JPH04112269A - Lattice searching system using difference of similarity between recognitive candidate - Google Patents

Abstract

PURPOSE:To search a correct word from a candidate lattice at a short time by using a lattice searching system using the difference of the similarity between the recognitive candidates. CONSTITUTION:When the difference of the similarity between the first candidate and the second candidate is more than 150, the recognitive candidates are divided into two groups 2 and 3 of the first candidate and the second candidate. When the difference of the similarity is less than 150, the candidates are divided into two groups 1 and 4 of the first candidate, and the second candidate, and the third candidate. Then, a vertical search or a horizontal search are executed in the groups, and when the candidates of input sentences can not be searched, a candidate lattice 5 to be searched is put from the next group 3 into the group whose difference is the minimum among the groups 2. Thus, correct candidate sentences or the correct words can be searched in the short time from the candidate lattice prepared by a recognizing device.

Description

DETAILED DESCRIPTION OF THE INVENTION [Summary] For example, a candidate word lattice is generated as a recognition result by a speech recognition device or a character recognition device.

Alternatively, regarding a search method for sequentially generating candidate sentences or candidate words from a candidate character lattice, etc., the present invention provides a candidate search method that can search for correct candidate sentences or words from a candidate lattice in a short time. With the aim of If the number is larger than a certain number (for example, 150), the first recognition candidate,
Divide the recognition candidates into two groups ■ and ■ with other candidates, and if the difference is smaller than the above specified number, divide the recognition candidates into the first and second candidates and two groups below the third candidate ■ and ■. A vertical search or a horizontal search is performed between the glues 10 and 0 and within the group, and the search process finds candidate words for the input word or input sentence.

Alternatively, if a candidate sentence cannot be searched, a candidate from the next group ■ is selected from among the groups ■ in which the difference in similarity between the first candidate and the second candidate is less than or equal to the specified number. Incorporating the lattice■ into the search target, input words,
Alternatively, it is configured to search for input sentences.

[Industrial Application Field] The present invention is a method for sequentially generating candidate sentences or candidate words from, for example, a candidate word lattice or a candidate character lattice generated as a recognition result by a speech recognition device or a character recognition device. Regarding the search method for generation.

Speech recognition devices and character recognition devices, which have been put into practical use for a long time, make the interface between humans and machines more familiar to humans, but in that case,
It is necessary to improve recognition accuracy and reduce the effort required by humans to correct recognition errors.

In the past, for example, in voice recognition devices, the voices that could be registered were numbers from '0゛ to '9'' and words in the hundreds, but with the improvement in the processing power of recent computer systems. Along with this, the recognition performance of the speech recognition device has improved and the number of words that can be registered has increased, so the number of candidates (also called elements) that make up the lattice output by the speech recognition device has also increased. Therefore, the speech recognition system is required to efficiently generate and display candidate sentences or candidate words.

Furthermore, a search method for sequentially generating candidate sentences or candidate words from a candidate word lattice or a candidate character lattice generated as a recognition result by the speech recognition device or character recognition device is a method using a voice input word processor or a Since it is an elemental technology of handwritten character input word processor, in order to improve its man-machine interface,
As mentioned above, it is necessary to efficiently generate and display candidate sentences or candidate words.

[Prior art and problems to be solved by the invention] FIG. 3 is a diagram explaining a conventional candidate search method for speech recognition/character recognition, in which (a) shows an example of a lattice in units of phrases;
b) shows an example of a character-by-character lattice, (c) shows a candidate search method based on the vertical format priority method, and (d) shows a candidate search method using the horizontal format priority method.

In a recognition device for a voice recognition system or a character recognition system (not shown), for example, the recognition distance (similarity) with a single character of a word registered in a dictionary is calculated,
The candidates with the shortest distance are selected, and the recognition results for the input speech or characters are (a);
(b) Output a lattice as shown in the figure. here,
The number in parentheses is the recognition distance (similarity) between each element and the input.

Depending on the unit of separation, there are lattices for characters, words, phrases, etc. Figure (a) shows an example of a lattice for phrases, and (b)
) shows an example of a character-based lattice.

In this lattice, the horizontal direction represents the passage of time from left to right, and the vertical direction represents the ranking of candidates from top to bottom.

In other words, in the example shown in (a), for the input "you are", the recognition device has the highest probability of saying "with you", and the second highest probability of saying "you are" (the same applies hereafter).
This shows the result of outputting the recognition result. Similarly,
In response to the input "book," the recognition device has the highest probability of saying "book", followed by "book too" (
The same applies hereafter) is shown.

That is, the recognition device does not necessarily include the correct answer in the first place in the candidate ranking ((a), (b)
In the figure, the correct answer is indicated by "*") Therefore, the candidate sentence selection unit (search mechanism) of the recognition system connects the candidates in the lattice to generate a candidate sentence or a candidate word.

At this time, regarding the order in which candidates are selected, for example, there is a vertical search method.

For example, Figure (c) shows a method of searching for candidates with priority given to the vertical format for the character unit lattice shown in Figure (b).

To search for the input word "similarity" from the candidate character lattice shown in figure (b), the vertical priority search method
Twenty-three searches were required.

In other words, in figure (b), the candidate words "gan, li, tang" are given "111" (1st place, 1st place, 1st place), and the candidate words "similar, i, keigo" are given "222" (2nd place, 2nd place).・2nd place), etc., the correct answer "similarity, similarity, degree" will be written as "213".

The search order using the above vertical search (also called depth-first) is:
Using the above notation, if the depth of the lattice is up to the fourth position as shown in figure (b), then the search method shown in figure (c) is
Since the search order is (1) → (2) → (3) → (4)), 111.112,113,114°121, 122. 123. 124°131.13
2,133,134°141.142,143,144°211 212.213,214°221.222,213°.

Therefore, the above correct answer "213" will be found in the 23rd search.

In addition, the search order by the horizontal search method (also called breadth priority) is shown in the above notation as the search method shown in Figure (d),
That is, if "111" is selected as the top candidate using this notation method, the next step is to select the one in which the rank of any one of the three elements of the selected candidate has changed (increased). Since this is a method of searching by repeating operations, if the duplicate selections are removed and displayed, the following will be displayed: 111°112.121,211°113.122,212°131.221°311°114.123,213. ... becomes.

Therefore, the above correct answer '213J' will be found in the 13th search.

In reality, for the candidates searched as above, for example, a dictionary matching mechanism performs dictionary matching, checks whether the candidate word exists in the dictionary, and outputs only the candidate words that exist in the dictionary. I will do it. In addition, when a search is performed on the clause lattice shown in (a), for example, a parsing mechanism parses each of the candidate sentences to determine if the candidate sentence is grammatical. Only those that are correct will be output as those that have been successfully parsed.

The drawbacks of the conventional vertical-first search method and horizontal-first search method described above will be described below.

In a lattice generated by a speech recognition device or a character recognition device, if the recognition distance values of adjacent elements in the candidate order from top to bottom are close, the probability that each is correct is close. 0 For example, Figure 3 (
In b), the recognition distance between "request" and "r class", a4s1 degree, is each *r2458"r2497J, and the difference is "
39'', while the difference between ``4g'' and ``or'' is r3.
89 J is also available.

In conventional vertical-first search and horizontal-first search, candidates who are colleagues are not considered as one, and r 1' 11
From searching for J to searching for r213J, 23 searches are required in the vertical-first search and 13 searches are required in the -horizontal-first search. However, it is desirable that the number of searches from rl 11J to r213J is small.

In view of the above-mentioned conventional drawbacks, the present invention provides a speech recognition device.

Alternatively, in order to sequentially generate candidate sentences or candidate words from a candidate word lattice or a candidate character lattice generated as a recognition result by a character recognition device, elements with small differences in similarity between recognition candidates may be selected. The object of the present invention is to provide a candidate search method that allows search opportunities to be brought closer together, that is, by grouping, and as a result, correct words can be searched from a candidate lattice in a short time.

[Means to solve the problem]

FIG. 1 is a diagram showing the principle of the present invention, and (al) is a diagram showing the principle of the present invention.
For example, when the input word is recognized as the first candidate, the difference in similarity between the first candidate and the second candidate is analyzed for 20 words, and (a2) shows that the input word is recognized as the second candidate. and the first and second candidates when recognized between disciple candidates, ~
The difference in similarity between 20 words is analyzed, (b
) is a graphical representation of the above analysis results, and (cl) ~
(c3) shows an example of grouping according to the present invention and an example of the vertical search method.

The above problem can be solved by a lattice search method that utilizes the difference in similarity between recognition candidates configured as follows.

(1) For the input word or input sentence, the recognition device 2
calculates the difference in similarity between the first and second candidate words or candidate characters as recognition results, and if the difference is greater than a certain number, for example 150, the first recognition candidate Divide into two groups, ■ and ■, candidates and others, and if the difference between them is smaller than the specified number above, select the recognition candidates as the first group,
Divide the second candidate into two groups ■ and ■ below the third candidate, perform vertical search or horizontal search between the groups 10 to 0 and within the groups, and in the above search process, input If a word or candidate word or candidate sentence for the input sentence cannot be searched, the first candidate and
Groups where the difference in similarity between second candidates is less than or equal to the above specified number ■
The candidate lattice (2) from the next group (2) is incorporated into the earliest group (2) as a search target, and the input word or input sentence is searched.

(2) In the lattice search method described in item 1 above, ′
The recognition device 2 is configured to be a speech recognition device.

(3) In the lattice search method described in item 1 above, the recognition device 2 is configured to be a character recognition device.

[Effect]

That is, when candidate sentences or candidate words are sequentially generated from a candidate word lattice or a candidate character lattice generated as a recognition result by a voice recognition device or a character recognition device, the voice recognition device is used.

Alternatively, the candidate lattice generated by the character recognition device generally includes 11 points between the candidates when the input word is recognized as the first place and the candidates when the input word is recognized as the second place or below.
As a result of analyzing the 1 degree difference for 20 words each, it was found that the difference in similarity between recognition candidates was 1 and had the following attributes.

1) When the input word is recognized as the first candidate, the average value of the difference in similarity between the first candidate and the second candidate is, for example, 24
It is 4. (See figure (al)) 2) If the input word is recognized between the second candidate and the disciple candidate, the average value of the differences of class A or higher between the first candidate and the second candidate is 50.9. ( (
a2) (See figure)) (b) in Figure 1 is the same as (al) above, (
a2) This is a graph of the difference in the average similarity values shown in the figure.

The present invention utilizes the above attributes to achieve the following:
The input is searched from among the candidate word lattice or candidate character lattice generated by the recognition device.

(1) First, calculate the difference in similarity between the first candidate and the second candidate, and if the difference is, for example, 150 or more, divide the recognition candidate into two groups: the first candidate and the second candidate. , divided into ■.

The difference in similarity between the first candidate and the second candidate calculated above is 15
If it is 0 or less, the first candidate and the second candidate.

Divide into two groups ■ and ■ below the third candidate.

(2) Repeat (1) to divide all candidate word lattices or candidate character lattices into several groups. - (See figure (cl)) (3) Execute vertical search or horizontal search between the groups divided as described above.

(4) If no candidate for the input sentence can be found in the above search, select the group with the smallest difference among the groups ■ in which the difference in similarity between the first candidate and the second candidate is 150 or less. Insert candidate lattice ■ to be searched from the next group ■.

(See figure (c3)) (5) Repeat step (3) above until candidates for the input sentence can be searched.

Repeat the process in (4).

This search method has the potential to perform a search in a short time by searching for an input word from a candidate character lattice, using the difference in similarity between recognition candidates, and grouping similar elements together.

For example, if the present invention is applied to perform a vertical priority search between groups and within a group, the number of searches from the above-mentioned search for rl12J to the search for r213J is 5,
Or three times.

As described above, the candidate search method according to the present invention has the advantage that correct candidate sentences or correct words can be searched for in a short time from the candidate lattice generated by the recognition device.

(Example〕 Embodiments of the present invention will be described in detail below with reference to the drawings.

The above-mentioned FIG. 1 is a diagram explaining the present invention in detail, and FIG.
The figures show an embodiment of the present invention, in which (a) shows a case where the present invention is applied to a speech recognition system to search for candidate sentences from a lattice of phrases, and (b) shows a case where the present invention is applied to a character recognition system. (C) shows the case where candidate sentences are searched from the lattice of phrase units, and (C) shows the case of searching for candidate words from the lattice of character units. To search for candidate words or candidate sentences from the lattice of the lattice or the lattice of the bunsetsu unit, for example, the difference between the first candidate and the second candidate in the candidate ranking is calculated, and if the difference is, for example, 150 or more, In this case, the recognition candidates are divided into two groups: the first candidate and the second candidate ■
, ■, and if the difference is less than 150, then the first candidate, second candidate, and two groups below the third candidate ■
, ■ between groups.

If a vertical priority search or horizontal priority search is executed within a group, and candidates cannot be found in the above search process,
The present invention provides means for including candidate lattice ■ from the next group ■ into the group ■ having the smallest difference among the groups ■ in which the difference in similarity between the first candidate and the second candidate is 150 or less as a search target. It is a necessary means to carry out. Note that the same reference numerals indicate the same objects throughout the figures.

Hereinafter, while referring to FIG. 1, the I! between recognition candidates of the present invention will be explained using FIG. 2. A lattice search method using degree differences will be explained.

Even if the present invention is implemented, the recognition means in the recognition device.

And from the lattice output by the recognition device 2, for example,
The basic operation when performing vertical-first search does not change in particular, so the details are omitted.Here, candidate sentences or candidate words are The explanation will focus on the means for generating .

First, the example shown in FIG. 2(a) is an example in which the present invention is applied to a voice recognition system, in which the voice input from the voice input microphone 1 is passed through a voice pattern registration dictionary (not shown) in the voice recognition device 2. After searching, the recognition distance (l! similarity) is calculated, and, for example, a phrase unit lattice or a word unit lattice is output in descending order of distance.

The lattice output from the speech recognition device 2 is processed by a microprocessor (not shown) in the personal computer 3.
1 (cl) by executing a specific search routine (search) 4.

Search based on the search procedure shown in (c3), that is, for example, calculate the difference between the first candidate and the second candidate in the candidate ranking, and if the difference is, for example, 150 or more, select the recognition candidate as the first candidate. Divide into two groups ■, ■ below the first candidate and second candidate, and if the difference between them is less than 150, divide into two groups ■, ■ below the first candidate, second candidate, and third candidate. , vertical-first search or horizontal-first search is performed between groups or within groups, and if no candidates can be found in the above search process, the first and second candidates are searched! ! Among the groups (2) with a difference in efficiency of 150 or less, candidate lattice (2) from the next group (2) is included in the search target for the group (2) with the smallest difference, and a vertical search is performed.

That is, a vertical-first search is performed between groups 10 and 0, and then a vertical-first search is performed within groups ■ and ■ to find candidate sentences,
Or, generate candidate words.

Next, the microprocessor (MPU) executes a language processing routine (language processing) 5 to parse the syntax of each of the candidate sentences, and determines which sentences are grammatically correct. only is output and displayed on the display 6.

By selecting the correct sentence from the displayed candidate sentences, the user, that is, a human being, can search for the correct candidate sentence in a shorter time than in the conventional method.

(b) Regarding the character recognition system shown in the figure, except that the input sentence is input from the input tablet 1, (a
) Since it has the same configuration as the speech recognition system shown in the figure, the character recognition device 22 outputs, for example, a word-by-word lattice for the sentence input from the input tablet 1 using the same procedure as in the case of speech recognition. , a search routine (search) executed by the microprocessor (1'1PU) of the personal computer 3
4, the word unit lattice is grouped.

Candidate sentences are generated by performing a vertical-first search between groups ■ and a vertical-first search within group ■.Next, a syntactic analysis is performed in a language processing routine (language processing) 5. Successful candidate sentences are displayed on the display 6.

(c) The embodiment shown in the figure is another example of the above-mentioned character recognition system, in which the character recognition device 2 outputs, for example, a lattice of characters for a word input from the input tablet 1. It shows.

In this case, the microprocessor (MPU) of PC 3
When a candidate word is searched for by the search routine (search) 4 executed by the dictionary matching routine (dictionary matching)-5a
The candidate words that are successfully matched are compared with words registered in a dictionary (not shown).
It operates to display on the display 6.

As described above, the present invention searches for candidate words or candidate sentences from the character-based lattice, word-based lattice, and phrase-based lattice generated by the recognition device as a recognition result. Calculate the difference between the candidate and the second candidate,
If the difference is, for example, 150 or more, the recognition candidates are divided into two groups, ■ and ■, the first candidate and the second candidate. Divide into two groups: the second candidate and the third candidate ■, ■.
If a vertical priority search or a horizontal priority search is performed between groups or within a group, and a candidate cannot be found in the above search process, the group in which the difference in II rank between the first candidate and the second candidate is 150 or less is Among (2), the candidate lattice (2) from the next group (2) is included in the search target for the group (2) with the smallest difference, and the above candidate search is performed.

〔Effect of the invention〕

As described above in detail, the candidate search method for speech recognition/character recognition of the present invention uses an input word.

Alternatively, the recognition device calculates the difference in similarity between the first and second candidate words or candidate characters as a recognition result for the input sentence, and sets the difference to a specific number, for example, 150. If the difference is smaller than the above specific number, divide the recognition candidates into two groups ■, ■, the first candidate and the others, and if the difference is smaller than the above specific number, divide the recognition candidates into the first, second, and Divide into two groups of three candidates or less■,■, and the above group■
In ~■, a vertical search or a horizontal search is performed, and if a candidate word or candidate sentence for the input word or input sentence cannot be searched in the above search process, the above first candidate and, The candidate lattice ■ is included in the search target from the next group ■ to the group ■ with the lowest similarity among the groups ■ in which the difference in similarity of the second candidate is less than or equal to the above specified number,
Since the input word or input sentence is searched, the correct answer candidate sentence or correct answer word can be searched from the candidate lattice in a short time, and the man-machine interface of the speech/character recognition system is improved. It has the effect of

[Brief explanation of drawings]

FIG. 1 is a diagram explaining the present invention in detail. FIG. 2 is a diagram showing an embodiment of the present invention. FIG. 3 is a diagram explaining a conventional candidate search method for speech recognition/character recognition. It is. the bottom two groups. ■ is a candidate lattice to be incorporated. are shown respectively. In the drawing, 1 is an input microphone or an input tablet. 2 is a voice recognition device, a character recognition device, or a recognition device. 3 is a computer. 4 is a search processing routine (search). 5 is a language processing routine (language processing). 5a is a dictionary matching routine (dictionary matching); 6 is the display. ■ and ■ are two groups in which the difference in similarity between the first candidate and the second candidate is 150 or more. ■, ■ are the first candidate, second candidate, and third candidate when the difference in similarity between the first candidate and the second candidate is 150 or less. Figure 1 (Part 1) Figure 1 (Part 2) Figure (Part 3) Figure (Part 1) Figure (Part 4) Figure (Part 2) Example: Input sentence: Are you reading a book? (a) Example: Input word: Similarity recognition result (* indicates correct answer) Figure (Part 1) (C) +23 Horizontal search order: (lH2H3)Bet(d) V to explain Figure (Part 1) 2)

Claims (3)

[Claims] (1) The recognition device (
2) calculates the difference in similarity between the first and second candidate words or candidate characters that are the recognition results, and if the difference is greater than a certain number, , and the others ([1], [4]), and if the difference between them is smaller than the specified number above, the recognition candidates are divided into the first,
The second candidate and the two groups below the third candidate ([2],
[3]), between the above groups ([1], [4]),
Then, a vertical search or a horizontal search is performed within the group, and if a candidate word or a candidate sentence for the input word or input sentence cannot be found in the above search process, the above first candidate is used. , the candidate lattice ([5]) from the next group ([3]) is assigned to the smallest group ([2]) among the groups ([2]) in which the difference in similarity between the second candidates is less than or equal to the specified number. ]) as a search target and searches for input words or sentences. (2) The lattice search method according to claim 1, wherein the recognition device (2) is a speech recognition device, using a difference in similarity between recognition candidates. (3) The lattice search method according to claim 1, wherein the recognition device (2) is a character recognition device, using a difference in similarity between recognition candidates.