JP2002221985A - Voice recognition device, morpheme analyzer, kana (japanese syllabary)/kanji (chinese character) converter, its method and recording medium which records its program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a voice recognition device or the like which precisely distinguishes a specialty field input from a general field input. SOLUTION: The device is provided with an inputting means 101 which is used to input voice, a phoneme probability computing means 102 which converts voice signals outputted by the means 101 into phonemes, computes a phoneme occurring probability corresponding to each phoneme and outputs phoneme string candidates, a word probability computing means 103 which refers to a phoneme string of an object language, an word mentioned string corresponding to the phoneme string and a phoneme n-gram which stores the occurring probability and computes the word occurring probability of each word candidate corresponding to the phoneme string candidates outputted by the means 102 and a word outputting means 104 which computes a word string candidate that is similar to the voice inputted by the means 101 employing the phoneme occurring probability computed by the means 102 and the word occurring probability computed by the means 103. The phoneme n-gram has specialty and general field phoneme n-gram 106 and 107, respectively.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to speech recognition, morphological analysis, or kana-kanji conversion based on n-gram, which is the probability of occurrence of a character or word in a target language, using statistics of a natural language. The present invention relates to an apparatus, a method, and a storage medium for performing the method. Here, the n-gram is an N-gram model as described in "Probabilistic Language Model", Kenji Kita, published by The University of Tokyo Press on November 12, 1999, (Prior Art Document 1).

[0002]

2. Description of the Related Art Analysis techniques using statistics of natural languages are applied to many document processes. For example, Japanese input by voice recognition is useful as a means for inputting a document, and further improvement in recognition accuracy is desired. In order to accurately recognize speech, a method using n-gram as a language model has attracted attention. However, since the constraint in the n-gram is affected by the order N, the constraint becomes weaker as N becomes smaller. vice versa
Increasing the degree N of the n-gram has a serious problem that the frequency counting table becomes huge, and a very large collection of example sentences is required to secure reliable statistics. There was a problem of becoming.

[0003] In order to solve the increase in the number of n-gram tables, a compression method as disclosed in Japanese Patent Application Publication No. 10-501078 has been proposed.

Hereinafter, a conventional example of the present invention will be described.
FIG. 26 is a block diagram showing an example of this type of conventional device. 101 is a microphone, 102 is a phoneme probability calculation device, 1
04 is an output device, 105 is a RAM for storing information, 40
1 is a word prediction device, and 402 is a 3-gram table.

[0005] The generation of word string candidates is based on the probability P (W | Y) of the word string.
Is obtained by calculating W that maximizes The probability of the word string is obtained from the following equation.

[0006]

(Equation 1)

At this time, W is an uttered word string, and Y
Is a phoneme sequence. Since it is sufficient to find W that maximizes P (W | Y), P (Y) common to word strings on the right side can be omitted, and W that maximizes P (Y | W) P (W) Should be obtained. P (Y | W)
Is the probability of appearance of the phoneme sequence when the word sequence W is given, and P (W) is the probability of appearance of the word sequence. Time t = 1,2, ...,
When the phoneme sequence corresponding to the word sequence W in L is determined by the following equation,

[0008]

(Equation 2)

P (Y | W) can be calculated from the phoneme probability by the following equation.

[0010]

(Equation 3)

The appearance probability P (W) of a word string is expressed as follows when the word string W of m words is determined by the following equation.

[0012]

(Equation 4)

Approximately from the following equation (probability of word 3-gram) independently of the phoneme probability.

[0014]

(Equation 5)

According to the above calculation, 3-gr
For words having a word string in the am index, W that maximizes the word string probability P (W | Y) is calculated. The appearance probability of each word is calculated based on the frequency values stored in advance in the 3-gram table of the word.

FIG. 27 shows an example of a 3-gram table. A structure in which three-word notation, phoneme notation, kana notation, and part of speech are pointed to using a three-word phoneme notation sequence as an index.
A frequency value is stored for each word in the index (see n-gram table 402).

The calculated word string W is output from the output device 104 as a recognition result.

Next, the process of recognizing speech using a conventional apparatus will be described with reference to a schematic flowchart of an operation and a specific example. FIG. 28 is a schematic flowchart of the operation of the conventional speech recognition apparatus.

Speak to the microphone 101 to start processing (ST3)
001). The microphone converts the input voice to an electrical signal (S
T3002, ST3003).

The phoneme probability calculating device 102 converts the electric signal into an A / D signal.
After conversion, quantization, and spectrum analysis, the recognition results separated into syllable units are concatenated and stored in the RAM 1 as phoneme string candidates.
05 (ST3004).

The word prediction device 401 takes out one phoneme string candidate from the RAM 105 and initializes the first word string (S
T3005).

The 3-gram information corresponding to the search key is
A search is made from the gram table 402 (ST3006).

Based on the retrieved 3-gram information, a probability value of three chains of words is calculated (ST3007).

Based on the probability value, the word string having the highest probability for the corresponding phoneme string candidate is stored in the RAM 105 (S
T3008).

The above calculation is performed for all phoneme string candidates, and the most probable word string and phoneme string candidates are selected (S
T3009).

In this way, a word string having a high probability of being similar to the utterance is obtained.

[0027]

A conventional device of this kind has been constructed as described above. By the way, in consideration of practical processing in a specialized field, not all inputs are in a specialized field. For example, even in the case of medical records in medical practice, the preamble is specialized, such as "I suspected photosensitivity epilepsy and take EEG and CT. It seems that I became conscious while playing a video game at home." Against
The latter sentence is a general field expression. To cope with such an example, conventionally, it was necessary to mix a specialized n-gram and a general n-gram.

However, in the conventional technique described above, n-gr
am was the only one created, so it was not possible to accurately discriminate between inputs in specialized fields and general fields.

This will be described with reference to the following specific example. In speech recognition, medical input "photosensitive epilepsy"
Consider a case in which the phrase “transformation of industrial structure” is input separately in a general field. The n-gram is extracted from the corpus, and the order of N is assumed to be 2 in order to obtain statistical reliability. For simplicity, the syllable recognition result and the morpheme delimiter are "Hikarikabinsei-no-epilepsy". It is assumed that "photosensitivity" is correctly analyzed. Assuming that "epilepsy" is selected using "n-gram" for "epilepsy" next to "no", the "transformation" part of the input such as "transformation of industrial structure" is similarly "no epilepsy" As a result, the only n-gram makes a mistake as "epilepsy".

As described above, since only the n-gram is created in the prior art, there is a problem that it is not possible to accurately discriminate an input between a specialized field and a general field.

The present invention has been made in order to solve the above-mentioned problems, and has been developed to provide a speech recognition device, a morphological analysis device, a kana-kanji conversion device, and the like which can accurately discriminate an input between a specialized field and a general field. The purpose is to provide.

[0032]

SUMMARY OF THE INVENTION In view of the above-mentioned object, the present invention provides an input means for inputting voice, and converts a voice signal output from the input means into phonemes, and calculates a phoneme occurrence probability corresponding to each phoneme. A phoneme probability calculating means for calculating and outputting a phoneme sequence candidate; a phoneme n-gram storing a phoneme sequence of a target language, a word notation sequence corresponding to the phoneme sequence, and an occurrence probability; and the phoneme n-gram.
With reference to the word probability calculation means for calculating the word occurrence probability of each word candidate corresponding to the phoneme sequence candidate output by the phoneme probability calculation means, the phoneme occurrence probability calculated by the phoneme probability calculation means, A word output means for calculating a word string candidate similar to the voice input from the input means, using the word occurrence probability calculated by the word probability calculation means, wherein the phoneme n A speech recognition device characterized in that it has a phoneme n-gram in both a specialized field and a general field.

The speech recognition apparatus according to claim 1, wherein the word probability calculation means is configured to be able to adjust the weight of the probability between the specialized field and the general field.

Further, the word output means is configured such that, when calculating the word string candidates, all fields in the phoneme n-gram corresponding to a series of voices match.
Or the speech recognition device according to 2.

An input means for inputting a kana-kanji mixed character string, a kana-kanji mixed character string of a target language, a word notation string corresponding to the kana-kanji mixed character string, and a kanji n-gram for storing the occurrence probability and the kanji character With reference to the n-gram, a word probability calculating means for calculating a word occurrence probability of each word candidate corresponding to the kanji-mixed character string output by the input means, and a word occurrence probability calculated by the word probability calculating means. make use of,
A word output means for calculating a word string candidate that matches the character string input from the input means, wherein in the kanji n-gram, the kanji n in both the specialized field and the general field A morphological analyzer characterized by having a -gram.

A morphological analyzer according to claim 4, wherein said word probability calculating means is configured to be able to adjust the weight of the probability between the specialized field and the general field.

The word output means is configured such that, when calculating the word string candidate, all fields in the kanji n-gram corresponding to a series of kana-kanji mixed character strings match. Or the morphological analyzer according to 5.

Input means for inputting a kana character string;
The kana character string of the target language, the word notation string corresponding to the kana character string, and the kana n-gram storing the occurrence probability and the kana n-gram are referred to, and each of the kana character strings output by the input means is referred to. Using a word probability calculating means for calculating a word occurrence probability of a word candidate, and a word occurrence probability calculated by the word probability calculating means, calculating a word string candidate that matches the kana character string input from the input means A kana-kanji conversion device comprising: a kana-kanji conversion device comprising: a kana-kanji conversion device having word output means.

The kana-kanji conversion device according to claim 7, wherein the word probability calculation means is configured to be able to adjust the weight of the probability between the specialty field and the general field.

Further, when calculating the word string candidate, the word output means is configured so that all fields in the kana n-gram corresponding to a series of kana character strings match. In the kana-kanji conversion device described in (1).

When a voice is input, the voice signal is converted into a phoneme, a phoneme occurrence probability corresponding to each phoneme is calculated, and a phoneme sequence candidate is output. Memorize the corresponding word expression string and occurrence probability, and
With reference to the n-gram and the phoneme n-gram, a step of calculating the word occurrence probability of each word candidate corresponding to the phoneme sequence candidate, and using the phoneme occurrence probability and the word occurrence probability, Calculating a similar word string candidate, which is characterized in that the phoneme n-gram has phonemes n-grams in both a specialized field and a general field.

Further, the weight of the probability can be adjusted for each field when calculating the word candidate.
0.

12. The speech recognition method according to claim 10, wherein, when the word string candidates are calculated, all fields in a phoneme n-gram corresponding to a series of speeches match. is there.

When a kana-kanji mixed character string is input, a kana-kanji mixed character string of the target language, a word notation string corresponding to the kana-kanji mixed character string, and the occurrence probability are stored.
With reference to the n-gram and the kanji n-gram, a step of calculating the word occurrence probability of each word candidate corresponding to the kanji mixed character string, and using the calculated word occurrence probability, Calculating a matching word string candidate, wherein the kanji n-gram has kanji n-grams in both a specialized field and a general field. .

Further, the weight of the probability can be adjusted for each field when calculating the word candidate.
3. The morphological analysis method according to 3.

Further, at the time of calculating said word string candidate, all fields in the kanji n-gram corresponding to a series of kana-kanji mixed character strings are made to match.
Or a morphological analysis method according to item 14.

When the kana character string is input, the kana character string is referred to by referring to the kana n-gram and the kana n-gram storing the kana character string of the target language, the word notation string corresponding to the kana character string, and the occurrence probability. Calculating a word occurrence probability of each word candidate corresponding to the sequence, and calculating a word string candidate that matches the input kana character string using the calculated word occurrence probability,
, Wherein the kana n-gram has a kana n-gram in both a specialized field and a general field.

Further, when calculating the word candidates, the weight of the probability can be adjusted for each field.
6 in the kana-kanji conversion method.

18. The method according to claim 16, wherein, when calculating the word string candidates, all fields in a kana n-gram corresponding to a series of kana character strings match.
In the Kana-Kanji conversion method described in (1).

An input function for inputting a voice, and a phoneme probability calculating function for converting a voice signal output from the input function into a phoneme, calculating a phoneme occurrence probability corresponding to each phoneme, and outputting a phoneme string candidate With reference to the phoneme sequence of the target language, the word notation sequence corresponding to the phoneme sequence, and the phoneme n-gram storing the occurrence probability and the phoneme n-gram, the phoneme sequence candidate output by the phoneme probability calculation function is supported. Using the word probability calculation function to calculate the word occurrence probability of each word candidate to be calculated, the phoneme occurrence probability calculated by the phoneme probability calculation function, and the word occurrence probability calculated by the word probability calculation function, And a word output function of calculating a word string candidate similar to the input speech, and wherein the phoneme n-gram has a phoneme n-gram of both a specialized field and a general field.
It is on a computer-readable recording medium on which a speech recognition program is recorded.

20. The recording medium according to claim 19, wherein in the word probability calculating function, the weight of the probability can be adjusted for each classification.

A word output function is configured so that all fields in a phoneme n-gram corresponding to a series of voices coincide with each other when the word string candidates are calculated.
The recording medium according to 9 or 20.

An input function for inputting a kana-kanji mixed character string, a kana-kanji mixed character string of a target language, a word notation string corresponding to the kana-kanji mixed character string, and a kanji n-gram for storing the occurrence probability and the kanji character With reference to the n-gram, a word probability calculation function for calculating the word occurrence probability of each word candidate corresponding to the kanji mixed character string output by the input function, and a word occurrence probability calculated by the word probability calculation function make use of,
A word output function of calculating a word string candidate that matches the character string input from the input function.
A gram has kanji n-grams in both a specialty field and a general field, and is on a computer-readable recording medium recording a morphological analysis program.

23. The recording medium according to claim 22, wherein the word probability calculation function is configured so that the weight of the probability can be adjusted for each field.

23. A word output function according to claim 22, wherein said word output function is configured such that when calculating said word string candidates, all fields in the kanji n-gram corresponding to a series of kana-kanji mixed character strings match. Or the recording medium according to 23.

An input function for inputting a kana character string;
The kana character string of the target language, the word notation string corresponding to the kana character string and the kana n-gram storing the occurrence probability and the kana n-gram are referred to, and each of the kana character strings output by the input function is referred to. Using a word probability calculation function for calculating a word occurrence probability of a word candidate and a word occurrence probability calculated by the word probability calculation function, a word string candidate that matches the kana character string input from the input function is calculated. And a word output function, wherein the kana n-gram has a kana n-gram in both a specialized field and a general field, and is in a computer-readable recording medium that records a kana-kanji conversion program. .

26. The recording medium according to claim 25, wherein in the word probability calculating function, the weight of the probability can be adjusted for each field.

27. The word output function according to claim 25, wherein the word output function is configured such that all fields in the kana n-gram corresponding to a series of kana character strings match when the word string candidates are calculated. Above.

[0059]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be described below according to each embodiment. Embodiment 1 FIG. FIG. 1 is a block diagram showing a speech recognition apparatus according to Embodiment 1 of the present invention. In the figure, 10
1 is a microphone, 102 is a phoneme probability calculation device, 103 is a word probability calculation device, 104 is an output device, 105 is a RAM for storing information, and 106 is a phoneme n-
A gram 107 is a phoneme n-gram in a general field. This device can be configured by a computer, for example.
Reference numerals 105 to 107 denote memories and databases;
4 can be composed of a program. FIG. 17 is a schematic flowchart of the operation of the voice recognition device of the present embodiment.

The generation of word string candidates is based on the word string probability P (W | Y)
Is obtained by calculating W that maximizes The probability of a word string is obtained from the following equation.

[0061]

(Equation 6)

At this time, W is an uttered word string, and Y
Is a phoneme sequence. Since it is sufficient to find W that maximizes P (W | Y), P (Y) common to word strings on the right side can be omitted, and W that maximizes P (Y | W) P (W) Should be obtained. P (Y | W)
Is the probability of appearance of the phoneme sequence when the word sequence W is given, and P (W) is the probability of appearance of the word sequence. Time t = 1,2, ...,
When the phoneme sequence corresponding to the word sequence W in L is determined by the following equation,

[0063]

(Equation 7)

P (Y | W) can be calculated from the phoneme probability by the following equation.

[0065]

(Equation 8)

Further, the appearance probability P (W) of the word string is obtained when the word string W of m words is determined by the following equation.

[0067]

(Equation 9)

Independently of the phoneme probability, it is obtained from the following equation (probability of phoneme n-gram).

[0069]

(Equation 10)

According to the above calculation, the phoneme of the phoneme sequence candidates
Word string probability for words with word strings in n-gram
Calculate W that maximizes P (W | Y). The calculation of the combination is described in, for example, "Speech Recognition by Probabilistic Model", Viterbi shown by Seichi Nakagawa, (prior art document 2).
Alternatively, it may be performed at a high speed using a method of stack decoding. Also, the calculation formula may be calculated as a sum using the probability as a logarithmic probability.)
-Use the value stored in the gram in advance.

Here, FIG. 3 shows an example of the phoneme n-gram 109 used for analyzing the example sentence 108 in FIG.

As shown in FIG. 3, the phoneme n-gram records an n-gram in the field of baseball and an n-gram in the general field as specialized fields. There are 2-gram and 1-gram n-grams, and the first phoneme sequence is a key for search. In 2-gram,
The leading morpheme, the trailing morpheme, and the probability are recorded for this key. The probability recorded here is the probability that the next morpheme connects after the preceding morpheme, and
It corresponds to the occurrence probability of gram. In the 1-gram, the information of the morpheme connected directly next is recorded. The 1-gram probability is the occurrence probability of the morpheme itself. Note that a morpheme is represented by a set of notation, phoneme notation, headline reading, and part of speech.

The output device 104 outputs the calculated word string W as a recognition result.

Next, a process of extracting a recognized word string using the present apparatus will be described with reference to a schematic flowchart of FIG. 17 and a specific example.

First, the sound is converted from the microphone 101 into an electric signal and taken in as analog data (ST3001, ST300
2, ST3003).

The phoneme probability calculating device 102 performs A / D conversion on the analog data captured by the microphone, quantizes the analog data, and then performs spectrum analysis (for details, for example, by various methods described in the prior art document 2) to obtain a phoneme unit. Are output as phoneme string candidates. The phoneme string candidates represent the likelihood of each phoneme corresponding to the analog data taken in from the microphone by a probability value, output as a pair of a chained phoneme chain and the acoustic likelihood of the chain, and store this in the RAM 105 ( ST3004).

In this embodiment, it is assumed that the following are output as the phoneme chain and the acoustic likelihood.

# SaNgooaacinoseNsee # 0.9 # saNgooacinoseNsee # 0.1 (Log likelihood may be used instead of probability for acoustic likelihood. For phoneme chains, an efficient storage method such as lattice may be used.)

The word probability calculation device 103 stores the phoneme sequence candidates and the acoustic likelihood output from the phoneme probability calculation device 102 in the RAM 1.
05, and an initialization process is performed. As initialization processing, the null word “{#### sentence start}” and its probability value “1” are stored in the RAM 105 as the initial language likelihood value of the preceding word string candidate (ST3005). Here, “# saNgooaacinoseNsee #” is first extracted as a phoneme sequence candidate.

The word probability calculating device 103 checks whether all preceding word string candidates correspond to the terminal phonemes of the phoneme string candidates, and if all correspond, goes to ST4007, and if not, performs the following processing. (ST4001).

One preceding word string candidate is extracted from the RAM 105 (ST4002). In the present embodiment, first, “@ ##
# Sentence # ”is extracted as a preceding word string candidate.

The n-gram phoneme index 106 is searched based on phoneme string information of the preceding word string candidate. In the case of the present embodiment, first, an initial preceding word string “{#### sentence start}” is searched. It is checked whether there is a backward word that matches forward in the substring of the phoneme string candidate after the searched preceding word. If there is no backward matching word, the process proceeds to ST4001, and if there is a backward matching word, the following process is performed (ST4003).

In the present embodiment, the preceding word string “$ ##
# Sentence n-gram as the backward word of "
A word in which the phoneme sequence from the beginning of “saNgooaa ...” following “#” partially matches is searched for and is set as a backward word. In 2-gram, "#saNgoo" matches the phoneme sequence "#saNgooaa ..." in front, so this 2-gram suffix morpheme "specialty: No. 3 saNgoo sangu noun" is one of the backward word candidates. I do. Also, 1
-Gram's “｛specialty: No. 3 saNgoo sango noun｝” is a candidate because it matches forward with the back phoneme sequence. In addition, “｛General: No.3 saNgoo Sango Noun｝” is also a candidate.

Although partial matching is used for simplicity in the present embodiment, DP matching processing or "2nd considering the difference tendency between the optimal solution and the correct solution in the first stage" is used for similarity search with an ambiguous phoneme chain. Abe et al., Ongaku Lecture, 1-R-15, 1998.9
A technique as shown in the publication (prior art document 3) may be used.

The likelihood is similarly calculated for each backward word and stored in the RAM, and the backward word is connected to the preceding word string. At this time, in the case of 2-gram, the field is made to be the same as the preceding morpheme, and in the case of 1-gram, since there is no connection, the field may be switched. A new preceding word string is stored in RAM (ST400
Four).

If the entire phoneme sequence corresponds to the preceding word sequence, the maximum likelihood and the preceding word sequence of the solution are stored in the RAM.
Return to ST4001 (ST4005 ST4006).

In the present embodiment, the preceding word string “$ ##
Replace “# sentence｝” with “｛special: # ## sentence｝ ｛specialty: No. 3 saNgoo sango noun｝”. The linguistic likelihood is the probability 1 of the preceding word string “｛#### sentence 文” and the specialty field “specialty: #### sentence｝ ｛
It is calculated by the following formula from the probability 0.01 of 2-gram of "No. 3 saNgoo sango noun｝".

[Equation 11]

If no matching word has been obtained for all phoneme string candidates, the process proceeds to ST3005 and the same processing is repeated. If obtained, the following processing is performed (ST400).
7).

In this embodiment, as described above, “{###### sentence start} {specialization: No. 3 corresponding to phoneme sequence candidates
saNgoo sangu noun｝ ｛specialty: arch aaci aachi noun｝ ｛specialty: no no particle｝… ”

Next, the word string of the solution having the maximum likelihood stored in the RAM is read. The maximum likelihood is the maximum value of the product of the linguistic likelihood and the acoustic likelihood (ST4008).

In the present embodiment, as a result of the calculation, the phoneme string candidate “# saNgooachinoseNsei #” is discarded because the corresponding phoneme n-gram does not exist. Phoneme string candidate `` #saNgooaachino
For “seNsei #”, “｛# # # sentence head｝ ｛
No. 3 saNgoo sango noun {specialist: arch aaci Aichi ordinary noun} {specialty: no no connecting particle} {specialty: pre-emptive seNsee sensei sa variant noun}, the maximum likelihood is 5.4 × according to the above equation (10). 10 to the -9th power (acoustic likelihood 0.9
Language likelihood 6 × 10 to the power of -9).

Only the notation is extracted from the word string of the solution, output from the output device, and the process ends (ST4009, ST3014).
In the present embodiment, “the pre-emption of No. 3 arch” is obtained.

In this way, the "no preemptive"
Despite having a higher 2-gram probability of "no-sensei" than that, it is possible to construct an n-gram with strong language constraints without increasing the order of the n-gram by separating the fields and collecting statistics. High-accuracy speech recognition can be configured.

Embodiment 2 Next, another embodiment of the present invention will be described. This embodiment is an example in which a morphological analysis device is configured by configuring a kanji n-gram. FIG. 4 is a block diagram showing a morphological analyzer according to this embodiment. The same or corresponding parts as those in the above embodiment are denoted by the same reference numerals (the same applies hereinafter). 104 is an output device,
201 is a file input device, 202 is a morpheme probability calculation device, 203 is a kanji n-gram in a specialized field, and 204 is a kanji n-gram in a general field. Note that this device can also be constituted by, for example, a computer.
04 is a memory or database, 201, 202, 104
Can be composed of programs. FIG. 18 is a schematic flowchart of the operation of the morphological analyzer according to the present embodiment.

The generation of the word string candidate is obtained by calculating W that maximizes the probability P (W) of the word string.

At this time, W is an input word string, and P
(W) is the appearance probability of the word string. Also, the appearance probability P (W) of the word string, when the word string W of m words is determined by the following equation,

[0097]

(Equation 12)

It is obtained from the following equation (using the probability of the kanji n-gram).

[0099]

(Equation 13)

The word sequence probability P (W) is maximized for a word sequence existing in the kanji n-gram by the above calculation.
Calculate W. For the calculation of the combination, for example,
The processing may be performed at high speed using the Viterbi method described in "Natural Language Processing", written by Makoto Nagao, Iwanami Shoten (Prior Art Document 4). Further, the calculation formula may be calculated as a sum with the probability being a logarithmic probability. The appearance probability of each word is calculated based on the probability value stored in advance in the kanji n-gram of the word.

FIG. 5 shows an example of the kanji n-gram and the corpus 205 created based on the example sentence 108 of FIG.

As shown in FIG. 5, an n-gram in a specialized field and an n-gram in a general field are recorded in the kanji n-gram. n-
There are 2-gram and 1-gram grams, and the first kanji string is the key for search. In the 2-gram, the leading morpheme, the trailing morpheme, and the probability are recorded for this key. The probability recorded here is the probability of the connection of the next morpheme after the previous morpheme, and corresponds to the occurrence probability of the 2-gram. In the 1-gram, information on the morpheme connected directly next is recorded. The 1-gram probability is the occurrence probability of the morpheme itself. Note that a morpheme is represented by a set of notation, phoneme notation, headline reading, and part of speech.

The output device 104 outputs the calculated word string W as a recognition result.

Next, the process of extracting a recognized word string using the present apparatus will be described with reference to the flowchart of FIG. 18 and a specific example.

First, a kanji character string is input from the file input device 201 (ST3001, ST5001).

The morpheme probability calculation device 202 receives the kanji character string captured by the file input device 201 as an input,
This is stored in M105 (ST5002). In the present embodiment, it is assumed that the following has been input as a kanji character string.

Teacher of No. 3 arch

The morpheme probability calculation device 202 retrieves the output kanji string candidates from the RAM and performs an initialization process. As the initialization process, the null word “｛####
The sentence head｝ ”and its probability value“ 1 ”are stored in the RAM as initial values of preceding word string candidates (ST5003). Here, first,
“# 3 arch teacher #” is extracted as a kanji string candidate.

The morpheme probability calculating device 202 checks whether all preceding word string candidates correspond to the terminal kanji of the kanji string candidate, and if all correspond, goes to ST4008, and if not, performs the following processing ( ST5004).

Take out one preceding word string candidate from RAM
(ST5005). In the present embodiment, “｛####” first
Sentence｝ ”is extracted as a preceding word string candidate.

The n-gram kanji index 203 is searched by the kanji string information of the preceding word string candidate. In the case of the present embodiment, first, an initial preceding word string “{#### sentence start}” is searched. It is checked whether there is a backward word that matches forward in the partial string of the kanji string candidate after the searched preceding word. If there is no backward word that matches the beginning, the process proceeds to ST5004, and if there is a backward word that matches the beginning, the following process is performed (ST5006).

In the present embodiment, the preceding word string “$ ##
An n-gram index is searched for as the backward word of "# sentence #", and a word in which the kanji string from the beginning of "3 no...." 2-gr
In am, "# 3 No." matches the kanji string of "# 3 No ....", so the morpheme of this 2-gram "specialty: No. 3 saNgoo
"Sango noun" is one of the candidates for the backward word. Also, the 1-gram “｛specialty: No. 3 saNgoo sango noun 前方” is a candidate because it matches the front of the back kanji string.
In addition, “｛General: No.3 saNgoo Sango Noun｝” is also a candidate.

Similarly, the likelihood is calculated for each backward word and stored in the RAM, and the backward word is connected to the preceding word string. At this time, in the case of 2-gram, the field is set to be the same as the preceding morpheme, and in the case of 1-gram, switching of the field is allowed because there is no connection. Then, it is newly stored in the RAM as a preceding word string (ST5007).

If the entire kanji string corresponds to the preceding word string, the maximum likelihood and the preceding word string of the solution are stored in the RAM.
Return to ST5004 (ST5008 ST5009).

In the present embodiment, the preceding word string "@ # #
Replace “# sentence｝” with “｛special: # ## sentence｝ ｛specialty: No. 3 saNgoo sango noun｝”. The linguistic likelihood is the probability of 1 for the preceding word string “｛#### sentence｝” and the probability of 2-gram for the specialty field “｛#｝, ｛3｝”
It is calculated by the following formula from 0.01.

[0116]

[Equation 14]

If no matching word has been obtained for all the preceding matrix candidates, the process moves to ST5005 and the same process is repeated. If it has been obtained, the following processing is performed (ST5004).

In the present embodiment, as described above, "{#### sentence head} {specialization: No. 3 corresponding to phoneme sequence candidates]
saNgoo sangu noun｝ ｛specialty: arch aaci aachi noun｝ ｛specialty: no no particle｝… ”

Next, the word string of the solution having the maximum likelihood stored in the RAM is read (ST4008). In the present embodiment, the kanji character string candidate “# 3
# Sentence head｝ ｛No. 3 saNgoo sango noun｝ ｛Arch a
aci aichi noun｝ no no no connecting particle｝ ｛preemptive seNs
The result of the morphological analysis of `` ee Sensei noun｝ '' shows that the maximum likelihood is 5.4 × 10 to the -10th power (the acoustic likelihood 0.
9 Language likelihood 6 × 10 to the power of -9).

The morpheme sequence of the solution is taken out, output from the output device, and the process is terminated (ST4009, ST3014).

In this way, it is possible to calculate that the part-of-speech is in the meaning of "preemptive" even with the ambiguous notation "partially". By separating the fields and obtaining statistics as described above, an n-gram with a strong language constraint can be constructed without increasing the order of the n-gram, and a highly accurate morphological analysis can be constructed.

Embodiment 3 FIG. Next, another embodiment of the present invention will be described. This embodiment is an example in which a kana-kanji conversion device is configured by configuring a kana n-gram. FIG. 6 is a block diagram showing a kana-kanji conversion device according to this embodiment. Reference numeral 301 denotes an input device, for example, a keyboard (actually including an I / O portion on the computer side); 302, a kanji probability calculation device; 303, a kana n-gram in the field of baseball as a specialty; gram
It is. This device can also be constituted by a computer, for example, and 105, 303, 304 can be constituted by a memory or a database, and 302, 104 can be constituted by a program. FIG. 19 is a schematic flowchart of the operation of the kana-kanji conversion device of the present embodiment.

Generation of a word string candidate is obtained by calculating W that maximizes the probability P (W) of the word string.

At this time, W is an input word string, and P
(W) is the appearance probability of the word string. Also, the appearance probability P (W) of the word string, when the word string W of m words is determined by the following equation,

[0125]

(Equation 15)

It is obtained from the following equation (using the probability of the pseudonym n-gram).

[0127]

(Equation 16)

The word string probability P (W) is maximized for a word string existing in the kana n-gram by the above calculation.
Calculate W. For the calculation of the combination, for example,
It may be performed at high speed using the Viterbi method disclosed in the above-mentioned prior art document 4. Further, the calculation formula may be calculated as a sum with the probability being a logarithmic probability. The appearance probability of each word is calculated based on a probability value stored in advance in the kana n-gram of the word.

Here, FIG. 7 shows an example of a pseudonym n-gram created based on the example sentence 108 of FIG.

As shown in FIG. 7, an n-gram in a specialized field and an n-gram in a general field are recorded in the pseudonym n-gram. n-
There are 2-gram and 1-gram grams, and the first kanji string is the key for search. In the 2-gram, the leading morpheme, the trailing morpheme, and the probability are recorded for this key. The probability recorded here is the probability of the connection of the next morpheme after the previous morpheme, and corresponds to the occurrence probability of the 2-gram. In the 1-gram, information on the morpheme connected directly next is recorded. The 1-gram probability is the occurrence probability of the morpheme itself. Note that a morpheme is represented by a set of notation, phoneme notation, headline reading, and part of speech.

Using the calculated word string W as a recognition result (Kanji)
Output from the output device 104.

Next, the process of extracting a converted word string using the present apparatus will be described with reference to the schematic flowchart of the operation shown in FIG. 19 and a specific example.

First, for example, a kana character string is input from the keyboard 301 (ST3001, ST6001).

The kanji probability calculation device 202 uses the keyboard 3
The kana character string fetched from 01 is input and stored in the RAM (ST6002). In the present embodiment, it is assumed that the following are input to the kana chain.

[0135] Teacher of Sango Art

The kanji probability calculation device 302 extracts the kana character string output from the file input device, that is, the keyboard 301, from the RAM and performs an initialization process. In the initialization process, the null word “{#### sentence}} and its probability value“ 1 ”are set as the initial value of the preceding word string candidate in the RAM.
(ST6003). Here, first, “# Sango Uchi no Sensei #” is extracted as a kana character string.

The kanji probability calculating device 302 checks whether all preceding word string candidates correspond to the terminal kana of the kana character string, and if all correspond, goes to ST4008, and if not, performs the following processing. (ST6004).

Retrieve one preceding word string candidate from RAM
(ST6005). In the present embodiment, “｛####” first
Sentence｝ ”is extracted as a preceding word string candidate.

The n-gram kana index 303 is searched by the kana string information of the preceding word string candidate. In the case of the present embodiment, first, an initial preceding word string “{#### sentence start}” is searched. It is checked whether there is a backward word that matches forward in the partial string of the kana string candidate after the searched preceding word. If there is no backward word that matches the beginning, the process proceeds to ST6004, and if there is a backward word that matches the beginning, the following process is performed (ST6006).

In the present embodiment, the preceding word string "@ # #
# Sentence head｝ ”is searched for n-gram as the backward word,
A word in which the kana string from the beginning of "#sangoua -..." following "#" partially matches is searched for and is set as a backward word. 2-gr
In am, "#Sangou" matches the kana sequence of "#Sangouaa ...", so the 2-gram suffix morpheme "｛Specialization: No. 3 saNgoo Sangou Noun｝" is the backward word. Is one of the candidates. Also, 1-gram “の Specialty: No. 3 saNgoo
Sangu Noun｝ ”matches the beginning of the back kanji string, so this is also a candidate. In addition, “｛General: No.3 saNgoo Sango Noun｝” is also a candidate.

For each backward word, the likelihood is calculated and R
In addition to storing in the AM, a backward word is connected to the preceding word string, and the new word is stored in the RAM as a preceding word string (ST600).
7).

In this embodiment, the preceding word string “$ ##
Replace “# sentence｝” with “｛special: # ## sentence｝ ｛specialty: No. 3 saNgoo sango noun｝”. The linguistic likelihood is the 2-gram probability of the preceding word string "｛#### sentence｝" and the specialty field "｛specialty: #### sentencehead｝ specialty: No. 3 saNgoo sango noun｝". 0.0
It is calculated from 1 by the following formula.

[0143]

[Equation 17]

If no matching word is obtained for all phoneme string candidates, the process moves to ST6004 and the same process is repeated. If obtained, the maximum likelihood and the preceding word string of the solution are stored in the RAM and the following processing is performed (ST6008 ST608
09).

In the present embodiment, as described above, “@ #### sentence head} {# 3 saNgo
o cormorant noun｝｝ arch aaci noun｝
A preceding word string candidate is obtained in the order of {no no particle {...} ".

The word string of the solution having the maximum likelihood stored in the RAM is read. The maximum likelihood is the maximum value of the product of the linguistic likelihood and the acoustic likelihood (ST4008).

In the present embodiment, the kana string candidate “# sangoaichi-no-sensei #” is replaced by “｛# # # sentence head｝
｛No.3 saNgoo sango noun｝ ｛arch aaci Aichi ordinary noun｝ no no no connecting particle｝ ｛precedent seNsee
The maximum likelihood is 5.4 × 10 to the -9th power (Acoustic likelihood 0.9 Language likelihood 6 × 10
-9).

The kanji notation is extracted from the morpheme sequence of the solution.
Output from the output device and end the process (ST4009, ST301
Four).

In this way, by separating the fields and taking the statistics, an n-gram with a strong language constraint can be constructed without increasing the degree of the n-gram, and a highly accurate kana-kanji conversion can be constructed.

Embodiment 4 Next, Embodiment 4 of the present invention
Will be described. FIG. 8 shows the configuration of the present embodiment.
This embodiment is different from the speech recognition device according to the first embodiment in FIG. 1 in that the word probability calculation device (2) 401a and the phoneme n-gram
(Field: baseball) 402, and phonological n-gram (field: general) 4
03 is different, and the other parts have the same configuration. Also,
FIG. 9 shows an example of a phoneme n-gram. FIG. 20 shows a schematic flowchart of the operation.

In the first embodiment, the phoneme n-gram is a 2-gram
And 1-gram, both are used in the present embodiment.
As shown in Fig. 2, only 2-gram is used. And ST70 of FIG.
Check for a match of only 2-grams as shown in 01. As a result, a series of morphemes for one utterance are morphemes in the same field, so that it is possible to prevent another field from intersecting during utterance.

Embodiment 5 FIG. Next, Embodiment 5 of the present invention
Will be described. FIG. 10 shows the configuration of the present embodiment. This embodiment is different from the morphological analyzer of the second embodiment shown in FIG.
n-gram (field: baseball) 502 and kanji n-gram (field:
General) 503 is different, and the other configuration is the same. FIG. 11 shows an example of a kanji n-gram. FIG. 21 shows a schematic flowchart of the operation.

In the second embodiment, the kanji n-gram is a 2-gram
And 1-gram were used, but in this embodiment, FIG.
As shown in FIG. 1, only 2-gram is used. And ST in FIG.
Check for a match of only 2-grams as shown in 8001. As a result, a series of morphemes for one utterance are morphemes in the same field, so that it is possible to eliminate crossing of other fields.

Embodiment 6 FIG. Next, Embodiment 6 of the present invention
Will be described. FIG. 12 shows the configuration of this embodiment. This embodiment is different from the kana-kanji conversion device of the third embodiment shown in FIG.
n-gram (field: baseball) 602 and kana n-gram (field:
General) 603 is different, and the other configuration is the same. FIG. 13 shows an example of the pseudonym n-gram. FIG. 22 shows a schematic flowchart of the operation.

In the third embodiment, the pseudonym n-gram is a 2-gram
And 1-gram were used, but in this embodiment, FIG.
As shown in FIG. 3, only 2-gram is used. And ST in FIG.
Check for a 2-gram only match as shown in 9001. As a result, a series of morphemes for one utterance are morphemes in the same field, so that it is possible to eliminate crossing of other fields.

Embodiment 7 FIG. Next, Embodiment 7 of the present invention
Will be described. FIG. 14 shows the configuration of the present embodiment. This embodiment differs from the speech recognition apparatus according to the fourth embodiment in FIG. 8 only in the word probability calculating device (3) 701, and has the same configuration as the other components. FIG. 23 shows a schematic flowchart of the operation.

In the fourth embodiment, the probability of the phoneme n-gram is used as it is, but in the present embodiment, ST1000 in FIG.
By applying the 2-gram probability weight to each field as shown in FIG. 1, the appearance probability can be adjusted for each field.

Embodiment 8 FIG. Next, Embodiment 8 of the present invention
Will be described. FIG. 15 shows the configuration of the present embodiment. This embodiment differs from the morphological analyzer of the fifth embodiment in FIG. 10 only in the morphological probability calculating device (3) 801 and has the same configuration as the other. FIG. 24 shows a schematic flowchart of the operation.

In the fifth embodiment, the probabilities of the kanji n-gram are used as they are, but in this embodiment, the ST1100 of FIG.
By applying the 2-gram probability weight to each field as shown in FIG. 1, the appearance probability can be adjusted for each field.

Embodiment 9 FIG. Next, Embodiment 9 of the present invention
Will be described. FIG. 16 shows the configuration of the present embodiment. This embodiment differs from the kana-kanji conversion device of Embodiment 6 in FIG. 12 only in the kanji probability calculation device (3) 901 and has the same configuration as the rest. FIG. 25 shows a schematic flowchart of the operation.

In the sixth embodiment, the probability of the pseudonym n-gram is used as it is, but in the present embodiment, ST1200 of FIG.
By applying the 2-gram probability weight to each field as shown in FIG. 1, the appearance probability can be adjusted for each field.

[0162]

As described above, the first and second aspects of the present invention are described.
0,19 speech recognition devices, methods and storage media
In n-gram, phonological n-gr of both specialized and general fields
With am, it is possible to accurately discriminate between inputs in specialized fields and general fields, and to achieve high-accuracy speech recognition.

Further, in the morphological analysis apparatus, method and storage medium according to claims 4, 13 and 22 of the present invention, the kanji n-gram has kanji n-grams in both the specialized field and the general field. Therefore, inputs in a specialized field and a general field can be discriminated with high accuracy, and highly accurate morphological analysis can be performed.

In the apparatus, method and storage medium for kana-kanji conversion according to claims 7, 16 and 25 of the present invention, the kana-n-gram
In the above, the kana n-gram of both the specialized field and the general field is provided, so that the input in the specialized field and the general field can be discriminated with high accuracy, and the kana-kanji conversion with high precision can be performed.

[0165] Further, the present invention is characterized in that:
The devices, methods, and storage media of 1, 14, 17, 20, 23, and 26 are configured so that the weight of the probability can be adjusted for each specialized field and general field, or for each field, so that the appearance probability can be adjusted for each field. Can be.

[0166] Further, claims 3, 6, 9, 1 of the present invention.
In the devices, methods, and storage media of 2,15,18,21,24,27, when calculating a word string candidate, all fields in the n-gram of phonemes, kanji or kana corresponding to a series of voices or character strings are used. Since the word output means is configured so as to match, it is possible to prevent the other fields from intersecting.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a speech recognition device according to a first embodiment of the present invention.

FIG. 2 is a diagram showing an example of an example sentence to be analyzed.

FIG. 3 shows a phoneme n-gram according to the first embodiment of the present invention.
It is a figure showing an example of.

FIG. 4 is a block diagram showing a morphological analyzer according to Embodiment 2 of the present invention;

FIG. 5 Kanji n-gram in Embodiment 2 of the present invention.
It is a figure showing an example of.

FIG. 6 is a block diagram showing a kana-kanji conversion device according to Embodiment 3 of the present invention.

FIG. 7 is a pseudonym n-gram according to Embodiment 3 of the present invention.
It is a figure showing an example of.

FIG. 8 is a block diagram showing a voice recognition device according to a fourth embodiment of the present invention.

FIG. 9 shows a phoneme n-gram according to Embodiment 4 of the present invention.
It is a figure showing an example of.

FIG. 10 is a block diagram showing a morphological analyzer according to Embodiment 5 of the present invention.

FIG. 11 shows a kanji n-gr according to the fifth embodiment of the present invention.
It is a figure showing an example of am.

FIG. 12 is a block diagram showing a kana-kanji conversion device according to Embodiment 6 of the present invention.

FIG. 13 shows a pseudonym n-gr according to the sixth embodiment of the present invention.
It is a figure showing an example of am.

FIG. 14 is a block diagram showing a speech recognition apparatus according to Embodiment 7 of the present invention.

FIG. 15 is a block diagram showing a morphological analyzer according to an eighth embodiment of the present invention.

FIG. 16 is a block diagram showing a kana-kanji conversion device according to Embodiment 9 of the present invention.

FIG. 17 is an operation schematic flowchart of the voice recognition device according to the first embodiment of the present invention;

FIG. 18 is a schematic flowchart of the operation of the morphological analyzer according to the second embodiment of the present invention.

FIG. 19 is a schematic flowchart of the operation of the kana-kanji conversion device according to the third embodiment of the present invention.

FIG. 20 is a schematic flowchart of the operation of the voice recognition device according to the fourth embodiment of the present invention.

FIG. 21 is a schematic flowchart of the operation of the morphological analyzer according to the fifth embodiment of the present invention;

FIG. 22 is a schematic flowchart of the operation of the kana-kanji conversion device according to the sixth embodiment of the present invention.

FIG. 23 is a schematic flowchart of the operation of the voice recognition device according to the seventh embodiment of the present invention.

FIG. 24 is a schematic flowchart of the operation of the morphological analysis device according to the eighth embodiment of the present invention.

FIG. 25 is an operation schematic flowchart of a kana-kanji conversion device according to Embodiment 9 of the present invention;

FIG. 26 is a block diagram showing an example of this type of conventional device.

FIG. 27 is a diagram illustrating an example of a 3-gram table.

FIG. 28 is a schematic flowchart of the operation of the conventional voice recognition device.

[Description of Signs] 101 microphone, 102 phoneme probability calculation device, 103
Word probability calculation device, 104 output device, 105 RAM,
106, 107 phoneme n-gram, 201 file input device, 202 morpheme probability calculation device, 203, 204 kanji n-gram, 301 keyboard, 302 kanji probability calculation device, 303, 304 kana n-gram.

──────────────────────────────────────────────────の Continued on the front page (51) Int.Cl. ⁷ Identification symbol FI Theme coat ゛ (Reference) G06F 17/30 310 G06F 17/30 330C 330 350C 350 G10L 3/00 537D 537J (72) Inventor Yuzo Maruta Tokyo 2-3-2 Marunouchi 2-chome, Chiyoda-ku, Tokyo F-term (reference) 5B009 KB00 KC00 MA00 ME12 QA01 5B075 ND03 NK34 NR12 PP07 PQ02 PR04 PR06 QM08 5D015 HH23

Claims (27)

[Claims] An input means for inputting a voice, and a phoneme probability calculating means for converting a voice signal output from the input means into a phoneme, calculating a phoneme occurrence probability corresponding to each phoneme, and outputting a phoneme sequence candidate. With reference to the phoneme sequence of the target language, the word notation sequence corresponding to the phoneme sequence, and the phoneme n-gram storing the occurrence probability and the phoneme n-gram, the phoneme sequence candidate output by the phoneme probability calculation means is corresponded. Using the word probability calculating means for calculating the word occurrence probability of each word candidate to be input, the phoneme occurrence probability calculated by the phoneme probability calculating means, and the word occurrence probability calculated by the word probability calculating means, And a word output means for calculating a word string candidate similar to the input speech, wherein the phoneme n-gram has a phoneme n-gram of both a specialized field and a general field. Sound characterized by Recognition device. 2. The speech recognition apparatus according to claim 1, wherein said word probability calculation means is configured to be able to adjust the weight of the probability in the specialty field and the general field. 3. The word output means according to claim 1, wherein said word output means is configured such that all fields in a phoneme n-gram corresponding to a series of voices coincide when said word string candidates are calculated. Voice recognition device. 4. An input means for inputting a kana-kanji mixed character string, a kana-kanji mixed character string of a target language, a word notation string corresponding to the kana-kanji mixed character string, and a kanji n-gr for storing an occurrence probability.
am and the kanji n-gram, the word probability calculation means for calculating the word occurrence probability of each word candidate corresponding to the kanji-mixed character string output by the input means, and the word probability calculation means A word output unit that calculates a word string candidate that matches the character string input from the input unit using the word occurrence probability, and a morphological analysis device comprising: A morphological analyzer characterized by having both kanji n-grams in the general field. 5. The morphological analyzer according to claim 4, wherein said word probability calculating means is configured to be able to adjust the weights of the probabilities of the specialized field and the general field. 6. The word output means is configured so that all fields in a kanji n-gram corresponding to a series of kana-kanji mixed character strings match at the time of calculating the word string candidates. Or the morphological analyzer according to 5. 7. An input means for inputting a kana character string, a kana n-gram for storing a kana character string in a target language, a word notation string corresponding to the kana character string, and an occurrence probability, and referring to the kana n-gram. A word probability calculating means for calculating a word occurrence probability of each word candidate corresponding to a kana character string output by the input means; anda word occurrence probability calculated by the word probability calculating means, Word output means for calculating a word string candidate that matches the input kana character string.In the kana-kanji conversion device, the kana n-gram includes, in the kana n-gram, the kana n-gram of both the specialized field and the general field. A kana-kanji conversion device characterized by having. 8. The kana-kanji conversion device according to claim 7, wherein the word probability calculating means is configured to be able to adjust the weight of the probability in the specialty field and the general field. 9. The word output means is configured so that, when calculating the word string candidates, all fields in a kana n-gram corresponding to a series of kana character strings match. The kana-kanji conversion device according to 1. 10. When a voice is input, the voice signal is converted into a phoneme, a phoneme occurrence probability corresponding to each phoneme is calculated, and a phoneme sequence candidate is output. Storing the corresponding word notation sequence and the occurrence probability, referring to the phoneme n-gram and the phoneme n-gram, and calculating the word occurrence probability of each word candidate corresponding to the phoneme sequence candidate; And calculating a word sequence candidate similar to the input speech using the word occurrence probability and a word occurrence probability, wherein the phoneme n-gram is a phoneme n-gram in both the specialized field and the general field. A speech recognition method characterized by having: 11. The method according to claim 10, wherein the weight of the probability can be adjusted for each field when calculating the word candidates.
Voice recognition method described in. 12. The speech recognition method according to claim 10, wherein, when calculating the word string candidates, all fields in a phoneme n-gram corresponding to a series of speeches match. 13. When a character string containing kana and kanji is input,
The kana-kanji mixed character string of the target language and the word notation string corresponding to the kana-kanji mixed character string and the occurrence probability are stored, and the kanji n-gr
a step of calculating a word occurrence probability of each word candidate corresponding to a character string mixed with kanji by referring to am and the above-mentioned kanji n-gram, and using the calculated word occurrence probability to match the input character string A morphological analysis method comprising: calculating a word string candidate; and wherein the kanji n-gram has kanji n-grams in both a specialized field and a general field. 14. The method according to claim 13, wherein the weight of the probability can be adjusted for each field when calculating the word candidates.
A morphological analysis method according to item 1. 15. The method according to claim 13, wherein, when calculating the word string candidates, all fields in a kanji n-gram corresponding to a series of kana-kanji mixed character strings match each other. Morphological analysis method. 16. When a kana character string is input, a kana character string is referred to by referring to a kana n-gram and a kana n-gram storing a kana character string of the target language, a word notation string corresponding to the kana character string, and an occurrence probability. Calculating a word occurrence probability of each word candidate corresponding to the sequence; and calculating a word string candidate that matches the input kana character string using the calculated word occurrence probability. A kana-kanji conversion method, wherein the kana n-gram has both a specialty field and a general field kana n-gram. 17. The method according to claim 16, wherein the weight of the probability can be adjusted for each field when calculating the word candidates.
The kana-kanji conversion method described in. 18. The kana-kanji character according to claim 16, wherein all the fields in the kana n-gram corresponding to the series of kana character strings match when the word string candidates are calculated. Conversion method. 19. An input function for inputting voice, and a phoneme probability calculating function for converting a voice signal output by the input function into phonemes, calculating phoneme occurrence probabilities corresponding to each phoneme, and outputting phoneme sequence candidates. With reference to the phoneme sequence of the target language, the word notation sequence corresponding to the phoneme sequence, and the phoneme n-gram storing the occurrence probability and the phoneme n-gram, the phoneme sequence candidate output by the phoneme probability calculation function is supported. Using the word probability calculation function to calculate the word occurrence probability of each word candidate to be executed, the phoneme occurrence probability calculated by the phoneme probability calculation function, and the word occurrence probability calculated by the word probability calculation function, And a word output function for calculating a word string candidate similar to the input speech from the input device, wherein the phoneme n-gram has phonemes n-grams in both a specialized field and a general field. Record recognition program Computer readable recording medium. 20. The recording medium according to claim 19, wherein in the word probability calculation function, the weight of the probability can be adjusted for each classification. 21. The word output function according to claim 19, wherein at the time of calculating the word string candidate, a word output function is configured such that all fields in a phoneme n-gram corresponding to a series of voices match.
Or the recording medium of 20. 22. An input function for inputting a kana-kanji mixed character string, a kana-kanji n-gr storing a kana-kanji mixed character string of a target language, a word notation string corresponding to the kana-kanji mixed character string, and an occurrence probability.
With reference to am and the kanji n-gram, a word probability calculation function for calculating the word occurrence probability of each word candidate corresponding to the kanji mixed character string output by the input function, and a word probability calculation function calculated by the word probability calculation function A word output function of calculating a word string candidate that matches the character string input from the input function using the word occurrence probability, and realizing that the phonological n-gram is a kanji character n in both the specialized field and the general field. A computer-readable recording medium storing a morphological analysis program, characterized by having a -gram. 23. The recording medium according to claim 22, wherein in the word probability calculation function, the weight of the probability is adjustable for each field. 24. The word output function according to claim 22, wherein at the time of calculating the word string candidate, a word output function is configured such that all fields in a kanji n-gram corresponding to a series of kana kanji mixed character strings match. Or the recording medium according to 23. 25. An input function for inputting a kana character string, a kana n-gram storing a kana character string in a target language, a word notation string corresponding to the kana character string, and an occurrence probability, and referring to the kana n-gram. A word probability calculation function for calculating a word occurrence probability of each word candidate corresponding to the kana character string output by the input function; and a word occurrence probability calculated by the word probability calculation function. A word output function of calculating a word string candidate that matches the input kana character string, and realizing that the kana n-gram has both a specialty field and a general field kana n-gram. A computer-readable recording medium that records a kana-kanji conversion program. 26. The recording medium according to claim 25, wherein the weight of the probability can be adjusted for each field in the word probability calculation function. 27. The word output function according to claim 25 or 26, wherein at the time of calculating the word string candidate, a word output function is configured such that all fields in a kana n-gram corresponding to a series of kana character strings match. A recording medium according to claim 1.