and the pose setting rule (pause priority rule), but the semantic information may also be described. good.

Further, in this embodiment, the punctuation mark is not applied to the combining rule and the pause setting rule but ignored, but it may be applied to the combining rule and the pause setting rule. That is, in the present embodiment, for example, when the morphological analysis result as shown in FIG. 5 is obtained, for example, the punctuation mark “,” preceding “Marunouchi” is ignored, and “are” and “Marunouchi” are separated. Although the joining rule (pause setting rule) is applied in between, it is possible to apply the joining rule (pause setting rule) between “,” and “Marunouchi”.

[0193]

As described above, according to the natural language processing method of the present invention, by combining morphemes, additional information can be given to the entire combined morphemes, and the additional information can be used. Since the pose setting rule and the combining condition are described, it is possible to set the pose according to the combination of the morphemes to be combined. That is, it is possible to insert a pause at an appropriate position in the input sentence without particularly limiting the input sentence.

Further, according to the speech synthesizer of the present invention, a natural and easy-to-understand synthesized speech can be obtained without particularly limiting the input sentence.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of an embodiment of a speech synthesizer to which the present invention is applied.

FIG. 2 is a diagram showing a part-of-speech information dictionary.

FIG. 3 is a diagram showing a part-of-speech information dictionary (a diagram following FIG. 2).

FIG. 4 is a diagram showing a part-of-speech information dictionary (a diagram following FIG. 3).

FIG. 5 is a diagram showing a result of morphological analysis.

FIG. 6 is a diagram showing a combination rule.

FIG. 7 is a diagram showing a joining rule (a diagram following FIG. 6).

FIG. 8 is a diagram showing a pause priority rule.

FIG. 9 is a diagram showing a result of a combination process and a pause priority process.

FIG. 10 is a flowchart illustrating operations of a combination processing unit 5 and a pose setting processing unit 6 in FIG.

FIG. 11 is a diagram for explaining a dynamic classification method.

FIG. 12 is a diagram for explaining a dynamic classification method.

FIG. 13 is a flowchart illustrating the details of the process of step S11 of FIG.

FIG. 14 is a flowchart (flowchart following the flowchart in FIG. 13) for explaining the details of the processing in step S11 in FIG.

[Explanation of symbols]

1 arithmetic unit 2 memory devices 3 Language Processing Department 4 Morphological analysis section 5 Join processing unit 6 Pose setting processing section 7 Phonetic symbol generator 8 Prosody processing section 9 Prosodic control model parameter generator 10 Prosody data generator 11 Sound processing unit

Claims (13)

(57) [Claims] 1. A morphological analysis of an input sentence in Japanese, and a compound word is identified from a morphological analysis result of the input sentence based on a compound word rule defining a grammatical relationship between morphemes forming a compound word. Then, from the identification result of the compound word, the bunsetsu is identified based on the bunsetsu rule that defines the grammatical relationship between the morphemes forming the bunsetsu, and from the bunsetsu identification result, the grammatical relationship between the morphemes forming the consecutive bunsetsu Pose that is inserted into the sentence, which is statistically obtained, for the continuous sentence clause that is composed of one or more morphemes forming the input sentence obtained by identifying the consecutive sentence clause based on the continuous sentence clause rule that defines the relation Is a natural language processing method in which a pose setting rule that defines the position of a phrase is applied to set the position of a pose to be inserted between the consecutive clauses, wherein the compound word rule, the clause rule, or the consecutive clause rule is And binding conditions for binding, compound words obtained by combining the morphemes each other that satisfy the join condition, and a additional information to be added to clause or Renbunsetsu, the pose rules and binding conditions, optionally,
A natural language processing method characterized by being described using the additional information. 2. The link clause rule includes a link modifier rule that defines a relationship between morphemes that form a link modifier link clause, and a link modifier rule that defines a relationship between morphemes that form a link modifier link clause. 2. The natural language processing method according to claim 1, further comprising identifying, from the identification result of (1), a linked modified linked sentence clause or a linked linked modified clause, based on the linked modification rule or the linked modifier rule. 3. The morpheme analysis result includes at least a morpheme character string and part-of-speech information, and the compound word rule, bunsetsu rule, continuous bunsetsu rule, or pause setting rule is described using the morpheme analysis result. The natural language processing method according to claim 1 or 2. 4. The natural language processing method according to claim 3, wherein the part-of-speech information includes, in addition to the part-of-speech, a higher-level classification that integrates the part-of-speech and a lower-level classification that subdivides the part-of-speech. 5. The natural language processing method according to claim 4, wherein the sub-class includes at least an inflectional form. 6. The natural language processing according to claim 1, wherein the additional information is a high-level classification in which the POS of a morpheme is integrated or a low-level classification in which the POS of a morpheme is subdivided. Method. 7. The compound word rule, bunsetsu rule, or continuous bunsetsu rule further describes a separation condition for separating two combined morphemes, and separates the combined morphemes according to the separation condition. The natural language processing method according to any one of claims 1 to 6, characterized in that. 8. Between the compound word composed of a plurality of the morphemes and an adjacent morpheme which is an adjacent morpheme,
When the compound word rule is applied, among the plurality of morphemes forming the compound word, the compound word rule is applied between the adjacent morpheme adjacent to the adjacent morpheme and the adjacent morpheme. The natural language processing method according to any one of claims 1 to 7, which is characterized. 9. The clause including a plurality of the morphemes,
Between the adjacent morpheme and the adjacent morpheme, if the phrase rule is applied, between the morpheme adjacent to the adjacent morpheme among the plurality of morphemes constituting the bunsetsu, between the adjacent morpheme The natural language processing method according to any one of claims 1 to 7, wherein the clause rule is applied. 10. Between the conjunctive phrase consisting of a plurality of the morphemes and an adjacent morpheme which is a morpheme adjacent to it,
When the consecutive phrase clause or the pause setting rule is applied, among the plurality of morphemes forming the consecutive phrase, between the adjacent morpheme and the adjacent morpheme, the consecutive phrase clause or the pose setting rule is The natural language processing method according to any one of claims 1 to 7, which is applied. 11. The pause setting rule sets a pause between a morpheme at the end of the morphemes forming the preceding syllable and a head morpheme of the morphemes forming the subsequent syllable in two adjacent syllables. The natural language processing method according to any one of claims 1 to 10, wherein the insertion frequency is created by statistically obtaining from the reading result of the actual sentence. 12. The pose setting rule is described with setting conditions for setting a pause between the last morpheme and the first morpheme added with a priority representing a priority order for setting a pause between them. Includes a priority rule and a narrowing-down rule in which a narrowing-down condition for narrowing down the setting position of the pose is described, based on the priority rule, a position candidate for setting the pose is determined, and from among the candidates, The natural language processing method according to claim 11, wherein a final position for setting a pose is determined based on the narrowing-down rule. 13. Generation means for obtaining phonological information and prosodic information corresponding to the input sentence by natural language processing the input sentence in Japanese, and to the input sentence based on the phonological information and prosody information. The synthesizing unit for synthesizing corresponding voices, the prosody information includes a position of a pose inserted in the input sentence, and the generating unit determines the position of the pose.
2. A voice synthesis device characterized by being obtained by the natural language processing method according to any one of 2.