KR100766058B1 - method and apparatus for exceptional case handling in spoken dialog system - Google Patents

Abstract

The present invention relates to a method and apparatus for handling an exception in an interactive voice interface system. The present invention extracts an exception utterance and speech recognition error by the user from the speech conversation corpus, and makes an exception handling target DB among them, and then creates exception resolution information and rules for solving such an exception situation, and actual system In case exception spoken is input from, we want to improve user's satisfaction and improve system performance by handling exception situation using pre-established DB, information and rules.

Interactive Voice Interface System, Exception Handling, Dialog Modeling

Description

Method and apparatus for exceptional case handling in spoken dialog system

1 is a block diagram showing an interactive voice interface device according to the present invention

2 is a configuration diagram showing in detail the exception processing unit in the interactive voice interface system according to the present invention;

3 is a flowchart illustrating an exception handling method according to the present invention.

* Explanation of symbols for main parts of the drawings

100: interactive voice interface device 110: voice recognition unit

120: conversation processing unit 121: exception processing unit

123: input sentence analysis unit 125: conversation management unit

127: task management unit 129: output statement generation unit

130: speech synthesis unit 200: speech conversation corpus

210: exception situation corpus 220: exception handling target DB

221: Service not supported expression DB 223: Chinese expression DB

225: Speech recognition error DB 227: Error discrimination DB

230: error resolution processing unit

TECHNICAL FIELD The present invention relates to a voice chat processing technique used in an interactive voice interface system, and more particularly, to a technique for dealing with an exception situation during a conversation.

In interactive speech interface systems, exceptions occur in speech recognition errors and in recognition result processing errors due to the user's lack of understanding of the system, and disrupt the normal flow of conversations, making the system unable to achieve its intended purpose. .

Prior art has mainly focused on resolving speech recognition errors to solve this exception. In other words, we tried to solve this problem by requesting recurrence when this value is below the threshold using language model reliability, acoustic model reliability, or parsing reliability of the recognition result of the speech recognizer. Further attempts have been made to resolve the error, such as restoring the expected section or changing the current position in the dialogue flow.

However, it has a limitation that it is impossible to discriminate an error when the reliability of the speech recognition result is higher than the threshold, and the error handling is performed under the assumption that the user's speech is basically made in accordance with the intention of the system. If an exception occurs frequently by an unexpected user, such as a system, there is a problem that an error remains in the processing result even though error processing is performed.

As the interactive voice interface system has not been fully practical to the general public yet, and has many limitations due to voice recognition performance and application area limitation, it is very difficult for the general user to use it naturally. In this case, it is not possible to effectively deal with the exceptions caused by users frequently occurring in the actual interactive voice interface system by applying only the countermeasures for the speech recognition error in a state in which the general consideration of the case is insufficient.

Therefore, the present invention has been made to solve the above problems, the dialogue flow proceeds to solve the exception naturally by using the pre-established DB and exception resolution information and rules to handle the exception situation of the system It is an object of the present invention to provide a method and apparatus for handling exceptions to achieve a goal.

The characteristic of the exception handling method for achieving the above object is (a) building a plurality of DB for exception handling, (b) creating and storing exception resolution information and rules in the plurality of DB And (c) processing the exception by using the prepared exception information and the rule when the expression corresponding to the exception is uttered by the user.

Preferably, the step (a) comprises: constructing an exception situation corpus indicating an exception situation and a speech recognition error in a speech conversation corpus formed of a speech file collected in a mock conversation manner and a speech transcription text and a speech recognition result text for the speech file. And constructing at least one of a service-unsupported expression DB, a median expression DB, a speech recognition error DB, and a speech error determination DB necessary for processing the exception based on the constructed exception situation corpus. .

Preferably, the step (b) may generate and store exception resolution information and rules for each exception situation stored in the service unsupported expression DB, the intermediate expression DB, the speech recognition error DB, and the speech error determination DB. .

Preferably, the step (c) may include generating words and characters that can be provided by the system to be serviced when the service-unsupported expression due to words and sentences that cannot be provided by the system is uttered, and using the generated words and characters. To present a supportable service.

Preferably, the step (c) includes re-determining whether it is possible to solve the problem by using the history when the speech is included with the neutral expression due to the word or sentence having the neutrality, and using the history as a result of the determination. Resolving neutrality if the resolution is possible, and resolving the neutrality by using the concomitant to resolve the neutrality if it is impossible to resolve the neutrality using the history of the determination result.

Preferably, the step (c) includes determining whether the voice recognition error can be solved using an error solving table when a voice recognition error exception situation due to the frequent voice recognition error word, sentence, or pattern is found, and the determination result. If it is possible to solve using the error elimination table to resolve the voice recognition error through this, and if it is impossible to solve using the error elimination table as a result of the determination, neutrality or multi-neutral by using the supplementary to solve the recognition error And switching to a modal input device.

Preferably, in the step (c), if a speech error that determines a user's answer as an error occurs when a sentence of a similar word order consisting of the same or similar words is found in the full-length chapter, the error is detected by using the incidental signal to solve the speech error. It is characterized by eliminating.

Features of the exceptional situation processing apparatus for achieving the above object are a speech recognition unit for converting a user's speech into text, a conversation processing unit for receiving the converted text to generate an output sentence for speech synthesis, and the generated An interactive voice interface device comprising a voice synthesizer for generating a synthesized sound based on an output sentence, wherein the voice recognition unit converts the voice file using at least one of a voice file collected by a simulated conversation method, a voice transcription sentence, and a collected voice file. A speech conversation corpus formed by adding the speech recognition result text, an exception context corpus composed of an exception situation and a speech recognition error corresponding to the speech conversation corpus, and specific information corresponding to the exception context using the configured exception speech corpus Exception handling target DB comprising a DB and the exception Using information comprised in the destination DB is characterized in that it comprises a processing unit for creating the error eliminated error resolution information and the rules for each representation.

Preferably, the voice conversation corpus is constructed in a simulated conversation method, which is a Wizard-of-Oz-Simulating (WOZ) method.

Preferably, the exception handling target DB is a service unsupported expression DB that collects words and sentences that a user frequently utters by intuition but cannot be provided by a system to be serviced, and a word or sentence having neutrality is collected. The user's response, such as a speech expression DB, a speech recognition error DB that collects frequent speech recognition error words, sentences, and patterns, and a sentence of a similar word order composed of the same or similar words as the previous sentence. It is characterized by consisting of a speech error determination DB that collects the rules that can be determined as an error.

Other objects, features and advantages of the present invention will become apparent from the following detailed description of embodiments with reference to the accompanying drawings.

A preferred embodiment of the exception handling method in the interactive voice interface system according to the present invention will be described below with reference to the accompanying drawings.

1 is a block diagram showing an interactive voice interface device according to the present invention.

As illustrated in FIG. 1, the interactive voice interface device 100 includes a voice recognition unit 110 that converts a user's speech into text, an exception situation processing target DB that is already built by receiving the converted text, and a solution for solving the same. The dialogue processing unit 120 generates an output sentence required for the speech synthesis in which the exception is processed using exception resolution information and rules, and the speech synthesis unit generating the synthesized sound based on the output sentence output through the exception processing in the dialogue processing unit. 130).

At this time, the conversation processing unit 120 extracts an exception speech and a voice recognition error by a user, and handles the exception using an exception processing target DB that is already built and exception resolution information and rules for solving the exception. The situation processing unit 121, an input sentence analysis unit 123 for identifying a user's speech intention based on the text output through the exception situation processing, and the dialog managing unit 125 for determining which conversation flow the current speech is in And a task management unit 127, and an output sentence generation unit 129 for generating an output sentence necessary for speech synthesis.

2 is a block diagram showing in detail the exception processing unit in the interactive voice interface system according to the present invention.

As shown in FIG. 2, the voice conversation corpus 200 configured by adding the speech recognition result text converted by the speech recognition unit 110 using the speech file and the voice transcription sentence collected by the simulation conversation method and the collected speech file. ), An exception situation corpus 210 composed of an exception situation and a speech recognition error corresponding to the speech conversation corpus 200, and specific information corresponding to the exception situation using the configured exception situation corpus 210 to a DB. It is composed of an exception situation processing target DB (220) configured and an error resolution processing unit (230) for creating error resolution information and rules for each expression by using the information configured in the exception situation processing target DB (220).

In this case, the voice conversation corpus 200 is preferably constructed by a simulation conversation method such as the Wizard-of-Oz-Simulating (WOZ) method.

In addition, the exception situation processing target DB 220 is a service unsupported expression DB 221 that collects words and sentences that a user frequently utters by intuition but cannot be provided in a system to be actually serviced, and a word having a neutrality or A idiom DB 223 that collects sentences, a VO of DB 225 that collects frequent speech recognition error words, sentences, and patterns, and a similar word order in which the user's answer is the same or similar to the previous sentence. It is configured to include a speech error determination DB 227 that collects rules for determining a user's answer as an error, such as when a sentence or the like appears.

In this case, other types of DBs may be added to the exception situation processing target DB 220 as necessary.

An operation according to the exceptional situation processing apparatus and the interactive voice interface apparatus using the same according to the present invention configured as described above will be described in detail with reference to the accompanying drawings.

Referring to FIG. 2, a voice conversation corpus 200 is constructed from a voice file and a voice transcription sentence collected by a simulated conversation method, and converted by the voice recognition unit 110 using the collected voice file. After the speech recognition result text is further added to the speech conversation corpus 200, an exception situation corpus 210 composed of an exception situation and a speech recognition error is constructed in response to the speech conversation corpus 200. FIG.

At this time, the voice conversation corpus 200 collects the conversation while providing only the same information to the user as when using the system to be actually constructed later.

The method of collecting the conversation naturally proceeds with the conversation at the performance level of the target system, while collecting the conversation by inducing the user to cause an exception in the user's speech.

The collected speech of the user is constructed using the voice recognizer to be applied to the interactive voice interface system as a voice conversation corpus 200 together. In addition, the constructed voice conversation corpus 200 may be reconstructed by an exception situation corpus 210 by annotating to determine a normal case, an exception situation by a user, and a case of a recognition error.

In this way, using the annotated exception situation corpus 210 to build a variety of DB for exception processing.

At this time, as the DB that can be constructed, first, the user frequently utters by intuition, but the 'service unsupported expression DB 221', which collects words and sentences that cannot be provided in a system to be actually serviced, and has a neutrality 'Intermediate expression DB (223)' that collects words or sentences, and 'Voice recognition error DB (225)' that collects frequent speech recognition error words, sentences, and patterns. There is a "speech error determination DB (227)" that collects rules for judging a user's answer as an error, such as when a sentence of similar word order composed of similar words appears, and other kinds of DBs are added as necessary. Can be.

When the exception situation processing target DB (220) construction is completed, the exception situation resolution information and rules are written to solve the exception situation using the built DB.

That is, when the corresponding expression is input using the 'service unsupported expression DB 221', the message 'service not supported' can be outputted and the related service that can be supported according to the input word type is prepared. .

If the neutrality can be resolved using the dialogue history of the corresponding expression using the 'intermediate expression DB 223', the identifier information is attached and the system then resolves the neutrality using the dialogue history. If this is not possible, the system should be configured to develop a system-driven error-resolving subdialog to resolve the neutrality of each expression.

Next, if a result misrecognized using the 'voice recognition error DB 225' is a semantically impossible expression, and if the same misrecognition is performed in the same situation, an error resolution table is created so that automatic correction is possible. Similarly, the unit can be deployed for deployment, or if multi-modal input equipment is available, it can be used to resolve the error. If it is expected that the expression present in the speech recognition error DB is likely to appear as a response, ask the user to speak in advance to improve the recognition performance or if a multimodal input device is available. You can also switch to this to write rules to prevent speech recognition errors at an early stage.

Further, by using the 'speech error determination DB 227', it is determined whether the corresponding speech of the user corresponds to an error according to the utterance error discrimination rule. Be prepared.

The exception resolution information and rules mentioned so far can be prepared in various ways depending on the applied system, but for the sake of understanding, errors that may occur assuming a weather-providing interactive voice interface system and representative methods for solving the exceptions Examples are shown in Table 1.

Exception Exception How to clear the exception Service Unsupported Expression DB  Can I wash my car? Is it hot this summer? When did Mt. Seorak fall? While the system's capabilities can only provide simple weather information, it asks for weather information that requires complex inference. In the case of frequent unsupported utterances, it is possible to recognize voices and solve exceptions by guiding the local weather and weather conditions that can be provided by the system when such utterances are input. Chinese expression DB  Please tell me how Seoul is I can't judge what you want to know in Seoul based on the current sentence. It can be seen that such expression is possible only in the presence of the above information request utterance, so that when the expression with the identifier information is input as input, the system uses the conversation history to resolve the neutrality.  What is the expected temperature today? It is not possible to determine whether the expected temperature asks for the expected minimum temperature or the expected maximum temperature. If such ignition comes into the input without additional information that can resolve the neutrality, it is possible to determine whether to ask for the expected minimum temperature or the expected maximum temperature. Speech Recognition Error DB  Jeju Island The speaker's intended speech is 'Jeju-do?' If the expression belonging to the frequently misunderstood object comes in as input, if 'Jeju Island' is always the same misrecognized expression, and if the misrecognized utterance does not have a meaning, it is automatically corrected by firing 'what is Jeju Island?' If you cannot, neutralize the problem Ignition error determination DB  ('Tell me the temperature today' after the ignition) If the user provided a response and the user re fired with the same or similar expression as the previous one. In case of such utterances, the second-level recognition result information is used to ask whether the user's intended utterance is the next-level question or to restart the conversation from the previous step. If a multi-modal input device is available, switch to the multi-modal input device, and if it is not available, resolve the exception by requesting to fire a pseudo-syntax.

3 is a flowchart illustrating an exception handling method according to the present invention, which is described below by applying the prepared exception handling target DB 220 and error solving information and rules as shown in Table 1 below.

First, the voice recognition unit 110 converts the user's speech into text to recognize the voice (S10).

Subsequently, the exception processing unit 121 checks whether there is an expression corresponding to each exception situation processing target DB 220 in which the input content as a result of the recognized speech recognition is pre-established before the utterance input from the user (S20). ).

If the corresponding expression exists as a result of the check, the error is resolved according to the error resolution information and rule written for the corresponding expression of each exception situation DB.

That is, if it is determined that the service is not supported speech due to the words and characters that cannot be provided in the system to be actually serviced (S30) (S30), the service is not supported and generates a supportable service (S40).

When it is determined that the check result (S20), the speech containing the neutral expression due to the word or sentence having a neutrality (S50), it is determined again whether or not the neutrality can be solved using the history (S60). If neutrality can be solved using history, neutrality is solved (S70). If neutrality cannot be solved using history, neutrality is solved using incidentalization according to neutralization.

In addition, if it is determined that the voice recognition error expected speech due to the check result (S20), frequent voice recognition error word, sentence, pattern (S90), it is determined again whether the voice recognition error can be solved using the error solving table (S100). . And if it is possible to solve using the error elimination table to solve the voice recognition error through this (S110), if it is not possible to solve using the error elimination table, neutral resolution or multi-modal input by using the side of the recognition error resolution Switch to the equipment (S120).

In addition, if the user response is determined to be a utterance error, such as when the check result (S20), a sentence of a similar word order composed of the same or similar words as the previous sentence, etc. (S130), the incidentalization according to the resolution of the utterance error Solve the error by using (S140). In addition, the situation in which the error is resolved is additionally stored in the utterance error determination DB, so that when the same utterance is input to the same situation later, the error can be solved without developing the additional unit (S150).

In this way, the dialogue processing is performed according to the result of the exception processing unit 121 to generate an output sentence (S160), and the speech synthesis unit 120 generates a synthesized sound based on the output sentence (S170).

Although the preferred embodiments of the present invention have been illustrated and described above, the present invention is not limited to the above-described embodiments, and the present invention is not limited to the above-described embodiments without departing from the spirit of the present invention as claimed in the claims. Various modifications may be made by those skilled in the art, and such modifications are intended to fall within the scope of the appended claims. Therefore, the true technical protection scope of the present invention will be defined by the technical spirit of the appended claims.

As described above, the method and apparatus for handling exceptions in the interactive voice interface system according to the present invention can effectively handle exceptions compared to the conventional method, thereby improving user satisfaction with the system and improving the success rate of the system. The height is effective.

Claims (10)

delete (a) constructing an exception situation corpus indicating an exception condition and a speech recognition error in a speech conversation corpus constructed of a speech file collected in a simulated conversation manner and a speech transcription text and a speech recognition result text for the speech file; (b) constructing at least one or more of a service unsupported expression DB, a median expression DB, a speech recognition error DB, and a speech error determination DB required for exception handling based on the constructed exception situation corpus; (c) generating and storing exception resolution information and rules in the plurality of DBs; (d) if the expression corresponding to the exception is uttered by the user, handling the exception using the generated exception resolution information and rule Exception handling method comprising a. The method of claim 2, wherein step (c) comprises: Generating and storing the exception resolution information and rule for each exception situation stored in the service unsupported expression DB, the expression expression DB, the speech recognition error DB, and the speech error determination DB, respectively. Exception handling method comprising a. The method of claim 2, wherein step (d) Generating words and characters that can be provided by the system when a service-unsupported expression due to words and sentences that cannot be provided by the system to be serviced is uttered; And Presenting a supportable service using the generated words and letters Exception handling method comprising a. The method of claim 2, wherein step (d) Determining whether the neutrality resolution is possible by using a history when a neutral expression including words or sentences having neutrality is spoken; Resolving the neutrality when the neutrality resolution is possible using the history as a result of the determination; And If it is impossible to resolve the neutrality by using the history as a result of the determination, resolving the neutrality by using the subsidiary according to the neutral resolution. Exception handling method comprising a. The method of claim 2, wherein step (d) Determining whether a speech recognition error can be solved using an error resolution table when a speech recognition error expected expression due to frequent speech recognition error words, sentences and patterns is uttered; Resolving the voice recognition error by using the error resolving table when it is possible to solve the voice recognition error using the error resolving table as a result of the determination; And If it is impossible to solve the speech recognition error using the error elimination table as a result of the determination, resolving the neutrality or switching to a multi-modal input device by using the auxiliary according to the resolution of the recognition error; Exception handling method comprising a. The method of claim 2, wherein step (d) If the user response is determined to be an error consisting of a sentence of similar word order consisting of the same or similar sentences to the previous sentence, using the incidental to solve the utterance error to resolve the error Exception handling method comprising a. An interactive voice including a speech recognition unit for converting a user's speech into text, a dialogue processing unit for receiving the converted text and generating an output sentence for speech synthesis, and a speech synthesis unit for generating synthesized sounds based on the generated output sentence. In the interface device, A speech conversation corpus formed by adding the speech recognition result text converted by the speech recognition unit by using one or more of the speech file collected by the simulated conversation method, the voice transcription, and the collected speech file; An exception situation corpus consisting of an exception and a speech recognition error in response to the speech conversation corpus; An exceptional situation processing target DB including specific information corresponding to the exceptional situation using the configured exception situation corpus, and a DB; And an error resolving processor for creating error resolving information and rules for each expression by using the information configured in the exception situation processing target DB. The method of claim 8, The speech conversation corpus is an exception handling apparatus, characterized in that it is constructed in a simulated conversation method of the Wizard-of-Oz-Simulating (WOZ) method. The method of claim 8, wherein the exception situation processing target DB, Service unsupported representation DB consisting of words and sentences that the user frequently speaks but cannot be provided in the system to be serviced; A neutral expression DB consisting of words or sentences having neutrality; Frequent speech recognition error DB consisting of a speech recognition error word, sentence and pattern; And Spoken error determination DB consisting of rules for judging the input sentence as an error if the user answer is a sentence of a similar word order composed of the same or similar words as the previous sentence An exception handling apparatus comprising a.

Images