TW202238435A - Natural language dialogue intention analysis method - Google Patents

Abstract

A natural language dialogue intention analysis method, the "keywords" interactive "interface" designed with "language" as the "direction", and the NLU engine conforming to the Chinese structure design. It is worth noting that the present invention is anchored in the [verb] establishment [The style model between entity] and [verb], and can extract Chinese [vocabulary]; set the verb position in the sentence pattern as an anchor point in the plural sentence structure, and establish the intention model between the entity vocabulary and the anchor point; The anchor point can be the reference point for the analysis of the natural language dialogue intention of the present invention; the system can output the analyzed sentence intention as a program code, and cooperate with the developer to establish their own business or application logic block calculations, and to compare their own business or The calculation result of the application is returned to the user, or other applications are performed.

Description

Natural language dialogue intent analysis method

The invention relates to the technical field of natural language processing, in particular to a semantic computing intent analysis method of deep structure of natural language.

Natural language processing (NLP) and natural language understanding (NLU) are important "human-computer interaction interface" technologies in the era of artificial intelligence. In the latest wave of AI boom, in addition to image recognition processing, the most important landing technology is NLP/NLU; however, since NLP/NLU is processing "natural language", it will of course be limited by "language data characteristics".

In the prior art, there is a Clickbot click-type chatbot that compares scripts and has very strict restrictions on scripts. This method requires a lot of script logic to distinguish intent associations, and more often simply lists all script options. In addition, the problem with the Keywordbot chatbot is that it requires a large number of keywords (including abbreviations). Users often get irrelevant replies because Keywordbot cannot match the keywords. Moreover, the longer the project life cycle, the more difficult it is to maintain. The NLU engine of the ML method also requires a large number of sentences to make character training models. It is not easy to adjust the model errors when using it, and due to the huge amount of calculation, it needs a large server to operate properly and is limited to online use.

In the field of many intelligent question answering systems that integrate existing technologies, most of them use large numbers Data, deep learning, AI to achieve natural language processing. Its disadvantage is that it is too rough and simple, which leads to poor accuracy of semantic analysis results, and requires a huge device capacity to store the database, which is difficult to download to the mobile phone for internal use.

In view of this, the present invention designs a model that does not require a large number of sentences and relies on syntactic structures. It is easy to trace and adjust model errors, and has the characteristics of extremely low computing requirements. It can also be installed in enterprises or on personal mobile phones. The invention can greatly reduce the capacity requirement of hardware equipment, only about 32MB of storage space is required, which is far superior to the storage space requirement of more than GB in the prior art; most of the semantic understanding systems on the market need to set up a third-party server, and the capacity requirement of the device of the present invention It is small so no third-party server is needed, and it can even be downloaded to the mobile phone in the form of an APP component to perform initial natural language conversations; since no third-party server is needed, the network requirement of this system is not a necessary condition.

The present invention is a method for analyzing the intent of a natural language dialogue. The "key word" interactive "interface" designed with "language" as the "guidance" and the NLU engine conforms to the Chinese structure design. It is worth noting that the present invention is anchored in [verbs] Create a pattern model between [entity] and [verb] for intent analysis/classification. Compared with other machine learning-based solutions, it also has the ability to additionally extract Chinese [vocabularies] for subsequent applications; its design aims to compare sentences and select vocabulary required for semantic calculation as parameters , so as to retain the language expression relationship of "sentence pattern-semantic" and also extract key words for the calculation interface.

The present invention divides the dialogue into three layers, which are: project name (Project): the scene where a certain group of intentions is applicable, such as the scene in a convenience store; intention name (Intent): a certain intention, For example, in the scene in the convenience store, there is the intention of payment and ticket purchase, but not the intention of accommodation; language expression (Utterance): a group of language expressions that can be used to express a certain intention in a certain scene, can be complete sentences or incomplete sentences.

A natural language dialogue intention analysis method of the present invention, wherein the steps of the method for establishing a plurality of syntactic structure models are as follows: S1. The manager first establishes an application scene as a context (context); S2. Establishes a plurality of intentions (intent ); S3. Further, under each intention, a plurality of example sentences that meet the intention in this application scenario are established in a sentence-by-example manner; S4. The sentence analysis module of this system performs lexical segmentation and part-of-speech tagging on these example sentences, Translate the example sentences marked with part of speech into plural sentence structures. Further, the sentence analysis module provides a physical vocabulary check screen according to the vocabulary and part of speech of these example sentences for the manager to check and match with the The entity vocabulary associated with the intent; S5. An intent analysis module binds the sentence structures in the application scene to their respective intents, encodes them into syntactic structure models, and stores them in one of the scene databases Application scenario model unit; S6. The system outputs a conversion module, which extracts the syntactic structure models of the application scenario model unit, and converts them into a program code output available for download.

The natural dialogue intention analysis method of the present invention, wherein the intention analysis method is as follows: S7. A user connects to a dialogue page of the application scene through a terminal device, and the user inputs a dialogue sentence on the dialogue page; S8. The sentence The analysis module extracts the dialogue sentence of the dialogue page, performs lexical segmentation and part-of-speech tagging on the dialogue sentence, sets the verb position as the anchor point, and translates the sentence marked with the part-of-speech into the sentence structure; S9. The intent analysis module links the application scenario model unit of the scenario database according to the application scenario, compares the sentence structure of the dialogue sentence with the syntactic structure models in the application scenario model unit, and further uses the anchor Sentence structure comparison is carried out at the point as the reference point; S10. If there is a group of model comparison results among the syntactic structure models that are consistent, then define that the dialogue sentence contains the intent bound by the group of models as a schematic diagram.

The natural language dialogue intention analysis method of the present invention can establish "mathematical application problems" as the project name (Project name), indicating that this group of intentions will be applicable to the language scene (Discourse) of mathematical application problems; and then establish the "addition and subtraction" method Intent name. Under this intention, all the sentences are to describe the language expression (Utterance) of "addition and subtraction". For example, "Dad ate two apples", "My sister broke three" or "My brother gave him two more pens"... etc.; use the sentence analysis module and intent analysis module to perform word segmentation, part-of-speech tagging and named entity recognition After processing, all sentences will be transformed into regular expressions that only retain tags for sentence pattern recognition; taking "Dad ate two apples" as an example, the complete model generation process is as follows:

In the above, the intent analysis module transcribes the training sentence "Dad ate two apples" into the following regular expression, and at the same time transcribes the verb into [eat] marked with square brackets, which can be extended and compatible Any verb that ends with "drop", or verbs and verb groups that contain "eat" expresses the intention of "reduce", and at the same time excludes the reverse semantics caused by the negative expression "not"; this design, You can use very little data to express (Utterance) sentences through reserved language Type of way to distinguish semantic intent, to obtain maximum compatibility.

Compared with the previous technology based on machine learning, such as LUIS, the natural language dialogue intention analysis method of the present invention, we also established a 18-group example sentence related to the exchange intention in this system and LUIS, as follows:

1. What was the exchange rate between US dollars and Taiwan dollars last Wednesday;

2. What is the exchange rate between US dollars and Taiwan dollars today;

3. I want USD 100;

4. I want 100 USD;

5. I want to buy USD 100;

6. I want to buy 100 USD;

7. How much is US$100 in Taiwan dollars;

8. How much is US$100 in Taiwan dollars;

9. How much is 100 US dollars in Taiwan dollars;

10. How much is 100 US dollars in Taiwan dollars;

11. How much is 100 US dollars in Taiwan dollars;

12. How much is 100 US dollars in Taiwan dollars;

13. How much is Taiwan dollars for 00 US dollars;

14. How much is 100 US dollars in Taiwan dollars;

15. How many Taiwan dollars can be exchanged for 100 US dollars;

16. How much can 100 US dollars be converted into Taiwan dollars;

17. How many Taiwan dollars can be exchanged for 100 US dollars;

18. How much can 100 US dollars be converted into Taiwan dollars;

In the above, the natural language dialogue intent analysis method of the present invention is compared with the prior art LUIS, and model training is performed after inputting 18 sets of example sentences; then we input related questions different from the example sentences to perform the LUIS test, and after the test, we can obtain and exchange currency Intention-related similarity scores, as shown in [Figure 4], we know that the similarity of the question sentence "I want US$100" that is the same as the third example sentence above is 0.985; the same intention "I want to exchange US$200 The similarity of “元” dropped to only 0.677. In the same sentence pattern, the same intention but only the amount of money is different, the similarity is so different. Therefore, using the same technology as LUIS requires a large number of example sentences to train the model; let’s look at Another question is "Is NT$15 for three masks and NT$7 for shipping expensive?" This example sentence has nothing to do with the intention of currency exchange at all, but its similarity is as high as 0.978; from the above we can know why the existing Language understanding technologies often lead to off-topic answers because they use mathematical principles to build machine learning models.

As mentioned above, the natural language dialogue intention analysis method of the present invention is compared with the prior art LUIS, and the above-mentioned 18 groups of example sentences are also input to train the model. The intent analysis of 元』 is as follows, we can find the same sentence structure in the example sentence model.

Based on the above, the analysis method of the natural language dialogue intention of the present invention is compared with the previous technology LUIS. The intention analysis of inputting "I want to buy US$200" in this system is as follows. We can also find the same sentence pattern in the example sentence model Structure, the system will not fail to compare the example sentence model because of the difference in the amount; this system uses the sentence structure comparison method, not the machine learning model based on mathematical principles, so as long as the sentence structure is consistent, the same sentence structure can be found example sentence model of .

Based on the above, compared with the prior art LUIS, the natural language dialogue intention analysis method of the present invention, this system adopts the sentence structure comparison method, and the intention analysis of inputting "I want to exchange US$20,000" in this system is as follows, We can also find the same sentence structure in the example sentence model.

Based on the above, compared with the previous technology LUIS, the natural language dialogue intention analysis method of the present invention, this system adopts the sentence structure comparison method, and enters "three masks, 15 NT dollars plus 7 NT dollars, is the shipping cost very expensive?" 』The intention analysis is as follows, the system cannot find the same sentence structure in the example sentence model.

As can be seen from the above comparison, the present invention has a very high language intent recognition ability, does not require a large amount of corpus for training, and does not require deep learning. It only needs to achieve intent analysis under the minimum limit of one example sentence for each sentence structure. These are all the different characteristics of the present invention and prior art.

The natural language dialogue intent analysis method of the present invention can also find out the focal person of a specific event for information extraction. , then the person must be suspected because of "doing something"; as shown in [Figure 8], we first process the news text of the training data, use the sentence analysis module of this system to mark the verbs in the text, and Save the collection of "verbs" into the data set; then, take the collection of verbs from [AML News] and delete the collection of verbs from [Non-AML News]; the result is the AML news classifier used by this system.

In the above, if a character is involved in the crime of "anti-money laundering (AML)", the pattern of the reporter's sentences describing this character is bound to be different from other innocent people; then, we use the sentence analysis module to [ All [sentences containing people's names] in AML news] are taken out, and These sentences are imported into the intent analysis module as a training model; the sentence analysis module is to learn [sentence patterns] rather than the distribution probability between characters; the sentence pattern structure is limited, but the combination is infinite, so such as [Zhang San was encountered Sentenced to 10 years and 6 months] and [Li Si was sentenced to 6 months] and [Li Si was sentenced to 6 months], the two sentences have few characters the same, but because the sentence patterns are the same, they are both "person + suffer + verb + time" structure, so the intent analysis module of this system will treat it as the same data, and only need to input it once. In this way, the demand for training data (model) will be greatly reduced, which is different from the existing ones that require a large amount of data to build Semantic analysis techniques for machine learning models.

It can be seen from the above that this system uses the sentence structure comparison method to overcome the previous technical problems and strengthen the recognition of sentence intent; further, developers who provide semantic analysis programs will be able to use the analysis results of the dialogue text by this system to their own Business or application logic, fill in the intent code into the corresponding block of its own code, and the development of the dialogue robot has been completed; further, the developer can install the program into the user's handheld device or its own server, There is no need to rent expensive network bandwidth, so that the chatbot can serve the most customers at the same time.

Preferably, since this system does not require a large number of training models, it can be deployed in the form of APP components on "the user's own mobile phone", so that most of the questions can be calculated and answered by the local APP, only The real business needs will be sent to the server in the cloud through the network in the form of instructions.

100: input module

200:Database module

300: Sentence analysis module

400: Intent analysis module

500: Output conversion module

S1~S8: Steps in the process of building a syntactic structure model

S100~S103: Process steps of intent analysis method

A1～A5: Embodiment 1 model building operation steps

B1~B8: Process Steps of Embodiment 2 Exceptional Circumstances

C3~C8: Embodiment 2 Process Steps for Innocent Persons

D3~D8: Steps in the suspected money laundering process of Embodiment 2

E1~E7: Steps in the application process of the model in Embodiment 2

[Figure 1] Schematic diagram of the natural language dialogue intent analysis system

[Figure 2] Schematic diagram of the process of building a syntactic structure model

[Figure 3] Schematic diagram of the flow chart of the intent analysis method

[Figure 4] Schematic diagram of the similarity score of the LUIS intent analysis method

[Figure 5] Schematic flow chart of the model building operation steps in the first embodiment

[Fig. 6] Example sentence input screen of embodiment one

[Figure 7] Embodiment 1 Entity Vocabulary Check Screen

[Figure 8] Schematic diagram of the analysis process of the focus person of a specific event

[Figure 9] Schematic flow chart of exceptions in the second embodiment

[Figure 10] Schematic diagram of the flow of the innocent person in the second embodiment

[Figure 11] Schematic diagram of the suspected money laundering process of Embodiment 2

[Figure 12] Schematic diagram of the application process of the model of the second embodiment

[Figure 13] Schematic diagram of the function event pool of the third embodiment

[Fig. 14a] ~ [Fig. 14d] Schematic diagram of the third embodiment of the Math project

In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in combination with specific embodiments and with reference to the accompanying drawings. It should be understood that these descriptions are exemplary only, and are not intended to limit the scope of the present invention.

A schematic diagram of a natural language dialogue intent analysis system of the present invention, as shown in [Fig. 1], includes: an input module 100, which allows a manager to create data in the system; a database module 200, connected to the input module Group 100 includes a verb database and a scene database, the verb database stores a plurality of verbs, the scene database stores a plurality of scenes; a sentence analysis module 300 connects the input module 100 and the database module Group 200, the sentence analysis module 300 can carry out lexical segmentation and mark the part of speech of a sentence or a text; an intention analysis module 400, connect the resource The database module 200 and the sentence analysis module 300, the intent analysis module 400 can summarize the sentence or the text into the sentence or the text containing an intent; an output conversion module 500, connected to the analysis module 400 and the database module 200, the output conversion module 500 converts the intent binding the statement or the text into a program code output.

The above-mentioned input module 100 further includes: a custom dictionary unit, which can be imported by the manager into a plurality of example sentences that meet an application scenario; an example sentence creation unit, where the manager can create these example sentences that meet the application scenario and Stored as a syntactic structure model.

A natural language dialogue intention analysis method of the present invention, wherein a plurality of syntactic structure models are established as shown in [Fig. 2], the method is as follows:

S1. The example sentence creation unit of the input module 100 of the system provides a manager to first create an mth application scenario;

S2. Further, the manager creates an nth intent data set belonging to the mth application scenario in the mth application scenario;

S3. Further, the manager creates a plurality of example sentences that meet the nth intent of the mth application scenario in a sentence-by-example manner under the nth intent, and sends them to a sentence analysis module 300;

S4. The sentence analysis module 300 performs lexical segmentation and part-of-speech tagging on these example sentences, and further, translates these example sentences marked with part-of-speech into plural sentence structure;

S5. An intent analysis module 400 binds the sentence structures in the mth application scenario to the nth intent, encodes these syntactic structure models and stores them in one of the mth application scenario models of the scenario database unit;

S6. Repeat steps S2~S5 until all intents belonging to the mth application scenario have been established;

S7. Repeat steps S1~S6 until all application scenarios are established;

S8. The system outputs the conversion module 500, which extracts the syntactic structure models of the m-th application scene model unit, and converts them one by one into a program code (sentence structure representation) available for download for output.

A preferred embodiment of the present invention is shown in [FIG. 7], the above step S4 further includes S4', the sentence analysis module 300 provides a physical vocabulary check screen for the manager according to the vocabulary and part of speech of these example sentences Check the entity vocabulary associated with the nth intent.

In the above-mentioned embodiments, the sentence structure sets the position of the verb in the sentence as an anchor point, and establishes the intention models between the entity vocabulary and the anchor point; it is worth noting that the anchor point is the anchor point A comparison benchmark for natural language conversational intent analysis.

The natural dialogue intent analysis method of the present invention, wherein the flow of the intent analysis method is shown in [Figure 3], and the steps are as follows:

S100. A user connects to a dialog page of the m-th application scene through a terminal device, the user inputs a dialog sentence with the dialog page, and the dialog page sends the dialog sentence to the sentence analysis module 300;

S101. The sentence analysis module 300 extracts the dialogue sentence on the dialogue page, performs lexical segmentation, part-of-speech tagging on the dialogue sentence, sets the verb position as the anchor point, and translates the sentence marked with the part-of-speech into the sentence pattern. structure, and further, transmit the sentence structure and the mth application scene code to the intent analysis module 400;

S102. The intent analysis module 400 links the mth application scenario model unit of the scenario database according to the mth application scenario, and further compares the sentence structure of the dialog sentence with the mth application scenario model unit These syntactic structure models are compared with the anchor point Sentence structure comparison;

S103. If there is a group of model comparison results among the syntactic structure models that are consistent, define the nth intent that the dialog sentence contains the group of model bindings as an expression graph.

In the above step S103, if the comparison results of the syntactic structure models are inconsistent, the dialogue sentence is stored in a data set without corresponding intent.

The above step S101 further includes that the sentence analysis module 300 translates the dialogue sentence into the sentence structure according to the vocabulary, the part of speech and the entity vocabulary.

In an embodiment of the present invention, the natural dialogue intent analysis method, the dialogue page and the input module 100 can also input a text, and the sentence analysis module 300 performs sentence segmentation, vocabulary segmentation, and part-of-speech tagging on the text.

In one embodiment of the present invention, the natural dialogue intent analysis method, the program code provides program developers to fill in their own business or application logic into the corresponding block of the program code, and the development of the dialogue robot has been completed; further, the developer can install the program to the user's handheld device, or to its own server.

According to an embodiment of the present invention, the natural dialogue intent analysis method, the diagram is sent to the logic block of the business or application program for calculation according to the structure selected by the user, and the developer sends back the calculation result of the business or application program users, or other applications.

Example 1, as shown in [Figure 5], an example of a model building operation with a foreign exchange intention:

A1. An example sentence establishment unit of this system receives a manager to establish a financial business scenario related to financial business here, and then establishes a currency exchange intention related to different currency exchange issues under the financial business scenario;

A2. Further, the system provides a plurality of example sentence input screens for the manager to input, and establishes one sentence structure and one example, and establishes possible question sentence patterns when the client inputs currency exchange questions, as shown in [Figure 6] shown;

A3. After these example sentences are input, execute a whole sentence analysis function, and notify a sentence analysis module 300 to segment these example sentences into vocabulary and mark the part of speech, and further, provide an entity vocabulary according to the principle of vocabulary and part of speech Check the screen for the manager to check the parameters required for intent analysis, as shown in [Figure 7];

A4. The manager checks a plurality of entity words related to the exchange intention, as shown in [Figure 7], the entity words related to the exchange intention are currency and amount;

A5. After the manager checks the required parameters of the intention and then executes a generation model, an intention analysis module 400 generates a plurality of syntactic structure models with the exchange intention. After these syntactic structure models are generated, they can be Perform the analysis of the natural language dialogue intent.

Continuing from the first example above, input a dialogue sentence "I want to buy 100 euros", even if the currency type is different from the example sentence in the model content, after the analysis of the natural language dialogue intention of this system, the same sentence structure can be found in the example sentence model (conform to [Fig. 6] the 16th example sentence syntactic structure model).

Continuing from the first example above, input the dialogue sentence "I want to buy RMB 200 Yuan", even if the amount number is different from the example sentence of the model content, the same sentence structure can be found in the example sentence model after the analysis of the natural language dialogue intention of the system (consistent with the syntactic structure model of the 16th example sentence in [Figure 6]).

Continuing from the first example above, input the dialogue sentence "Is it expensive to have three masks, NT$15 plus shipping fee of NT$7?", after the analysis of the natural language dialogue intention of this system, the same sentence structure cannot be found in the example sentence model , the system judges that the dialogue sentence has nothing to do with the currency exchange intention, thus presenting a natural language understanding effect different from the current machine learning model based on mathematical principles.

Preferably, the system fixedly uses "named entities (such as place names, currency amounts, addresses, etc.)", "time words", "quantifiers", "classifiers", "measurement words", location words, "quantitative verbs" " and "Sentence mark (direct sentence/interrogative sentence)" are automatically matched with anchored verbs to form a sentence model. Only "adjectives", "adverbs" and "nouns (including user-defined words)" can be built into the model at the developer's discretion, so there is a check box.

Embodiment 2: As shown in [Fig. 9] to [Fig. 12], a model-building operation embodiment of a suspect (focus person) involved in "anti-money laundering (AML)":

B1. The example sentence creation unit of this system receives the manager to create an AML project scenario;

B2. Establish the three intents belonging to the AML project under the AML project;

B3. Establish an exception of intent 01, as shown in [Figure 9];

B4. The input module 100 inputs plural texts of the exception, and the system performs lexical segmentation, part-of-speech tagging and setting the verb position as the anchor point for these texts through the sentence analysis module 300, and the tagged The sentence of the part of speech is translated into these sentence structures;

B4'. Provide an entity vocabulary check screen for the manager to check the name of the person with the exception as a parameter (non-essential step);

B5. The intent analysis module 400 binds the sentence structure and the AML project scenario to the exception intent, encodes the syntactic structure models, and stores them in one of the AML project model units of the scenario database;

B6. The output conversion module 500 generates the program code available for download according to the model content of the AML project model unit;

B7. The program developer downloads the program code, and stores the name of the person marked by the sentence analysis module 300 into an exception data set of the database module 200;

B8. The program developer writes the program into the server.

Continuing from the second embodiment above, step B2. Establish one of the three intentions and proceed as follows:

C3. Establish intent 02-innocent person, as shown in [Figure 10];

C4. The input module 100 inputs a plurality of texts of the innocent person, and the system takes these texts Carry out vocabulary segmentation, part-of-speech tagging and set the verb position as the anchor point through the sentence analysis module 300, and translate the sentence marked with part-of-speech into these sentence structures;

C4’. Provide the entity vocabulary check screen for the manager to check the name of the innocent person as a parameter (non-essential step);

C5. The intent analysis module 400 encodes the sentence structure and the intent of the innocent person bound to the AML project into these syntactic structure models, and stores them in the AML project model unit of the scene database;

C6. The output conversion module 500 generates the program code available for download according to the model content of the AML project model unit;

C7. The program developer downloads the program code, and stores the names marked by the sentence analysis module 300 into one of the innocent person data sets of the database module 200;

C8. The program developer writes the program into the server.

Continuing from the second embodiment above, step B2. Establish one of the three intentions and proceed as follows:

D3. Establish intention 03-money laundering suspicion, as shown in [Figure 11];

D4. The input module 100 inputs multiple texts of the suspected money laundering, and the system performs word segmentation and part-of-speech tagging on these texts through the sentence analysis module 300, and sets the position of the verb as the anchor point, and the tagged part-of-speech The sentence is translated into these sentence structures;

D4'. Provide the entity vocabulary check screen for the manager to check the name of the person suspected of money laundering as a parameter (non-essential step);

D5. The intent analysis module 400 binds the sentence structure and the AML project to the suspected intention of money laundering, encodes these syntactic structure models, and stores them in the AML of the scene database project model unit;

D6. The output conversion module 500 generates the program code available for download according to the model content of the AML project model unit;

D7. The program developer downloads the program code, and stores the names marked by the sentence analysis module 300 into one of the money laundering suspect data sets of the database module 200;

D8. The program developer writes the program into the server.

Continuing from the second embodiment above, a model application procedure involving a crime involving "anti-money laundering (AML)" is shown in [Figure 12]:

E1. A user connects to the AML project scene through a terminal device, and inputs a plurality of news contents in the input module 100, and sends them to the sentence analysis module 300;

E2. The sentence analysis module 300 performs lexical segmentation, part-of-speech tagging and setting the position of the verb as the anchor point for these news contents, and translates the sentence marked with part-of-speech into these sentence structures, and further, The sentence structure and the AML project code are sent to the intent analysis module 400;

E3. The intent analysis module 400 links the AML project model unit of the scene database according to the AML project, and further compares the sentence structures of the news content sentences with the AML project model unit These syntactic structure models are compared with the anchor point as the reference point for sentence structure comparison. If the comparison results of a group of models among these syntactic structure models are consistent, it is defined that the dialogue sentence contains a binding of this group of models. show intention;

E4. The table is intended to cooperate with the user to check the structure and send it to the application logic block for calculation;

E5. The sentence "1. Wang ○○ embezzled two million public funds..." in the news text conforms to the sentence structure in the intention of the money laundering suspicion. According to the checked result, the name [Wang ○○] is output and stored in It's time to wash Money suspect data collection;

E6. The sentence "2. Prosecutor Li ○○ points out that the plaintiff..." in the news text conforms to the sentence structure in the intention of the exception. According to the result of the check, the name [Li ○○] is output and stored in the exception data concentrated;

E7. The sentence in the news text "3. Lin ○○ was sentenced not guilty in the second instance..." conforms to the sentence structure in the intention of the innocent person. According to the result of the check, the name [Lin ○○] is output and stored in the innocent person's data set .

In the above-mentioned embodiment 2, ticking the person's name in the entity vocabulary means to the system that the person's name is a variable parameter, that is to say, the system will store [Xie ○○] embezzling $1.2 million as: [a certain Sentence pattern: [someone] embezzled $1.2 million; Sentence pattern: [someone] misappropriated $1.2 million, which will be further translated into a "sentence structure" as follows: [someone]-Verb (misappropriation)-Currency Come on, as long as you see the sentence pattern: [so-and-so]-Verb (misappropriation)-a certain amount of money next time, you will know that [so-and-so] is a [money laundering suspect].

Embodiment 3: The embodiment flow of an "addition event" and "subtraction event" is as follows:

F1. Establish a Math project scene, further establish an addition intention data set and a subtraction intention data set, and then establish these syntactic structure models according to steps S3~S8;

F2. A function event pool (Event Pool) that defines two events of "addition" and "subtraction" is shown in [Figure 13]. The comparison rule of "formula"; the text of the calculation intention, after the comparison with the event, is the mathematical operation that should be performed to generate the calculation intention text; for example: "My brother has X oranges, and give him Y oranges" X+Y And when "the older brother has X pencils and lends the younger brother Y pencils" "Addition event" and "subtraction event" of X-Y;

F3. Design the "addition event" as an independent function, and translate the event into a regular expression through the intent analysis module 400. The regular expression can be selected through the entity vocabulary (gift/give/purchase.. .) to illustrate that the semantics of these sentences all represent the function of "addition event";

F4. Design the "subtraction event" as an independent function, and translate the event into the regular expression through the intent analysis module 400, and the regular expression can be selected through the entity vocabulary (eat/borrow/sell ...) to illustrate that the semantics of these sentences all represent the function of "subtraction event";

F5. The system transmits the sentences in a math topic one sentence at a time to the Math project, and the sentences are processed by the sentence analysis module 300 into strings containing part-of-speech tags (POS/NER);

F6. After the intent analysis module 400 compares the syntactic structure models under the Math project, it is translated into a question sentence string that can represent the intention of "addition event" or "subtraction event";

F7. Extract the functions in the event pool according to the sentence string of the question, the entity vocabulary and different sentence structures, and then solve the problem.

In the third embodiment above, take the topic "There are three apples on the table, Xiao Ming eats one, and how many apples are left" as an example. Eat one" and "How many apples are left"; after comparing the Math project, "There are three apples on the table" will return the intention of "Definition" and the sentence pattern The lexical units that can be counted in are three arguments: "table (Possessor)", "apple (Entity)" and "three (Quantity) (Classifier)", as shown in [Figure 14a].

In the third embodiment above, the second sentence "Xiao Ming eats one" will return the intention of "Calculation", so the lexical units available for calculation in the sentence pattern are "Xiao Ming (Possessor)", "Apple ( Entity)", "Quantity (Classifier)" and other arguments, because "eat" is a "subtraction event", so "one" plus a minus sign becomes "-1", as shown in [Figure 14b] .

In the third embodiment above, the last sentence "how many apples are left" will be returned from the Math project with the intent of "question"; the object of the query is "(several) apples", and the entity of the query is " Apple”; because the owner is not mentioned in the sentence, this field is left blank; as shown in [Figure 14c], according to the solution instruction, the entity to be queried is “apple”, so start from “Definition Context Scenario” Start the calculation, get the initial definition of "apple = 3", then enter the calculation process, get the calculation process of "apple-1", because there is no other calculation process, so get 3-1=2, and there are 2 pieces left Apple's final answer.

In the third embodiment above, if the question is not asking "How many apples are left", but "How many apples did Xiao Ming eat in total", then "Question" will be as shown in [Figure 14d] Result: During the calculation process, because there is no record of "Xiao Ming holds an apple" in the "Definition" data, the definition context scene will be skipped and directly enter the "Calculation" In the data, get the record of "the number of apples held by Xiao Ming is -1", and because "eating" is a "subtraction event", so the obtained value must be added with a minus sign to get "-(-1)" to "1". Therefore, Xiao Ming ate the final answer of "1 apple" in total.

It should be understood that the above specific embodiments of the present invention are only used to illustrate or explain the principle of the present invention, and not to limit the present invention. Therefore, any modification, equivalent replacement, improvement, etc. made without departing from the spirit and scope of the present invention shall fall within the protection scope of the present invention. Furthermore, the appended claims hereto are intended to cover All changes and modifications within the scope and boundaries of the claims, or equivalents of such scope and boundaries are intended.

100: input module

200:Database module

300: Sentence analysis module

400: Intent analysis module

500: Output conversion module

Claims (10)

A method for analyzing the intention of a natural language dialogue, wherein the method for establishing a plurality of syntactic structure models is as follows: S1. An example sentence creation unit of an input module of the system provides a manager to create an mth application scenario; S2. Further, the manager creates a data set belonging to an nth intent in the mth application scenario in the mth application scenario; S3. Further, under the nth intent, the manager creates a plurality of example sentences that meet the nth intent of the mth application scenario in a sentence-by-example manner, and transmits them to a sentence analysis module; S4. The sentence analysis module performs lexical segmentation and part-of-speech tagging on these example sentences, and further, transfers these example sentences marked with part-of-speech into plural sentence structure; S5. An intent analysis module binds the sentence structures in the mth application scenario to the nth intent, encodes these syntactic structure models, and stores them in one of the mth application scenario model units of a scenario database ; S6. Repeat steps S2~S5 until all intents belonging to the mth application scenario have been established; S7. An output conversion module of the system extracts the syntactic structure models of the m-th application scene model unit and converts them into a program code for output. The method for analyzing the intention of a natural language dialog as described in Claim 1, wherein the step S4 further includes, the sentence analysis module provides a screen for checking entity words according to the words and parts of speech of the example sentences. The natural language dialogue intent analysis method as described in claim 1, wherein the input module The group further includes a custom dictionary unit, which can import these example sentences. The natural language dialogue intent analysis method as described in claim 1, wherein step S7 of the process is continued as follows: S8. A user connects to a dialog page of the m-th application scene from a terminal device, inputs a dialog sentence in the input module of the terminal device, and sends it to the sentence analysis module; S9. The sentence analysis module extracts the dialogue sentence on the dialogue page, performs lexical segmentation, part-of-speech tagging on the dialogue sentence, sets the verb position as the anchor point, and translates the dialogue sentence marked with the part-of-speech into the sentence pattern structure, encoding the sentence structure and the mth application scenario and sending them to the intent analysis module; S10. The intent analysis module links the mth application scenario model unit of the scenario database according to the mth application scenario, and compares the sentence structure of the dialogue sentence with the mth application scenario model units. Syntax structure model, using the anchor point as a reference point for sentence structure comparison; S11. If it conforms to a set of models among the syntactic structure models, then define the n-th intent bound by the set of models in the dialogue sentence as an expression map. The method for analyzing the intention of a natural language dialogue as described in Claim 1, wherein the sentence structure sets the position of the verb in the sentence as an anchor point, which is used as a reference point for sentence structure comparison. The method for analyzing the intent of a natural language dialogue as described in Claim 2, wherein the sentence analysis module translates the dialogue sentence into the sentence structure according to the vocabulary, part of speech and the entity vocabulary. In the method for analyzing the intent of a natural language dialog as described in claim 1 or claim 4, the input module can also be input in a text form. The natural language dialogue intent analysis method as described in claim item 4, wherein the expression means Graphs can be sent to logical blocks of multiple applications. The method for analyzing the intent of a natural language dialog as described in Claim 4, wherein the graph can be connected to a function event pool for mathematical operations. The method for analyzing the intent of a natural language dialogue as described in Claim 3, wherein the representation can be used for text classification, and can also be further used for name extraction or other information extraction.

Images