KR100292376B1 - Device and method for converting sentence - Google Patents

Abstract

PURPOSE: A device and method for converting a sentence are provided to merge a sentence as a voice easily by modeling and regulating an interpret of a morpheme and a speaking rule of a human with respect to an input sentence, respectively, without a meaning interpret and performing a disconnected reading based thereon. CONSTITUTION: An input unit(10) receives a sentence(data) from an exterior. A dictionary unit(20) stores a basic mode of a word. A disconnected reading model table(30) stores a disconnected reading model. An operation memory unit(40) temporarily stores data being processed until an inputted sentence is converted into a voice. A program storing unit(60) stores a processing program with respect to a disconnected reading. A control unit(70) performs a control for performing a data process with a peripheral device in accordance with the stored program. An output unit(80) outputs a converted sentence. In addition, a central processing unit(50) is provided.

Description

Translator and its method

1 is a block diagram of a sentence conversion apparatus according to the present invention.

2 is a flowchart schematically showing a sentence conversion method according to the present invention.

* Explanation of symbols for main parts of the drawings

10: input unit 20: dictionary unit

30: Cut off model table 40: Operation memory part

50: central processing unit 60: program storage unit

70: control unit 80: output unit

The present invention relates to a sentence converting apparatus for converting a sentence to be read in a predetermined unit, and a method thereof, and more particularly, sentence transformation to generate a sentence in a reading unit according to the morphological analysis of the sentence and human voice rules. An apparatus and a method thereof are provided.

Efforts to convert the sentences used by humans to speech through machines have been developed in various ways from the past to the present. In order to convert a sentence into speech, it is necessary to read the sentence like a human speech rule to accurately grasp the contents of the converted speech information and to secure naturalness.

However, in the case of humans, when a sentence is read, the sentence is interpreted at the same time and the sentence is spoken. However, at the current level of technology, a sentence is interpreted based on the semantic interpretation of the input sentence. It is almost impossible.

In addition, in the conversion method according to the interpretation of the morpheme used in the case of a general speech conversion device, since the voice conversion is performed by simply analyzing the partial morpheme without considering the whole sentence, it always has a constant broken reading pattern. There was a problem that is not suitable as a vocal model.

In more detail, the reading method, which has been used in the related art, is as follows.

First, in the case of broken reading by morphological interpretation only, if there is a corresponding morpheme among the morphemes defined in the morpheme table in reverse order for each spacing unit of input sentences (words) with a morpheme table containing data on morphemes The part of speech is determined and if it does not exist, the part of speech is the most appropriate. Based on the morphemes determined in this way, the grammar rules (e.g. access checks, narrative checks, etc.) were used to create and use appropriate reading rules.

However, this method has a problem that the error of morpheme interpretation occurs a lot, and the accuracy of the break-and-read rule is considerably lowered because the error thus formed has a significant effect on the break-through rule.

Second, in the case of cut-off reading by means of meaning interpretation, the dictionary of words and transformed endings is constructed in advance, and the meaning is analyzed and the cut-out rule is applied according to the meaning. Impossible and not complete, but at some level (here the reading results can be understood by the listener), the formation of dictionaries containing information on vast amounts of words and the devices that implement them are quite significant. It is difficult, and it is very difficult to apply to speech synthesis technology because it is not easy to perform real-time processing of immediately reading the input data and converting the result into speech.

Accordingly, an object of the present invention is to provide a sentence converting apparatus for avoiding semantic interpretation and morphological interpretation of input sentences and human speech rules, modeling and regularizing them, and breaking them based on them, thereby facilitating synthesis into speech. To provide a way.

In addition, another object of the present invention is to provide a sentence conversion apparatus and method capable of minimizing problems caused by morphological interpretation errors by variably determining the number of broken readings of a sentence.

In order to achieve the above object, the present invention provides an input means for receiving the sentence data from the outside;

A dictionary including basic form and part-of-speech information of words;

Control means for controlling the determination of the reading level between each word by analyzing the input sentence;

The sentence type is temporarily stored through the input means, and the sentence is separated from each word under the control of the control means, and the basic form of the word obtained by analyzing the morpheme for each word by searching the dictionary part An operation memory unit for storing the part-of-speech information of the word;

A cut-out model table that defines a cut-out level according to the part-of-speech information of a word in order to set the cut-off level of words stored in the operation memory unit; And an output unit configured to output an input sentence according to the cut-off level determined by the cut-out model table under the control of the controller.

In addition, the present invention is the morphological analysis step of separating the sentence data to be input to each word and search the dictionary for each of the separated words to extract the basic form and part-of-speech information of the phrase by morpheme analysis;

A truncated reading processing step of determining a truncated reading level of each word according to the relationship between the information obtained in the morphological analysis step and each of the adjacent words in the input sentence; And an output step of outputting the input sentence for each reading unit according to the level information output in the cut-out processing step.

Hereinafter, the present invention will be described in detail with reference to the accompanying drawings.

1 is a schematic block diagram of a sentence conversion apparatus according to the present invention. In FIG. 1, 10 is an input unit for receiving a sentence which is data from the outside, 20 is a dictionary unit in which a basic form of a word is stored, 30 is a break reading model table in which a break reading model is stored, and 40 is a voice for converting an input sentence into voice. An operation memory unit for temporarily storing data processed up to 50, a central processing unit (Central Processing Unit: CPU); 60 is a program storage unit for storing a processing program for spacing according to the present invention, 70 is a control unit for controlling data processing with a peripheral device according to the stored program, 80 is output the converted sentence Is an output section.

The operation of the sentence conversion apparatus having the above configuration will be described with reference to the flowchart of FIG.

When the operation is turned on together with the input of the power to the sentence converter, the control unit 70 sequentially executes the programs stored in the internal program storage unit 60. When the stored program is executed as described above, initialization for sentence conversion is performed.

In the initializing state, the data to be voice-converted is first input through the input unit 10 in order to convert it into a form suitable for reading in accordance with the present invention. In this case, the input data may be in the form of a data file including a sentence or a sentence input through a keyboard. When input through a keyboard, a specific sentence may be input on the screen, and data may be input through a return key, a character indicating the end of various sentences, or a separate command.

The sentence input in this way is first temporarily stored in the operation memory unit 40. The sentence stored in the operation memory section 40 is stored until the input data forms one complete sentence. What is stored until the input sentence is a complete sentence is to make the voice conversion more natural. This is because if the data is input and converted into voice and output, the meaning of the converted sentence cannot be accurately understood. At this time, a check to see if the input data is a complete sentence can be performed by detecting "."? "!" Indicating the end of the sentence, or when more spaces appear in succession than a given value (e.g., eight). Alternatively, this may be performed by an agreement such as when line feed (LF) is continuous.

When a completed sentence is made, it is separated for each sentence of the sentence. At this time, the separation of sentences can be known by the space (space) between words. When the words are separated by each word like this, the basic form for each separated word is searched and found in the dictionary unit 20. That is, morphological analysis of words is performed. In this case, the long form matching method is used as a method of searching the basic form (ie, if the longest word exists, the shortest word is defined as the longest word). As an example of such a search, if the word "love" is searched in a sentence, the following words may be searched in the dictionary unit 20.

"① ④ ② love ③ love"

At this time, "love" is the stem of "love". In general, in the case of verbs or adjectives, since the mother may vary depending on the utilization type, only the stem is recorded beforehand. When the word is searched as described above, "love" is determined by the longest matching method as mentioned above.

And if you look at the analysis of the morpheme mentioned above in detail, the syllable unit analysis that separates the morphemes that make up the words from the sentence is carried out to restore the prototype of the morphological transformation where the morphological transformation has occurred and to generate possible morphological candidates. An example of how to perform a restore is as follows.

For example, when the irregular adjective "beautiful" is used as "beautiful" in a sentence, the word is analyzed by morphological processing as "beautiful" ("beautiful-" + "-"). . This analysis first separates the word "beautiful" in the input sentence into "beautiful" ("beautiful-" + "-ㄴ") and checks the end of the stem from it if the end is "right" and the ending is "b". If you start with "우" to replace "ㅂ" look for "beautiful answer" in the dictionary. At this time, since the information about the 아름 -irregularity is contained in the dictionary about "beautiful answer-", the result of analysis ("beautiful answer-" + "-eun") comes out. The analysis of the morpheme described above is one example by applying one of several methods can be implemented to obtain the same results in other methods.

According to the morpheme analysis as described above, the result (basic type and part-of-speech information about the word) searched through the dictionary unit 20 is stored in the operation memory unit 40. The stored data is transferred to a three-read model table, and the three words are sequentially shifted by one word to perform a process for determining the level of the broken read. An example of such a performance is as follows.

"① I think ② I love ③ Young-hee ⑤. Is determined by the relationship between the words ①, ②, ②, and ③, and the reading level after ② is determined first. In other words, the relationship between the type of word that comes before it and the word that follows it determines the level. Since the reading level of word ③ is related to ② and ③, ③ and ④, it is applied to the table by shifting by one word.

The cut-out model table used in this way is composed of three tables (Table I, Table II, and Table III), and each table consists of 60 × 60 tables. The table contains the word ID and the type ID of the word. The part of speech refers to the part of speech, which consists of 60 types of parts of speech, and the part of speech ID was determined by considering the basic type and its application.) Indicates the IDs of the x-axis and the y-axis. This is recorded.

In practice, the model table may be structured as a three-dimensional table, that is, [previous] [current] [next], without dividing it into three, but this includes the problem that the size of the table becomes huge and the data is redundant. do. Therefore, a three-dimensional table of 60 × 60 × 60 is formed to form three 60 × 60 × 30 tables, that is, two 60 × 60 two-dimensional tables, so that the data size can be reduced and redundant data can be reduced. It was.

The current word and the previous word are applied to the horizontal and vertical axes of Table I to take the value of the corresponding column, that is, the value of the position determined by the row and column ID. Next, the current word and the next word are applied to the horizontal and vertical axes of Table II and are projected onto the horizontal and vertical axes of Table III. The corresponding words are stored as the current reading value of the current word.

When the cutoff value is determined (that is, the level for each word is determined through the breakreading process I), the total number of syllables (1 if there is no support and 1.5 if there is a support) is formed. The result is then multiplied by the previously stored cutoff value to store the result. The reason for multiplying the weight according to the total number of syllables in this process is to consider the case that the longer the syllable is, the shorter the breath becomes when a real person pronounces it.

In this way, the number of cut-outs is determined according to the total number of syllables forming the sentence, and the cut-out process II is performed to perform a more natural cut-out in a word having a value satisfying the set number. This process determines the first position (segment classification) and repeats the above process for each sphere and outputs the result to the output unit. In other words, once there is break level information, it is decided how many times to read according to the number of syllables forming a sentence, and adjusts the level of break level accordingly. Here, the number of cut-outs determines the number of syllables (average) that can be naturally read by one breath.

In other words, consider the following sentence as an example.

③ We turned the ③ historical ③ mission of ⑤ national revival ② and ③ was born ④ on the ground.

The above sentence is a sentence composed of 7 syllables and 23 syllables, and 13 to 15 syllables can be composed of one breath. Therefore, the number of broken readings is 2, and the information of the broken readings is processed as follows. Become together.

We ④ turned off the ② historic ① mission of national revival ④ and was born ② on the ground.

That is, it is appropriate to cut off the sentence at the level ⑦, which is the largest level, as described above, and then divide the text into two parts. Rather than reading the level ⑦ in a little smaller than the rest period is a very natural break reading.

Looking at the process in order, using the sentence "We were born on earth with the historical mission of national revival," we look at the whole process in order.

The input sentences are separated from each word to calculate the number of syllables.

We were born here on the historical mission of national revival.

① (3) ② (5) ③ (3) ④ (3) ⑤ (2) ⑥ (3) (4) ⑦

In this way, when the word is divided into input sentences, parts of speech information for each word are given by analyzing the morpheme.

We were born here on the historical mission of national revival.

10,16 3,2,19 3,18 2,17 34,48 12,3,20 34,43

Given the part-of-speech information, the level for each word is determined by the table.

We were born ② on the ground, ② with the ② historical ① mission of national revival.

The sentence above is determined by the level of each word, since the total number of syllables is 23 syllables, 13 to 15 syllables are the number of syllables that humans can produce with one breath. have. Therefore, the most important level is between "Thigo" and "On earth", so the reading of the sentence above is possible by reading "We have a historical mission of national revival" and "Born on earth". Since the first part of the two parts is a long sentence consisting of five words and fifteen syllables, the highest level of the sentence is between "we" and "ethnic revival". Subject reading is the most desirable result.

In general, the accuracy of the spacing of the input sentences in the case of converting the input data, that is, the sentences into speech, is determined by whether there is no sense of rejection when listening by transmitting the correct meaning of the converted speech. Table 1 shows the result of comparing the accuracy between the conversion method and the conventional methods, the time taken to convert the same sentence, and the complexity of the system for implementing the same.

TABLE 1

As shown in [Table 1], the voice conversion performed by the method of the existing I has advantages in consideration of execution time and system complexity, but there is a disadvantage of not being able to read the correct spacing. In this case, the accuracy can be improved, but it is extremely economically disadvantageous because the execution time by the semantic analysis increases rapidly or the amount of memory for processing the same increases. However, in the case of the spacing of the speech model according to the present invention, it can be seen that it is the most preferred embodiment in consideration of accuracy, execution time and system complexity.

The results obtained in [Table 1] are the results obtained by conducting the National Education Charter, 30 editorials consisting of 30 random sentences, Korean middle school language, and 500 sentences extracted from the contents of history books. This is an integer compared to the created spacing position.

In addition, it was applied to the speech synthesizer for the application of the method according to the invention and the listening test was performed to obtain the results as shown in Table 2.

TABLE 2

Clarity in Table 2 is the listening result of a sentence consisting of meaningless words, and naturality is the average value of the evaluation results given to 20 listeners. The evaluation criteria used are shown in <Table 3>.

TABLE 3

As shown in the description and evaluation test described above, the present invention has a large number of meanings for semantic interpretation because the speech conversion for the input data generates the reading unit according to the morphological analysis and the human speech rules. No dictionaries are required, making it easy to process in real time and simplifying system configuration. In addition, since the read-reading model does not bias the grammar and considers the actual human vocalization habit, there is a natural effect than that used in the past.

In addition, since the post-processing can be performed by the weight of the syllable number, the natural voice conversion can be performed in consideration of the respiratory characteristics of the human.

The number of cut-outs can be determined according to the total number of syllables, and accordingly the cut-out value of words can be prevented from always being a constant cut-out pattern.

Claims (6)

An input unit for inputting sentence data; A dictionary unit for storing basic form and part-of-speech information of words; A cut-out model table that defines a break-read level between words according to word-of-speech information; An operation memory unit for temporarily storing the input sentence data in one sentence unit; After dividing the stored sentences into word units, the dictionary unit is searched to analyze the morphemes of each word, the basic form and the part-of-speech information of each analyzed word are stored in the operation memory unit, and stored by referring to the broken reading model table. A control unit for determining a level of reading between words and selecting a final level of reading according to the number of readings in consideration of the total number of syllables of the input sentences among the reading levels; And an output unit which cuts out and outputs the stored sentence according to the final cut-off level. The sentence converting apparatus of claim 1, wherein the controller adds a weight in consideration of the number of syllables of each word to the cutoff level determined by referring to the cutout model table. The sentence transformation apparatus of claim 1, wherein the cut-out model table comprises three tables that form a two-dimensional array having a predetermined size. A morpheme analysis step of dividing an input sentence into word units, analyzing a morpheme for each word by searching a dictionary storing basic form and part-of-speech information of a word, and obtaining basic form and part-of-speech information of each word; According to the word-reading information of words, the level of word reading is determined by referring to the word reading model table that defines the level of word reading between words. The total number of syllables of sentences entered from the word reading levels is determined. A cut-out processing step of selecting a final cut-off level according to the number of cut-outs considering the cut-off reading; And an output step of reading and outputting an input sentence according to a final breaking read level. The method of claim 4, wherein the number of cut-outs is determined based on one of 13 syllables and 15 syllables as a reference syllable unit from all syllables of the input sentence. 5. The sentence of claim 4, wherein in the cut-out processing step, after determining the cut-off level of each word that has undergone the morphological analysis step, a weight in consideration of the number of syllables of each word is added to the determined cut-off level. How to convert.