JPS6365499A - Syntax recognition system - Google Patents

Abstract

(57) [Abstract] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention relates to a syntax recognition method, and more specifically, the present invention relates to a syntax recognition method, and more specifically, the syntax of a natural language sentence (for example, a Japanese sentence) inputted by voice is recognized by a voice parameter series (feature parameter series). ).

[Conventional technology]

Conventionally, a keyword spotting method is known as a method for recognizing input syntax at the stage of a speech parameter string. In this method, a string of audio parameters (for example, a string of correlation coefficients obtained by sampling the input audio signal at regular time intervals) of words (keywords) that characterize syntactic candidates is registered in advance in a dictionary, and the input By comparing the input voice parameter string with the voice parameter string of these keywords and performing evaluation points (scoring), it is determined whether or not the keyword of interest exists in the input voice parameter string, and the input voice parameter string is evaluated. It recognizes the syntax that is used. Regarding this, for example, W
, A. Lec wrote “The Value of 5pe”
ech Recognition Systems” TRENDS IN 5PEECHRECOG
NITION, 1980. PRENTICE-1-IA
Discussed in LL, INC, pp. 10-11.

The following is a keyword spotting (hereinafter referred to as ``keyword spotting'') assuming a ticket reservation question and answer system in Japanese.

A conventional syntax recognition method using word spotting (simply referred to as word spotting) will be explained below.

Figure 4 shows a ticket reservation question-and-answer system that applies the conventional syntax recognition method.To simplify the explanation, we will assume that the two syntax candidates are "'ticket reservation request" and "price verification." There is. The system consists of a keyword dictionary 1, a word spotting section 2, and a syntax determining section 3. In the keyword dictionary 1, speech parameter strings corresponding to keywords characterizing each syntactic candidate are registered in advance. The word spotting section 2 extracts keywords included in the input speech parameter string. The syntax determining section 3 determines syntax candidates based on the keywords extracted by the word spotting section 2. The input audio parameter string is represented by a string of correlation coefficients, LPG cepstrum, etc. obtained by sampling the input audio at regular intervals, for example. The operation of FIG. 4 will be explained below, focusing on the word spotting section 2.

(2) Take out the voice parameter string of a certain keyword from the keyword dictionary 1 and set it in the keyword register 23.

■ An audio parameter partial string whose length is equal to the length of the audio parameter string set in the keyword register 23 from the beginning of the target input audio parameter string is input to the input register 2.
2, the keyword scoring unit 24 performs a comparison between the two, obtains a score corresponding to the degree of similarity (usually a value obtained by a calculation called distance calculation), and indicates that the value is retained. R3 is the parameter value at position j of the input register 22, and S is the parameter value at position j of the keyword register 23.

■ Repeat the operation in ■ above while shifting the start position of the subsequences of the input audio parameter string one by one, and compare the similarities between all the subsequences of the input audio parameter string and the audio parameter string of the keyword of interest. A keyword of interest is determined to exist when the best score among them satisfies a certain criterion.

(2) The above operations (2) to (2) are performed for all keywords registered in the keyword dictionary 1 to detect keyword candidates that are considered to be included in the input voice parameter string.

■ The syntax determination unit 3 compares the keyword set information given in advance to characterize each syntax with the set of keywords detected by the word spotting unit 2, and uses prespecified criteria (for example, specified 50% of the specified keywords exist) are selected and output.

[Problem that the invention seeks to solve]

In the conventional syntax recognition method described above, when the number of target syntax candidates increases, 9. The number of keywords to be recorded increases, and the number of keywords with similar characteristic parameters increases accordingly. For this reason, the keyword scoring unit may increase the number of errors such as detecting that a word of interest exists in the target input sentence when it does not exist, or mistaking a word for another simple word and recognizing it. Moreover, since these errors occur across syntax, the probability of correctly recognizing the syntax (herein referred to as syntax recognition performance) decreases.

In addition, the syntax recognition method described above basically uses a digital detection method that deals with the presence or absence of keywords, so it cannot deal with the degree of similarity of the syntax candidate being focused on to the input sentence. No, ■By not focusing on information that characterizes the syntax, such as the position of the keywords or the order in which the keywords appear, it is difficult to distinguish between syntactic candidates that have the same keywords but different orders (e.g., It is considered difficult to accurately identify the correct sentence from a large number of sentence candidates including similar expressions. In fact, there are documents that position the present technique as a means for preliminary selection of syntax.

(For example, Niimi, "Speech Recognition," Kyoritsu Shuppan, p.) The purpose of the present invention is to maintain a certain level of syntactic recognition performance regardless of the number of syntactic candidates, and to recognize a large number of syntactic candidates including similar expressions. The purpose of this invention is to provide a new syntax recognition method that enables accurate detection of correct syntax candidates from among them.

[Means and effects for solving the problem] The main components of the syntax recognition method of the present invention are a keyword dictionary, a word spotting unit that uses the dictionary, and a quantitative method for determining the similarity between an input sentence and syntactic candidates. It consists of a scoring section that performs evaluation.

The keyword dictionary is constructed by dividing it exclusively into syntax candidate units, and even within the same syntax candidate, into keyword categories that characterize the syntax candidate. The word spotting section performs word spotting for each keyword category between the keyword registered in that category and the input voice parameter string. the result,
Since the size of the dictionary used in word spotting remains constant regardless of the number of syntactic candidates, a certain level of syntactic recognition performance is maintained. Moreover, because keyword dictionaries have the above-mentioned exclusivity, the impact of errors in word spotting is limited to a specific keyword dictionary, i.e., within a specific keyword category characterized by syntactic candidates; At most, it is localized to the extent that it is reflected in the score of that construct.

In addition, the syntax scoring unit determines which keywords should be included in each syntactic candidate, what range each keyword should be in, what order the keywords should appear in,
Syntactic knowledge such as this is given in advance, and the keyword score values (analog values) calculated by the word spotting unit are weighted and added based on this knowledge. Furthermore, considering the difference in the number of keywords due to differences in syntax candidates,
Take the relative ratio (normalized score value) between the obtained score value and the upper limit of the score that can be achieved for the relevant construct,
This is taken as the score value of the syntax candidate of interest. By introducing a syntactic scoring method with such characteristics, it becomes possible to accurately detect the correct answer from among a large number of syntactic candidates containing similar expressions.

〔Example〕

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

In addition, in the following example, assuming a ticket reservation question and answer system in Japanese, the two syntax candidates are ``ticket reservation request'' and ``price inquiry'', and the same example as in Fig. 4 will be used. .

FIG. 1 shows a system configuration of an embodiment of the present invention. This syntax recognition system consists of a keyword dictionary section 1 in which keywords are registered, a word spotting section 2 that performs word spotting, and a syntax scoring/judgment section that calculates the score of each syntax candidate and selects the syntax candidate with the best score. Consists of 3.

As shown in FIG. 2, the keyword dictionary 1 is constructed mutually exclusively for each syntax candidate and, within a syntax candidate, for each keyword category that characterizes the syntax. That is, in the embodiment shown in FIG. 1, regarding the syntactic candidates for requesting ticket reservation, there are two categories: ``reservation request'' which is a collection of keywords representing the intention to reserve a ticket, and ``ticket'' which is a collection of keywords meaning tickets. Two categories of keyword dictionaries 11
, 12 are configured. This assumes that, for example, the syntax "I would like to reserve a ticket" is input. Similarly, regarding syntactic candidates for price inquiries, there are two types of keyword dictionaries: the ``inquiry'' category, which collects keywords expressing the intention to inquire about prices, and the ``price'' category, which collects keywords expressing prices.1.3 .14 is constructed.This is intended to detect, for example, the sentence "I want to know the price of seat A."

The word spotting unit 2 includes a selection circuit 21 that selects a keyword category, which is a unit of word spotting.
, an input register 22 that sets a subsequence of an input voice parameter string, a keyword register 23 that sets a voice parameter string of keywords registered in the dictionary 1, and a subsequence most similar to the keyword set in the keyword register 23. It consists of a keyword score rig unit 24 that finds the keyword from the input voice parameter string and holds the score. In this word spotting section 2, a selection circuit 21 sequentially selects keyword categories one by one, and a keyword scoring section 24 performs word spotting on all keywords registered in the category and holds their scores. do. Note that the word spotting method is the same as that described in connection with the prior art shown in FIG.

As described above, the syntax scoring/judgment unit 3 characterizes the syntax for each syntax, such as which keyword should exist, in what range should each keyword exist, and what order should each keyword appear in? Information (syntax knowledge)
is given in advance, and its operation is as follows.

First, focusing on the first syntax candidate (in this case, ticket reservation request), the keyword score rig unit 24 reads the score value calculated for the specified keyword category (reservation request, ticket). Next, given syntactic knowledge,
(Then, these score values are weighted and added.

For example, a weighted calculation such as (ΣW1j-w2j-Aj)/W3 is performed. Here, Aj is the score of keyword category j, and Wlj is the weight that keyword category j has as a syntactic determining factor (for example, Wl 1 = 1,
Wl 2 = 0.5), W2j is the weight related to the position where keyword category j is detected (for example, w2j = 1 when detected within the specified range, W2j = when outside the specified range
0.3), W3 is the weight related to the order of keywords (for example, when the detected keywords are in the specified order, W3=
1, when it is not in the designated order, W3 = 0.2). Furthermore, calculate the relative ratio between the obtained weighted addition result and the reachable upper limit of the weighted addition result for the relevant syntax candidate, and use this as the syntactic score of the relevant syntax candidate. Perform these operations for all syntax candidates. Then, calculate the syntactic score for each syntactic candidate. Finally, the syntax candidate with the best score among these syntax scores is selected and output as the correct syntax candidate.

FIG. 3 shows a system configuration of another embodiment of the present invention. In the case of the configuration shown in FIG. 1, the calculation time is dominated by the wrinkle botting process. FIG. 3 shows an attempt to speed up this word spotting process. That is, the third
The figure shows the word spotting process of FIG. 1 parallelized for each keyword category, and the selection circuit 21 of FIG. 1 has been removed accordingly. In addition, with the aim of performing syntactic scoring on syntactic candidates in parallel,
The syntax scoring/judgment unit 3 shown in FIG. 1 has functions divided into a syntax scoring unit 61, 62 and a syntax determination unit 7.

In FIG. 3, word spotting and syntactic scoring are performed in the same manner as in FIG. The obtained keyword scores are stored in the keyword scoring section.

Next, the syntax scoring units 61 and 62 respectively perform in parallel,
The score value related to the specified keyword category is read from the corresponding keyword scoring section, weighted addition and normalization processing are performed to obtain the syntax score. Finally, the syntax determining unit 7 selects the syntax candidate with the best score among the syntax scores obtained for each syntax candidate, and outputs this as the correct answer.

Furthermore, by developing the idea shown in Figure 3, it is possible to parallelize the word spotting sections up to the maximum number of word spotting sections in which each word spotting section shares one keyword, further speeding up the word spotting process. becomes possible.

〔Effect of the invention〕

As explained above, according to the present invention, by using a keyword dictionary that is exclusively divided and configured, the size of the dictionary used in word spotting remains constant regardless of the number of syntactic candidates. , this maintains a certain level of syntactic recognition performance.

Furthermore, the exclusivity of keyword dictionaries limits the impact of errors in word spotting to within a particular keyword category, making errors more or less localized and more or less reflected in the syntax score. It becomes a fail-safe system.

In addition, by weighted addition of keyword score values based on syntactic knowledge and normalization of the weighted addition results using the upper limit of the score for each syntax candidate, it is possible to accurately detect the correct syntax from a large number of syntax candidates including similar expressions. It becomes possible. moreover,
If highly accurate syntax recognition becomes possible in this way, the use of natural language processing techniques such as syntactic analysis and semantic analysis can be significantly simplified or even made unnecessary. can be configured economically and efficiently (speeding up syntax recognition).

[Brief explanation of drawings]

FIG. 1 is a diagram showing an embodiment of the syntax recognition method of the present invention, FIG. 2 is a diagram for explaining the keyword dictionary structure according to the present invention, and FIG. 3 is a diagram showing an embodiment of the syntax recognition method of the present invention. FIG. 4 is a diagram for explaining a conventional syntax recognition method using word spotting. 1...Keyword dictionary, 2...Word spotting section, 3...Syntax scoring/judgment section, 5
1-54...Word spotting section. 61.62...Syntax scoring section. 7...Syntax determination section. Figure 2 Keyword categories

Claims (4)

[Claims] (1) In a method that directly recognizes the syntax of a natural language sentence input by voice from a voice parameter string, a keyword dictionary section stores a voice parameter string characterizing the keyword, and the input voice parameter string and the keyword dictionary are stored. A word spotting unit that collates the voice parameter strings and calculates a score representing the probability that the voice parameter string exists; and a syntactic scoring unit that evaluates whether or not a syntactic candidate exists based on a specified keyword, and the keyword dictionary unit divides and configures the syntactic candidates into keyword categories that characterize the syntactic candidates. A syntax recognition method characterized in that a keyword category stores a voice parameter string of a keyword to which the category belongs, and the word spotting section performs word spotting processing for each keyword category. (2) The syntactic scoring unit expresses the probability of the existence of the syntactic candidate by adding weighted scores of keyword categories that characterize the syntactic candidate, based on syntactic knowledge regarding the type of keyword category, order of appearance, and range of appearance. The syntax recognition method according to claim 1, characterized in that a syntax score is determined. (3) The syntax scoring unit calculates a syntax score representing the probability of the existence of the syntax candidate obtained based on the score of a keyword category characterizing the syntax candidate, and calculates the relative 2. The syntax recognition method according to claim 1, wherein the value is a syntax score of the syntax. (4) The word spotting unit is provided with at least one word spotting unit for each of the keyword categories characterizing syntactic candidates, and performs word spotting processing in parallel on input speech parameter sequences of interest. The syntax recognition method described in Scope 1.