JPH0827635B2 - Compound word processor used for sentence-speech converter - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Outline] In a sentence-speech conversion device for Western languages such as English, German, French-Western, etc., a phonological symbol string is generated from a sentence and a speech is output. At this time, high-frequency words and words with exceptional pronunciations are registered in the dictionary, and accents and pronunciations of other words are determined according to a predetermined rule.

When sentence-to-speech conversion of an arbitrary sentence is performed, a compound word in a sentence can be relatively easily divided if the words forming the compound word are registered in a dictionary. is there.

However, if the words that make up the compound word are not registered in the dictionary, the compound word is not divided into constituent words,
Treated as a word, the rules determine accent and pronunciation.

Therefore, in the conventional sentence-to-speech conversion of a compound word, if the words constituting the compound word are not registered in the dictionary, the pronunciation and accent of the output sound may be unnatural. There was a point. The present invention, in order to solve such a conventional problem, for the corresponding language system, the appearance probability of a character arranged next to a certain character is obtained in advance for a word not including a compound word, and Disclosed is a sentence-speech conversion device capable of obtaining a correct pronunciation and accented speech output by determining a compound word based on a value.

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to control of a sentence-speech conversion device for Western languages, and particularly to a compound word in which a plurality of words are combined when a phonological symbol is generated from a sentence. Concerning the processing method of.

[Prior Art] When sentence-to-speech conversion of Western languages (English, German, French, etc.) is performed, in order to generate a phonological symbol string from a sentence, high-frequency words or words with exceptional pronunciation are The words are registered in the dictionary, and the accents and pronunciations of other words are determined according to a predetermined rule.

When converting an arbitrary sentence into speech, it is necessary to divide the compound word contained in the sentence into constituent words, determine the accent and pronunciation for each word, and further determine the accent as the compound word. Is. Naturally, a normal sentence includes a compound word in which words are combined. In this case, if the words that make up the compound word are registered in the dictionary, it is possible to relatively easily divide them using the longest match method, but if they are not registered in the dictionary, It cannot be recognized as a compound word. Therefore, in the conventional sentence-speech conversion device, when a word forming a compound word is a word that is not registered in the dictionary, the compound word is not divided into the constituent words and is treated as one word, and according to a predetermined rule. Accent and pronunciation were decided.

[Problems to be Solved by the Invention] As described above, in the conventional sentence-speech conversion device,
If the words that make up the compound word are not registered in the dictionary, the compound word is not divided into constituent words and is treated as one word, and the accent and pronunciation are uniquely determined by a predetermined rule. Therefore, there is a problem that the pronunciation or accent of the output voice may be incorrect or unnatural.

For example, in the case of a compound noun, generally, the first accent is placed on the preceding word and the second accent is placed on the latter word, but if the compound noun is not divided and is treated as one word, the correct accent is obtained. The position cannot be determined.

This will be explained by taking the compound word "bedroom" as an example. When the compound word is recognized, the first accent is put in "bed" and the second accent is put in "room". However, if it is not divided into the constituent words, it is regarded as a word with the prefix "be" and the stem "droom", the second accent is placed on the second syllable, and the first syllable becomes a weak syllable without accent. .

In addition, if there is one character, the pronunciation is determined for each character, but there is a set of characters that gives a special pronunciation when the characters continue. Therefore, when a word ending with one character and a word starting with the other character are combined to form a compound word, it is considered that two characters are continuous unless it is divided into constituent words, You can't get the correct pronunciation.

For example, if a word that ends in't 'and a word that ends in'h' are combined into a compound word that is not split,
The pronunciation rule of'th 'determines the pronunciation as [θ] or [δ].

In view of such a conventional problem, the present invention can recognize a compound word as a compound word even if a word forming the compound word is not registered in a dictionary. It is an object of the present invention to provide a sentence-speech conversion device capable of correctly outputting the pronunciation and accent of a word.

[Means for Solving Problems] According to the present invention, as described above, the object is to generate a phoneme symbol string from a sentence described in characters or symbols to generate the phoneme symbol. A compound word processing device used in a sentence-speech conversion device for outputting speech based on a string, wherein a certain character or a character next to a character string statistically obtained in advance for a certain number of existing words is used. Alternatively, with respect to a storage unit that holds a value relating to the probability of occurrence of a character string, and a character string that forms a compound word included in the sentence, for each character or part of the character string in the character string, then The value relating to the probability that a character or a part of a character string adjacent to the position appears is searched for in the storage means, the value relating to the calculated probability is compared with a predetermined threshold value, and the result is less than the threshold value. Character in the small string Alternatively, a compound word processing unit that divides the compound word with a word forming the compound word between a part of the character string and a character or a character string adjacent to the succeeding position or a word boundary is provided. It is achieved by the compound word processing device.

[Operation] The probability (transition probability) p (character string 1 | character string 2) that a certain character string 1 transits to the next character string 2 is the number of appearances of the character string 1 in the character strings in many word sets. Then, the character string 1 and the character string 2 can be obtained by dividing the number of appearances of the states that occur simultaneously (successively).

For example, considering the transition of one character, the probability that's 'changes to't' can be obtained by dividing the number of appearances of'st 'by the number of appearances of's'.

Probability of transition of character string 1 to character string 2: By examining the spelling of each word in a word set that contains a large number of words (but not compound words),
The transition probability of each character string to the next character string is obtained and held as a table.

When processing a compound word, the transition probability of each character string in the word is extracted from the table, and a continuous portion of the character string having a transition probability that is less than or equal to a threshold value is selected as a candidate for a boundary of constituent words.

[Embodiment] FIG. 1 is a block diagram of an embodiment of the present invention, showing a logical configuration of a word-phonetic symbol converter in a sentence-speech converter of a Western language, where 1 is an exceptional word. Processing unit, 2
2 is an affix separation processing unit, 3 is a compound word processing unit, 4 is an accent determination unit, 5 is a character-phonetic symbol conversion unit, 6 is an exception word dictionary, 7 is a transition probability table, 8 is an accent rule, and 9 is pronunciation. Rule 10, the affix separation rule.

In FIG. 1, the affix separation processing unit 3 first separates the affix of the input word, and checks whether it is an exceptional word. If the word is an exceptional word, the word is converted into a phonetic symbol, and then the process ends. If it is not an exception word, compound word processing is performed on the stem portion. If it is a compound word, a symbol indicating a word boundary is inserted in the word. Furthermore, determine the accent of the word, and finally, convert the spelling to a phonetic symbol,
Finish the process.

For the transition probability table, a word set including a large number of words (not including compound words) was prepared, the transition probability of each character string was calculated for the words therein, and their values were stored. It is a thing.

The compound word processing unit 3 selects a candidate for a compound word division position by the control shown in the flowchart of FIG.

That is, when processing a compound word, the beginning of the word is set in the pointer ptr pointing to the word, and if the number of characters equal to or less than the character pointed to by ptr is less than n characters, the process ends. Otherwise, cut out the n characters following ptr.
Next, the transition probability of the cut out character string is extracted from the table. If the transition probability is less than or equal to the threshold value, the position is regarded as a word boundary candidate. Then, set the next character to ptr and return to point A again.

For example, suppose the English compound word "earthquake" is entered. The transition probability table used here is 1
The transition probability from one character to one character is calculated from about 20,000 words. The transition probabilities extracted from the table for each character "earthquake" are shown in FIG. The smaller the value, the weaker the bond between characters. In the figure, the transition probability from'h 'to'q' is 0, which indicates that this character sequence does not exist in one word. Therefore, a word boundary symbol is inserted between'h 'and'q' to perform compound word division.

In this case, the position where the transition probability is 0 is set as the word boundary. However, depending on how the threshold is taken, there may be a plurality of positions that can be word boundaries, or there may not be any. Therefore, the probability that the character string before the position of the boundary candidate transitions to the end of the word or from the beginning of the character string to the character string before the position of the boundary candidate is extracted from the table, and the value exceeds the threshold. If you select a candidate, you can further narrow down the candidates.

[Effects of the Invention] As described above, according to the compound word processing device of the present invention, even if a compound word in a sentence is composed of words that are not registered in the dictionary, the characters that make up the compound word By using the sequence transition probability, the compound word can be divided into correct constituent words. Therefore, by using the compound word processing device of the present invention in the sentence-speech conversion device, a correct accent and pronunciation sound can be obtained. There is an advantage that can be obtained.

[Brief description of drawings]

FIG. 1 is a block diagram of an embodiment of the present invention, FIG. 2 is a flow chart showing control of a compound word processing unit, and FIG. 3 is a diagram showing an example of transition probability. 1 ... Exception word processing unit, 2 ... Affix separation processing unit, 3 ... Compound word processing unit, 4 ... Accent determination unit, 5 ... Character-phonetic symbol conversion unit, 6 ... Exception word dictionary, 7 ... … Transition probability table, 8 …… Accent rule, 9 …… Pronunciation rule, 10 ……
Affix separation rules

Claims (1)

[Claims] 1. A compound word processing device used in a sentence-speech conversion device for generating a phoneme symbol string from a sentence described by characters or symbols and outputting a voice based on the phoneme symbol string, comprising: Storage means for holding a value relating to the probability that a certain character or character string will appear next to a certain character or character string statistically obtained in advance for the existing word, and included in the sentence. For the character string forming the compound word, the storage means is searched for a value relating to the probability that a character or a part of the character string adjacent to the succeeding character appears for each character or part of the character string in the character string. The obtained probability is compared with a predetermined threshold value, and the result is a character or a part of the character string in the character string whose result is smaller than the threshold value and a character or character string adjacent to the posterior position. Between the compound words That it comprises a compound word processing means for dividing the compound word as a word and a word boundary to be formed compound word processing apparatus according to claim.