JPH0580793A - Interactive understanding device with word predicting function - Google Patents

Abstract

PURPOSE:To reduce the calculation area of a speech recognition device and a syntax and semantic analytic device and further shorten the calculation time by actually predicting words by the interactive understanding device which performs speech recognition and obtaining the vocabulary of the speech recognition device. CONSTITUTION:A plan recognition part 5 gathers a plan to be operated next by using a plan storage part 6 and a knowledge part 7 and regards this as an estimated plan. A prediction part 8 converts the estimated plan into actual words by using a plan word relation table 9 to obtain estimated words. Those estimated words are used as the vocabulary of the speech recognition device 1 to limit the vocabulary quantity of the speech recognition device 1.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a dialogue understanding device in various question answering systems including a database search and telephone exchange system using natural language voice, a dialogue translation system and the like.

[0002]

2. Description of the Related Art Several methods have been conventionally proposed as a method for processing a dialogue. One of the methods is a method using plan recognition. In this conventional plan recognition method, the input dialogue is parsed and semantically analyzed for each input sentence, the analysis result (structure) is collated with the plan of the plan recognition means, and information is extracted from the analyzed structure. Store in the information storage place of. Also,
It is a method of representing the interpretation and history of ongoing dialogue by sequentially combining a plurality of plans created according to the progress of the dialogue according to the conditions in the plan. Cognitive Science, Vol.11, Number
2, 1987, ABLEX PUBLISHING CORPORATION, DJLitman,
JFAllen "A Plan Recognition Model for Subdailogu
e inConverations ", Pages 163-200 (hereinafter referred to as Document 1).

Using this plan recognition method, Information Processing Society of Japan Natural Language Processing Research Group, 78-16, 1990.7.20, Information Processing Society of Japan, Takayuki Yamaoka Hitoshi Iida, "Context-aware speech recognition result narrowing method", Pages 121-128 proposes a method of predicting the next dialogue for the speech recognition result (hereinafter referred to as reference 2).

FIG. 5 is a block diagram of an apparatus using the conventional plan recognition system. The dialogue understanding device is a voice recognition device 1.
The speech recognition result obtained by the plan recognition unit 1 using the information of the dictionary 3 in the syntax and semantic analysis device 2.
The plan storage unit 1 is converted by the standard form and the plan recognizing unit 15 which are converted into the standard form which is the input of No.
6 and the knowledge storage unit 17 are used to output the combination of plans representing the interpretation and history of the ongoing dialogue.

As shown in FIG. 6, in this plan prediction method, in the prediction unit 8, the plan recognition unit 15 has a plan and a combination of plans that represent the history and the interpretation of the dialogue output from the plan recognition unit 15. A part of the plan of the plan storage unit 16 that satisfies the conditions indicated by the interpretation of the current dialogue and the history by comparing with the plan of the storage unit 16 is set as the prediction plan. Next, with respect to the input sentence (speech), the processing result of the syntax and semantic analysis device 2 is converted into a plan using a table or the like, and the one that matches the plan and the expected plan is selected by the selecting unit
In, the method of selecting.

[0006]

However, in the above-mentioned conventional method, since the recognition result is output in the voice recognition processing and the prediction is selected for the result, the processing such as the voice recognition is also performed for those rejected by the prediction. I'm doing it. For this reason, with respect to the calculation areas used in the speech recognition processing, the syntactic and semantic analysis processing, many calculation areas must be prepared for the results rejected by the prediction, and extra calculation processing must be performed for that purpose. There was a problem if it didn't happen. The present invention has been made to solve these problems, and an object of the present invention is to provide a dialogue understanding device that reduces the calculation area and the calculation time by predicting the vocabulary of voice recognition.

[0007]

In order to solve the above problems, the present invention uses a dictionary information to convert a sentence input from a voice recognition device into a standard form of the predicate and case component of the sentence. A syntactic and semantic analysis device that performs syntactic and semantic analysis, a plan storage unit that stores a predetermined plan, a collation between the plan and the standard form, and determination of whether or not the dialogue history and conditions match A plan recognition unit, a knowledge storage unit that stores knowledge depending on the area of the dialogue, a plan word relation table that indicates the relationship between the plan and words, a determination result by the plan recognition unit, and a plan that satisfies the conditions as a prediction plan A prediction unit that determines the prediction plan and changes the prediction plan into a prediction word by using the plan word relation table, and a selection unit that limits the next input vocabulary to only the prediction word by using the prediction word are provided. It is characterized in.

[0008]

According to the present invention having the above configuration, the number of vocabularies in the speech recognition apparatus can be limited, so that the calculation area can be reduced and the calculation time can be shortened. That is, in a dialogue understanding device having a word prediction function related to a voice recognition device, a grammatical procedure used in a dialogue for output of a result of parsing and semantic analysis of a sentence input to a voice input device, Recognizing the response procedure between the interlocutors, the knowledge and the procedure depending on the area of the dialogue, obtaining the prediction plan, and converting the prediction plan into the prediction word corresponding to the actual word using the plan word relation table.

[0009]

DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 is a block diagram of an apparatus including a dialogue understanding apparatus with a word prediction function to which the present invention is applied. In the figure,
1 is a voice recognition device, 2 is a syntactic and semantic analysis device, 3 is a storage device, and a dictionary is used in the syntactic and semantic analysis device 2,
Reference numeral 5 is a plan recognition unit, 6 is a plan storage unit used in the plan recognition unit 5, 7 is a knowledge storage unit used in the plan recognition unit 5, 8 is a prediction unit of a plan recognition result, and 9 is a plan word indicating a relation between a plan and a word. A relation table 10 is a selection unit for limiting the vocabulary of the next input, 11 is a plan recognition unit 5, a plan storage unit 6, a knowledge storage unit 7, a prediction unit 8, a word table 9 and a dialogue with a word prediction function including a selection unit 10. Understanding device.

Next, the outline of the operation of this embodiment will be described according to the processing flow shown in FIG. 3. First, the speech recognition result of the speech recognition device 1 is converted into the dictionary 3 in the syntax and semantic analysis device 2.
The information is converted into a standard form which is input to the plan recognition unit 5. In the standard form, the plan recognition unit 5 uses the plan storage unit 6 and the knowledge storage unit 7 to output the combination of plans representing the interpretation and history of the ongoing dialogue. Then, the predicting unit 8 predicts a plan representing the understanding and history of the dialogue and a plan corresponding to the next predicted input sentence using the plan storage unit 6. In the plan word relation table 9, the prediction plan obtained by the prediction unit 8 is converted into an actual word. In the selection unit 10, the words obtained from the plan word relation table 9 are
The output of the dictionary 3 is limited so that the vocabulary of the voice recognition device 1 and the syntax and semantic analysis device 2 is used.

The detailed operation of the syntax and semantic analysis device 2 will be described according to the device processing flow shown in FIG.
The syntactic and semantic analysis device 2 uses the dictionary 3 to convert the sentence input from the speech recognition device 1 into the standard form of the predicate and case component of the sentence. For example, the input sentence "Please tell me about the General Affairs Department, Public Relations Office, and Ono Office." Using the syntax and semantic analysis device 2 and the dictionary 3, (I wish (object (departm
ent_name, General Affairs Department) (section_name, Public Relations Office) (person
_name, Ono) (title, office manager))) standard form.

The detailed operation of the plan recognition unit 5 with this standard form as an input will be described. First, the contents of the plan storage unit 6 used by the plan recognition unit 5, that is, the contents of the plan will be described with reference to FIG. A plan has class, which is the type of plan. In this class, according to Reference 1, it is a domain-plan that is a plan peculiar to the area handled by the dialogue understanding device, or a general plan for dialogue such as grammar.
Describe whether it is ourse-plan. HEADER is the name of the plan and has a plan-name and its variables. It also shows the actions taken when the plan is executed. constraints
Describes the conditions for executing this plan. The plan searches the information storage location, working-memory, and is executed when this condition is satisfied. decompositions
Describes a sub-plan that is a plan for realizing this plan. This plan is realized when the sub-plan is realized. prerequisites is the above constrai
As with nts, describe the conditions for executing this plan. However, unlike constraints, when a condition is not satisfied, another necessary plan is executed to satisfy the condition. delete list and add li
st is for operating working-memory which is an information storage location when this plan is realized. delete
list deletes and add list adds the described items.

Then, the plan recognition section 5 collates the standard form output from the syntax and semantic analysis device 2 with the plan. Further, it is determined whether or not the plan matches the history of the dialogues stored so far in the plan recognition unit 5 and conditions.
For example, plan (inform_perso
(n_name) is a plan that describes the person's name, and the utterance needs to be in the standard form in which the case element of the predicate "is" or "hope" indicates the person's name. In addition, this utterance is a plan (inform_pe
rson_name), and the plan executed immediately before is the plan (r
If it is equest_action), a plan (ask_to_connect) that the user knows the name of the called party is introduced.

The plan recognition section 5 outputs the history of the dialogue and the contents of the current plan. For example, it is confirmed from the history of dialogues stored so far in the same unit 5 that the plan executed immediately before is a plan (request_action) requesting request contents. Plan (request_action) (inform_pe
rson_name) is connected to the plan (ask_to_connect), and as a history of the dialogue in the plan recognition unit 5, (ask_to_connect)
It is stored in the form of (request_action) (inform_person_name). Each plan is connected to the standard form of the spoken sentence, so two utterances "Which do you want to use?"
"I would like General Affairs Department, Public Relations Office, Ono Office Manager." Is for each plan (request_action) (inform_person_name)
Can be described. The contents of the current plan are (inf
oirm_person_name), in the example, (inform_person_n
ame ((department_name, general affairs department) (section_name, public relations room) (person_name, Ono) (title, office manager)). In addition, from the dialogue history, the plan (request_action) (inform_pe
rson_name) indicates that the plan (ask_to_connect), which has already been executed and the name of the called party is known from the user, has ended.

The knowledge stored in the knowledge storage unit 7 is knowledge about the area in which the dialog understanding device is used. For example, in the extension telephone exchange device, the organization or affiliation in the company is described. An example of knowledge about "Ono" is ((department_name, General Affairs Department) (section_name, Public Relations Office) (person_name, Ono) (title, Director), and in the input sentence "I would like to ask Mr. Ono of General Affairs", It is used by the plan recognition unit 5 to get information that Ono belongs to the Public Relations Office and that the General Affairs Department has no other person named Ono.

In the process outline of the prediction unit 8, for the prediction plan, the current dialog is compared by comparing the plan interpretation and the combination of the plans representing the history of the dialog output from the plan recognition unit 5 with the contents of the plan storage unit 6. The plan that satisfies the conditions indicated by the interpretation and history is judged as the prediction plan. Regarding the predicted word, the predicted plan is converted into a predicted word using the plan word relation table 9 and output.

Next, an example of selecting a prediction plan will be described in detail. For the purpose of prediction, first, the current state of the plan is understood from the history of dialogues stored so far in the plan recognition unit 5. For example, when predicting the next utterance of "Do you want to belong to the Public Relations Office of the General Affairs Department?", Confirm the department name and section name from the dialogue history stored in the plan recognition unit 5. Plan for (comfirm_department_and_se
ction) is running and the plan (comfirm_department_
and_section) confirmation plan for success (comfi
rm_value) works and is a plan for the question sentence (inf
orm_if) works.

For the prediction, the plan storage unit 6 makes the above-mentioned judgment.
Of all the plans described above are compared with the history of the dialogues stored in the plan recognition unit 5, and the plan in the plan storage unit 6 that satisfies the conditions for operation (constraints) is checked. Get as a forecast plan. For example, the prediction plan for the utterance example "Do you want to belong to the Public Relations Office of the General Affairs Department" is a condition for operation (constraints)
Is satisfied, the plan (a
ffirmative) (for example, “Yes, yes” is the case), as a plan showing a negative answer, (negative)
(For example, "No, it is different" corresponds.),
A plan (unknown) indicating that it is unknown (for example, "I do not know the section name" corresponds), a plan indicating that information is transmitted (inform_value) (for example, "General Affairs Department of General Affairs Department" Corresponding.) Etc. are selected.

An example of conversion of predicted words will be described in detail. In order to obtain a prediction word from the prediction plan, the prediction unit 8 develops the prediction plan into a prediction word using the plan word relation table 9 or the like. The plan word relation table 9 is arranged in the order of plan name or variable type, representative form, and actual word.
For example, (inform_value, inform_value, [is, probably, I think, please, please, ...]) or (department_name, General Affairs Department, [General Affairs Department, General Affairs]) It

The expansion of the prediction plan into words is performed by two procedures: expanding the plan name into words and making variables in the plan into words. The expansion of the plan name into words converts the plan name into words in the predicted plan name / plan word relationship table 9. In order to obtain a prediction word from the prediction plan, the variables of the predicted plan are converted into words from the plan word relation table 9 according to the constraints.

As the variable, a representative form determined from the dialogues so far is obtained from the prediction plan, and this representative form is converted into a word in accordance with the plan word relation table 9. For variables that have not been determined, according to the variable type,
Desired. For example, a plan (inform-depa
(rtment-name DEPARTMENT_NAME) is the variable to expand into words when it is predicted as `` General Affairs Department, General Affairs Department, Person in Charge According to the plan word relation table 9, the constituent words of the noun phrase representing “general affairs department” such as “” are obtained. Also, if the variable has not been determined, the variable DEPARTMENT
Since _NAME is a variable representing a department name, the plan word relationship table 9 is used to obtain the constituent words of the noun phrase that becomes the department name, that is, "general affairs department, accounting department, sales department, ...". These procedures are performed for all forecast plans.

The selection unit 10 uses the predicted word output from the prediction unit 8 to limit the vocabulary of the dictionary 3 in the next voice input to only the predicted word. That is, for the next voice input, the predicted word is used as the vocabulary of the voice recognition device 1 and the syntactic and semantic analysis device 2.

[0023]

As described above, according to the present invention,
In the dialogue understanding device, the actual word prediction is performed and the vocabulary of the speech recognition device is used, so that the calculation region and the calculation time required for the speech recognition device and the syntactic and semantic analysis device in the conventional method can be compared with the calculation region. Can be reduced and the calculation time can be shortened.

[Brief description of drawings]

FIG. 1 is a schematic device configuration diagram of a device of the present invention.

FIG. 2 is a diagram showing an example of contents of a plan of the present invention.

FIG. 3 is a diagram illustrating a processing flow of an apparatus to which the present invention has been applied.

FIG. 4 is a diagram showing an example of prediction word selection to which the present invention is applied.

FIG. 5 is a schematic device configuration diagram of a conventional device.

FIG. 6 is a diagram exemplifying prediction word selection in the conventional method.

[Explanation of symbols]

 1 Speech Recognition Device 2 Syntax and Semantic Analysis Device 3 Dictionary 5 Plan Recognition Unit 6 Plan Storage Unit 7 Knowledge Storage Unit 8 Prediction Unit 9 Plan Word Relationship Table 10 Selection Unit 11 Dialogue Understanding Device

Claims (1)

[Claims] 1. A dialogue understanding device having a word predicting function related to a speech recognition device, which converts a sentence input from the speech recognition device into a standard form of a predicate and a case component of the sentence using information of a dictionary. A syntactic and semantic analysis device that performs syntactic and semantic analysis, a plan storage unit that stores a predetermined plan, a collation between the plan and the standard form, and determination of whether or not the dialogue history and conditions match A plan recognition unit, a knowledge storage unit that stores knowledge depending on the area of the dialogue, a plan word relation table that indicates the relationship between the plan and words, a determination result by the plan recognition unit, and a plan that satisfies the conditions as a prediction plan Judge,
And a prediction unit that changes the prediction plan into a prediction word by using the plan word relation table, and a selection unit that limits the vocabulary of the next input using the prediction word to only the prediction word. For the output of the result of parsing and semantic analysis of the input sentence, recognize the grammatical procedure used in the dialogue, the response procedure between the interlocutors, the knowledge and the procedure depending on the area of the dialogue, and make a prediction plan. A dialogue understanding device with a predictive word function, characterized by converting the predictive plan into a predictive word corresponding to an actual word using a plan word relation table.