KR20040025969A - Natural Language Processing Method Using Classification And Regression Trees - Google Patents

Abstract

PURPOSE: A method of processing a natural language using classification and regression trees (CART) is provided to analyze phrase information through probability and statistics to generate a control rule. CONSTITUTION: A corpus database required in the event of voice synthesis with respect to character information is constructed(S210). Phrases are modeled through CART using the constructed corpus database(S220). Specifically, a phrase break index is estimated using a decision-tree model. A phoneme period is estimated using a regression-tree model. Natural phrases are estimated on the basis of the modeled data when unknown character information is input(S230).

Description

Natural Language Processing Method Using Classification And Regression Trees}

The present invention relates to a natural language processing method, and more specifically, to develop a control rule by probabilistic statistical analysis of added rhyme information such as intonation, pause, length of voice and intensity for the development of a high quality speech synthesizer. A natural language processing method using a CART.

In general, natural language processing technology is a technique for uncovering the meaning and syntax of a vocabulary by analyzing a source language according to its purpose of use.

1 is a flow chart of a general natural language processing method, a grammatical analysis step (S110) of dividing a sentence without grammatical defects into a vocabulary by a word and a minimum meaning unit (S110), and a grammatical component in each sentence. Parsing step (S120) for distinguishing between, a semantic analysis step (S130) to pruning on the basis of the meaning in a plurality of syntax structures, and that is not solved in the semantic analysis step through the discourse analysis step (S140) To deal with.

This natural language processing method converts voice announcements of automobile navigation, text data of FM DARC (DAta Radio Channel), and e-mails from Auto PC into voices for more natural rhythm and sound quality. It is widely used in the development of speech synthesizers that provide.

However, in the conventional speech synthesis, the rhythm processing is merely a space reading, and there is a problem in that it cannot provide natural rhyme and sound quality of the synthesized sound due to the lack of prediction of phoneme duration.

Accordingly, an object of the present invention is to construct a corpus database for speech synthesis and statistically model various rhythmic phenomena using CART (Classification And Regression Trees), and then process the input sentences to predict natural rhyme phenomena. It is to provide a natural language processing method using a CART.

In order to achieve the above object, the natural language processing method using the CART according to the present invention comprises the steps of constructing a corpus database (Corpus Database) required for speech synthesis of information in the form of characters, and using the constructed corpus database Modeling the rhyme phenomenon through a CART, and predicting a natural rhyme phenomenon when inputting information in an unknown character form based on the modeled data.

1 is a flow chart of a general natural language processing method.

2 is a flowchart of a natural language processing method using CART according to the present invention;

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

2 is a flowchart of a natural language processing method using a CART according to the present invention.

Referring to FIG. 2, the natural language processing method according to the present invention constructs a pronunciation database for all phonemes for speech synthesis and connects them to generate a continuous voice. In this case, the size, length, and height of the voice are used. Adjust your back to synthesize natural voices.

First, a natural language processing apparatus (not shown) for speech synthesis that speaks text information or a symbol into a human voice constructs a corpus database required for speech synthesis of information or symbols in a text form (S210). .

In this case, by using a corpus-based TTS (Corpus-Based Text To Speech) technology, it solves the unnaturalness of voice connection caused by the use of another speech processing technology, a limited lexical speech synthesis technology.

For reference, the limited vocabulary speech synthesis technique is a technique for synthesizing speech for a limited vocabulary and sentence type, and is a technology for generating a continuous speech by pre-recording necessary pieces of speech and connecting them to an automatic response device (ATS) service It is mainly applied to, but it is technically very simple but the sentence form is limited and the voice connection site is unnatural disadvantages. On the other hand, TTS has no limitations on the vocabulary to be synthesized, and converts information in general text form into speech, which embodies the intonation, reading, etc. similar to the actual human voice for natural speech synthesis.

Therefore, for natural speech synthesis, a corpus speech database is constructed by selecting sentences that are most suitable for rhyme boundary strength prediction and sentences that are most suitable for phoneme duration prediction.

At this time, 400 sentences with accent phrases (AP) and accents (IP) checked by linguists are constructed by corpus speech database for prediction of rhyme boundary strength, and accent phrases are calculated by linguists for phoneme period prediction. AP), 65,000 phonemes in which the accent (IP) is checked, or about 900 sentences, are constructed as a corpus voice database.

When the construction of the corpus database is completed (S210), the rhythm phenomenon is modeled using the constructed and regression trees (CART) method using the constructed corpus database (S220). For reference, the CART method is a Decision-Tree procedure introduced in 1984 by several world-renowned UC Berkeley and Stanford statisticians.

In this case, when modeling the rhyme phenomenon through the CART, the boundary strength of the rhyme is predicted using a decision tree model, and the duration of the rhyme using a regression-tree model, that is, Predict phoneme duration.

The Phrase Break Index is a psychoacoustic parameter rather than an objective judgment as a rhythmic heterogeneity between words that a person feels when listening to a spoken voice as a parameter representing a degree of reading of a sentence. .

The rhyme boundary strength prediction process, which is the process of extracting the rhyme components necessary for natural speech synthesis, uses the CART method with 400 sentences in which accents (AP) and accents (IP) are checked by experts built in the corpus database. Use it to predict.

In addition, the rhyme boundary strength prediction by the CART method uses the parameters of Table 1 below as feature parameter elements corresponding to the number of words observed in 400 sentences of the database, that is, 4847 words.

parameter Commentary DPOS Representative parts of speech DLPOS The left-hand article of the word DPLPOS A left-handed article DPPLPOS The front article of the front word DNLPOS Next article DNNLPOS Next article DRPOS The postman of the word DPRPOS A postman DPPRPOS The postman of the front word DNRPOS Next word DNNRPOS Next postman C_LOC The position of the word in the sentence C_LOCFRBEG Word position from front of sentence C_LOCFREND Word position from behind the sentence

In this case, in the decision tree-based modeling used for predicting the rhyme boundary strength, the splicing method selects the GINI index method and the SE rule (Standard Error Rule) selects the minimum cost tree. do. In addition, by dividing its data into segments of the same size for testing purposes, and determining the optimal tree using the V-fold Cross-Validation method that provides one segment at a time, the rhyme boundary strength Predict.

On the other hand, the duration of the phone (Phone) is an important factor in determining the naturalness of the synthesized sound with the intonation, the process of controlling this is important because it is also affected by the speed of speech. For reference, the sound that participates in making a word is called a phoneme, and the phoneme may not only have a slight difference depending on the people who pronounce the word, but may also show a difference depending on the pronunciation environment. Therefore, the phoneme is the abstraction of such phonology into representative sounds.

The duration of the phoneme, that is, the phoneme period, is predicted using 900 sentences in which accents (AP) and accents (IP) are checked by an expert built in the corpus database. At this time, a sentence is often used as a database between 400 and 1000 sentences in predicting rhyme. In the case of phoneme prediction, a database of thousands of sentences was constructed, but due to the limitation of the database program, it has only 65,000 phonemes, that is, about 900 sentences. Predict.

In addition, 900 sentences of the corpus database are constructed by dividing into five categories: initial, neutral, and final phonemes, the last phoneme of the AP boundary, and the last phoneme of the IP boundary.

The average of the durations of the phonemes for each phoneme set divided into the five categories is shown in [Table 2].

Type of phoneme Average duration Initiality 69.62 msec neutrality 74.59 msec Jongseong 55.98 msec AP last phoneme 84.74 msec IP last phoneme 155.58 msec

In addition, the parameters used for the phoneme duration control at this time are as shown in Table 3 below.

parameter Commentary DLPHONE Phoneme in front of the observation phoneme DPHONE Observation DRPHONE Phoneme behind the phoneme DLOCEOJ Phoneme location within the word DLOCAP Phoneme location within the accent phrase DLOCIP Phoneme location within the accent DNUMEOJ The number of syllables in the word DAP Whether the word accented boundary DIP Whether the word is bordered by an accent

In this case, in the regression tree modeling used for the phoneme period prediction, a splicing method selects a GINI index method and a SE rule (Standard Error Rule) selects a minimum cost tree. In addition, the phoneme period is predicted by determining the optimal tree using the V-fold cross-validation method.

When the modeling of the rhyme phenomenon through the prediction of the rhyme boundary strength and the phoneme period is completed (S220), a natural rhyme phenomenon is predicted when inputting information in an unknown character form based on the modeled data (S230).

Therefore, in the natural language processing method using the CART according to the present invention, after constructing a corpus database for speech synthesis and using it to statistically model various rhyme phenomena using the CART method, the natural sentence phenomenon can be predicted by processing the input sentence. Can be.

What has been described above is to suggest a method for optimal rhyme prediction through an experimental method for the natural language processing method using the CART according to the present invention, any person having ordinary knowledge in the field of the present invention It will be said that the technical spirit of the present invention is to the extent that modifications can be made.

As described above, the natural language processing method using the CART according to the present invention statistically modeling the rhyme information, such as intonation, pause, length and strength of the tone through the CART method to generate a control rule and process the input sentence By doing so, it is possible to develop a high quality speech synthesizer capable of providing more natural sound quality instead of providing mechanical and unstable sound quality in the conventional speech synthesis.

Claims (5)

Constructing a corpus database required for speech synthesis of textual information; Modeling a rhyme phenomenon through classification and regression trees (CART) using the constructed corpus database; And Natural language processing method using the CART comprising the step of predicting a natural rhyme phenomenon when inputting information in the unknown character form based on the modeled data. The method of claim 1, wherein the building of the corpus database, Extract 400 sentences with the most suitable accent phrase (AP) and intonation (IP) for predicting the rhyme boundary strength, construct them into a database, and check the most suitable accent phrase (AP) and accent (IP) for phoneme period prediction. Natural language processing method using CART, which extracts 900 sentences and then divides them into initial, neutral, final, last phoneme of AP boundary and last phoneme of IP boundary and constructs them into database. The method of claim 1, wherein the modeling of the rhyme phenomenon through CART comprises: A rhyme boundary strength prediction process that predicts the boundary strength of rhyme using a decision-tree model; A natural language processing method using a CART, comprising: a phoneme duration prediction process for predicting a phoneme duration using a regression-tree model. The method of claim 2, wherein the rhyme boundary strength prediction process, A splicing method selects a GINI index and a SE rule selects a minimum cost tree. The method of claim 2, wherein the phoneme duration prediction process, The splicing method selects the GINI index and the SE rule selects the least cost tree.