KR20200072233A - Context-based Spoken language Understanding Method - Google Patents

Abstract

The present invention relates to a method for understanding a vocal language based on a context. The method comprises the steps of: classifying whether a context is changed by referring to a context change learning model when a result of voice recognition for utterance of a user is inputted; recognizing a vocabulary-based dialogue behavior of an input sentence by referring to a vocabulary-based dialogue behavior learning model when classified as a context information change; changing current context information to the recognized vocabulary-based dialogue behavior; recognizing a semantic object of a current input sentence based on the changed context information; determining a dialogue behavior of the input sentence by referring to a meaning-based dialogue action learning model; and recognizing slot information of the input sentence by referring to a slot tagging learning model.

Description

Context-based spoken language understanding method

The present invention relates to a context-based speech language understanding method, and more specifically, a context-based speech language capable of removing the ambiguity of the meaning and the ambiguity of a slot appearing in a speech language sentence in a speech processing system using speech language understanding technology. It's about how to understand.

In general, a language understanding technique is a technique for recognizing a speaker's speech language and grasping the speaker's intention in the recognized speech language.

These language understanding technologies are becoming more important as services such as intelligent assistants, interactive navigation systems, and artificial intelligence speakers have emerged.

Conventional language understanding techniques have used traditional rules or semantic parsing based methods, but recently machine learning based methods are mainly used.

However, in the conventional language understanding technology, there is a lot of ambiguity when a large object dictionary such as a person's name, a place name, and a title needs to be applied, and in the case of a machine learning base, there is a limitation in processing object and slot ambiguities.

Therefore, in order to deal with the above ambiguity, the use of context information is inevitable. The method mainly used in the past is to refer to the previous speech information.

However, since the previous speech information does not always provide the most information for ambiguity processing, there is a problem in that it is necessary to define which context to use and also a method for recognition.

As such, the conventional language understanding technology has a problem in that object recognition and object semantic ambiguity processing are difficult when a large object dictionary is applied.

The present invention has been devised to solve the conventional problems, and provides a context-based speech language understanding method capable of processing semantic ambiguity and slot ambiguity appearing in a speech language sentence using context.

The objects of the present invention are not limited to the above-mentioned objects, and other objects not mentioned will be clearly understood by those skilled in the art from the following description.

A context-based speech language understanding method according to an embodiment of the present invention for achieving the above object is to classify whether context is changed by referring to a context change learning model when a speech recognition result for a user's speech is input; Classifying the context information, recognizing the vocabulary-based conversational behavior of the input sentence with reference to the vocabulary-based conversational learning model; Changing context information to a recognized vocabulary-based conversational action; Recognizing a meaning object of the input sentence based on the changed context information; Determining a conversation behavior of the input sentence with reference to the semantic-based conversation behavior learning model; And recognizing slot information of the input sentence with reference to the slot tagging learning model.

Meanwhile, in the step of recognizing the slot information, slot tagging is performed by adding currently set context information and semantic information as qualities.

Here, the slot tagging is performed by successive tagging of vocabulary feature information, semantic feature information, and context feature information through a BIO method.

In addition, the conversation behavior information is recognized using morpheme vocabulary, part-of-speech and prototype information.

In the step of recognizing the conversation behavior, it is preferable to classify the conversation behavior using one or more of vocabulary feature information, semantic feature information, and context feature information.

A context-based speech language understanding apparatus according to an embodiment of the present invention includes a context change classification unit that classifies whether a context is changed by referring to a context change learning model when a voice recognition result for a user's speech is input; If classified as contextual information change, the dialogue act recognition unit recognizing the vocabulary-based dialogue act of the input sentence with reference to the vocabulary-based dialogue act learning model; A contextual information changing unit that changes the current contextual information with the recognized vocabulary-based dialogue action; A semantic object recognition unit recognizing a semantic object of the input sentence based on the changed context information; Determining a conversation behavior of the input sentence with reference to the semantic-based conversation behavior learning model; And a slot recognition unit for recognizing slot information of the input sentence with reference to the slot tagging learning model.

According to an embodiment of the present invention, it is possible to resolve the ambiguity of object recognition and meaning by recognizing the context information of the input sentence and applying the dictionary meaning of a large object according to the classification of the context change.

In addition, according to an embodiment of the present invention, there is an effect capable of improving speech language understanding performance by resolving the ambiguity of conversation and slots by using the recognized object semantic information and context information.

1 is a diagram illustrating a configuration block of a context-based speech language understanding apparatus according to an embodiment of the present invention.
2 is a flow chart for explaining a context-based speech language understanding method according to an embodiment of the present invention.
Figure 3 is a flow chart for explaining the detailed process of the step of recognizing the semantic object of the input sentence according to an embodiment of the present invention.
4 is a flowchart illustrating a method for generating a learning process for generating learning information of a context-based speech language understanding apparatus according to an embodiment of the present invention.

Advantages and features of the present invention, and methods for achieving them will be clarified with reference to embodiments described below in detail together with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, but will be implemented in various different forms, and only the embodiments allow the disclosure of the present invention to be complete, and the ordinary knowledge in the technical field to which the present invention pertains. It is provided to fully inform the holder of the scope of the invention, and the invention is only defined by the scope of the claims. Meanwhile, the terms used in the present specification are for explaining the embodiments and are not intended to limit the present invention. In the present specification, the singular form also includes the plural form unless otherwise specified in the phrase. As used herein, "comprises" and/or "comprising" refers to the components, steps, operations and/or elements mentioned above, the presence of one or more other components, steps, operations and/or elements. Or do not exclude additions.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. 1 is a view for explaining a configuration block of a context-based speech language understanding apparatus according to an embodiment of the present invention.

As illustrated in FIG. 1, a context-based speech language understanding apparatus according to an embodiment of the present invention includes a context change classification unit 100, a conversation behavior recognition unit 200, a context information changing unit 300, and a meaning object It includes a recognition unit 400, a conversation action determining unit 500 and a slot recognition unit 600.

The context change classification unit 100 classifies whether the context is changed by referring to the context change learning model when a voice recognition result for the user's speech is input. That is, the context change classification unit 100 first classifies whether to use the existing contextual information for the currently input input sentence or to change to the conversation action of the currently input input sentence. At this time, the context change classification unit 100 refers to a context change learning model database in which information such as vocabulary, part-of-speech, and prototype is stored.

Language understanding refers to the generation of intentional expressions that can be understood by machines such as computers from natural language sentences used by humans. Here, the intention expression may be generally composed of a conversation action and a slot. The conversation act refers to the user's original intention or purpose, and the slot corresponds to an item with information in the sentence. For convenience, express the intention in the form of “conversation (slot name = slot value)”.

For example, in the vehicle navigation system, “Let's go to KBS near Yeouido Station” can be expressed as “navigate(landmark=“Yeouido Station”,dest=“KBS”)”.

To this end, a learning corpus with tagged semantic expression is constructed to automatically extract or learn knowledge.

<landmark=Yeouido Station/poi> Let's go <dest= KBS/poi> #navigate(landmark=“Yeouido Station”, dest= “KBS”)

In the above, “landmark” and “dest” are slot names, and poi (point of interest) is a semantic code name. As above, even if it has the same poi meaning, it can be used in various slots such as “landmark, dest”.

In general, slots are a sequential tagging method, and conversation recognition can be treated as a classification problem.

However, when learning and applying on a sentence basis, unambiguity occurs.

For example, when uttering only “KBS”, it is ambiguous whether “KBS” is used as a poi or a channel such as DMB.

In this case, you must necessarily refer to the previous context, and ambiguity can be dealt with mainly by knowing what the previous utterance is.

Therefore, the context for understanding the language is defined as "the dialogue act having the most information to analyze the slot of the current input sentence among the previous dialogue acts including the present".

For example, if the system utters “What is the destination #request(dest)” and the user utters “KBS #navigate(dest=”KBS”)”, the context is “request(dest)”. In the above utterance, the expression after “# is not an actual utterance, but represents a conversational action for each utterance. For reference, a slot in which a value is not allocated among the slots in the user intention result is included in the conversation action. Therefore, in the case of inform(dest=”KBS”), the conversation action is “inform”, and in “request(dest)”, the conversation action is “request(dest)”.

As another example, if the system utters “Which channel to #request(dmb.channel)” and then the user utters “KBS #play(dmb.channel=”KBS”)”, the context is “ request(dmb.channel)”. However, in general, the previous dialogue act becomes the most powerful context information, but the previous dialogue act may become the context.

The following is a description of an example in which the previous dialogue act becomes the contextual information.

If the system utters “Which channel to play #request(dmb.channel)”, then the user utters “MBS # play(dmb.channel=”MBS”)” and then the system “sorry. , There is no MBS channel. Please repeat, if #repeat()” is spoken, and the user speaks “MBC # play(dmb.channel=”MBC”)” again, the previous conversation is not “repeat()”. "Request(dmb.channel)" is the context.

On the other hand, an example of a case in which the conversation topic is changed will be described.

First, when the system utters “What is the destination #request(dest)”, and then the user utters “KBS or turn #play(channel=”KBS”)”, the conversation action corresponding to the current input is slotted. It provides the most contextual information for recognition, so the context becomes “play”.

When the context change classification result for “KBS or turn on” uttered by the user in the above embodiment is classified as context change, the conversation behavior recognition unit 200 refers to the vocabulary-based conversation behavior learning model to base the vocabulary of the input sentence. Recognize the conversation behavior (S300). The conversation activity recognition unit 200 recognizes the conversation action of the current input sentence as "play" by referring to the vocabulary-based conversation behavior learning model.

Subsequently, the contextual information changing unit 300 changes the current contextual information to the recognized vocabulary-based dialogue action.

Then, the semantic object recognition unit 400 may recognize the semantic object of the current input sentence based on the changed context information. Here, the semantic object is an item having a meaning, and may be any name specified by a user dictionary or a rule by a user (Name Entity, NE). That is, proper nouns, common nouns, time, date, and amount are included. Whether these semantic objects are applied depends on the context. For example, if the context is “request(dest)”, it can be applied with objects such as poi (branch), city (city, province), gu (gu, gun), dong (dong, myeon, ri), etc. In the case of "play", objects such as channel (channel), title (music, program), etc. may be applied.

Thereafter, the dialogue action determining unit 500 determines the dialogue action of the input sentence by referring to the semantic-based conversation behavior learning model.

Meanwhile, the conversation action determining unit 500 determines the conversation action based on the currently set context information and the quality information including the previously recognized semantic object.

For example, if the input sentence is "KBS", the context information is @request(dmb.channel), and the semantic information is channel, the vocabulary feature may be generated as "L=KBS_NN L2=KBS_$ L2=$_KBS". And the semantic qualities are created as "S=channel S2=channel_$ S2=$_channel", and the context qualities are L=KBS_NN@request(dmb.channel) L2=KBS_$@request(dmb.channel) L2=$ _KBS@request(dmb.channel) S=channel@request(dmb.channel) S2=channel_$@request(dmb.channel) S2=$_channel@request(dmb.channel).

Here, the types of qualities are L: unigram vocabulary information and part-of-speech information, L2: bigram vocabulary information, S: unigram meaning information, S2: bigram meaning information, and $: information indicating the beginning or end of a sentence.

Here, the final feature is the sum of all vocabulary/meaning/context features, and the context feature can be seen by adding the feature that combines the context information to the existing feature. In this case, it is for recognizing dialogue act information by vocabulary and semantic qualities.

Based on the above qualities, it is possible to recognize conversation behavior information through learning and application. As you can see in the above example, there is ambiguity in the dialogue act information only with the vocabulary information for the utterance "KBS", but the dialogue act information is "play" through the semantic information "channel" or context information "request(dmb.channel)". Disambiguation is possible.

Accordingly, the slot recognition unit 600 recognizes the slot information of the input sentence by referring to the slot tagging learning model. Slot recognition is performed through sequential tagging of the BIO method.

For example, if the input sentence is "KBS", the context information is @request(dmb.channel), and the semantic information is channel, the vocabulary is "L-10=$_KBS L0=KBS S0=channel PL0=NN_KBS L01= KBS_$", semantic qualities are created as "S01=channel_$", context qualities are L-10=$_KBS@request(dmb.channel) L0=KBS@request(dmb.channel) S0= channel@request(dmb.channel) PL0=NN_KBS@request(dmb.channel) L01=KBS_$@request(dmb.channel) S01=channel_$@request(dmb.channel).

Here, P means part-of-speech qualities, and numbers after L, S, and P mean relative positions based on the present. That is, "L-10" means the bigram of the previous vocabulary and the current vocabulary.

Then, it is tagged as “KBS/B-dmb.channel”, and finally “KBS” is recognized as a slot called “dmb-channel”.

Accordingly, according to an embodiment of the present invention, the ambiguity of the slot is resolved by the semantic and contextual qualities, and thus there is an effect of clearly tagging the conversation.

2 is a flow chart for explaining a context-based speech language understanding method according to an embodiment of the present invention.

As illustrated in FIG. 2, the context-based speech language understanding method according to an embodiment of the present invention is preferably performed by the detailed configuration of the speech language understanding apparatus.

When the voice recognition result for the user's speech is input, the context change classification unit 100 classifies whether the context is changed by referring to the context change learning model (S100). That is, the context change classification unit 100 first classifies whether to use the existing contextual information for the currently input input sentence or to change to the conversation action of the currently input input sentence. At this time, the context change classification unit 100 refers to a context change learning model database in which information such as vocabulary, part-of-speech, and prototype is stored.

Thereafter, it is determined whether the user's input sentence is classified as a context change (S200). If the user's input sentence is classified as a context change in the determination step (S200) (YES), the conversation behavior recognition unit 200 recognizes the vocabulary-based conversation behavior of the input sentence by referring to the vocabulary-based conversation behavior learning model (S300). ). For example, when looking at "Play KBS", context information is classified as change (change=1), and the conversation behavior recognition unit 200 refers to the vocabulary-based conversation behavior learning model to display the conversation behavior of the current input sentence. Recognize as "play".

Thereafter, the contextual information changing unit 300 currently changes to the recognized vocabulary-based dialogue action (S400). In the present embodiment, "play", which is conversation action information, becomes context information.

Subsequently, the semantic object recognition unit 400 recognizes the semantic object of the current input sentence based on the changed context information (S500).

Here, the semantic object is an item having a meaning, and may be any name specified by a user dictionary or a rule by a user (Name Entity, NE). That is, proper nouns, common nouns, time, date, and amount are included. Whether these semantic objects are applied depends on the context. For example, when the context is request(dest), objects such as poi (branch), city (city, province), gu (gu, county), dong (dong, myeon, ri), etc. can be applied, and the context is In the case of "play", objects such as channel (channel), title (music, program), etc. may be applied.

Thereafter, the conversation action determining unit 500 determines the conversation action of the input sentence by referring to the semantic-based conversation action learning model (S600).

Meanwhile, the conversation action determining unit 500 determines the conversation action based on the currently set context information and the context information including the previously recognized semantic object.

To this end, the conversation action determining unit 500 generates vocabulary quality information, semantic quality information, and contextual quality information.

For example, if the input sentence is "KBS", the context information is @request(dmb.channel), and the semantic information is channel, the vocabulary feature may be generated as "L=KBS_NN L2=KBS_$ L2=$_KBS". And the semantic qualities are created as "S=channel S2=channel_$ S2=$_channel", and the context qualities are L=KBS_NN@request(dmb.channel) L2=KBS_$@request(dmb.channel) L2=$ _KBS@request(dmb.channel) S=channel@request(dmb.channel) S2=channel_$@request(dmb.channel) S2=$_channel@request(dmb.channel).

Here, the types of qualities are L: unigram vocabulary information and part-of-speech information, L2: bigram vocabulary information, S: unigram meaning information, S2: bigram meaning information, and $: information indicating the beginning or end of a sentence.

The final qualities are the sum of all vocabulary/meaning/context qualities, and the context qualities can be seen by adding the qualities combined with the context information to the existing qualities. Is to recognize.

Based on the above qualities, conversation can be recognized through learning and application. As you can see in the above example, there is ambiguity in the dialogue act information only with the vocabulary information for the utterance "KBS", but the dialogue act information is "play" through the semantic information "channel" or context information "request(dmb.channel)". Disambiguation is possible.

Subsequently, the slot recognition unit 600 recognizes the slot information of the input sentence by referring to the slot tagging learning model (S700).

For example, if the utterance is "KBS", the semantic information and context information can be recognized as KBS/channel @request(dmb.channel).

At this time, the slot recognition unit 600 tags in the form of "KBS/B-dmb.channel" in a BIO manner.

Here, the vocabulary feature is generated as "L-10=$_KBS L0=KBS S0=channel PL0=NN_KBS L01=KBS_$", the semantic feature is created as "S01=channel_$", and the context feature is L- 10=$_KBS@request(dmb.channel) L0=KBS@request(dmb.channel) S0=channel@request(dmb.channel) PL0=NN_KBS@request(dmb.channel) L01=KBS_$@request(dmb. channel) S01=channel_$@request(dmb.channel).

Here, P means part-of-speech qualities, and the number after L,S,P means the position relative to the current. That is, "L-10" means the bigram of the previous vocabulary and the current vocabulary.

On the other hand, if the user's speech in the determination step (S200) is not classified as a context change (NO), the semantic object recognition unit 400 recognizes the semantic object of the current input sentence based on the unchanged context information (S500). Performs from the process.

Accordingly, according to an embodiment of the present invention, the ambiguity of the slot is resolved by semantic and contextual qualities, and thus there is an effect of clearly recognizing a conversational action.

3 is a flow chart for explaining the detailed process of the step of recognizing the meaning object of the input sentence according to an embodiment of the present invention.

As illustrated in FIG. 3, the semantic object recognition unit 400 applies a semantic dictionary based on the changed context information (S510 ). The semantic object recognition unit 400 refers to a dictionary of semantic objects according to context information. Here, the result of the morpheme analysis may be different depending on the influence of the dictionary of semantic objects.

For example, while an embodiment of the present invention is applied to a document retrieval system, the following conversation may be made between the system and the user.

If the system utters “Which sort do you want to sort the document on?”, then the user utters “Lookup order,” and if “Lookup order” is set to “Lookup order/person” in the semantic object dictionary, “Sort the document by which criteria. Shall we?" is recognized as "#request(sorting_criteria)", and the context of the user's utterance "Lookup Soon" is recognized as "@request(sorting_criteria)".

Then, even if the semantic code of "lookup order" is set to "person" in the semantic object dictionary, "lookup order" is not recognized as "person" because person is not used as a semantic object in this context.

Again, when the system utters “Who is the author?” and when the user utters “Lookup order,” the conversation of the system's utterance “Who is the author?” is recognized as “request(author)”, and the user's The context of the utterance "lookup order" is recognized as "request(author)".

In this context, "lookup order" is recognized as "person" because person is used as a semantic object.

Subsequently, the semantic object recognition unit 400 recognizes the semantic object based on the rule (S520 ). To this end, in the semantic object recognition unit 400, the semantic object, time, date, and amount expressions previously recognized are added with corresponding semantic codes. Here, the rule base recognizes patterns, such as time, date, and amount, using rules such as regular expression.

Thereafter, the semantic object recognition unit 400 tags the semantic object on a learning basis (S530). That is, the semantic object recognition unit 400 adds the morpheme information and context recognized in the previous step to perform semantic object tagging on a learning basis.

For example, if the learning corpus has a sentence such as "<author=lookup order/person>", the context-based semantic object tagging learning unit 803 is "#inform(author="lookup order")@request(author) "Is used as information for tagging "person" corresponding to the semantic tag of the meaning object, and generates BIO information for sequential tagging as follows.

"Lookup Soon / B-person is / O is / O"

Then, the semantic object recognition unit 400 may perform semantic object tagging through feature generation and tagging for BIO tagging.

Accordingly, according to an embodiment of the present invention, through the tagging process as described above, there is an effect of preventing over-applicable that may occur when a semantic object that is not in the dictionary is recognized or applied based on a simple dictionary.

4 is a flowchart illustrating a method for generating a learning process for generating learning information of a context-based speech language understanding apparatus according to an embodiment of the present invention.

As illustrated in FIG. 4, a case where the learning sentence is "<author=query order/person>. #inform(author="query order")@request(author)" will be described as an embodiment.

The meaning of the above learning sentence means that the user's speech is “inquiry order”, where the meaning code of “inquiry order” is “person”, the slot name is “author”, the conversation action is “inform”, and the context is “request(author). ”

First, when a learning sentence such as "<author=query order/person>. #inform(author="query order")@request(author)" is input, the context change learning unit 801 is required for learning whether the context is changed. A learning sentence such as "query/NN net/FS is/O is/O -> change=0" is generated (S801).

Here, the method for creating whether or not to change the context is changed if the conversation action of the current input sentence is the same as the conversation action of the context (change=1), and if it is different, it is not changed (change=0).

For example, in the case of “KBS -> #navigate(dest="KBS")@request(dest)", classify as "change=0", and turn on "KBS == #play(dmb.channel="KBS In case of ")@play", it is classified as "change=1".

Then, the vocabulary-based conversation behavior learning unit 802 generates a learning sentence such as "query/NN order/FS is/O is/O -> inform" required for vocabulary-based conversation behavior learning (S802).

Then, the context-based semantic object tagging learning unit 803 generates a learning sentence such as “lookup order/B-person is/O is/O @request(author)” required for semantic object tagging (S803).

Subsequently, the context-based slot tagging learning unit 805 generates a “query order/B-person is/O is/O @request(author)” required for context-based slot tagging learning as a learning sentence (S805).

Subsequently, the semantic tag extracting unit for each context 806 adds the “person” semantic tag to the context semantic tag list in the context “@request(author)” (S806).

The configuration of the present invention has been described in detail with reference to the accompanying drawings, but this is only an example, and those skilled in the art to which the present invention pertains have various modifications and changes within the scope of the technical spirit of the present invention. Of course this is possible. Therefore, the protection scope of the present invention should not be limited to the above-described embodiments and should be defined by the following claims.

100: context change classification unit 200: conversation behavior recognition unit
300: Context information change unit 400: Meaning object recognition unit
500: dialogue action determining unit 600: slot recognition unit

Claims (1)

Classifying whether the context is changed by referring to the context change learning model when the voice recognition result for the user's speech is input;
Classifying the context information, recognizing the vocabulary-based conversational behavior of the input sentence with reference to the vocabulary-based conversational learning model;
Changing the current context information to the recognized vocabulary-based conversational action;
Recognizing a semantic object of the current input sentence based on the changed context information;
Determining a conversation behavior of the input sentence with reference to the semantic-based conversation behavior learning model; And
Recognizing the slot information of the input sentence with reference to the slot tagging learning model; context-based speech language understanding method comprising a.

Images