JP3099845B2 - Post-processing of pattern recognition - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern recognition post-processing system capable of improving recognition accuracy.

[0002]

BACKGROUND OF THE INVENTION The polarization of the total number of categories to be recognized as a m, to the respective category as C ₁ -C _m. Let L be the length of an input category sequence, ω be a set of all possible category sequences, and k be the number of elements of ω.

[0003] In this case, for a certain input category column feature vector X ₁ ~X _L of, I think that you choose a high category column of the probability of the most correct answers. It has been derived from statistical determinism that this method is to calculate an evaluation value represented by the following equation (1) for each category column and to find a candidate word having the maximum value.

[0004]

(Equation 1) In the above equation, W is an arbitrary category sequence in ω, P
(W) is the appearance probability of W, _Cwi is the i-th category of W,
P (X _i | C _wi ) is a probability that the feature vector X _i of the i-th pattern is generated from the category C _wi . in this case,
The occurrence probability P (W) of the category sequence can be relatively easily obtained.

[0005]

[SUMMARY OF THE INVENTION] However, conventionally it is fitted the feature vector extracted by the feature extraction process of the pattern recognition processing as a feature vector X _i, a feature vector X _i from the category of the case C _i is occurring P
It is very difficult to find (X _i | C _wi ). The reliability of a candidate category output in general pattern recognition is an amount, such as distance and similarity, that is difficult to directly convert into a probability P (X _i | C _i ). Therefore, it was difficult to calculate the above-mentioned evaluation value using the conventional pattern recognition technology.

[0006] The present invention has been made in view of the above,
An object of the present invention is to provide a pattern recognition post-processing method in which a category string having the highest correct answer probability is selected and recognition accuracy is improved.

[0007]

In order to achieve the above object, a pattern recognition post-processing method according to the present invention recognizes an input pattern such as a voice or a character, and obtains a plurality of candidate categories and a distance between the candidate categories as a recognition result. A candidate category sequence set creating means for creating a set of candidate category sequences, and a predetermined rule for each candidate category sequence of the candidate category sequence set output from the candidate category sequence set creating method. A category string occurrence probability calculating means for assigning the occurrence probability of the category string; and a candidate comprising all two candidate category strings from a set of candidate category strings to which the occurrence probability calculating means is assigned by the category string occurrence probability calculating means. Candidate category string pair set creating means for creating a set of category string pairs, and a candidate created by the candidate category string pair set creating means. The evaluation value is calculated using the candidate category column pair determination means for all candidate category column pairs in the category column pair set, and the candidate category column having a higher evaluation value than the candidate category column pairs other than the own is checked to determine the maximum. A maximum likelihood candidate category string selection unit that outputs as a likelihood candidate category string;
A set of partial recognition results including categories corresponding to each other and a distance between the categories is created from an arbitrary candidate category string pair, and the occurrence probability of the partial recognition result set is calculated using occurrence probability calculation means. And a candidate category string pair determining means for calculating an evaluation value from the occurrence probability of the candidate category string, and for any two partial categories included in the partial recognition result, the partial recognition result is determined from each category. And an occurrence probability calculating means for obtaining and outputting the generated occurrence probability.

[0008]

In the pattern recognition post-processing method of the present invention, the number of all categories to be recognized is m, and each category is C ₁
To C _m . Let L be the length of an input category sequence, ω be a set of all possible category sequences, and k be the number of elements of ω. Here, the evaluation for a certain input category column feature vector X ₁ to X _L of the ability to select high category column probability of the most correct, for each category string included in the ω represented by the formula (1) A method of calculating a value and finding a candidate category sequence having the maximum value is realized.

An arbitrary feature vector X from a certain category C
It is difficult to calculate a probability P (X | C) at which a feature vector extracted in a feature extraction process of a pattern recognition process as a feature vector X as in the related art. Therefore, a vector composed of distances for all categories as a result of recognition of the input pattern is adopted as the feature vector X, and this is called a distance vector. Here, the serial numbers (1 to m) of each category when all the categories to be recognized are arranged in order are referred to as category numbers, and the category whose category number is j is represented by C _j . When the distance of the input pattern to category number j is represented by r _j , the distance vector is (r ₁ , r ₂ ,...,
r _m ), and P (X | C) is represented by the following equation.

P (X | C) = P ((r ₁ , r ₂ ,..., R _m ) | C) (2) Here, the category that can be correctly answered is limited to two by some method. Consider the case. When this category as C _u and C _v, since it is r _u and r _v if the most effective elements to identify both in the distance vector choosing two new feature vector from the two elements ( r _u , r
_v ) and use this instead of the distance vector in equation (2). This vector is called a partial distance vector.
In this case, equation (2) becomes as follows.

[0011] _{P (X | C u) =} P ((r u, r v) | C u) P (X | C v) = P ((r u, r v) | C v) In this case, any of the the probability of occurrence of partial distance vector for between 2 categories (P in the above description ((r _u, r _v)
| C _u) and _{P ((r u, r v} ) | C v)) becomes a need for a means for obtaining a, which is either theoretically derived from the identification function is used in the pattern recognition, or a large amount of recognition This can be realized by collecting the results and statistically obtaining the distribution of the partial distance vector between all two categories.

[0012]

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing a configuration of a pattern recognition post-processing system according to an embodiment of the present invention. FIG.
, 1 is a candidate category string set creating means, 2 is a category string occurrence probability calculating means, 3 is a candidate category string pair creating means, 4 is a maximum likelihood candidate category string selecting section, 5 is a candidate category string pair determining means, 6 Is an occurrence probability calculation means.

For the purpose of explanation, alphabets A to G are used as categories and characters are used as patterns. The serial numbers when the categories are arranged in order are called category numbers and are shown in FIG.

First, the pattern recognition result is input to the candidate category sequence set creating means 1. FIG. 3 shows an example of a recognition result in which the correct answer is “FC”.

The candidate category sequence set creating means 1 creates a candidate category sequence set. This method can be a method of using a dictionary in which all possible candidate category strings are prepared in advance, or a method of creating by combining candidate categories of pattern recognition results. A method using a category string dictionary in which all category strings are prepared in advance is used. FIG. 4 shows an example of the category string dictionary. In this case, the processing result of the candidate category sequence set creating means 1 is {FG, FC, AD}
It is.

Next, the occurrence probability of each category string included in the set of candidate category strings is calculated and assigned by the category string occurrence probability calculating means 2. As this method, for the sake of explanation, the occurrence probability is recorded in advance in each category column of the category column dictionary shown in FIG. The processing result in this case is ｛(FG, 0.3), (FC, 0.2), (AD, 0.
5)｝ (() indicates a category sequence and an occurrence probability).

Next, in response to the processing result of the category string occurrence probability calculation means 2, the candidate category string pair set creation means 3 creates a candidate category string pair set. FIG. 5 shows the processing result in this case.

Next, the maximum likelihood candidate category sequence selecting unit 4 takes out the elements of the candidate category sequence pair set one by one and sends them to the candidate category sequence pair determining means 5 to determine the candidate category sequence included in each candidate category sequence pair. An evaluation value is calculated, and a candidate category string having an evaluation value higher than that of all candidate category strings other than the subject is set as the maximum likelihood candidate category string from the results. In the case of the example, the results of sending three candidate category string pairs included in the candidate category string pair set of FIG. 5 to the candidate category string pair determining means 5 are shown in FIGS. 6 (a), (b) and (c). . From this result, the candidate category string “FC” is judged to be the maximum likelihood candidate category string because the evaluation value is higher than all other candidate category strings (FIGS. 6A, 6B, and 6C). The higher evaluation value is indicated by “◎”).

Next, the operation of the candidate category string pair determining means 5 will be described.

As an example, the candidate category string pair NO. The operation when 1 is sent will be described. First, a partial distance vector composed of the category and the distance of the corresponding character position is created for all character positions from the candidate category string pair, and is sent one by one to the occurrence probability calculation means 6 to be included in the partial distance vector. The probability that the partial distance vector will occur from one category is calculated. In this case, the partial distance vector is (18 (E), 25
(F)) and (32 (G), 16 (C)) are sent.

The occurrence probability calculation means 6 calculates the probability of occurrence of the partial distance vector for each of the two categories included in the received partial distance vector.

In the example of FIG. 6A, the probability that this partial vector will occur from the category E for the partial distance vector (18 (E), 25 (F)) is 0.2, and the category F
Is calculated to be 0.1.

[0024]

As described above, according to the present invention,
Since the evaluation value calculated by Expression (1) can be used as an evaluation scale for selecting the maximum likelihood candidate category sequence from a set of candidate category sequences by comparing the pattern recognition result with a word dictionary, etc. Can be selected, and the recognition accuracy can be improved.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a pattern recognition post-processing method according to an embodiment of the present invention.

FIG. 2 is a diagram showing category numbers of categories used for explaining the operation of the pattern recognition post-processing method of FIG. 1;

FIG. 3 is a diagram showing a sequence of recognition results used for explaining the operation of the pattern recognition post-processing method of FIG. 1;

FIG. 4 is a diagram showing a category string dictionary used to show the operation of a candidate category string set creating means and a category string occurrence probability calculating means used in the pattern recognition post-processing method of FIG. 1;

FIG. 5 is a diagram illustrating an example of a processing result of a candidate category string pair set creating unit used in the pattern recognition post-processing method of FIG. 1;

FIG. 6 is an explanatory diagram showing a processing process of a maximum likelihood candidate category sequence selecting means in FIG. 1;

[Explanation of symbols]

 1 candidate category string set creating means 2 category string occurrence probability calculating means 3 candidate category string pair creating means 4 maximum likelihood candidate category string selecting means 5 candidate category string pair determining means 6 occurrence category calculating means

────────────────────────────────────────────────── ─── Continuation of the front page (56) References JP-A-62-182982 (JP, A) JP-A-1-204196 (JP, A) JP-A-60-135999 (JP, A) Magazine, Vol. 30 No. 11 p. 1394-1401 (1989) (58) Fields investigated (Int. Cl. ⁷ , DB name) G06K 9/72 G06T 7/00 G10L 3/00

Claims (1)

(57) [Claims] 1. Recognizing an input pattern such as a voice or a character,
In a pattern recognition method that generates a plurality of candidate categories and a distance between the candidate categories as a recognition result, a candidate category sequence set creating unit that creates a set of candidate category sequences, and a candidate category sequence output from the candidate category sequence set creating unit. A category string occurrence probability calculating means for assigning the occurrence probability of the category string to each candidate category string of the set according to a certain rule; and a candidate category string set assigned the occurrence probability calculating means by the category string occurrence probability calculating means. Means for creating a set of candidate category string pairs composed of all the two candidate category strings from, and a candidate category string pair set created by the candidate category string pair set creating means. The evaluation value is calculated for all candidate category string pairs using the candidate category string pair determination means,
A maximum likelihood candidate category column selecting unit that examines a candidate category column having a higher evaluation value than a candidate category column pair other than the user and outputs the same as a maximum likelihood candidate category column, and a category corresponding to each other from an arbitrary candidate category column pair. A set of partial recognition results consisting of category distances is created, an occurrence probability of the partial recognition result set is calculated using an occurrence probability calculation unit, and a candidate for calculating an evaluation value from the occurrence probability and the occurrence probability of the candidate category string is calculated. Category string pair determining means, and occurrence probability calculating means for obtaining and outputting an occurrence probability of generating a partial recognition result from each of two candidate categories included in the partial recognition result for an arbitrary partial recognition result And a pattern recognition post-processing method comprising: