JPH05281989A - Speech language interface device - Google Patents

Abstract

PURPOSE:To provide the speech language interface device of morpheme level equipped with a function for adjusting the difference between grammar for speech recognition and grammar for language processing. CONSTITUTION:A speech recognizing device 1 recognizes an input speech by using the grammar for speech recognition stored in a memory 4 and outputs a morpheme series to a morpheme adjustment part 2. The morpheme adjustment part 2 converts the morpheme series outputted by the speech recognizing device 1 into a morpheme series to be inputted to a syntax analyzing device 3 by using a rule for morpheme adjustment stored in a memory 5. The syntax analyzing device 3 inputs the morpheme series outputted by the morpheme adjustment part 2 analyzes the syntax by using grammar for syntax analysis stored in a memory 6, and outputs the meaning expression of the input sentence.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech language interface device, and more particularly, to a speech recognition device and a syntax analysis device.
When different grammars optimized for each application are used, the speech recognition device is adjusted while adjusting the background linguistic framework, vocabulary division, part-of-speech system, etc., between the two grammars. The present invention relates to a speech language interface device that reduces the load on the entire speech language system by avoiding redundant morphological analysis processing between the speech recognition device and the syntactic analysis device by passing the morpheme sequence from the syntactic analysis device to the syntactic analysis device. .

[0002]

2. Description of the Related Art Conventionally, a speech language system has been required to be processed in two phases: "speech recognition" and "language processing". The speech recognition generates symbolic data such as a phoneme string, a character string, or a word string from speech data uttered by acoustic processing. On the other hand, the language processing generates a syntactic structure or a semantic structure from the symbolic data output by the speech recognition by the symbol processing.

However, in recent years, in order to improve the accuracy of speech recognition, in addition to the conventional acoustic processing, speech processing has come to perform language processing utilizing knowledge about morphology and syntax. As a result, the voice recognition device is now able to output a morpheme string to which grammatical information such as a part-of-speech or a conjugation is added. On the other hand, a language processing device is generally composed of a (broadly defined) syntactic analysis part and a semantic analysis part, although it is somewhat different depending on its use (machine translation, dialogue interface, etc.). The syntactic analysis unit includes a morphological analysis unit that handles morphological information and a (narrowly defined) syntactic analysis unit that handles syntactic information.

[0004]

However, the grammar for speech recognition and the grammar for language processing are different in their linguistic frameworks, vocabulary delimiters, part-of-speech systems, etc., because they are used for different purposes. Therefore, conventionally, the morpheme information output from the speech recognition apparatus cannot be used as it is by the language processing apparatus, and the language processing apparatus must eventually redo the analysis from the level of the phoneme string or the character string. That is, the conventional interface system of the speech recognition device and the language processing device has a problem that it is redundant and inefficient because the morphological analysis processing overlaps between the speech recognition device and the language processing device.

The reason why the syntactic analysis units of the speech recognition device and the language processing device adopt different grammars is that the requirements for the grammars (language models) of the two are different. While speech recognition uses a language model as a means to efficiently limit the search space and improve recognition accuracy, syntactic analysis aims to obtain a linguistically valid analytic structure. Also,
The grammar for speech recognition is used for prediction or generation, while the grammar for language processing is used for analysis. Therefore, there is the same difficulty in integrating the two as in the case of bidirectional grammar. That is, under the present circumstances, if the same grammar is used for both purposes, either the execution efficiency of the speech recognizer or the generality of the language processor grammar must be sacrificed.

Further, in the grammar for speech recognition, since the description of the connection constraint between morphemes is the main focus, short unit words tend to be used as vocabulary items, whereas in the grammar for language processing, , Since the purpose is to obtain a semantic expression, a long unit word that represents a semantic unit is often adopted. For example, the end-of-sentence expression "must-have" is replaced with "no / te / wa / nari / mase /"
However, the language processing unit considers it as one auxiliary verb that represents the mood of the subject. In this way, there is a slight deviation between the two morpheme units.

Therefore, a main object of the present invention is to provide a function of adjusting a difference (grammatical framework, vocabulary segmentation, part-of-speech system, etc.) between a speech recognition grammar and a language processing grammar, and a speech recognition device and a language. It is an object of the present invention to provide a spoken language interface device that can avoid duplication of morphological analysis in a processing device.

[0008]

SUMMARY OF THE INVENTION The present invention is a speech language interface device for delivering a morpheme string of an input speech recognized by a continuous speech recognition means to a syntax analysis means in the form of input data. Storage means for storing the morpheme adjustment rule for converting the morpheme sequence output from the morpheme sequence into the morpheme sequence of the syntactic analysis means, and the morpheme adjustment rule stored in the storage means for the morpheme sequence output from the continuous speech recognition means. With reference to, morphological adjustment means for converting grammatical information such as vocabulary delimiters, parts of speech, and conjugations to match the vocabulary and items of the syntactic analysis means.

[0009]

In the speech language interface apparatus according to the present invention, when there is a difference between the speech recognition grammar and the language processing grammar regarding the grammatical framework, the vocabulary segmentation, the part-of-speech system, etc., the morpheme adjusting means performs the morpheme adjustment rule. According to the above, by converting the morpheme data output from the continuous speech recognition device into a format that can be treated as an input of the language processing device,
Avoid duplication of morphological analysis processing between the continuous speech recognizer and the language processor.

[0010]

1 is a schematic block diagram of an embodiment of the present invention. First, the configuration of an embodiment of the present invention will be described with reference to FIG. The input voice is recognized by the voice recognition device 1 using the voice recognition grammar stored in the memory 4, and a morpheme string is output to the morpheme adjustment unit 2.
The morpheme adjustment unit 2 uses the morpheme adjustment rule stored in the memory 5 to convert the morpheme sequence output by the speech recognition device 1 into a morpheme sequence that is input to the syntax analysis device 3.
The syntactic analysis device 3 receives the morpheme string output by the morpheme adjusting unit 2 as input, performs syntactic analysis using the syntactic analysis grammar stored in the memory 6, and outputs the semantic representation of the input sentence.

2 to 4 are views showing a concrete example of a voice language interface system according to an embodiment of the present invention, and in particular, FIG. 2 is a sentence saying "You must send us a registration form". Is a description of the morpheme sequence output by the speech recognition apparatus for the sentence, FIG. 3 is an example of the description of the morpheme adjustment rule applied to this sentence, and FIG. It is a morpheme string converted to

Next, the specific operation of the embodiment of the present invention will be described with reference to FIGS. The voice recognition device 1 recognizes a continuous voice uttered in phrase units and outputs a sequence of voice recognition candidates. For example, the sentence "You must send us your registration form" is divided into two clauses, "Please send us your registration form" and "You must send us". The recognition result as shown is obtained. The data structure of FIG. 2 has a lattice structure so that a plurality of candidates can be handled for the same voice section.
The character string enclosed in "" is the recognized morpheme:
Pos represents the part of speech of the morpheme, and: prob represents the likelihood of speech recognition.

The morpheme adjustment unit 2 applies the morpheme adjustment rule stored in the memory 5 to the morpheme sequence output by the speech recognition apparatus 1 to perform conversion on a morpheme unit basis. For example, in FIG.
If there is a morpheme adjustment rule as shown in, the morpheme adjusting unit 2 converts the part of speech of the grammar for speech recognition stored in the memory 4 into the part of speech of the grammar for parsing stored in the memory 6 at the same time. , The morpheme sequence of sentence end expression "n / te / ha / nari / masen / n" is put together into one auxiliary verb. Also, the bunsetsu boundary is removed, and a morpheme string as shown in FIG. 4 is output. The morpheme adjustment rule stored in the memory 5 is an annotated regular grammar (regular grammar).
The rewriting rules described in (grammar) are handled. The left side of this rule is the vocabulary item of the syntax analysis device 3, and the right side is the annotated regular expression of the vocabulary item of the speech recognition device 1. Notations, parts of speech, inflectional forms, and inflectional forms can be written in the annotation section. Since the rewriting rule described in the range of the regular grammar can be converted into the deterministic finite state automaton, the morpheme adjusting unit 2 operates at high speed. The syntactic analysis device 3 receives the morpheme string output by the morpheme adjustment unit 2 as an input, and stores it in the memory 6
Parsing is performed using the parsing grammar stored in, and the semantic representation of the input sentence is output.

[0014]

As described above, according to the present invention, in the interface portion between the speech recognition means and the language processing device, the morpheme adjusting means outputs the morpheme data output by the continuous speech recognition device according to the morpheme adjustment rule to the language processing device. Since the conversion is made so that it can be used as an input, it is possible to realize an efficient speech language interface in which the morphological analysis processing does not overlap between the speech recognition apparatus and the language processing apparatus.

[Brief description of drawings]

FIG. 1 is a schematic block diagram of an embodiment of the present invention.

[Fig.2] "You must send us a registration form."
It is a figure which shows the example which described the morpheme sequence which the voice recognition device outputs with respect to the sentence.

3 is a diagram showing an example in which a morpheme adjustment rule applied to the sentence shown in FIG. 2 is described.

FIG. 4 is a diagram showing a morpheme string converted by a morpheme adjusting unit so as to be input to a syntax analysis device.

[Explanation of symbols]

 1 voice recognition device 2 morpheme adjustment unit 3 syntax analysis device 4, 5, 6 memory

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Toshiyuki Takezawa Kyoto, Soraku-gun Seika-cho, Osamu Osamu, Osamu Osamu, 5 Hiratani, Arai Co., Ltd. Automatic translation telephone laboratory (72) Inventor, Takuma Morimoto Kyoto, Soraku Gunma Seika-cho, Osamu Inui, Osamu, 5 Mihiratani Co., Ltd.

Claims (1)

[Claims] 1. A speech language interface device for delivering a morpheme string of input speech recognized by a continuous speech recognition means to a syntax analysis means in the form of input data of the syntax analysis means, wherein the continuous speech recognition means. A storage unit that stores a morpheme adjustment rule for converting the morpheme sequence output from the morpheme sequence into the morpheme sequence of the syntax analysis unit, and a morpheme sequence output from the continuous speech recognition unit are stored in the storage unit. With reference to the morpheme adjustment rule, the vocabulary delimiter, grammatical information such as part of speech, and conjugation are used in the vocabulary of the parsing means
A spoken language interface device comprising morpheme adjusting means for converting so as to match an item.