JP3241556U - language data processing system - Google Patents

Abstract

A linguistic data processing system with a wider range of applications is provided. A language data processing system (1) includes a processing unit (11) and a storage unit (12). The storage unit 12 stores a language processing model M realized by machine learning technology and including a plurality of intention tags d1. The processing unit 11 obtains a semantic analysis result by performing semantic analysis processing on the text data corresponding to the speech input using the language processing model M, and based on the semantic analysis result, the intention matched in the intention tag d1 Determining whether there is a tag, and if it is determined that there is no matching intention tag, operate based on the semantic analysis result and N language processing histories related to the text data using the language processing model M By executing the prediction process, a target operation procedure corresponding to the text data is selected from the operation procedure d2 and executed, and N is an integer of 1 or more. [Selection drawing] Fig. 1

Description

The present invention relates to data processing systems and, more particularly, to linguistic data processing systems that process user dictated content.

With the development of language processing technology, when trying to operate a voice-operable electronic device, the user is not limited to a specific voice command, and can issue commands to the electronic device in more everyday expressions. can be done.

Taiwan Patent Application Publication No. 201916002

However, there are various expressions of commands that express the same request, and there is a possibility that the conventional technique cannot determine the user's intention for a command whose expression of the request is not clear. There is room.

SUMMARY OF THE INVENTION It is therefore an object of the present invention to provide a linguistic data processing system that alleviates at least one of the disadvantages of the prior art.

The language data processing system includes a processing unit and a storage unit electrically connected to the processing unit.

The storage unit stores a language processing model realized by machine learning technology. A language processing model includes multiple intent tags. Each of the intent tags corresponds to at least one of a plurality of action steps to be performed by the processing unit.

The processing unit obtains text data corresponding to the voice input, performs semantic analysis processing on the text data using the language processing model, obtains a semantic analysis result corresponding to the text data, and performs semantic analysis based on the semantic analysis result. If it is determined that there is no intention tag matched in the intention tag, the semantic analysis result and the text data are related using the language processing model. A target motion procedure corresponding to the text data is selected from the motion procedures by executing motion prediction processing based on N language processing histories, where N is an integer equal to or greater than 1, and the target motion procedure is executed. be configured to

When it is determined that there is no intention tag matched in the intention tag (that is, when the user's intention expressed by the text data corresponding to the voice input cannot be clearly determined), the processing unit outputs the language processing result and A target action procedure corresponding to the text data is selected from the action procedures by executing the action prediction process based on the language processing history. In other words, when it is determined that the user's intention cannot be determined only from the content of the user's speech, the processing unit determines the intention further based on the content of the user's recent speech, and selects an operation procedure to be executed. Therefore, even if the expression of the user's intention is not complete, or even if the user speaks while inserting content unrelated to the intention between the expressions of the intention, the language data processing system of the present invention can judge the intention. It is possible to expand the applicable range of voice operation.

Other features and advantages of the present invention will become apparent in the following detailed description of embodiments which refers to the accompanying drawings.

1 is a block diagram illustratively illustrating one embodiment of a language data processing system of the present invention; FIG. 1 is a flow chart exemplifying a language data processing method executed by a language data processing system; FIG. 3 is a diagram showing an example of a conversation between a user and a service device; FIG.

Before describing the present invention in more detail, as used herein, the term "coupling" or "connection" refers to a direct connection between a plurality of electrical devices/equipment/equipment by means of conductive material (e.g., electrical wires). or that there is an indirect connection between two electrical appliances/devices/installations by one or more other appliances/devices/installations or wireless communication. Also, as used herein, the term "unit" refers to computer hardware, for example, "control unit" refers to computer hardware or a combination thereof having control functions.

Referring to FIG. 1, one embodiment of a language data processing system 1 is shown. Language data processing system 1 can be electrically connected to one or more service devices 2 via a network. Here, for convenience of explanation, the language data processing system 1 electrically connected to one service device 2 will be explained as an example. Therefore, in FIG. 1, one service device 2 is depicted.

In this embodiment, the service device 2 is installed in a guest room of an accommodation facility (for example, a hotel), and includes a microphone (not shown), a speaker (not shown), and a display (not shown). ,including. As a result, the service device 2 picks up the voice of a user (for example, a guest in a guest room) via a microphone, transmits corresponding voice data using a speaker, and responds using a display for the user's reference. information. Note that the service device 2 is not limited to this embodiment. In other words, the location of application of the present invention is not limited to lodging facilities.

In this embodiment, the language data processing system 1 is a server device and includes a processing unit 11 and a storage unit 12 electrically connected to the processing unit 11 . The processing unit 11 is electrically connected to the service device 2 through a network. In this embodiment, the processing unit 11 is a single central processing unit having data calculation and processing functions. In terms of form, the processing unit 11 may be a combination of multiple central processing units, and is not limited to this embodiment. In addition, in this embodiment, the storage unit 12 is one data storage device for storing digital data, specifically, for example, a hard disk. or a combination of multiple computer-readable storage media, and is not limited to this embodiment. Also, in another embodiment, the language data processing system 1 may be a plurality of server devices electrically connected to each other, and is not limited to this embodiment.

The storage unit 12 stores a language processing model M that is realized by machine learning technology and used by the processing unit 11 . In this embodiment, the language processing model M includes a semantic analysis submodel m1 and a motion prediction submodel m2.

More specifically, the semantic analysis submodel m1 includes a plurality of preset intention tags d1. The intention tag d1 indicates an intention and may include, but is not limited to, those indicating "turn off lights", "set alarm", "turn on TV", and the like. In this embodiment, the semantic analysis sub-model m1 is a neural network, which is trained by machine learning using a plurality of phrase data as training data after setting the intention tag d1. Phrase data is a phrase directly or indirectly expressing the request in natural language, and may include, but is not limited to, for example, "The TV in the room does not work. It seems to be broken."

Thereby, the trained semantic analysis sub-model m1 is capable of performing semantic analysis processing on the input data, performing the semantic analysis processing to infer intent from the input data, and at least one Extract one or more characters that are one point.

The semantic analysis sub-model m1 may also determine whether the user's intent represented by the input data matches at least one of the intent tags d1. For example, for the input data "Turn off the lights in the room", the semantic analysis submodel m1 executes semantic analysis processing on the input data to determine the intention expressed by the input data. It is determined that the intention tag d1 matches the intention tag d1 indicating "turn off the lights", and "turn off the lights" and "turn off" are extracted from the input data.

The motion prediction sub-model m2 includes a plurality of motion procedures d2 that are set in advance and executed by the processing unit 11. FIG. The processing unit 11 controls the service device 2 and causes the service device 2 to output data by executing the operation procedure d2. For example, the electronic device (for example, (lights, televisions, electric curtains, etc. in the same room) to transmit control signals.

The action procedure d2 corresponds to the intention tag d1 included in the semantic analysis submodel m1. More specifically, each intent tag d1 corresponds to at least one action procedure d2. For example, the intention tag d1 indicating "turn off the lights" includes an operation procedure d2 for causing the service device 2 to transmit a control signal to the lamp to turn off the lights, and a voice indicating that the service device 2 has turned off the lights. and an operation procedure d2 for causing data to flow.

In this embodiment, the motion prediction sub-model m2 is a neural network, which is trained by machine learning using a plurality of conversation data as training data after the motion procedure d2 is set. Each piece of conversation data indicates the content of multiple utterances made by two people taking turns in conversation, and includes a plurality of request phrases, a plurality of reply phrases corresponding to the request phrases, and a plurality of requests related to the request phrases. It contains an intent and multiple request point parts associated with the request phrase. The plurality of request phrases and the plurality of reply phrases are ordered, and each of the reply phrases follows the corresponding request phrase, in other words, the plurality of request phrases and the plurality of reply phrases are interlaced. More specifically, a request phrase is a colloquial expression of a request by a person to whom a service is provided (eg, a hotel guest) using natural language. A reply phrase is a colloquial expression of a response to a request by a service provider (eg, a hotel employee) using natural language. Each request intent corresponds to one of the request phrases and represents the intent of the corresponding request phrase. The request intent may also be generated by performing a semantic analysis process on the corresponding request phrase using the semantic analysis sub-model m1, or set by a person. A request point portion is one or more characters that correspond to one of the request phrases and are points in the corresponding request phrase. Also, the request point part may be generated by executing semantic analysis processing on the corresponding request phrase using the semantic analysis sub-model m1, or may be set by a person.

This allows the trained motion prediction sub-model m2 to perform motion prediction processing based on the input data, and select one of the motion procedures d2 by performing the motion prediction processing.

Note that the action procedure d2 is not necessarily included in the action prediction sub-model m2. For example, in another embodiment, the action prediction sub-model m2 includes a plurality of action tags respectively corresponding to action procedures d2, and the processing unit 11 is configured to execute the corresponding action procedures d2 based on the action tags. be done.

Referring to FIG. 2, the language data processing method executed by the language data processing system 1 is exemplified below.

In step S1, the processing unit 11 obtains text data corresponding to speech input from the service device 2, and uses the semantic analysis submodel m1 of the language processing model M to perform semantic analysis processing on the text data. , to obtain a semantic analysis result corresponding to the text data. That is, text data is input to the semantic analysis sub-model m1 as input data, and the semantic analysis result is output data generated based on the text data by the semantic analysis sub-model m1.

In this embodiment, the semantic analysis result includes inferred intent data indicating the intent inferred from the text data, and at least one point portion including at least one character each serving as a point extracted from the text data. Inferred intent data is a multi-dimensional semantic vector. That is, the inferred intent data indicates the intent represented by the text data in the form of vectors. Specifically, a plurality of component vectors of the inferred intention data, which is a semantic vector, correspond to a plurality of intention tags d1, respectively, and have component vector magnitudes (for example, in the range of 0 to 1, but limited to this). not) indicates the degree of matching with the inferred intention data of the corresponding intention tag d1. Each of the point parts is at least one character determined to be highly relevant to the intent in the text data by the semantic analysis submodel m1.

Also, the semantic analysis sub-model m1 performs semantic analysis processing and generates semantic analysis results by performing calculations on text data using its own parameters. Calculations for the parameters and text data of the semantic analysis sub-model m1 are the result of training the semantic analysis sub-model m1 by machine learning, and the details are not the point of this specification, so they will not be described in detail.

In this embodiment, the voice input is an electrical signal of a voice picked up by the service device 2, for example a single phrase spoken by a user (eg, a guest) to express a request. After obtaining the voice input, the service device 2 uses voice recognition technology to process the voice input, generates corresponding text data and sends it to the processing unit 11 of the language data processing system 1 . In another embodiment, after receiving the voice input, the service device 2 directly transmits the voice input to the processing unit 11 of the language data processing system 1, and the processing unit 11 uses voice recognition technology to recognize the voice input. may be processed to generate corresponding text data.

In step S2, based on the semantic analysis result, the processing unit 11 determines whether there is an intention tag matched in the intention tag d1. More specifically, the processing unit 11 determines whether there is at least one intention tag d1 whose degree of matching with the inferred intention data is equal to or greater than the matching threshold in the intention tag d1, thereby determining whether the matched intention tag is determine whether there is When it is determined that there is at least one intention tag d1 whose degree of matching with the inferred intention data is equal to or greater than the matching threshold among the intention tags d1, the processing unit 11 selects the intent tag d1 with the highest degree of matching with the inferred intention data. is a matched intention tag, and it is determined that there is a matched intention tag in the intention tag d1. When it is determined that there is no intention tag d1 whose degree of matching with the inferred intention data is equal to or greater than the matching threshold in the intention tag d1, the processing unit 11 determines that there is no intention tag matched in the intention tag d1.

If the processing unit 11 determines that there is a matched intent tag, the flow proceeds to step S3, otherwise the flow proceeds to step S4.

When it is determined that there is a matched intent tag, in step S3 the processing unit 11 performs at least one action procedure d2 corresponding to the matched intent tag. For example, when the matched intention tag is the intention tag d1 indicating "turn off", the processing unit 11, in this embodiment, causes the service device 2 to send a control signal to the lamp to turn off the light. and an operation procedure d2 for outputting voice data indicating that the service device 2 has turned off the lights, but the present invention is not limited to this.

After executing step S3, the flow proceeds to step S6.

When it is determined that there is no matching intention tag, in step S4, the processing unit 11 performs semantic analysis results and N language processing related to the text data using the motion prediction submodel m2 of the language processing model M. A target operation procedure corresponding to the text data is selected from the operation procedure d2 by executing the operation prediction process based on the history. More specifically, the processing unit 11 generates a semantic integration result based on the semantic analysis result and the N language processing histories, and selects the operation procedure d2 that has the highest degree of matching with the semantic integration result in the operation procedure d2. is selected as the target operating procedure. That is, the semantic analysis result and the N language processing histories are input to the action prediction submodel m2 as input data, and the target action procedure is applied to the semantic analysis result and the N language processing histories by the action prediction submodel m2. is the output data generated based on In this embodiment, N is 2, but in other embodiments, N may be any integer greater than or equal to 1 and is not limited to this embodiment.

Also, the N language processing histories are N text data ( hereinafter referred to as N pieces of past text data). The N pieces of past text data respectively correspond to the N pieces of speech input that were input before the speech input was input to the service device 2 (hereinafter referred to as N pieces of past speech input), that is, Each corresponds to N phrases spoken by the user within the fixed period of time.
In the present embodiment, the N language processing histories are processed by the language data processing method for N past text data before the processing unit 11 executes the current language data processing method for the text data. is executed, semantic analysis processing is performed on each of N pieces of past text data to generate corresponding N pieces of semantic analysis results (hereinafter referred to as N pieces of past semantic analysis results) , are generated based on at least N past semantic analysis results and stored in the storage unit 12 . The language processing history includes past speculative intent data and at least one past point portion. The past guessed intention data is a multidimensional semantic vector, and is the guessed intention data of the past semantic analysis result of the correspondence. Each of the past point parts is one of the point parts of the past semantic analysis result of the correspondence, which is in the past text data of the correspondence and determined by the semantic analysis sub-model m1 to be highly relevant to the intent.

Furthermore, the motion prediction submodel m2 uses its own parameters to perform calculations on the semantic analysis results and the N language processing histories, respectively, so as to correspond to the semantic analysis results and the N language processing histories ( N+1) semantic vectors are generated, and (N+1) semantic vectors are concatenated to generate a semantic matrix, and the semantic matrix is taken as a semantic integration result. Then, the action prediction submodel m2 uses its own parameters to calculate the degree of matching with the semantic integration result of each action procedure d2, and the action procedure d2 has the highest degree of matching with the semantic integration result. Select d2 as the target operating procedure. The calculations for the parameters and semantic analysis results of the action prediction submodel m2 and the N language processing histories are the results of training the action prediction submodel m2 by machine learning. I won't go into detail because I don't.

In step S5, the processing unit 11 executes the target operation procedure.

In step S6 following steps S3 and S5, the processing unit 11 generates another language processing history (hereinafter referred to as the latest language processing history) based on at least the semantic analysis result, and the latest language processing history The latest language processing history is stored in the storage unit 12 so that the history can be used in future executions of the language data processing method. The latest language processing history is the same as the N language processing histories described above, so the description will not be repeated here.

Referring to FIG. 3, the technical effect of the language data processing system 1 of the present invention will be described by taking a conversation between the user and the service device 2 as an example.
First, in Conversation 1, the user inputs a first voice input msg1 (ie, "Is there a restaurant in the hotel?" in FIG. 3) into the service device 2, and the service device 2 responds to the first voice input msg1. Then, voice data saying "Yes, the restaurant is on the second floor" was output. Then, in the following conversation 2, for example, the user felt that the volume of the voice data output by the service device 2 was too loud, so the user sent the second voice input msg2 to the service device 2 (that is, "Turn down the volume" in FIG. 3). ”), and the service device 2 outputs the voice data “Yes, the volume is turned down” in response to the second voice input msg2. Subsequently, the user entered a third voice input msg3 (ie "How much is it?" in FIG. 3) into the service device 2 .

When the third voice input msg3 is viewed alone, it is unclear what the user asked about the price, and the expression is incomplete. When performing the language data processing method on the input msg3 (that is, in step S1 the processing unit 11 obtained the text data corresponding to the third speech input msg3), in step S2 the processing unit 11: It is determined that there is no intention tag matched in the intention tag d1, and the process proceeds to step S4. That is, the processing unit 11 cannot determine the user's intention only from the text data corresponding to the third voice input msg3. In step S4, the processing unit 11 generates the semantic analysis result associated with the third speech input msg3, the language processing history associated with the first speech input msg1, and the language processing history associated with the second speech input msg2. to select a target motion procedure related to the third voice input msg3 from the motion procedure d2. The language processing history related to the first speech input msg1 includes past inference intention data indicating that the intention inferred from the text data corresponding to the first speech input msg1 is "ask about the restaurant"; and a past point part indicating "Restaurant" extracted from the text data corresponding to the speech input msg1 of . The language processing history related to the second speech input msg2 includes past inferred intention data indicating that the intention inferred from the text data corresponding to the second speech input msg2 is "Turn down the volume of the speech"; and two past point portions respectively indicating "volume" and "lower" extracted from the text data corresponding to the second voice input msg2. In other words, the processing unit 11 analyzes the user's intention based on the semantic relevance among the three parties and selects the target action procedure. In this example, with the machine learning trained motion prediction sub-model m2, the processing unit 11 determines that the third speech input msg3 is more semantically related to the first speech input msg1. , that the user is asking about restaurant prices, and therefore the target action sequence selected is, for example, to cause the service device 2 to output restaurant price information.

It should be understood that steps S1 to S6 and the flowchart of FIG. 2 merely exemplify the language data processing method. Even if steps S1 to S6 are combined, divided, or changed in order, if the same effect can be obtained by substantially the same method as the language data processing method of the present invention, the scope of rights of the present invention include. Accordingly, steps S1 through S6 and the flowchart of FIG. 2 are not intended to limit the present invention.

In summary, when determining that there is no intention tag matched in the intention tag d1 (that is, when the user's intention represented by the text data corresponding to the voice input cannot be clearly determined), the processing unit 11 , a target operation procedure corresponding to the text data is selected from the operation procedure d2 by executing the operation prediction process based on the language processing result and the language processing history. In other words, when it is determined that the user's intention cannot be determined based only on the content of the user's speech, the processing unit 11 further determines the intention based on the content of the user's recent speech, and selects the operation procedure d2 to be executed. do. Therefore, even if the expression of the user's intention is not complete, or even if the user speaks while inserting a content separate from the request, the language data processing system 1 of the present invention can determine the intention, and the voice operation can be performed. You can expand the scope of application.

In the above description, for purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of the embodiments. However, it will be apparent to one skilled in the art that one or more other embodiments may be practiced without the specific details. In addition, in the descriptions indicating "one embodiment" and "one embodiment" in this specification, all descriptions accompanied by indications such as ordinal numbers are included in specific implementations of the present invention having specific aspects, structures, and features. It should be understood that Further, in this specification, at times multiple variations may be incorporated into a single embodiment, drawing, or description thereof for the purpose of streamlining the specification and recognizing the versatility of the invention. It is intended to be understood that one or more features or specific examples of one embodiment may, where appropriate, be applied to one or more of the other embodiments in the practice of the invention. It may be implemented with more features or with specific embodiments.

Although the embodiments and variations of the present invention have been described above, the present invention is not limited to these, and includes all modifications and equivalent configurations as various configurations included within the spirit and scope of the broadest interpretation. shall include

1 language data processing system 11 processing unit 12 storage unit M language processing model m1 semantic analysis submodel m2 action prediction submodel d1 intention tag d2 action procedure 2 service device msg1 first voice input msg2 second voice input msg3 third Voice input S1-S6 step

Claims (7)

a processing unit;
a storage unit electrically connected to the processing unit;
The storage unit stores a language processing model realized by machine learning technology,
The language processing model includes a plurality of intent tags,
each of the intent tags corresponds to at least one of a plurality of operational procedures performed by the processing unit;
The processing unit is
Obtaining text data corresponding to voice input and performing semantic analysis processing on the text data using the language processing model to obtain a semantic analysis result corresponding to the text data,
determining whether there is an intention tag matched in the intention tag based on the semantic analysis result;
When it is determined that there is no matching intention tag in the intention tag, motion prediction processing based on the semantic analysis result and N language processing histories related to the text data using the language processing model. to select a target operation procedure corresponding to the text data from the operation procedures, N being an integer equal to or greater than 1;
configured to perform the target operating procedure;
Language data processing system. the processing unit is electrically connected to a service device;
the voice input is input through the service device;
The N language processing histories respectively correspond to N past text data obtained before obtaining the text data, and the N past text data are obtained before the speech input is input. 2. A language data processing system according to claim 1, corresponding to each of the N input past speech inputs. The action prediction process generates a semantic integration result based on the semantic analysis result and the N language processing histories, and selects the operation procedure having the highest degree of matching with the semantic integration result in the operation procedure. 2. The language data processing system according to claim 1, wherein the language data processing system is selected as a target operation procedure. the semantic analysis result includes inferred intention data corresponding to the intention tag;
4. The language data processing system of claim 3, wherein each of said language processing histories includes past inferred intention data corresponding to said intention tag. the semantic analysis result further includes at least one point portion extracted from the text data using the language processing model;
5. The language processing history of claim 4, wherein each of said language processing histories further includes at least one past point portion extracted from a corresponding one of said N past text data using said language processing model. Language data processing system. the semantic analysis result includes inferred intention data corresponding to the intention tag;
The processing unit determines whether the intention tag includes the intention tag whose degree of matching with the inferred intention data is equal to or greater than a matching threshold, and the degree of matching with the inferred intention data is equal to or greater than the matching threshold. When it is determined that the intention tag exists, the intention tag having the highest degree of matching with the inferred intention data is regarded as the matched intention tag, and it is determined that the intention tag has the matched intention tag. 2. The language data processing system of claim 1, further configured to: The processing unit is further configured to perform at least one of the action procedures corresponding to the matched intent tag when it is determined that the matched intent tag exists in the intent tag. Item 1. The language data processing system according to item 1.

Images