JP2003233391A - Language processor - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To realize robust, highly precise and fast language processing. <P>SOLUTION: In a word dictionary 1, a symbol $ is added to a word which is possibly a slot value (keyword) of a facility name, a symbol % is added to a word which is possibly a slot value of the type of industry, and a symbol Ð is added to a word which is possibly a slot value of an address. A speech recognition part does not impart a category symbol (Ð/$/%,...) to a reading of each registered word in the word dictionary 1, but imparts a category symbol to only word notation (the left side of the word dictionary) so as to perform recognition processing according to the reading (phoneme string) of each word by using an acoustic model, the word dictionary 1, and a statistic language model. The speech recognition part performs the speech recognition processing by using such a word dictionary 1 and then a speech recognition result character string such as 'Well, it is % a restaurant in Ð Chikusa Ward.' is obtained. Hence the calculation cost for category identification processing, etc., by a meaning understanding part can effectively be reduced. <P>COPYRIGHT: (C)2003,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition unit that recognizes a speaker's voice as a character string, and a candidate that is a candidate for a desired keyword from a voice recognition character string that is output information from the voice recognition unit. The present invention relates to a language processing device having a meaning understanding unit that extracts words, identifies categories of candidate words, and specifies the keywords. Therefore, the present invention can be applied to, for example, a car navigation system mounted on a vehicle, and a voice dialogue for performing dialogue for filling information corresponding to a so-called “slot” such as a facility name or an address of a destination. It can be applied to devices and the like.

[0002]

2. Description of the Related Art As a language processing device having a meaning understanding part for specifying a keyword from a voice recognition character string which is output information of a voice recognition part, for example, Japanese Patent Laid-Open Publication No. Hei 7-262190: Language Processing Device. ("Reference 1" below)
Say ), The device described in Japanese Patent Laid-Open No. 8-263494: Language analysis device (hereinafter referred to as "Document 2"), and the like are generally known.

[0003] Document 1 described above describes a device intended to perform a semantic analysis on an input sentence including various phrases. In this conventional device, the input sentence is morpheme before the semantic analysis process. The semantic feature information of each word is analyzed and determined by referring to the semantic feature knowledge storage means. Further, the latter document 2 describes a device intended for linguistic analysis so as to obtain each structure based on a character string of a spoken voice sentence. In this conventional device, the case structure of an input sentence ( Morphological analysis is performed to obtain morphological information, syntactic information, and semantic information of each word included in the input sentence in order to obtain (syntactic structure).

In the field of speech recognition, a speech recognition method for recognizing only predetermined keywords from speech (so-called "word spotting technology") or a language that defines (limits) the arrangement of words or categories based on a fixed grammar. A voice recognition method using a model is often used. However, in a conventional device using a voice recognition method based on these techniques and grammars, for example, in the conventional device as described above, it is determined at which position in a recognized word string a certain word belonging to a certain category appears. , It is possible to recognize keywords and extract keywords and identify categories at the same time.
It is not possible to robustly recognize the language of the user's various expressions (including grammatically inappropriate sentences). Therefore, at present, generally, a speech recognition method using a statistical language model is generally used as a speech recognition method for such a phrase.

[0005]

However, in a voice dialog device using a statistical language model, it is not clear which word appears in which position in a word string obtained as a recognition result. Therefore, to determine which word is a keyword and which category it belongs to in the recognized word string, prepare a separate word category table for reference processing, or Since it becomes necessary to perform morphological analysis, in these processes, it is necessary to perform a process with high calculation cost other than voice recognition for keyword extraction and category identification. That is, in the conventional meaning comprehension process of identifying the category of a word using the morphological analysis of a word string recognized by speech, the calculation cost (time, memory capacity) required for the process is large.

For example, as described above, in the conventional apparatus, since the processing based on the morpheme analysis is performed as the semantic analysis processing of the free utterance, the extraction of the word as the keyword and the category (in the above invention, the semantic feature and the meaning are defined). High computational cost is required for the identification of (information). Further, the conventional technology is premised on that the input word string forms a grammatically correct sentence, but the accuracy of the current actual voice recognition technology is high enough to correctly recognize all natural utterances. Therefore, it is difficult to obtain a guarantee that the expected semantic processing can be reliably executed in the above-described conventional technology.

The present invention has been made to solve the above problems, and an object thereof is to realize robust, highly accurate and high-speed language processing.

[0008]

Means for Solving the Problems, and Functions and Effects of the Invention In order to solve the above problems, the following means are effective. That is, the first means extracts a candidate word that is a candidate for a desired keyword from a voice recognition unit that recognizes a speaker's voice as a character string and a voice recognition character string that is output information of the voice recognition unit. Then, in the language processing device having a meaning understanding part for identifying the category of the candidate word and specifying the keyword, the above-mentioned speech recognition part, for each notation character string of each registered word that can be a candidate word, for each registered word Based on the word dictionary to which each category symbol indicating the category of, is provided with a character string generation means for generating a voice recognition character string, further, the meaning understanding unit, the presence or absence of the category symbol in the voice recognition character string or It is to have a symbol search means for searching a value.

As a result, the meaning understanding unit in the subsequent stage can easily extract the above-mentioned keyword from the recognized word string (that is, the voice recognition character string which is the output information of the voice recognition unit) at a low calculation cost, and You can identify the category. That is, according to the means of the present invention, since the recognized word itself is provided with the symbol indicating the category, the category identification can be performed simultaneously with the keyword extraction, so that the calculation cost in the semantic processing can be reduced.

In addition, in the current speech recognition technology using a statistical language model, a grammatically inadequate sentence (and the utterance itself may be grammatically inadequate) is recognized as a recognition result due to erroneous recognition or the like. Often obtained. Such a grammatically inadequate sentence cannot be robustly and accurately analyzed by the current morphological analysis technique. However, according to the above-mentioned means of the present invention, the meaning understanding process is performed based on the above "category symbol" given to the recognized word without performing the language process such as morphological analysis. It is possible to robustly and accurately perform desired language processing even on an unrecognized recognized word string.

The second means of the present invention is the above-mentioned first means.
This means is provided with a dialogue control means for determining the type, configuration or meaning of the response screen or response sentence corresponding to the content of the voice recognition character string. A third means is that, in the above-mentioned second means, it is provided with a response sentence generating means for generating a response sentence as a series of output character strings. The fourth means is that the second or third means is provided with a voice output means for converting the response sentence into voice and outputting the voice.

That is, as can be seen from the above-mentioned principle of operation,
The present invention is characterized by the procedure for determining keywords in language processing, and is not necessarily premised on voice output. Therefore, the present invention can be applied to a language processing device such as an automatic translation device or an automatic minutes generating device which is not premised on a real-time response or an interactive response to the user. The operation of the present invention described above
The effect is effective not only for Japanese processing but for any natural language processing. By the means of the present invention described above, the above problems can be effectively or rationally solved.

[0013]

BEST MODE FOR CARRYING OUT THE INVENTION The present invention will be described below based on specific embodiments. However, the present invention is not limited to the examples shown below. [Embodiment] In this embodiment, a destination setting task will be described as an example. FIG. 1 is a system configuration diagram illustrating a logical configuration of the voice interaction device 10 of the present embodiment. In the destination setting task according to the present embodiment, there are three types of destinations, “facility name”, “industry”, and “address”, while interacting with the user.
Suppose you need to fill one slot (the keyword tray).

FIG. 2 and FIG. 3 are schematic diagrams illustrating the configuration of the word dictionary 1 and the statistical language model 2 used by the voice dialog device 10, respectively. In these word dictionary 1 and statistical language model 2, a word that is a slot value (keyword) of a facility name is preceded (or back) by a $ sign, and a word that is a slot value of an industry is preceded (or after) by a symbol. %symbol,
It is assumed that the @ symbol is added to the word that becomes the slot value of the address.

In the speech recognition, the recognition process is executed according to the reading of the word (phoneme string; the right side of the description example of the word dictionary in FIG. 2) using the acoustic model, the word dictionary 1, and the statistical language model 2. Therefore, in the above-mentioned word dictionary 1, a symbol is not given to reading, and a symbol (category symbol) indicating a category only in the word notation (on the left side of the description example of the word dictionary).
Shall be given.

As a result of the voice recognition processing performed by the voice recognition unit 11 using the above-mentioned word dictionary 1 (FIG. 2) and the statistical language model 2 (FIG. 3), the following <actual dialogue A> The following voice recognition result <voice recognition character string B> is obtained in response to the “user side response”. <Actual dialogue A> Voice dialogue device 10: Please tell us the address of the shop and the type of business. User Response: Well, this is a restaurant in Chikusa Ward. <Voice recognition character string B> User response: Eh @ @ Chikusa-ku% Restaurant.

A language processing procedure in the meaning understanding unit 12 after the voice recognition result <voice recognition character string B> as described above is obtained will be illustrated below. FIG. 4 is a general flowchart illustrating a language processing procedure of the meaning understanding unit 12 of the voice interaction device 10.

In this flowchart, first, the above-mentioned <voice recognition character string B> is input from the voice recognition unit 11. Next, in step 123, the leading one character of each word in the recognized word string (voice recognition character string B) is examined to extract a keyword candidate word. That is, when the first character at the beginning of each word in the word string is a $ symbol, a% symbol, or an @ symbol, those words are extracted as keyword candidate words. At this time,
The other words are not used in the subsequent processing and need not be retained. Therefore, the following keyword string (keyword candidate word) is held by this processing.

<Extracted keyword string> (User response): @ Chikusa-ku% Restaurant

Next, in step 126, the category of each candidate word is identified based on the category symbol (first character: @ /% / $) of each candidate word. For example, when the above <voice recognition character string B> is obtained, if the first character of the extracted keyword is the $ symbol, the slot value is the facility name, and if the% symbol is the slot value of the industry,
In case of @ symbol, it is the slot value of the address. As a result, each slot is filled as follows.

<Resolution result of each slot value> Industry Slot Value = Restaurant Address slot = Chikusa Ward

For example, as a result of the above-mentioned processing, the voice recognition character string is input, the word extraction unit extracts only the keywords to which the category symbols are added, and the word category identification unit identifies the keyword categories according to the category symbols. can do.

In the above embodiment, special characters (for example, @ /% / $ ..., etc.) are used as category symbols.
Any characters can be used for these category symbols depending on the type of language to be analyzed. Also,
In the above embodiment, the category symbol is limited to one character, but the number of characters of the category symbol may be arbitrary.

Further, the entire categories themselves are hierarchized,
For example, in the case of classifying into three layers, the category symbols may be similarly divided into three layers. As a result, it becomes easy to handle complicated categories (concepts) in a finely hierarchical manner. For example, when representing an address, “@ 1” is applied as a category symbol to prefecture names, “@ 2” is applied to city names as a category symbol, and “@ 3” is assigned to category names as a category symbol. In some cases, it may be effective to implement layering such as applying as.

[Brief description of drawings]

FIG. 1 is a system configuration diagram illustrating a logical configuration of a voice dialog device 10 according to an embodiment of the present invention.

FIG. 2 is a schematic diagram illustrating the configuration of a word dictionary 1 used by the voice interaction device 10.

FIG. 3 is a schematic diagram illustrating a configuration mode of a statistical language model 2 used by the voice interaction device 10.

FIG. 4 is a general flowchart illustrating a language processing procedure of the meaning understanding unit 12 of the voice dialog device 10.

[Explanation of symbols]

1… Word dictionary 2… Statistical language model 10 ... Spoken dialogue device (language processing device) 11 ... Voice recognition unit 12 ... Meaning understanding 123 ... Word extraction unit 126 ... Word category identification unit 13 ... Dialogue control unit 14 ... Voice output section $… Category code (store name or facility name) %… Category symbol (type of industry or product name) @… Category symbol (place name or address)

Front page continuation (51) Int.Cl. ⁷ Identification code FI theme code (reference) G10L 15/18 G10L 3/00 R 15/22 15/28

Claims (4)

[Claims] 1. A voice recognition unit that recognizes a voice of a speaker as a character string, and a candidate word that is a candidate for a desired keyword is extracted from a voice recognition character string that is output information of the voice recognition unit, In a language processing device having a meaning understanding unit that identifies a category of candidate words and specifies the keyword, the voice recognition unit, in each notation character string of each registered word that can be the candidate word,
Based on a word dictionary to which a category symbol indicating the category of each of the registered words is respectively added, it has a character string generation means for generating the voice recognition character string, the meaning understanding unit, the meaning in the voice recognition character string A language processing apparatus having a symbol search means for searching for the presence or value of a category symbol. 2. The language processing apparatus according to claim 1, further comprising a dialogue control unit that determines a type, a structure, or a meaning of a response screen or a response sentence corresponding to the content of the voice recognition character string. 3. A response sentence generating means for generating the response sentence as a series of output character strings is provided.
The language processing device according to. 4. The language processing apparatus according to claim 2, further comprising a voice output unit that converts the response sentence into voice and outputs the voice.