CN114756646A

CN114756646A - Conversation method, conversation device and intelligent equipment

Info

Publication number: CN114756646A
Application number: CN202210283666.7A
Authority: CN
Inventors: 罗沛鹏; 谭欢
Original assignee: Shenzhen Ubtech Technology Co ltd
Current assignee: Shenzhen Ubtech Technology Co ltd
Priority date: 2022-03-22
Filing date: 2022-03-22
Publication date: 2022-07-15

Abstract

The application discloses a conversation method, a conversation device, intelligent equipment and a computer readable storage medium. Wherein, the method comprises the following steps: extracting key information of all news which occur in the current day through the trained news identification model; generating candidate open field white sentences according to the key information; extracting a user portrait of the user through the trained user portrait model; determining a target open field white sentence according to the user portrait in the candidate open field white sentences; and initiating a conversation to the user based on the target open field white statement. Through the scheme, the intelligent device can actively have meaningful conversation with the user.

Description

Conversation method, conversation device and intelligent equipment

Technical Field

The present application belongs to the technical field of artificial intelligence, and in particular, relates to a dialogue method, a dialogue device, an intelligent device, and a computer-readable storage medium.

Background

With the mature popularization of the artificial intelligence technology, the dialogue system is gradually applied to various industries, and various fields such as vehicle-mounted, smart home, smart wearing and smart sound boxes are overturned. However, most of the existing dialog systems are passive dialog systems, i.e. each time the user wakes up, the user actively asks the device to answer. Because the user does not know which range of conversations the device can handle, the conversations between the user and the device often only stay on common topics such as weather queries or road condition queries. Over time, the user's conversation enthusiasm may be affected and the device's conversation system may be difficult to represent its true use value.

Disclosure of Invention

The application provides a conversation method, a conversation device, an intelligent device and a computer readable storage medium, which can enable the intelligent device to actively carry out meaningful conversation with a user.

In a first aspect, the present application provides a dialog method, including:

extracting key information of all news which occur in the current day through the trained news identification model;

generating candidate open field white sentences according to the key information;

extracting a user portrait of the user through the trained user portrait model;

determining a target open field white sentence according to the user portrait in the candidate open field white sentences;

and initiating a conversation to the user based on the target field opening sentence.

In a second aspect, the present application provides a dialog device, comprising:

the first extraction module is used for extracting key information of all news which occur on the current day through the trained news identification model;

the generating module is used for generating candidate open field white sentences according to the key information;

the second extraction module is used for extracting the user portrait of the user through the trained user portrait model;

a determining module, configured to determine a target open field white sentence according to the user portrait from the candidate open field white sentences;

And the conversation initiating module is used for initiating a conversation to the user based on the target field opening sentence.

In a third aspect, the present application provides a smart device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method according to the first aspect when executing the computer program.

In a fourth aspect, the present application provides a computer readable storage medium having stored thereon a computer program which, when executed by a processor, performs the steps of the method of the first aspect as described above.

In a fifth aspect, the present application provides a computer program product comprising a computer program which, when executed by one or more processors, performs the steps of the method of the first aspect as described above.

Compared with the prior art, the beneficial effect that this application exists is: the intelligent equipment extracts key information of all news which occur on the same day through a trained news identification model, then generates candidate open field white sentences according to the key information, extracts user figures of users through a trained user figure model, determines target open field white sentences according to the user figures in the candidate open field white sentences, and finally initiates a conversation to the users based on the target open field white sentences. Through the process, the intelligent device can actively talk with the user, and topics during active talking have the following characteristics: on one hand, the news is generated based on the news which occurs in the current day, so that the user can obtain the latest information; on the other hand, the user images are screened, so that the output topics can catch the interests of the user. Therefore, the intelligent device can actively carry out meaningful conversation with the user, and the initiative of the conversation between the user and the intelligent device is fully aroused. It is understood that the beneficial effects of the second aspect to the fifth aspect can be referred to the related description of the first aspect, and are not described herein again.

Drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments or the prior art descriptions will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present application, and it is obvious for those skilled in the art to obtain other drawings without creative efforts.

Fig. 1 is a schematic flow chart illustrating an implementation of a dialog method according to an embodiment of the present application;

FIG. 2 is a diagram illustrating an example of an output of a news recognition model in a dialog method according to an embodiment of the present application;

FIG. 3 is a diagram illustrating an exemplary output of a user representation model in a dialog method according to an embodiment of the application;

fig. 4 is a block diagram of a dialog device according to an embodiment of the present application;

fig. 5 is a schematic structural diagram of an intelligent device provided in an embodiment of the present application.

Detailed Description

In the following description, for purposes of explanation and not limitation, specific details are set forth, such as particular system structures, techniques, etc. in order to provide a thorough understanding of the embodiments of the present application. It will be apparent, however, to one skilled in the art that the present application may be practiced in other embodiments that depart from these specific details. In other instances, detailed descriptions of well-known systems, devices, circuits, and methods are omitted so as not to obscure the description of the present application with unnecessary detail.

In order to explain the technical solution proposed in the present application, the following description is given by way of specific examples.

The following describes a dialogue method proposed in an embodiment of the present application, taking an example in which an intelligent device is a robot. Certainly, the smart device may also be a smart phone, a tablet computer, a smart home, or a smart sound box, and the type of the smart device is not limited herein. Referring to fig. 1, the implementation flow of the dialog method is detailed as follows:

step 101, extracting key information of all news occurring in the current day through the trained news identification model.

In this embodiment, the robot may perform offline processing on all news that occur on the same day at a specified time point or time period, and obtain key information of each news that has occurred up to the specified time point or time period on the same day. Wherein the offline processing is mainly performed depending on the trained news recognition model. By way of example only, the key information of news extracted by the news recognition model includes the following items: entities, events, actors and acceptors. The actors and respondents also belong to entities, and both can be considered as special entities related to the event.

The entity may express specific things such as time, place, or people, for example: the team of China.

Where an event may express an entity-triggered event or an entity-occurred event, such as: the Chinese team seizes the crown.

In some embodiments, for any news item, the robot may specifically input the title of the news item, rather than the text, into the trained news recognition model to extract the key information of the news item. The number of words of the news title is far less than that of the text of the news, and the news title is obtained by summarizing the core information of the news, so that the processing pressure of a news identification model can be greatly reduced, and the extraction speed of the key information of the news is effectively improved.

And 102, generating candidate open field white statements according to the key information.

In the embodiment of the present application, the robot may generate candidate open-air sentences according to the key information of each news item of the current day extracted in step 101. Generally, each news item may generate at least one candidate open-field white sentence. It can be understood that the candidate open field white sentence provided in the embodiment of the present application is actually a news-related topic actively proposed by the robot, and may be a statement sentence or a question sentence, and the representation form of the candidate open field white sentence is not limited herein.

And 103, extracting the user portrait of the user through the trained user portrait model.

In the embodiment of the application, the robot can acquire the historical chat records of the robot and the user, and then the historical chat records are input into the trained user portrait model, so that the user portrait model can output the user portrait of the user.

Of course, after obtaining the user image of the user output by the user image model, the robot may also adjust the user image based on the user information of the user to obtain a more accurate user image, which is not limited herein.

And step 104, determining a target open field white sentence according to the user portrait in the candidate open field white sentences.

In embodiments of the application, from the user representation, the bot may determine which aspects of news are of interest to the user. In order to be able to invoke the user's positivity in the subsequently initiated conversation, the robot may consider that, among the candidate open-white sentences, news that may be of interest to the user are determined from the user portrait, and select the target open-white sentence among the candidate open-white sentences generated by these news.

And 105, initiating a conversation to the user based on the target field opening sentence.

In the embodiment of the application, when the target open field white sentence is a complete sentence, the robot can directly output the target open field white sentence as the open field white of the dialogue. It can be understood that, according to the interaction form adopted by the current robot, the robot can initiate a conversation in a voice form; alternatively, the robot may initiate the dialog in a text form, which is not limited herein.

In some embodiments, to obtain a trained news recognition model, the robot may do the following:

first, data is collected, such as: collecting historical news, and storing titles, descriptions and/or texts and the like in the historical news in a database in a structured form according to categories; besides, the entities (such as names of people, places, organizations and the like) and events mentioned in the historical news can be stored in the database in a structured form according to categories. Specifically, the title, description, and/or text of the historical news may be stored in one table in the database, and each entity of the historical news and each event may be stored in another table in the database, and the two tables may be associated by a primary key (primary key) -foreign key (foreign key). Of course, the various items of data related to the historical news may be stored in the form of files, etc., instead of being stored in the database, and the present invention is not limited thereto.

Then, the model to be used is selected. For example only, the robot may employ an existing entity recognition Model as the news recognition Model, including but not limited to Hidden Markov Model (HMM), Conditional Random Field (CRF) Model, Bi-directional Long Short-Term Memory (BiLSTM), and BiLSTM + CRF; alternatively, the robot may also use a pre-trained model with better effect but larger computational cost as the news recognition model, including but not limited to BERT (bidirectional Encoder reproduction from transformers), BERT + CRF, and albert (a Lite BERT), etc., without limitation.

Next, two types of outputs of the model and corresponding two loss functions are specified.

Wherein, one kind of output is used for identifying the entity and the event, and the other kind of output is used for identifying the executing party and the receiving party. Both types of outputs are labeled in BIO format.

When the entity and the event are identified, the labeling mode is specifically as follows: o represents a word other than an entity and an event; b _ e represents the first character of the entity; i _ e represents the non-first character of the entity; b _ a represents the first character of the event; i _ a represents the non-first word of the event.

When identifying the executing party and the receiving party, the labeling mode is as follows: o represents a word except the executing party and the receiving party; b _ s represents the first character of the seller; i _ s represents the non-first character of the seller; b _ o is the first word of the victim; i _ o is the non-first word of the victim.

Referring to fig. 2, fig. 2 shows an example of two different types of output based on the same input text. As can be seen from FIG. 2, the user inputs "surprise! Capturing european cup in italy, "the first category output of the news recognition model will recognize: "surprise", "! "and" is a word other than an entity and an event; "meaning" in "italy" is the first letter of an entity, "big" and "profit" are the non-first letters of an entity; "Europe" in "European cup" is the first character of the entity, "continent" and "cup" are the non-first characters of the entity; the "grab" in "grabbing the crown" is the first letter of the event and the "crown" is the second letter of the event. The second type of output of the news recognition model will recognize: "surprise", "! "," grab "," crown "and" has "are words other than the executing party and the receiving party; "Italian" means "the first letter of the Party," Large "and" Lily "means the second letter of the Party; "Europe" in "European cup" is the first letter of the receiver, and "continent" and "cup" are the non-first letters of the receiver.

For the first kind of output (i.e., output for identifying entities and events), a first Loss function is adopted, and the calculated Loss value is denoted as Loss 1; for the second type of output (i.e., the output used to identify the actor and victim), a second penalty function is used, and the calculated penalty value is denoted as Loss 2. The first loss function and the second loss function may each label the loss function with some sequences of the current main flow, such as cross entropy and the like. It should be noted that Loss1 is completely independent of Loss2, i.e., Loss1 is only associated with the first type of output and Loss2 is only associated with the second type of output. In general, for the same input text, the robot may consider the Loss value of the average of Loss1 and Loss2 as a whole; alternatively, the weights of both of the Loss1 and the Loss2 may be set according to the importance of the Loss 3878 and the Loss2, and the weighted average of the Loss1 and the Loss2 is obtained as the overall Loss value, for example, the overall Loss value Loss ═ is (α ═ Loss1+ β ×) Loss2)/2, where α is the weight of the Loss1, β is the weight of the Loss2, and α + β is 1.

It can be understood that after the model is selected and the output and loss functions are specified, a large number of titles of the historical news need to be labeled in a labeling mode specified by the output of the model to make labels.

And finally, training the model, and obtaining the trained news recognition model when the overall loss value reaches convergence or the training iteration number meets a preset number threshold. The trained news recognition model can be put into use. The training and tuning mode of the model is similar to the current mainstream deep learning training mode, and is not repeated here.

In some embodiments, the robot may randomly select one of the models that may be used by the news recognition models shown above (e.g., HMM, CRF, BilSTM + CRF, BERT + CRF, and ALBERT, etc.) as the user profile model, but unlike the news recognition models, the user profile model has only one type of output that functions as: for identifying people, entities and relationships. Where a person is said to belong to a particular class of entities, e.g. you, me or he. Specifically, the person names are used for distinguishing and judging whether a certain relation or entity is related to the user; relationships are used to express the type of user representation, such as: expressing the preference, interest, hobby, idol, hometown and the like of the positive relationship; entities are used to express specific attributes of a user representation, such as: military news. The output of the type is labeled by adopting a BIO format, and the labeling mode is specifically as follows: b _ p represents the first character of a person; i _ p represents the non-first character of the person; b _ r represents the first character of the relationship; i _ r represents the non-first character of the relationship respectively; b _ e represents the first character of the (non-human) entity; i _ e represents a non-first character of the (non-human) entity; o represents a word outside of an entity and a relationship.

Turning to FIG. 3, FIG. 3 illustrates an example of the output of a user representation model. As can be seen in FIG. 3, for the input "I like military News, I, the output of the user representation model will recognize: "very" and "woollen" are words outside the entity and relationship; "I" is the first word of the name; the 'liking' in the 'liking' is the first character of the relationship, and the 'huan' is the non-first character of the relationship; the "military" in "military news" is the first character of the entity, and the "affair", "new" and "smelling" are the non-first characters of the entity.

It can be understood that the training and tuning method of the user portrait model is similar to the current mainstream deep learning training method, and the details are not repeated herein. After the trained user profile model is obtained, the trained user profile model may be put into use. During application, the trained user representation model is input as sentences in the user's historical chat history. The trained user profile model may extract entities, relationships, and names from the statements in the historical chat history, and based on this information, a user profile may be obtained.

Specifically, the robot may determine whether the extracted relationship and the entity are related to the user according to the person name; if the user is related, the robot may add the entity to the user representation table based on the extracted relationship to form a user representation of the user. For example, according to a history chat record of a certain user that "i like military news tweed", a person name "i", a relation "like" and an entity "military news" are extracted; by the person being 'me', the relationship 'like' and the entity 'military news' are determined to be relevant to the user; thereafter, the entity "military news" is added to the "preferences" column in the user profile table for the user by the relationship "like" and the entity "military news".

In some embodiments, based on the news recognition model described above, the key information extracted in step 101 includes the following four categories: entities, events, actors, and victims. Then, for each piece of news that has occurred on the current day, step 102 specifically includes:

and A1, combining the entities, the events, the actors and the acceptors of the news to obtain all possible combined results.

And A2, generating candidate open-air sentences related to the news according to the combination result and the preset template corresponding to the combination result.

For ease of understanding, the following description is given in specific examples:

1. candidate open-field white sentences are generated based on the single entity and a first preset template, wherein the first preset template can be ' entity today ' upper head bar '. Such as: today's Zhou Ji head bar.

2. Candidate open-white sentences are generated based on the individual events and a second preset template, which may be "event information up-to-date". Such as: there is the latest captivation information.

3. Candidate open-white sentences are generated based on separate actors and a third preset template, which may be "[ actors ] today do one. Such as: one big thing is done today in italy.

4. Candidate open-white sentences are generated based on the independent victim and a fourth preset template, wherein the fourth preset template can be 'the victim' stands up for a point. Such as: england today stands out of points.

5. And generating candidate open-field white sentences based on the plurality of entities and a fifth preset template, wherein the fifth preset template can be 'entity 1' and 'entity 2'. Such as: italy and england have melons.

Of course, there are more combinations and corresponding default templates, which are not described herein for brevity. As can be seen, the robot can generally generate a plurality of candidate open-field white sentences through the key information extracted from a news.

In some embodiments, in order to ensure the reasonableness of the target open-field white statement selected by the robot when dealing with various types of users, step 104 may include:

b1, determine whether the user image is empty.

The user representation model is input as a historical chat history of the user. If the user is a new user who has not had any interaction with the bot, or if the user is a temporarily logged-in guest user, then the user's historical chat history must be empty. Obviously, when the history chat record of the user is empty, the output of the user portrait model is also empty inevitably because the input of the user portrait model is empty. That is, in an application scenario where the user is a new user or a guest user, the robot cannot acquire an effective user representation of the user.

And B2, if the user portrait is empty, determining a target open field white sentence in the candidate open field white sentences according to the hot news.

When the user portrait is empty, the user is known as a new user or a guest user. Because the portrait of the user is empty, the user's preference on news cannot be known, and therefore personalized and targeted conversation with the user is not considered any more. At this time, hot news can be screened out from all news occurring on the current day, and a target open field white sentence is determined according to the hot news, which specifically comprises the following steps: and determining any sentence in the candidate open white sentences generated based on the hot news as a target open white sentence.

Wherein, the hot news may be: accumulating N news with the highest click rate in all news which occur in the same day, wherein N is a preset positive integer; or, in all news which occur on the same day, accumulating news of which the click rate exceeds the first preset click rate; or, the N news with the highest click rate in the latest unit time (for example, the latest hour) among all news that have occurred on the day; or, the news with the click rate exceeding the second preset click rate in the latest unit time is all news which occur on the same day. The determination of the hot news is not limited herein.

B3, if the user image is not empty, according to the user's preference indicated by the user image, determining the target open white sentence in the candidate open white sentences.

When the user portrait is not empty, the user is known as an old user. In order to arouse the enthusiasm of the user for interacting with the robot, the robot can select news which the user may be interested in according to the preference of the user indicated by the user portrait, and determine any sentence in candidate open-field white sentences generated based on the news as a target open-field white sentence.

It will be appreciated that the target open field white statement may contain only one candidate open field white statement. For example, the target open field white statement may be: italy today recapitulates. Alternatively, the target open field white statement may also contain two or more candidate open field white statements. When the target open white sentence contains two or more candidate open white sentences, the two or more candidate open white sentences may originate from the same news (i.e., generated based on the same news), or the two or more candidate open white sentences may originate from different news (i.e., generated based on different news). For example, the target open field white statement may be: italy today has taken over, and baiden today visits china.

In some embodiments, step B3 may specifically include:

b31, detecting whether a candidate open field white sentence which is matched with the preference of the user exists.

In some embodiments, based on the user profile model described above, the user's preferences typically include one, two, or more entities. Based on this, the bot may match entities contained in the user's preferences with news items that have occurred on the current day. To improve the matching efficiency, the robot may extract the summaries of the news, and then match the entities included in the preferences of the user with the summaries of the news that have occurred in the current day. If the matching is successful, the corresponding news is the news matched with the preference of the user; correspondingly, the candidate open-field white sentences generated based on the news are the candidate open-field white sentences matched with the preference of the user.

And B32, if the candidate open white sentences matched with the preference of the user exist, determining the target open white sentences in the candidate open white sentences matched with the preference of the user.

And B33, if the candidate open white sentences matched with the preference of the user do not exist, determining the target open white sentences in the candidate open white sentences which have the association relation with the preference of the user.

The robot can find out topics which are related to the preference of the user through a data mining mode or a knowledge graph mode, and news matched with the topics are news which are related to the preference of the user; correspondingly, the candidate open white sentences generated based on the news are the candidate open white sentences associated with the preference of the user.

The following explains the manner of data mining:

the robot can utilize association rules in data mining to perform entity identification on a large number of news, regard each news as a transaction (transaction), and number each news to obtain a transaction sequence number; for each news item, all the entities contained in the content of the news piece are pieced together into one data item (item). Thus, a data mining table is constructed, as shown in table 1 below:

transaction sequence number	Data item
		1	Australia, iron ore, chessmen, … …
2	Harden, basket net, durant, … …
		……	……

TABLE 1

Based on the constructed data mining table, the Confidence (Confidence) of an entity B relative to another entity A can be calculated, specifically to represent the frequency of the simultaneous occurrence of the entities B when the entities A occur, and is denoted as { A → B }. In other words, the confidence of a certain entity B with respect to another entity a refers to: the ratio of the number of transactions containing both A and B terms to the number of transactions containing A terms. This formula can be expressed simply as: confidence of { a → B }, P (a | B)/P (a). For example only, the confidence may be calculated by Apriori and other algorithms, which are not described herein.

Based on the concept of confidence given above, the robot can perform association rule analysis based on data mining tables to calculate the confidence of an entity relative to another entity. If the confidence is sufficiently high, then the entity may also be added to the relationship to which the other entity belongs in the user representation. Such as: the Yaoming and the rocket team commonly appear in a large amount of news, the confidence coefficient of the rocket team relative to the Yaoming is calculated through the previous text, and the user likes the Yaoming with high probability because the user likes the Yaoming. In this way, news that is associated with the user's preference can be found. For example, the user's preference indicated in the user profile is "yaoming," but there is no news associated with "yaoming" on the day; through data mining, the confidence coefficient of the rocket team relative to the Yaoming is high, and news related to the rocket team exists on the day, so that the news related to the rocket team is news which has an association relation with the preference of the user; thereafter, a target open-field white sentence may be determined among candidate open-field white sentences generated based on news related to the "rocket team".

The following explains the way of the knowledge graph:

a plurality of knowledge maps are generated in advance in the robot. In a knowledge graph, each entity may be linked to other entities by a given relationship. When the robot extracts the names, the relations and the entities from the historical chat records, if the relations meet the relations shown in the knowledge graph, information can be transmitted through the knowledge graph to find other entities having the association relations with the entities preferred by the user, news corresponding to the other entities are news having the association relations with the preferences of the user, and the robot can immediately determine target open-field white sentences in candidate open-field white sentences generated by the news.

In some embodiments, in order to avoid the difficulty in understanding the responses of the user by the robot, when designing the dialogs used by the candidate open-field white sentences generated by the robot, the dialogs should be designed as much as possible so as to guide the user to make interesting or uninteresting responses and avoid the responses of the user from being too open. For example, the candidate open-field sentence may be designed to guide the user to get a news introduction or summary. For the dialog actively initiated by the robot based on the target open-field sentence, the responses of the user can be divided into the following types: positive, meaning interested, wishing to chat down to get details of the news; neutral, meaning not interesting, wish to switch to another news; negative, meaning not interesting, does not wish to chat about news. Based on this, the robot can train a classifier in advance, and this classifier is provided with three categorizations, is respectively: positive, neutral and negative. The robot may comb through a large number of corpora that these three categories may contain and train the classifier using commonly used text classification algorithms. Based on the classifier, after step 105, the dialog method further comprises:

C1, after receiving the answer of the user to the target open field white sentence, classifying the answer.

It is understood that after the robot initiates a dialog based on the target open text, the user can reply to the target open text. After receiving the reply, the robot may classify the reply by the classifier set forth above to determine the attitude of the user to the target open-field white sentence.

C2, if the answer is positive, outputting the abstract of the target news.

If the user's response is classified as a positive response, it indicates that the user currently wishes to drill down into the target news, which refers to: and opening news corresponding to the white sentences of the target. The robot may extract a summary of the target news and output the summary of the target news to the user when the user gives a positive response as feedback by the robot to the user's positive response. By the method, when the user is interested in the actively initiated news topic, the abstract can be simply and clearly replied to the user, so that the feeling of long-length and big-theory can not be generated, and the effect of natural conversation with the user is achieved.

The abstract of the target news can be extracted in the following way:

A text abstract model is preset in the robot, wherein the text abstract model is input into a certain paragraph of news and output as an abstract of the paragraph. The text abstract model adopts a seq2seq structure and can be a plurality of network models as follows: recurrent Neural Network (RNN), LSTM, and transformer, etc., without limitation. By way of example only, when using the transformer network model, the encoder is responsible for extracting features of an input paragraph and the decoder is responsible for generating a summary relating to the features of the paragraph.

C3, if the answer is neutral, switching the target open-field white sentence and returning to execute the step 105 and the following steps based on the switched target open-field white sentence.

If the user's response is classified as a neutral response, it indicates that the user currently wishes to switch to other news topics. At this time, the robot may find out other news that match the preference of the user, or may also find out other news that have an association relationship with the preference of the user in the above-described knowledge graph manner or data mining manner, and redetermine a new target open white sentence from candidate open white sentences generated by the other news, so as to implement switching of the target open white sentence (that is, a news topic). The robot may then re-initiate a new dialog to the user based on the switched target open field white statement.

C4, if the answer is negative, triggering the chatting system to continue the dialogue with the user.

If the user's response is classified as a negative response, it indicates that the user does not currently wish to talk to the bot about the news. At this time, the robot may trigger the operation of the chat system, and perform chat with the user based on the chat system.

C5, if the answer can not be classified, analyzing the answer and continuing the dialogue with the user based on the analysis result.

If the user fails to respond as guided by the preset dialect, the response is likely to be unable to be classified by the classifier as any of a positive response, a neutral response, and a negative response. For the case where such a response cannot be classified, the robot may consider the response as an open response. At this point, the robot needs to further analyze the reply and determine what dialog strategy should be employed to continue the dialog with the user based on the analysis results. Based on this, step C5 specifically includes:

and C51, analyzing the relevance of the target news and the answer.

For example only, the bot may determine the relevance based on the frequency of occurrence of keywords in the response in the target news. Obviously, the higher the frequency of occurrence, the higher the degree of association. That is, the frequency of occurrence is proportional to the degree of correlation.

Alternatively, the robot may determine the association based on the similarity of the reply to the digest of the target news, and obviously, the higher the similarity, the higher the association. That is, the similarity is in direct proportion to the correlation.

C52, if the association degree is higher than a preset association degree threshold value, triggering the multi-turn dialogue system to continue dialogue with the user;

the robot presets a threshold of relevance. When the relevance calculated in step C51 is higher than the relevance threshold, the answer is considered to be the answer given by the user around the target open field sentence, i.e. the user is still interested in the target news. Considering that the response is an open response, the bot may trigger the multi-turn dialog system to continue the dialog with the user in order to achieve a natural dialog effect.

And C53, if the relevance is equal to or lower than the relevance threshold, triggering the chatting system to continue the conversation with the user.

When the degree of association calculated at step C51 is not higher than the degree of association threshold, the response is considered to be substantially unrelated to the target news. That is, the user is not concerned with the specific content of the target news. At this point, the robot may trigger the chat system to feed back the user's response. Specifically, the chat system includes a QA pair matching library and a generative model, which are not described herein again.

In some embodiments, the multi-turn dialog system works as follows:

the core in a multi-turn dialog system is a trained language model. The robot can splice the abstract of the target news, the sentences actively initiated by the robot and the responses of the user into a conversation sequence, and then the conversation sequence is input into the language model for prediction. The language model may use RNN or GPT (generic Pre-Training) algorithm, but is not limited thereto. It can be understood that in the embodiment of the present application, the structure of the language model itself is not improved; the actual improvement part is: the input of the language model specifically comprises the following steps: the input dialog sequence contains a summary of the target news. This may enable the robot to give feedback around the target news.

It can be understood that the sentence actively initiated by the robot in the first round is actually the target open field white sentence; therefore, after receiving a response from the user to the target open-field sentence, if the response is an open response and the response is highly associated with the target news, the robot enters a multi-turn dialog state, and the multi-turn dialog system can input a dialog sequence to the language model based on the first turn of dialog, where the dialog sequence is specifically:

[ CLS ] Abstract of target News [ SEP ] target open-field white sentence [ SEP ] user's answer [ SEP ]

It is understood that in the dialog sequence, the target open field sentence is the sentence output by the robot in the first round, and the answer of the user is the answer of the user in the first round. Based on the dialog sequence, the language model may predict the sentence that the robot will output in the second round.

Specifically, the language model may predict the first word after the dialog sequence and continue predicting the next word after splicing the predicted first word into the dialog sequence. By analogy, the operations of concatenating and predicting the next word are repeated until the next word is predicted as [ SEP ], or a termination rule is reached (e.g., the predicted statement contains a preset number of words that has reached a preset number of words threshold). Therefore, the robot can output the predicted sentence as the output of the current round (namely, the second round) of the robot.

Similarly, when the multi-turn dialogue system predicts the sentence to be output by the nth turn of the dialogue, the input dialogue sequence is:

abstract [ SEP ] target open field white sentence [ SEP ] of [ CLS ] target news user answer [ SEP ] output by first round of [ SEP ] robot answer [ SEP ] user second round of [ SEP ] answer [ SEP ] output by second round of [ SEP ] robot answer [ SEP ] user third round of [ SEP ] … … answer [ SEP ] output by N-1 round of robot answer [ SEP ] user N-1 round of answer [ SEP ]

In some embodiments, the robot also needs to train the language model before it is put into use. The training process is briefly as follows:

firstly, a large number of corpora belonging to sequence types are constructed in advance, such as:

[ SEP ] is the answer of the [ SEP ] user of the dialogue [ SEP ] actively initiated by the [ CLS ] news summary [ SEP ] robot to continue the [ SEP ] user to continue the [ SEP ].

Then, the next possible word can be predicted according to the existing sequence, and the operation logic of the next possible word is the same as that of the language model in the application process. Because the robot constructs a large number of corpora belonging to the sequence type in advance, the language model can learn the rule of the type of conversation through continuous training.

In some embodiments, the timing of the operation of the user representation model is not limited. For example, it may be that, regardless of the response speed, all users are subjected to real-time processing, which refers to: and in the process of the conversation between the robot and the user, inputting the latest chat records into the user portrait model in real time to realize the real-time updating of the user portrait. Alternatively, the above-described real-time processing may be performed only for the new user and the guest user, and the offline processing may be performed for the old user, where the offline processing refers to: after the current conversation between the user and the robot is finished, the current chat record (namely the latest historical chat record) is input into the user portrait model, and the offline updating of the user portrait is realized. Of course, all user images of all users may be processed offline, and this is not limitative.

Therefore, according to the embodiment of the application, the intelligent device extracts the key information of all news which occur on the same day through the trained news identification model, then generates the candidate open field white sentences according to the key information, extracts the user portrait of the user through the trained user portrait model, determines the target open field white sentences according to the user portrait in the candidate open field white sentences, and finally initiates the dialogue to the user based on the target open field white sentences. Through the process, the intelligent device can actively talk with the user, and topics during active talking have the following characteristics: on one hand, the news is generated based on the news which occurs in the current day, so that the user can obtain the latest information; on the other hand, the user portrait is screened, so that the output topics can catch the interests of the user. Therefore, the intelligent device can actively carry out meaningful conversation with the user, and the enthusiasm of the user for conversation with the intelligent device is fully aroused.

Corresponding to the dialogue method provided above, the embodiment of the present application further provides a dialogue device. As shown in fig. 4, the dialogue device 400 includes:

the first extraction module 401 is configured to extract, through the trained news identification model, key information of all news that have occurred on the current day;

A generating module 402, configured to generate a candidate open field white statement according to the key information;

a second extraction module 403, configured to extract a user portrait of the user through the trained user portrait model;

a determining module 404, configured to determine a target open field white sentence according to the user portrait from the candidate open field white sentences;

a dialog initiating module 405, configured to initiate a dialog to the user based on the target field opening statement.

Optionally, the key information includes: entities, events, actors, and victims; for each piece of news that has occurred on the current day, the generating module 402 includes:

the key information combination unit is used for combining the entities, the events, the actors and the respondents of the news to obtain all possible combination results;

and the candidate open-field white sentence generating unit is used for generating candidate open-field white sentences related to the news according to the combination result and a preset template corresponding to the combination result.

Optionally, the determining module 404 includes:

a judging unit for judging whether the user portrait is empty;

a first determining unit, configured to determine, according to hot news, the target open field white sentence in the candidate open field white sentences if the user portrait is empty;

And a second determining unit, configured to determine the target open white sentence in the candidate open white sentences according to the user preference indicated by the user picture if the user picture is not empty.

Optionally, the second determining unit includes:

a detecting subunit, configured to detect whether there is the candidate open field white sentence matching the preference of the user;

a first determining subunit configured to determine, if there is the candidate open white sentence matching the preference of the user, the target open white sentence from the candidate open white sentences matching the preference of the user;

a second determining subunit, configured to determine, if there is no candidate open white sentence that matches the preference of the user, the target open white sentence from the candidate open white sentences that have an association relationship with the preference of the user.

Optionally, the dialog device 400 further includes:

a classification module, configured to classify a reply of the user to the target open field white sentence after receiving the reply;

a processing module, configured to output an abstract of a target news if the answer is a positive answer, where the target news is: news corresponding to the target open field white sentence; if the answer is a neutral answer, switching the target open field white sentence and triggering the operation of the dialogue initiating module 405 again; if the answer is a negative answer, triggering the chatting system to continue to have a conversation with the user; if the responses cannot be classified, the responses are analyzed and a dialog with the user is continued based on the analysis result.

The processing module is specifically configured to, in a case where the response cannot be classified, analyze a degree of association between the target news and the response; if the association degree is higher than a preset association degree threshold value, triggering a multi-turn dialogue system to continue to have dialogue with the user; and if the association degree is equal to or lower than the association degree threshold, triggering the chatting system to continue to have a conversation with the user.

Optionally, the analyzing the association between the target news and the response includes: determining the degree of association based on the frequency of occurrence of the keywords in the response in the target news; alternatively, the association degree is determined based on the similarity between the reply and the digest of the target news.

Therefore, according to the embodiment of the application, the intelligent device extracts the key information of all news which occur on the same day through the trained news identification model, then generates the candidate open field white sentences according to the key information, extracts the user portrait of the user through the trained user portrait model, determines the target open field white sentences according to the user portrait in the candidate open field white sentences, and finally initiates the dialogue to the user based on the target open field white sentences. Through the process, the intelligent device can actively talk with the user, and topics during active talk have the following characteristics: on one hand, the news is generated based on news which occurs on the same day, so that the user can obtain the latest information; on the other hand, the user portrait is screened, so that the output topics can catch the interests of the user. Therefore, the intelligent device can actively carry out meaningful conversation with the user, and the enthusiasm of the user for conversation with the intelligent device is fully aroused.

Corresponding to the dialogue method provided above, the embodiment of the present application further provides an intelligent device. Referring to fig. 5, the intelligent device 5 in the embodiment of the present application includes: a memory 501, one or more processors 502 (only one shown in fig. 5), and a computer program stored on the memory 501 and executable on the processors. Wherein: the memory 501 is used for storing software programs and units, and the processor 502 executes various functional applications and diagnoses by running the software programs and units stored in the memory 501, so as to obtain resources corresponding to the preset events. Specifically, the processor 502 realizes the following steps by running the above-described computer program stored in the memory 501:

extracting a user portrait of the user through the trained user portrait model;

Assuming that the above is the first possible implementation manner, in a second possible implementation manner provided on the basis of the first possible implementation manner, the key information includes: entities, events, actors, and victims; for each piece of news that has occurred on the current day, the generating of the candidate open-air sentences according to the key information includes:

Combining the entities, the events, the parties doing the affairs and the parties receiving the affairs of the news to obtain all possible combined results;

and generating candidate open-air sentences relevant to the news according to the combination result and a preset template corresponding to the combination result.

In a third possible embodiment based on the first possible embodiment, the determining a target open field white sentence from the user picture in the open field white sentence candidates includes:

judging whether the user portrait is empty;

if the user portrait is empty, determining the target open field white sentence in the candidate open field white sentences according to hot news;

and if the user image is not empty, determining the target open white sentence in the candidate open white sentences according to the user preference indicated by the user image.

In a fourth possible embodiment based on the third possible embodiment, the determining the target open text in the candidate open text according to the user preference indicated by the user picture includes:

detecting whether the candidate open field white sentences matched with the preference of the user exist or not;

If the candidate open white sentence which is matched with the preference of the user exists, determining the target open white sentence in the candidate open white sentence which is matched with the preference of the user;

and if the candidate open white sentence which is matched with the preference of the user does not exist, determining the target open white sentence in the candidate open white sentences which are in association with the preference of the user.

In a fifth possible implementation manner provided as a basis for the first possible implementation manner, after initiating a dialog to the user based on the target open field statement, the processor 502 further implements the following steps when running the computer program stored in the memory 501:

after receiving a reply of the user to the target open field white sentence, classifying the reply;

if the answer is a positive answer, outputting a summary of the target news, wherein the target news is: news corresponding to the target open field white sentence;

if the answer is a neutral answer, switching the target open-field white sentence, returning to execute the target open-field white sentence based on the switched target open-field white sentence, and initiating a conversation to the user;

If the answer is a negative answer, triggering the chatting system to continue to have a conversation with the user;

if the responses cannot be classified, the responses are analyzed and a dialog with the user is continued based on the analysis result.

In a sixth possible embodiment based on the fifth possible embodiment, the analyzing the response and continuing the dialog with the user based on the analysis result includes:

analyzing the relevance between the target news and the answer;

if the association degree is higher than a preset association degree threshold value, triggering a multi-turn dialogue system to continue dialogue with the user;

and if the relevance is equal to or lower than the relevance threshold, triggering the chatting system to continue the conversation with the user.

In a seventh possible implementation manner provided based on the sixth possible implementation manner, the analyzing a relevance between the target news and the response includes:

determining the degree of association based on the frequency of occurrence of the keywords in the response in the target news;

or,

the association is determined based on the similarity of the reply to the summary of the target news.

It should be understood that in the embodiments of the present Application, the Processor 502 may be a Central Processing Unit (CPU), and the Processor may be other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field-Programmable Gate arrays (FPGAs) or other Programmable logic devices, discrete Gate or transistor logic devices, discrete hardware components, and the like. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like.

Memory 501 may include both read-only memory and random access memory and provides instructions and data to processor 502. Some or all of the memory 501 may also include non-volatile random access memory. For example, the memory 501 may also store information of device classes.

It should be clear to those skilled in the art that, for convenience and simplicity of description, the foregoing division of the functional units and modules is only used for illustration, and in practical applications, the above functions may be distributed as different functional units and modules according to needs, that is, the internal structure of the apparatus may be divided into different functional units or modules to implement all or part of the above described functions. Each functional unit and module in the embodiments may be integrated in one processing unit, or each unit may exist alone physically, or two or more units are integrated in one unit, and the integrated unit may be implemented in a form of hardware, or in a form of software functional unit. In addition, specific names of the functional units and modules are only used for distinguishing one functional unit from another, and are not used for limiting the protection scope of the present application. For the specific working processes of the units and modules in the system, reference may be made to the corresponding processes in the foregoing method embodiments, which are not described herein again.

In the above embodiments, the description of each embodiment has its own emphasis, and reference may be made to the related description of other embodiments for parts that are not described or recited in any embodiment.

Those of ordinary skill in the art will appreciate that the various illustrative elements and algorithm steps described in connection with the embodiments disclosed herein may be implemented as electronic hardware, or combinations of external device software and electronic hardware. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the technical solution. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present application.

In the embodiments provided in the present application, it should be understood that the disclosed apparatus and method may be implemented in other ways. For example, the above-described system embodiments are merely illustrative, and for example, the division of the above modules or units is only one type of logical functional division, and other divisions may be realized in practice, for example, multiple units or components may be combined or integrated into another system, or some features may be omitted, or not executed. In addition, the shown or discussed mutual coupling or direct coupling or communication connection may be through some interfaces, indirect coupling or communication connection of devices or units, and may be in an electrical, mechanical or other form.

The units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one position, or may be distributed on multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of the embodiment.

The integrated unit may be stored in a computer-readable storage medium if it is implemented in the form of a software functional unit and sold or used as a separate product. Based on such understanding, all or part of the flow in the method of the embodiments described above can be realized by a computer program, which can be stored in a computer-readable storage medium and can realize the steps of the embodiments of the methods described above when the computer program is executed by a processor. The computer program includes computer program code, and the computer program code may be in a source code form, an object code form, an executable file or some intermediate form. The computer-readable storage medium may include: any entity or device capable of carrying the above-described computer program code, recording medium, usb disk, removable hard disk, magnetic disk, optical disk, computer readable Memory, Read-Only Memory (ROM), Random Access Memory (RAM), electrical carrier wave signal, telecommunication signal, software distribution medium, etc. It should be noted that the computer readable storage medium may contain other contents which can be appropriately increased or decreased according to the requirements of the legislation and the patent practice in the jurisdiction, for example, in some jurisdictions, the computer readable storage medium does not include an electrical carrier signal and a telecommunication signal according to the legislation and the patent practice.

The above embodiments are only used to illustrate the technical solutions of the present application, and not to limit the same; although the present application has been described in detail with reference to the foregoing embodiments, it should be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; such modifications and substitutions do not substantially depart from the spirit and scope of the embodiments of the present application and are intended to be included within the scope of the present application.

Claims

1. A method of dialogues, comprising:

extracting key information of all news which occur on the same day through a trained news identification model;

extracting a user portrait of the user through the trained user portrait model;

and initiating a conversation to the user based on the target open field white statement.

2. The dialog method of claim 1, wherein the key information comprises: entities, events, actors and acceptors; for each piece of news which occurs on the current day, generating candidate open-field white sentences according to the key information comprises the following steps:

and generating candidate open-scene white sentences related to the news according to the combined result and a preset template corresponding to the combined result.

3. The dialog method of claim 1 wherein said determining a target open field statement from said user representation among said candidate open field statements comprises:

judging whether the user portrait is empty or not;

and if the user portrait is not empty, determining the target open white sentence in the candidate open white sentences according to the user preference indicated by the user portrait.

4. The dialog method of claim 3 wherein said determining said target open white sentence in said candidate open white sentences in accordance with said user's preferences indicated by said user representation comprises:

detecting whether the candidate open-field white sentences matched with the preference of the user exist or not;

if the candidate open white sentences matched with the preference of the user exist, determining the target open white sentences in the candidate open white sentences matched with the preference of the user;

And if the candidate open white sentences matched with the preference of the user do not exist, determining the target open white sentences in the candidate open white sentences which have incidence relation with the preference of the user.

5. The dialog method of claim 1, wherein after said initiating a dialog to the user based on the target open field statement, the dialog method further comprises:

if the answer is a positive answer, outputting an abstract of target news, wherein the target news is as follows: news corresponding to the target open field white sentence;

if the answer is a neutral answer, switching the target open field white sentence, returning to execute the step of initiating a conversation to the user based on the target open field white sentence and the subsequent steps based on the switched target open field white sentence;

if the response cannot be classified, the response is analyzed and a dialog is continued with the user based on the analysis result.

6. The dialog method of claim 5 wherein said analyzing said response and continuing dialog with said user based on the results of the analysis comprises:

analyzing the relevancy of the target news and the reply;

if the association degree is higher than a preset association degree threshold value, triggering a multi-turn dialogue system to continue to have dialogue with the user;

and if the association degree is equal to or lower than the association degree threshold value, triggering a chatting system to continue to have a conversation with the user.

7. The dialog method of claim 6, wherein said analyzing the relevance of the target news to the response comprises:

determining the relevancy based on the frequency of occurrence of keywords in the response in the target news;

or,

determining the relevance based on a similarity of the reply to the summary of the target news.

8. A dialogue apparatus, comprising:

The determining module is used for determining a target open field white sentence according to the user portrait in the candidate open field white sentences;

9. A smart device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that the processor implements the method according to any of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, in which a computer program is stored which, when being executed by a processor, carries out the method according to any one of claims 1 to 7.