KR19980047177A - Korean document analyzer for voice conversion system - Google Patents

Abstract

A statistical language processing technique is introduced in a document speech conversion system for converting a given input document into speech, thereby providing a more accurate phonetic representation and syntax structure for a rhythm generation module that determines the naturalness of a synthetic sound. A preprocessing unit that extracts a document one sentence at a time and transforms non-Hangeul characters into Hangeul characters using an amorphous finite automata; morphological analyzing means for obtaining all possible morpheme analysis results in one sentence phrase applied by the preprocessing unit; , The morphologically analyzed input sentence is selected based on the probability information and the most probable morphological analysis result is selected to extract the optimal morpheme analysis string and the result of the morpheme analysis string applied in the morpheme tagger Pronunciation of each word And a syntactic analysis means for outputting the phonetic representation of the phrase for the input document, the morpheme part sequence, and the dependency tree as the final result is obtained by using the phonetic transcription conversion means and the probability dependent grammar to obtain the dominance- The accuracy of the parser is 85.6% and the accuracy of the parser is 78.68%. Especially, in the case of the part - of - speech tag, it provides a high accuracy of 96.46% when there is no unicorded word in the document being processed.

Description

Korean document analyzer for voice conversion system

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a text-to-speech system for converting a given input document into speech, and more particularly, to a prosody generation module for introducing a statistical language processing technique, And more particularly to a document analyzer for a Korean document speech conversion system that provides accurate pronunciation notation and syntax structure.

In general, a document speech conversion system is a system for converting a given input document into speech, and has many applications such as a blind reader.

This document speech-conversion system consists of a document analyzer, a rhyme generation and a signal synthesis module, and the document analyzer has a close relationship between the phonetic representation of a word and the rhyme of the word, It is used to analyze this.

Traditional Korean document analyzers extract only one result by using the longest matching method in the morpheme analysis stage, and perform phonetic transformation and syntax analysis using the same.

Recently, a phonetic transcription module based on a two-level model has been developed and many studies are under way to extract more accurate information.

However, since the document speech conversion system as described above analyzes the part of speech information of a document inputted as a part of speech which is set by the programmer, if the same syllable has different parts of speech, There is a problem that the reliability of the document speech conversion is deteriorated.

SUMMARY OF THE INVENTION The present invention has been made in view of the above problems, and its object is to provide a rhythm generation module that determines the naturalness of a synthetic sound by determining the phonetic representation of a word, Notation and syntax structure to provide reliable document speech conversion.

In order to accomplish the above object, the present invention provides a speech recognition system, comprising: a preprocessing unit for extracting a document input for speech conversion, one sentence by one, and converting non-Korean characters into Hangul characters using an amorphous finite automata; Morphology analyzing means for obtaining all possible morpheme analysis results in one sentence word dictionary applied by the preprocessing means; A morphological analyzer for extracting an optimal morpheme analysis sequence by selecting the input morpheme analysis result based on probability information; A phonetic transcription conversion unit that obtains phonetic transcription of each of the phrases using the morphological analysis string applied to the morphological phrase analysis unit; And a syntax analyzing means for obtaining a dominant-dependency relationship of the phrases by using a probability-dependent grammar and outputting a phonetic representation of the input document, a morpheme phrase, and a dependency tree as a final result.

1 is a block diagram of a document analyzer for a Korean document speech conversion system according to the present invention,

2 is a schematic diagram of an automata for preprocessing in the present invention.

3 is a block diagram of the morpheme analyzer of the present invention shown in Fig. 1,

FIG. 4 shows a morphological lattice configuration for I in the present invention,

5 is an algorithm for the lattice configuration in the present invention.

6 is a morphological analysis lattice configuration of reporting a new report in the present invention,

FIG. 7 is a flowchart showing the configuration of a dependency tree according to the present invention.

FIG. 8 is a diagram showing the structure of a tree structure of the present invention.

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

1, the preprocessing module 10 in the document analyzer for the Korean document speech conversion system according to the present invention converts a document input for speech conversion into a non-Korean character using a nondeterministic finite automata And one sentence is provided to the morpheme analyzer 20 side.

The morpheme analyzer 20 obtains all possible morpheme analysis results. If the morpheme analysis fails, the morpheme analyzer 20 determines that there is an unrecorded word in the analyzed word and evaluates the unrecorded word. Then, using the linguistic heuristic, Reduce the number.

In the part-of-speech tagger (30), the optimal morpheme analysis sequence is extracted from the input sentence through morphological analysis based on the probability information obtained from the corpus.

In the phonetic transcription conversion module 40, phonetic transcription of each of the phrases is obtained by using the result of the morphological analysis column applied in the speech pacifier 30.

The parser 50 uses the probability-dependent grammar to obtain the dominance-dependence relationship of the phrases, and outputs the phonetic representation of the phrase for the input document, the morpheme part sequence, and the dependency tree as the final result.

In addition, the speech tagger 30 and the syntax analyzer 50 are provided with probability information 70 for speech tagging and probability information 80 for classification analysis, respectively, in order to perform deconvolution.

The speech conversion operation of the Korean document inputted in the present invention having the functions as described above will be described as follows.

When a document for speech conversion is input into the preprocessing module 10, an indefinite finite automata consisting of 48 states is used, and 12 final states are pre-processed by different Hangul conversion module The module 10 extracts the sentences one by one from the input document and at the same time converts all non-Korean characters in the sentence into Hangul.

At this time, non-Korean characters are distinguished by English, English abbreviation, number, telephone number, year, and time, and they are translated into Korean by considering the following Hangul characters.

For example, Mr. When a document called Lee's telephone number is inputted, as can be seen from the automata for the preprocessing of FIG. 2, the state of No. 4 is reached by 'Mr.', and 'Mr.' is registered in advance If it is, it is localized as 'Mr.', Otherwise it is changed to 'M-al'.

At this time, 'Mr.' is an English abbreviation registered in advance, so it is translated into 'Mr.'.

In the case of 'Lee', 'Lee' is changed to 'Lee' in the state 5, and 'Hang' is processed in the state 2 by re-entering '

In the process, Mr. Lee's phone number gets the final result.

Thereafter, the morpheme analyzer 20 constructs a morpheme lattice by searching for morphemes from the right to left of the sentence phrase to be preprocessed through the operation as described above. As shown in FIG. 3, The position estimation module 22 previously calculates the irregular and shortened development positions for the speech recognition module 21 and performs the morphological analysis by the algorithm set in the lattice configuration module 23. [

The morpheme analysis result is expressed as a morpheme lattice. For example, the final analysis result of I is as shown in FIG. 4, and all possible paths from 'INI' to 'FIN' represent different analysis results.

In order to express the morpheme lattice as shown in FIG. 4, the set L shown in [Table 1] below is defined.

The elements (k, w, t, I) of the set L represent one node in the morpheme lattice and the edges to the right of the node, and the set L corresponding to FIG. 4 is shown in Table 2 below.

[Table 1]

k: index of l

w: morpheme

t: Part of speech

i: the index of the elements pointed to by l

[Table 2]

L = {(0, FIN, NULL, {})

(1, b, prime end, {0})

(2,, malleable end, {0})

(3,, subject schedules, {0})

(4, day, verb, {2})

(5, I, auxiliary verb, {2})

(6, I, verb, {2})

(7, I, personal pronoun, {3})

(8, INI, NULL, {7, 6, 5, 4})}

On the other hand, the lattice construction algorithm included in the lattice configuration module 23 first increases the number of steps starting from 0 for each algorithm, do.

For example, if you set the step to 6, I will be '0 ㄴ 6 ㄱ 12 ㄴ 18 ㅡ 24 ㄴ 30'.

5, the w _{i, j} denotes a row of digits between the digits i and j, and the function LookupDict is a set of digits from the dictionary Finds possible parts of the input sequence w _{i, j} .

The function searchL functions to receive the set of indexes of elements that (w _{i, j} , t) in the set L can connect to, and this is done by using pre-constructed part-of-speech tables.

The set J used in the algorithm is a collection of the left numbers of the nodes with paths from the completed grid to the 'FIN' node, which is used to reduce unnecessary dictionary searches.

The irregular part is based on the irregular position information obtained beforehand. In the case of I, 'd' can be eliminated at 12 and 24, so if 'j' is 12 or 24, 'd' The variable h is decreased from j to 0, and w _{i, j} is found in the dictionary.

So (4, day, verb, {2}) in the set L is added when j is 12.

On the other hand, if the analyzed phrase includes an unregistered word, all possible unregistered word candidates are generated from the incomplete morpheme lattice to complete the lattice.

At this time, assuming that all the functional words such as the survey and the mother are registered in advance, only the contents such as a censor or a verb can become unregistered words, so the unregistered word always appears in the left part of the word.

Therefore, the assumption of the unregistered word is that the left column is allocated to the part of speech that can be connected to the part of the node at the right. In this way, a morpheme lattice is generated by estimating the unregistered word.

However, since the number of possible morphological analyzes is too large, the number of candidates can be reduced by using syllable information and clue morpheme to prevent the next step of speech tagging. .

First, the method of using syllable information is to exclude the case where the last syllable of the unregistered word is not used as the last syllable of the estimated part-of-speech.

For example, using the fact that the last syllable in Korean is only six words, the last syllable is 'n'. If the unspoken part of speech is a verb and the last syllable of the unregistered word is 'n' .

The method using the clue morpheme removes the rest from the lattice, leaving only the unrecognized word in front of the morpheme when a morpheme that is very frequent in the incomplete morpheme lattice is found to be sufficient to estimate its composition.

For example, if the 'detour' is an unregistered word in the word 'detour', unregistered candidates will have 'detour / dongjang noun', 'detour / verb', 'detour / adjective' 'Is a clue morphea, so' bypass / donghyeongseong noun 'is left only, and all the rest will be removed.

Using this method, the number of morpheme analysis per word can be reduced from 22.01 to 10.83 for a word containing an unregistered word.

When the morpheme analysis is completed on the input sentence, the puzzle tagger 30 focuses on the method of processing unregistered words to find the optimal morpheme analysis result among the morphological analysis candidates of the respective phrases, .

Probability-based tagging of part-of-speech is a sentence consisting of n word-phrases, ie, word heat w ₁ w ₂ w _{3 ...} w _n , w _{1 ‥ n} , the morphological analysis result of the ith word is expressed as a pair of the morpheme column m _i and the partial column t _i , so that the part-of-speech tagging function ψ (w _{1 ... n} ).

[Equation 1]

Equation 1 is expressed by the following equation using the Bayes rule and the first-order Markov assumption.

&Quot; (2) "

Expression 2 finally finds the largest value of the morphological analysis column when the probabilities of transition from the part-of-speech string of the previous word to the part-of-speech string of the current word and the probabilities of occurrence of arbitrary morpheme strings in the part- It makes sense.

It assigns P (m _i | t _i ) to the node and P (t _i | t _i-1 ) to the edge with the edge between the morpheme analysis result nodes and the morpheme analysis results of adjacent words It can be seen as taking the path that gives the highest probability.

For example, a word line for reporting a new report can be represented as shown in FIG. 6 attached, and a problem of this kind is obtained by a Viterbi algorithm.

In this case, nc is a normal noun, po is a subject verb, vb is a verb, ex is an auxiliary connection mother, ec is a connection mother, xj is an adjective derived suffix, en is a noun qualitative mother, na is a normal noun, Suffix, and vx: auxiliary verb.

However, since the shortage of data may occur when Equation 2 is used as it is, the number of parameters of the model is reduced by approximating it to a probability value of smaller units.

First, in P (t _i | t _i-1 ) in Equation 2, t _i means the morpheme part of speech, .

here Is the jth morpheme part of the morpheme analysis of the _ith word and N _i is the number of parts of the morpheme used in the morpheme analysis.

For example, if t _i is vb, ex, vx, en Is vx, and N _i is 4.

In this case, vb: verb, ex: auxiliary connection ending, vx: auxiliary ending, en: noun ending.

Using this method, P (t _i | t _i-1 ) can be expanded as follows.

Equation 4 is derived from the assumption that the lexical strings in the current process are dependent only on the last morpheme word in the previous word, which is again developed by the chain rule in Equation 5, and the first-order Markov assumption is applied To obtain equation (6).

That is, it is approximated that the probability of transit of the part of speech is multiplied by the probability of transiting from the last morpheme part of the previous word to the first morpheme part of the current processing word and the parts of speech transition probabilities in the word.

P (m _i | t _i ) in Eq. (2) is the probability of morpheme heat occurring in the part-of-speech sequence. This may be an unregistered word.

therefore A new variable that means either this register or unregistered word The following equation is redefined.

In Equation 7 silver When 1 is assigned to this register and 0 is assigned when it is not registered, it is developed as follows.

In Equation 9, P ( | T _i ) is the probability that an arbitrary morpheme string is given, of which the first part of speech is an unregistered word or a registered word, which can be approximated as follows using the independent assumption.

On the other hand, P (m _i | t _i , ) Can be expanded as follows.

Equation (15) is obtained by using the following equations (16) and (17), and finally Affects only the first term of Eq. (15).

The first term in Equation 15 The value of < RTI ID = 0.0 > In the case of 1, the approximate expression of Equation 18 is used, and when it is 0, the same assumption as that of Equation 19 is used.

Equation (19) is a kind of verbal heuristic assuming that the estimation of unregistered words in Korean depends on morphemes and parts of speech that follow the unregistered words.

Using the above two equations, P (m _i | , ti) The following approximate expression is obtained.

The above two equations and Equations 6 and 9 can be put in Equation 2 to get the following partly equations.

In Equation 23, P ( | ) Is the probability that an unregistered word or registrar will occur when given any part of speech, and these probabilities are obtained from the corpus containing unregistered words.

In particular, these values depend on the number of headwords in the current dictionary, and the optimal speech marking result is obtained stochastically regardless of the dictionary size.

The expression thus far developed is shown in Fig. 6 And P ( | ), Respectively, and can be solved by a bi-terbi algorithm.

6, it can be seen that the nodes have released the morpheme lattice resulting from the morpheme analyzer, so that the results can be obtained more quickly by connecting the morpheme lattice of the phrases and performing the bi-directional algorithm.

Upon completion of the morphological analysis of each of the phrases by the above operation, the phonetic transcription conversion module 40 searches for the phonetic transcription of the given sentence based on the result of the morpheme tag to accurately recognize the usage of all the phrases constituting the sentence, Even if it is a word, it is processed so that it can be pronounced differently according to the use.

The pronunciation conversion rules used at this time apply the standard pronunciation method notified by the Ministry of Education.

The method of pronunciation notation conversion is based on the results of the first part marking tagging, assigning parts of speech to each of the input poems of the input word, and writing the position of the morpheme analysis result.

For example, the results of allocating parts of speech to report a god are shown in [Table 3].

[Table 3]

On the other hand, since the standard pronunciation method of Korean generates the pronunciation rule by the neutral 'ㅕ', 'ㅖ', and 'ㅢ', the pronunciation rule of three neutrality and the pronunciation rule of 27 utterance are created See the neutrality and neutrality of the verse and implement the rules.

For example, the rule of 'ㄱ' can occur as follows, as in [Table 4], the following transitions occur, depending on the part of the beginning.

[Table 4]

Sounding: The first sentence of the mother combined with a backstop 'b' is pronounced as a, c, g, i.

Consonant assimilation: 'b' is pronounced [d] before or after 'd'.

YEARS RULES: If a single pedestal or a pair of pedals is combined with a search, an ending, or a suffix that begins with a vowel, move to the next syllable with the original sound.

Therefore, in the case of the above example, the former 'pronounced' is pronounced as 'Sinoko' and the latter 'pronounced' is pronounced as 'notify' because no phenomenon occurs.

When the part-of-speech information of the morpheme analysis string extracted through the above-described processing is applied to the phonetic transcription conversion module 40, the phonetic transcription module 40 processes most pronunciation transitional phenomenon by rules, Exceptional pronunciation is defined as a word that corresponds to a phenomenon (eg, a door knob), a phenomenon of sound addition (for example, a cushion [sombibit]), or a circumstance phenomenon in a Chinese language (eg, thirst).

In this way, the pronunciation of the input sentence is obtained by using the part-of-speech information, but a semantic processing step is required to obtain the accurate pronunciation of the phrase.

Thus, although the pronunciation of the input sentence is obtained, the meaning processing step is required to obtain the correct pronunciation of the phrase.

Dependent grammar is used to find the dominance - dependence relationship of words that make sentences, unlike the sentence structure grammar, which can be expressed in general sentence structure grammar.

Therefore, we can use the parsing method using the old structure grammar as it is.

Since the parser 50 needs a nonterminal symbol that can represent the phrases when parsing, the parser 50 generates phrases for each of the phrases based on the tagged results.

The method of generating the phrase part is based on the connection between the leftmost and the rightmost part of the input word and if there are any sentence symbols such as a comma on the right side, the symbols are added.

For example, the word "bloom" is morphemically analyzed as p / vb + / ex + vx + + /ef+./se, and the part of speech is vbefse.

On the other hand, there are cases where two morpheme parts are added and become morpheme parts of another morpheme. For example, in the case of studying, the morpheme analysis is performed by study / na + ha / xv + da /ef+./se. Because it acts as a verb, it makes vbsfse a part of speech.

The rules for changing two morpheme parts into one morpheme part are as follows.

1. State Noun + Adjective Derived Suffix → Adjective

(Eg health / ns + ha / xj + da /ef+./se → vjefse)

2. Conjugation Noun + Noun Derived suffix → verb

(Eg study / na + ha / xv + da /ef+./se → vbefse)

3. Normal noun + noun Suffix → Normal noun

(For example, person / nc + s / xn → nc)

4. Stateful noun + adverb derivative suffix → adverb

(Eg simple / ns + l / xa → ad)

5. Verb + adverb derivative suffix → adverb

(Eg, precious / ns + ha / xj + cra / xa → ad)

After obtaining all the parts of the phrase for the sentences, the probabilistically optimal syntactic structure is searched using the CYK table.

In general, a dependency tree can be represented by a phrase structure tree. In particular, Korean is characterized by a principle of a postfix rule, so it can be represented by a simpler form of a phrase structure tree.

For example, the dependency tree for declaring a declaration of a sentence is shown in FIG. 7, and the corresponding phrase structure tree is as shown in FIG. 8 attached.

In this case, it can be observed that the rules used in the structure tree of FIG. 8 correspond to one of the following three modes.

1. S → A

(S: start symbol, A: part of speech)

2. A → BA

(A, B: part of speech)

3. A → a

(A: the word of the word, a: the word of the word)

The probability of obtaining the syntax tree shown in FIG. 8 is represented by the product of the rule probabilities of all the rules used in the syntax tree. In the implemented parser, both the first rule type and the third rule type have a probability of 1.0 Is used.

This is because the last word of the sentence always becomes the dominant part of the sentence, so that it does not affect the search for the optimal syntax tree. In addition, Will have no effect on the search.

Therefore, the syntax tree probability of the attached FIG. 8 becomes P (vbefse? Vbec vbefse) P (vbec? Ncpo vbec).

In the document analyzer executed with the functions as described above, the results of the performance evaluation of the pachinko tiger are as follows.

To evaluate the performance of the pacha tagger (30), the model was trained with 49,506 manually tagged learning corpus and tested for 4,729 words.

In the 49,506 phrases, 3,652 words have probability P ( | ). When a dictionary is constructed based on the remaining 45,854 words, if a word not found in the dictionary is found for 3,652 words, it is regarded as an unregistered word.

In this case, the performance of the speech tagger is as shown in [Table 5], and in [Table 5], the registered word and the unregistered word phrase refer to a word phrase including an unregistered word and an unregistered word, respectively.

[Table 5]

As described above, it can be observed that the accuracy rate of the registration word in Table 5 is not so high as 88.03%, which means that if the implemented tagger does not constitute a morphological lattice, It is due to assumptions.

For example, in a sentence that says to report a god, the correct tagging result of 'god' is 'god / normal god + subject / investigation'. If there is only information of god in dictionary, Verb + d / archetypal morpheme ', and it is assumed that there is no unregistered word, and then the tag is tagged as it is.

This kind of error has made the tagging results not very high for a word that does not include an unregistered word.

As a result of tagging after adding the words used in the experimental corpus, the accuracy rate was 96.46% for the word and 98.01% for the morpheme.

From the results of this experiment, the higher the number of headwords in the dictionary, the higher the performance of the tagger.

The performance evaluation test results of the parser 50 are as follows.

The model was learned based on the syntactic trees of 498 sentences which were created manually, and then 100 sentences were tested.

100 sentences consist of 738 words, and the last word of the sentence always determines itself as the dominant place, so there are 638 dominance - dependent relationships.

In the first case, the performance of the parser itself was evaluated and the result of the experiment using the result of the pragmatic tagger was used. In the former case, the accuracy was 80.87%, and in the latter case, the accuracy was 78.68% .

For example, for a sentence of about 11 words, eight words in the verse except the last verse can be regarded as the correct choice of their own dominance.

As described above, the document analyzer according to the present invention performs preprocessing using an amorphous finite automata for an input document, and then obtains all possible morpheme analysis results of the word in the morpheme analysis step, After selecting the most probable morphological analysis result, the selected result is assigned to all the prosodic phrases through the phonetic transcription module, and then the rules of three neutrality and 27 verbs are applied. After obtaining the phonetic notation, we use the probability-dependent grammar of the parser to find the dominance-dependence relationship of input phrases and generate a rhyme to provide reliable document speech conversion.

The document analyzer according to the present invention shows a precision of 85.11% in the parser tag and a precision of 78.68% in the parser, and in particular, in case of the parser tag, it provides a high accuracy of 96.46% when there is no unlisted word in the document being processed .

In addition, the document analyzer according to the present invention can be used to construct a phonetic dictionary for a speech recognition system in addition to a document speech conversion system.

Claims (16)

A document speech conversion system comprising: preprocessing means for extracting a document inputted for speech conversion one sentence at a time and converting non-Korean characters into Hangul characters using an amorphous finite automata; Morphology analyzing means for obtaining all possible morpheme analysis results in one sentence word dictionary applied by the preprocessing means; A morphing tiger for extracting an optimal morpheme analysis sequence by selecting the morpheme analysis result as the most likely morphological analysis result based on probability information; A phonetic transcription conversion unit that obtains phonetic transcription of each of the phrases using the morphological analysis string applied to the morphological phrase analysis unit; And a syntactic analysis means for obtaining a dominant-dependence relationship of the phrases by using a probability-dependent grammar and outputting a phonetic representation of a word of the input document and a morpheme part sequence and a dependency tree as a final result, Document analyzer for document speech system. [2] The document analyzer according to claim 1, wherein the parsing tag and parsing means are provided with probability information for parsing tagging, and probability information for parsing each parsing tag to perform parsing solving. [2] The method according to claim 1, wherein the preprocessing unit divides non-Korean characters into English, an English abbreviation, a number, a telephone number, a year, and a time, Document analyzer for voice conversion system. [Claim 2] The method according to claim 1, wherein the preprocessing unit is implemented by an amorphous finite automata consisting of 48 states, and 12 of them are composed of non-Korean Hangul conversion modules having different final states Document analyzer. 2. The system of claim 1, wherein the morpheme analyzing means directly forwards the phonetic transcript to the phonetic transcription means when the phonetic transcription means refers to a word read from the input document, Document analyzer. The morpheme analyzing apparatus according to claim 1, wherein the morpheme analyzing means judges that there is an unregistered word in the analyzed word if the morpheme analysis fails on the sentence applied by the preprocessing means, estimates the unregistered word, and minimizes the candidate using the linguistic heuristic A document analyzer for a Korean document speech conversion system. The document analyzer according to claim 1, wherein the lattice structure algorithm set in the morpheme analysis means is as follows. [2] The apparatus according to claim 1, wherein the morpheme analyzing means calculates the irregularity and shortening phenomenon positions of the input word in advance and then morphologically analyzes it according to a lattice construction algorithm and expresses the analyzed result as a morphological lattice. Document analyzer for. The document analyzer according to claim 1, wherein the morpheme analyzing unit forms a morpheme lattice by finding a morpheme from the right side to the left side of the input sentence word. [Claim 2] The method according to claim 1, wherein the morpheme analyzing means uses the syllable information to exclude the node from the grid if the last syllable of the unregistered word is not used as the last syllable of the estimated part-of- Document analyzer. The morpheme analyzing apparatus according to claim 1, wherein the morpheme analyzing means uses a clue morpheme to detect a morpheme that is highly incoincident in an incomplete morpheme lattice and is considered to be sufficient to estimate the composition of the morpheme, And removing the rest from the grid. The document analyzer according to claim 1, wherein the part-of-speech tagger is implemented as follows. [Claim 2] The method according to claim 1, wherein the phonetic transcription conversion means assigns parts of speech to all the phrases constituting the phrase, and then obtains the phonetic transcription of the phrases as a rule based on three neutrality and 27 utterances, Document analyzer for system. 2. The system according to claim 1, wherein the phonetic transcription conversion means defines words corresponding to a phenomenon of recurring phenomenon, a phenomenon of addition of sound, and a recurring phenomenon in a Chinese language to an exceptional pronunciation word in a compound word Analyzer. The document analyzer according to claim 1, wherein the parsing means generates a basic word phrase by connecting a morpheme word positioned at the leftmost position of the input word and a morpheme word positioned at the rightmost position, . The document analyzer according to claim 1, wherein the parsing unit generates a phrase part by adding a symbol to a punctuation mark such as a comma on the right side.