CN115408508A - Dialogue recommendation and model training method and device, electronic equipment and storage medium - Google Patents

Abstract

The embodiment of the invention provides a dialogue recommendation and model training method and device, electronic equipment and a storage medium. The conversation recommendation method comprises the following steps: determining a plurality of recommendation elements in the current dialogue statement; querying the incidence relation among the plurality of recommended elements in a pre-constructed recommended object map, wherein the entity of the recommended object map indicates the recommended elements, and the entity relation of the recommended object map indicates the incidence relation among different recommended elements; constructing a preference representation of a current topic of the current dialog statement based on at least the contextual semantic association between the plurality of recommended elements and their associations; predicting a next topic for the current topic based on the preferred representation of the current topic; recommending the recommended object matched with the next theme. The proposal of the embodiment of the invention ensures the accuracy of recommendation and further improves the recommendation efficiency.

Description

Dialogue recommendation and model training method, device, electronic device and storage medium

technical field

The embodiments of the present invention relate to the field of computer technology, and in particular, to a dialogue recommendation and model training method, device, electronic device and storage medium.

Background technique

The Conversational Recommender System (CRS) can realize the recommendation process of recommended objects such as commodities in the process of dialogue with users. Traditional dialogue recommendation tasks often determine user needs through some rules or models, and then analyze user preferences based on user dialogue information, and then recommend recommended objects such as commodities to users.

With the development of the dialogue recommendation system, a dialogue strategy module is introduced, which can make a judgment on topic guidance or direct recommendation according to the current dialogue state with the user, which further improves the fluency of user experience and the accuracy of recommendation. However, the current Dialogue recommendation systems are not accurate enough to estimate user preferences or intentions in the topic guidance process or recommendation process, resulting in room for improvement in recommendation efficiency.

Contents of the invention

In view of this, embodiments of the present invention provide a dialog recommendation and model training method, device, electronic device, and storage medium to at least partially solve the above problems.

According to the first aspect of an embodiment of the present invention, a dialogue recommendation method is provided, including: determining a plurality of recommended elements in the current dialogue sentence; Association relationship, the entity of the recommended object graph indicates the recommended element, and the entity relationship of the recommended object graph indicates the association relationship between different recommended elements; at least based on the contextual semantic association between the plurality of recommended elements and their associations Constructing a preference representation of the current topic of the current dialogue sentence; predicting a next topic of the current topic based on the preference representation of the current topic; recommending a recommendation object matching the next topic.

In another implementation manner of the present invention, the recommending a recommended object that matches the next topic includes: when the next topic matches an object tag, recommending a recommended object to which the object tag belongs.

In another implementation manner of the present invention, the recommending the recommended object to which the object tag belongs includes: determining a plurality of candidate recommended objects with the object tag; sorting the plurality of candidate recommendation objects based on the similarity of the preference representation of the current topic; selecting a recommendation object among the plurality of candidate recommendation objects for recommendation based on the ranking of the plurality of candidate recommendation objects, and The sorting ordinal number of the recommended object among the plurality of candidate recommended objects is less than a preset ordinal number.

In another implementation manner of the present invention, the method further includes: when the next topic does not match the object label, based on the preference representation of the current topic and the next topic, generating the The reply statement for the current conversation statement.

In another implementation manner of the present invention, the constructing the preference representation of the current topic of the current dialogue sentence based at least on the contextual semantic association between the plurality of recommended elements and their associations includes: respectively generating the Multiple initial vector representations of multiple recommended elements; based on the association relationship of the multiple recommended elements, perform contextual semantic processing on the multiple initial vector representations to obtain a first contextual semantic representation of the multiple recommended elements; at least Based on the first contextual semantic representation, a preference representation of the current topic of the current dialogue sentence is constructed.

In another implementation manner of the present invention, based on the association relationship of the plurality of recommended elements, the context semantic processing is performed on the plurality of initial vector representations to obtain the first context semantic representation of the plurality of recommended elements, Including: constructing an initial matrix representation of the plurality of initial vector representations; constructing a relationship matrix and a degree matrix among the plurality of initial vectors; encoding and inputting the initial matrix representation, the relationship matrix and the degree matrix In the graph convolutional network, the first contextual semantic representation of the plurality of recommended elements is obtained, and the graph convolutional network is pre-trained through the initial matrix representation of the graph training data, the relationship matrix, the degree matrix and its classification labels.

In another implementation manner of the present invention, the method further includes: generating respective second contextual semantic representations of at least one of the historical dialog sentences of the current dialog sentence, their historical topics, and dialog user preferences. The constructing the preference representation of the current topic of the current dialogue sentence based at least on the first contextual semantic representation includes: inputting the first contextual semantic representation and each second contextual semantic representation to a pre-trained automatic In the attention layer, the preference representation of the current topic of the current dialog sentence is obtained, and the self-attention layer is used to perform contextual representation based on the dialog topic for the first contextual semantic representation and each second contextual semantic representation .

In another implementation of the present invention, the output of the self-attention layer is connected to the input of the topic guidance layer, and the self-attention layer and the topic guidance layer form a guidance topic guidance model, and the topic guidance model passes The topic-guided relations of different dialogue sentences are trained. The predicting the next topic of the current topic based on the preference representation of the current topic includes: inputting the preference representation of the current topic into a topic guidance layer to obtain the next topic of the current topic.

In another implementation manner of the present invention, the method further includes: generating a reply sentence to the current dialogue sentence based on the preference expression of the current topic and the recommended object.

According to the second aspect of the embodiment of the present invention, there is provided a model training method, including: acquiring a dialogue sentence sample and the guiding theme of the dialogue sentence sample; determining a plurality of recommended elements in the dialogue sentence sample; In the constructed recommendation object graph, the association relationship between the plurality of recommended elements is queried, the entity of the recommendation object graph indicates the recommended element, and the entity relationship of the recommended object graph indicates the association relationship between different recommended elements; at least Based on the contextual semantic association between the plurality of recommended elements and their associations as input, the topic guidance model is trained with the guidance topic as a supervision condition.

According to the third aspect of the embodiment of the present invention, there is provided a dialogue recommendation device, including: a determination module, which determines a plurality of recommendation elements in the current dialogue sentence; a query module, which queries the multiple elements in the pre-built recommendation object map An association relationship between recommended elements, the entity of the recommended object graph indicates the recommended element, and the entity relationship of the recommended object graph indicates the association relationship between different recommended elements; the building module is at least based on the plurality of recommended elements and the recommended elements. The contextual semantic association between its associations constructs the preference representation of the current topic of the current dialogue sentence; the prediction module predicts the next topic of the current topic based on the preference representation of the current topic; the recommendation module recommends and A recommendation object matched by the next topic.

According to a fourth aspect of the embodiments of the present invention, there is provided a model training device, including: an acquisition module, which acquires a dialogue statement sample and the guiding theme of the dialogue statement sample; a determination module, which determines the number of dialogue statement samples recommended elements; the query module, in the pre-built recommended object map, query the association relationship between the plurality of recommended elements, the entity of the recommended object map indicates the recommended element, and the entity relationship of the recommended object map indicates different The association relationship between recommended elements; the training module is at least based on the context semantic association between the plurality of recommended elements and their association relationship as input, and uses the guidance theme as a supervision condition to train the topic guidance model.

According to a fifth aspect of the embodiments of the present invention, there is provided an electronic device, including: a processor, a memory, a communication interface, and a communication bus, and the processor, the memory, and the communication interface complete mutual communication through the communication bus The communication among them; the memory is used to store at least one executable instruction, and the executable instruction causes the processor to execute the operation corresponding to the method described in the first aspect or the second aspect.

According to a sixth aspect of the embodiments of the present invention, there is provided a computer storage medium, on which a computer program is stored, and when the program is executed by a processor, the method as described in the first aspect or the second aspect is implemented.

In the scheme of the embodiment of the present invention, in the pre-built recommendation object map, query the relationship between multiple recommended elements of the current dialogue sentence, and recall more information that is conducive to recommendation, so that the constructed current dialogue sentence The preference representation of the current topic can more accurately reflect the user's preference and intention of the current dialogue statement, and then the preference representation based on the current topic can more accurately predict the next topic, while ensuring the accuracy of the recommendation and further improving the Improve recommendation efficiency.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only These are some embodiments described in the embodiments of the present invention, and those skilled in the art can also obtain other drawings based on these drawings.

Fig. 1 is a schematic block diagram of a dialog recommendation system according to an example.

Fig. 2 is a flowchart of steps of a dialog recommendation method according to an embodiment of the present invention.

Fig. 3 is a flowchart of steps of a model training method according to an embodiment of the present invention.

FIG. 4 is a schematic block diagram of a dialogue recommendation system to which the embodiments of FIG. 2 and FIG. 3 are applied.

Fig. 5 is a structural block diagram of a dialogue recommendation device according to another embodiment of the present invention.

Fig. 6 is a structural block diagram of a model training device according to another embodiment of the present invention.

FIG. 7 is a schematic structural diagram of an electronic device according to another embodiment of the present invention.

Detailed ways

In order to enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and detailedly described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described The embodiments are only some of the embodiments of the present invention, but not all of them. All other embodiments obtained by persons of ordinary skill in the art based on the embodiments in the embodiments of the present invention shall fall within the protection scope of the embodiments of the present invention.

The specific implementation of the embodiments of the present invention will be further described below in conjunction with the accompanying drawings of the embodiments of the present invention.

Fig. 1 is a schematic block diagram of a dialog recommendation system according to an example. The dialogue recommendation system includes a user interface 110 , a dialogue strategy module 120 , a recommendation module 130 , a topic guidance module 140 and a dialogue generation module 150 .

Specifically, the user interface 110 is used to obtain the natural language sentence input by the user, that is, the current dialog sentence, and is also used to give the reply sentence of the dialog recommendation system. The dialog strategy module 120 is used to make a decision on how to reply according to the current dialog sentence, such as whether to continue to inquire about the next topic or make a recommendation directly. The recommendation module 130 is used to give a corresponding recommendation list or a single recommendation result according to user preferences or intentions. The topic guidance module 140 is used to continue the query to the next topic. The dialog generation module 150 is used to generate reply sentences according to the recommendation list or a single recommendation result or the next topic. The current dialogue recommendation system is not accurate enough to estimate user preferences or intentions during the topic guidance process or recommendation process, resulting in room for improvement in recommendation efficiency.

Fig. 2 shows a dialogue recommendation method according to an embodiment of the present invention. The solution of this embodiment can be applied to any suitable electronic device with data processing capabilities, including but not limited to: servers, mobile terminals (such as mobile phones, PADs, etc.), and PCs. For example, in the model training stage of the topic-guided model, a computing device (for example, a data center) configured with a CPU (an example of a processing unit)+GPU (an example of an acceleration unit) architecture can be utilized to decode the encoder based on training samples model for training. Computing devices such as data centers may be deployed in cloud servers such as dedicated clouds, private clouds, or hybrid clouds. Correspondingly, in the inference stage of the dialogue recommendation method, a computing device configured with a CPU (an example of a processing unit)+GPU (an example of an acceleration unit) architecture can also be used to perform inference operations. The dialog recommendation method of this embodiment includes:

S210: Determine multiple recommended elements in the current dialogue sentence.

It should be understood that the multiple recommended elements may be elements associated with candidate recommended objects. When the recommended object is a product, the recommended elements include but are not limited to price, sales volume, brand, store, and popularity. When the recommendation object is a knowledge point, the recommended elements include but are not limited to experts and fields related to the knowledge point, expert-related information such as nationality, gender, age, etc., books, citation resources, etc.

It should also be understood that the multiple recommended elements may be recommended elements matching the keywords of the current dialogue sentence, or recommended elements matching the subject of the current dialogue sentence. When identifying multiple recommended elements that appear in the current dialogue sentence, if the current dialogue sentence is "I want to buy a cheap electronic product", then the keywords matching "electronic product" and "cheap" are recommended elements. In addition, the topic Matching "mobile phones", "bracelets", "VR glasses", and "cost-effective" are also recommended elements.

S220: In the pre-built recommendation object map, query the association relationship between multiple recommendation elements, the entity of the recommendation object map indicates the recommendation element, and the entity relationship of the recommendation object map indicates the association relationship between different recommendation elements.

It should be understood that the recommended object graph can be stored in different ways, and the recommended object graph includes entities and internship relationships. There may or may not be an entity relationship between different entities. For example, "mobile phone" in the above example may have a certain relationship with "value-for-money products", for example, a certain mobile phone brand is famous for value-for-money products. For another example, "mobile phone" and "VR glasses" may not have an association relationship. For another example, there is a certain relationship between "bracelet" and "VR glasses". For example, a wearable device supplier produces or sells "bracelet" and "VR glasses". To sum up, "a certain mobile phone brand is famous for its cost-effective products" and "a certain wearable device supplier produces or sells bracelets and VR glasses" do not exist in the current dialogue sentence, but are related to the intention or Prefer closely related information, that is, more and more implicit intent information or preference information is recalled through the recommended object graph.

S230: Construct a preference representation of the current topic of the current dialogue sentence based at least on the contextual semantic association between the plurality of recommended elements and their associations.

It should be understood that the pre-trained self-attention matrix can be used to represent the contextual semantic association between multiple recommended elements and their associations, and the preference representation of the current topic of the current dialogue sentence can be obtained as a reliable and comprehensive intermediate representation. Such an intermediate Representation is conducive to learning model parameters with stronger generalization ability from supervised conditions.

S240: Based on the preference representation of the current topic, predict the next topic of the current topic.

It should be understood that the next topic may be a subtopic of the current topic, or a topic of the same level as the current topic, but belonging to a different field or subfield. For example, the current topic is "electronic products", and the subordinate topics of the current topic are "mobile phone" or "laptop computer" or "smart bracelet". When accurately specifying the recommended object, it is advantageous to more accurately determine the recommended object when the next topic indicates "mobile phone" or the like. In addition, when the current topic and the next topic belong to the same level, for example, when the current topic is "mobile phone", "mobile phone" is not the main recommended scene due to the limited number of alternative recommendation objects (for example, merchants using dialogue recommendation system Not a mobile phone seller or manufacturer), the next topic is "smart speaker" or "VR glasses", that is, the current topic and the next topic belong to the same level, both belong to "electronic products", and the next topic and the current topic belong to In different product fields, if a merchant using a dialogue recommendation system can sell "smart speakers" or "VR glasses", then "smart speakers" or "VR glasses" can be recommended as recommendation objects. Alternatively, if the merchant using the dialogue recommendation system does not sell "smart speakers" or "VR glasses", the dialogue strategy module (which can call the recommendation module and prediction module in Figure 5) can predict the next topic from the next topic The next topic until the next topic matches the products sold by merchants using the dialogue recommendation system.

S250: Recommend a recommendation object matching the next topic.

It should be understood that the dialogue strategy module may invoke a natural language generation model (for example, GPT or GPT-2) to generate a reply sentence for recommendation or a reply sentence for guiding the next topic.

In the scheme of the embodiment of the present invention, in the pre-built recommendation object map, query the relationship between multiple recommended elements of the current dialogue sentence, and recall more information that is conducive to recommendation, so that the constructed current dialogue sentence The preference representation of the current topic can more accurately reflect the user's preference and intention of the current dialogue statement, and then the preference representation based on the current topic can more accurately predict the next topic, while ensuring the accuracy of the recommendation and further improving the Improve recommendation efficiency.

In some other examples, constructing a preference representation of the current topic of the current dialogue sentence based at least on the contextual semantic association between multiple recommended elements and their associations includes: respectively generating multiple initial vector representations of multiple recommended elements; based on The association relationship of multiple recommended elements, contextual semantic processing is performed on multiple initial vector representations, and the first contextual semantic representation of multiple recommended elements is obtained; at least based on the first contextual semantic representation, a preference representation of the current topic of the current dialogue sentence is constructed.

In some other examples, based on the association relationship of multiple recommended elements, contextual semantic processing is performed on multiple initial vector representations to obtain the first contextual semantic representation of multiple recommended elements, including: constructing an initial matrix representation of multiple initial vector representations ;Construct the relationship matrix and degree matrix between multiple initial vectors; encode the initial matrix representation, relationship matrix and degree matrix into the graph convolution network to obtain the first context semantic representation of multiple recommended elements, graph convolution The network is pre-trained by the initial matrix representation of the graph training data, relation matrix and degree matrix and their classification labels.

In some other examples, the dialog recommendation method further includes: generating each second contextual semantic representation of at least one of the current dialog sentence's historical dialog sentences and their historical topics, and dialog user preferences. Constructing the preference representation of the current topic of the current dialogue sentence based on at least the first context semantic representation, including: inputting the first context semantic representation and each second context semantic representation into a pre-trained self-attention layer to obtain the current dialogue sentence The preference representation of the current topic, the self-attention layer is used for the context representation based on the conversation topic for the first context semantic representation and each second context semantic representation.

In other examples, the output of the self-attention layer is connected to the input of the topic-guiding layer, and the self-attention layer and the topic-guiding layer form a guiding topic-guiding model. The topic-guiding model is obtained by training the topic-guiding relationship of different dialogue sentences; based on the current The preference representation of the topic predicts the next topic of the current topic, including: inputting the preference representation of the current topic into the topic guidance layer to obtain the next topic of the current topic.

Fig. 3 is a flowchart of steps of a model training method according to an embodiment of the present invention. In the model training (training) stage of the topic-guided model, a computing device (for example, a data center) configured with a CPU (an example of a processing unit) + GPU (an example of an acceleration unit) architecture can be used to train the encoder-decoder model based on training samples to train. Computing devices such as data centers may be deployed in cloud servers such as dedicated clouds, private clouds, or hybrid clouds.

The model training method of this embodiment includes:

S310: Acquiring the dialogue statement sample and the guide topic of the dialogue statement sample.

S320: Determine multiple recommended elements in the dialogue sentence sample.

S330: In the pre-built recommendation object map, query the association relationship between multiple recommendation elements, the entity of the recommendation object map indicates the recommendation element, and the entity relationship of the recommendation object map indicates the association relationship between different recommendation elements.

S340: Based on at least the contextual semantic association between multiple recommended elements and their associations as input, and using the guiding topic as a supervision condition, train the topic guiding model.

In the solution of the embodiment of the present invention, in the pre-built recommendation object map, query the relationship between multiple recommended elements of the dialogue statement sample, and recall more information that is conducive to recommendation, so that multiple recommended elements and their The contextual semantic association between association relations can more accurately reflect the user's preferences and intentions of the dialogue sentence samples, so that the topic guidance model obtained after training can perform topic guidance more accurately, thereby improving the recommendation efficiency.

It should be understood that in the end-to-end training phase of the topic-guided model, the input of the topic-guided model includes the contextual semantic association between multiple recommended elements and their associations, and the guiding topic used as the supervision condition is the current topic corresponding to the dialogue sentence sample The next topic of , the next topic can be a subtopic of the current topic, or a topic of the same level as the current topic, but belongs to a different field or subfield.

FIG. 4 is a schematic block diagram of a dialogue recommendation system to which the embodiments of FIG. 2 and FIG. 3 are applied. The dialog recommendation model of Figure 4 includes a graph vector encoding module 410 such as a graph encoder, a text encoding model 420 such as a text vector encoder, a topic guidance model 430, a recommendation module 440, and a natural language generation (Natural Language Generation, NLG) module 450.

The graph encoding module 410 encodes the sub-graph through graph encoders of various graph convolutional networks (Graph Convolutional Networks, GCN) to obtain the first context representation. For example, subgraphs can be queried from the recommended object graph. The recommended object graph as a knowledge graph includes entities and entity relationships. The entities in the recommended object graph indicate recommended elements, and the entity relationships in the recommended object graph indicate the relationship between different recommended elements. There may or may not be an entity relationship between different entities. The recommended object graph can be constructed based on empirical knowledge or data related to recommended objects.

Specifically, each entity in the recommended object graph may be an object related to the recommended object itself or related to the recommended object. For example, when the recommendation object is a product, each entity includes but not limited to logo, brand feature, category feature, store information, popularity index, factory information, price feature, etc. Correspondingly, each entity relationship includes, but is not limited to, product identification, product brand features, product category features, product store information, product popularity indicators, product ex-factory information, product price features, and the like. Generally speaking, if one mentions a product, it is easy to think of a certain theme feature of the product, which means that the product has a high degree of correlation with the theme feature. For example, if one mentions a mobile phone, people easily think of a certain brand and price of the mobile phone, which means that the theme characteristics of the brand and price are closely related to the mobile phone.

Each entity in the sub-graph is a subset of each entity in the recommended object graph, and can be queried in the recommended object graph through the graph query module 415 . For example, firstly, multiple recommended elements appearing in the current dialogue sentence are identified. For example, if the current dialogue sentence is "I want to buy a cheap electronic product", then "electronic product" and "cheap" are recommended elements. Correspondingly, in the graph of recommended objects, query entities that match (keyword match or topic match) with each recommended element.

For example, if "electronic product" is in the entity of the recommended object graph, it belongs to keyword matching, and accordingly, the entity "electronic product" is extracted from the recommended object graph. Alternatively, if "electronic product" is not in the entity of the recommended object graph, and "mobile phone" is in the entity of the recommended object graph, because "mobile phone" belongs to "electronic product", it belongs to the topic matching, correspondingly, from the recommended object In the map, extract the entity "mobile phone". Alternatively, if both "electronic product" and "mobile phone" are entities in the recommended object graph, then "electronic product" and "mobile phone" are extracted from the recommended object graph.

It should be understood that multiple recommended elements may be determined from all entities in the recommended object graph, and the multiple recommended elements may be more than the number of recommended elements identified from the current dialogue sentence.

Then, from the recommended object graph, the association relationship between multiple recommended elements is obtained (recalled). For example, "I want to buy a cheap electronic product" does not include a specific mobile phone brand, but the "mobile phone" as an entity and the "cheap" "brand" related to "mobile phone" can be obtained from the recommended object map, Therefore, more intention preference information than the current dialogue sentence itself is improved through the recommended object map, and more intention preference information is conducive to more accurate recommendations.

In addition, the text encoding model 420 uses a text encoder such as a BERT encoder to encode at least one of the historical dialogue sentences of the current dialogue statement, its historical topics, and dialogue user preferences to obtain respective second context representations. For example, input the historical dialog sentences into the BERT encoder to obtain the second context representation of the historical dialog sentences; input the dialog user preferences into the BERT encoder to obtain the second context representation of the dialog user preferences; input the historical topics of the dialog user preferences to the BERT encoder to obtain a second contextual representation of historical topics.

In a dialog example, a GCN encoder as an example of a graph encoder can be encoded using (Equation 1):

其中，A为邻接矩阵，D为节点的度的矩阵，

I_N是N阶单位矩阵，

是

对应的度矩阵，

是可学习的线性转换器。应理解，GCN编码器的线性转换器采用训练样本训练得到。

Among them, A is the adjacency matrix, D is the matrix of the degree of the node,

I _N is an identity matrix of order N,

Yes

The corresponding degree matrix,

is a learnable linear converter. It should be understood that the linear converter of the GCN encoder is trained using training samples.

Further, the first contextual representation of the graph coding module 410 and the respective second contextual representations of the textual coding module 420 are input to the topic guidance model 430 . The topic guidance model 430 may be a classification model obtained through end-to-end training. For example, the topic guidance model 430 includes a self-attention layer and a topic guidance layer, and the self-attention layer is used to concatenate the first context representation and each second context representation to obtain a concatenation matrix representing the concatenation context representation (an example of the current topic preference representation ), and then, based on the pre-trained self-attention matrix, contextual semantic processing is performed on the concatenation matrix to obtain a more reliable semantic intermediate representation. Each element in the self-attention matrix is a weight value, and the weight value of each element represents each Semantic correlation between two characters corresponding to elements.

Then, the concatenated matrix is processed by a topic-guided layer such as a decoder to get the next topic.

Judging by the dialog strategy module whether there is an object tag matching the next topic in the object tag collection, if there is, then in the recommendation module 440, the recommended object described in the object tag is obtained, and then the recommended object (for example, the recommended object's text matrix or text vector) and the splicing matrix of preference representation (for example, splicing the two again), input in the natural language generation module 450, obtain the reply sentence (recommended sentence) of current dialogue sentence, the natural language generation module 450 here It can be implemented with a pre-trained model such as GPT or GPT-2.

If there is no object label matching the next topic in the object label set, the next topic (e.g., the text matrix or text vector of the next topic) and the concatenation matrix of the preference representation (e.g., concatenate the two again) , input into the natural language generation module 450 to obtain the reply sentence of the current dialog sentence.

Further, the recommendation module 440 can determine multiple candidate object tags to which the corresponding tag belongs in the set of candidate recommended objects, and then, based on the similarity between the multiple candidate recommended objects and the preference representation of the current topic, The recommended objects are sorted, and then, based on the sorting of multiple candidate recommended objects, the recommended objects among the multiple candidate recommended objects are selected for recommendation. , to recommend the first candidate recommendation object, so as to make recommendations more accurately by sorting based on similarity.

In a more general case, as an example of recommending a recommendation object that matches the next topic, when the next topic matches the object tag, the recommendation object to which the object tag belongs can be recommended, and the next topic can be the current topic The sub-topics are recommended when the tags of the sub-topics match the object tags, which improves the efficiency of recommendation.

Alternatively, when the next topic does not match the object label, a reply sentence to the current dialogue sentence is generated based on the preference representation of the current topic and the next topic. For example, if the next topic is a subordinate topic of the current topic, it means that the label of the subordinate topic still does not match the object label. Object tags match. For another example, if the next topic is at the same level as the current topic, and the next topic and the current topic belong to different fields or subfields, then the direction of the conversation topic is changed through the next topic, which is more conducive to matching with an object label .

It should be understood that when constructing the object label set, the object labels of all candidate recommended objects in the candidate recommended object set can be marked first, each candidate recommended object can include at least one object label, and then all the candidate recommended objects can be counted All object labels of , remove duplicate object labels, and obtain a set of object labels.

Further, Fig. 5 shows a dialog recommendation device according to another embodiment of the present invention. The dialogue recommendation device in this embodiment corresponds to the dialogue recommendation method in Figure 2, including:

A determining module 510, determining multiple recommended elements in the current dialogue sentence.

The query module 520, in the pre-built recommended object graph, queries the association relationship between the multiple recommended elements, the entity of the recommended object graph indicates the recommended element, and the entity relationship of the recommended object graph indicates the relationship between different recommended elements. relationship between.

The construction module 530 constructs a preference representation of the current topic of the current dialogue sentence based at least on the contextual semantic association between the plurality of recommended elements and their associations.

Prediction module 540, based on the preference representation of the current topic, predict the next topic of the current topic;

A recommending module 550, recommending recommended objects matching the next topic.

In the scheme of the embodiment of the present invention, in the pre-built recommendation object map, query the relationship between multiple recommended elements of the current dialogue sentence, and recall more information that is conducive to recommendation, so that the constructed current dialogue sentence The preference representation of the current topic can more accurately reflect the user's preference and intention of the current dialogue statement, and then the preference representation based on the current topic can more accurately predict the next topic, while ensuring the accuracy of the recommendation and further improving the Improve recommendation efficiency.

In some other examples, the recommendation module is specifically configured to: when the next topic matches the object tag, recommend the recommended object to which the object tag belongs.

In some other examples, the recommendation module is specifically configured to: determine a plurality of candidate recommended objects with the object label; Sorting a plurality of candidate recommendation objects; based on the ranking of the plurality of candidate recommendation objects, selecting a recommendation object among the plurality of candidate recommendation objects for recommendation, and the recommendation object among the plurality of candidate recommendation objects The sort ordinal is less than the default ordinal.

In some other examples, the dialog recommendation device further includes: a dialog generation module that generates the The reply statement for the current conversation statement.

In some other examples, the construction module is specifically configured to: respectively generate multiple initial vector representations of the multiple recommended elements; perform context semantic processing on the multiple initial vector representations based on the association relationship of the multiple recommended elements , obtaining a first contextual semantic representation of the plurality of recommended elements; constructing a preference representation of a current topic of the current dialogue sentence based at least on the first contextual semantic representation.

In some other examples, the construction module is specifically configured to: construct an initial matrix representation represented by the multiple initial vectors; construct a relationship matrix and a degree matrix among the multiple initial vectors; construct the initial matrix representation, the The relationship matrix and the degree matrix are encoded and input into the graph convolutional network to obtain the first contextual semantic representation of the plurality of recommended elements. The graph convolutional network uses the initial matrix representation of the graph training data, the relationship matrix and the The matrix and its classification labels are pre-trained.

In some other examples, the building module is further configured to: generate each second contextual semantic representation of at least one of the historical dialog sentences of the current dialog sentence, their historical topics, and dialog user preferences. The building module is specifically used for: inputting the first contextual semantic representation and the second contextual semantic representations into a pre-trained self-attention layer to obtain the preference representation of the current topic of the current dialog statement, and the self-attention layer The attention layer is used to perform contextual representation based on dialogue topics for the first contextual semantic representation and each of the second contextual semantic representations.

In some other examples, the output of the self-attention layer is connected to the input of the topic guidance layer, and the self-attention layer and the topic guidance layer form a guidance topic guidance model, and the topic guidance model passes through different dialogue sentences. Topic-guided relationship training is obtained. The prediction module is specifically configured to: input the preference representation of the current topic to the topic guidance layer to obtain the next topic of the current topic.

In some other examples, the dialogue generation module further includes: generating a reply statement to the current dialogue statement based on the preference expression of the current topic and the recommended object.

The device of this embodiment is used to implement the corresponding methods in the foregoing multiple method embodiments, and has the beneficial effects of the corresponding method embodiments, so details are not repeated here. In addition, for the function implementation of each module in the apparatus of this embodiment, reference may be made to the description of corresponding parts in the foregoing method embodiments, and details are not repeated here.

Fig. 6 is a structural block diagram of a model training device according to another embodiment of the present invention.

The model training device of this embodiment corresponds to the model training method in Figure 3, including:

Obtaining module 610, acquiring a dialogue statement sample and a leading topic of the dialogue statement sample.

The determination module 620 is to determine a plurality of recommended elements in the dialogue sentence sample.

The query module 630, in the pre-built recommended object graph, queries the association relationship between the multiple recommended elements, the entity of the recommended object graph indicates the recommended element, and the entity relationship of the recommended object graph indicates the relationship between different recommended elements relationship between.

The training module 640 is at least based on the contextual semantic association between the plurality of recommended elements and their associations as input, and uses the guiding topic as a supervision condition to train the topic guidance model.

In the solution of the embodiment of the present invention, in the pre-built recommendation object map, query the relationship between multiple recommended elements of the dialogue statement sample, and recall more information that is conducive to recommendation, so that multiple recommended elements and their The contextual semantic association between association relations can more accurately reflect the user's preferences and intentions of the dialogue sentence samples, so that the topic guidance model obtained after training can perform topic guidance more accurately, thereby improving the recommendation efficiency.

Referring to FIG. 7 , it shows a schematic structural diagram of an electronic device according to another embodiment of the present invention. The specific embodiment of the present invention does not limit the specific implementation of the electronic device.

As shown in FIG. 7 , the electronic device may include: a processor (processor) 702 , a communication interface (Communications Interface) 704 , a memory (memory) 706 storing a program 710 , and a communication bus 708 .

The processor, the communication interface, and the memory communicate with each other through the communication bus.

The communication interface is used for communicating with other electronic devices or servers.

The processor is configured to execute the program, and specifically may execute the relevant steps in the foregoing method embodiments.

Specifically, the program may include program code, and the program code includes at least one executable instruction.

The processor may be a CPU, or an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of the present invention. The one or more processors included in the smart device may be of the same type, such as one or more CPUs, or may be different types of processors, such as one or more CPUs and one or more ASICs.

Memory for storing programs. The memory may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), such as at least one disk memory.

Specifically, the program can be used to make the processor execute the following dialog recommendation method in FIG. 2 or the model training method in FIG. 3 .

In addition, for the specific implementation of each step in the program, refer to the corresponding descriptions in the corresponding steps and units in the above method embodiments, and details are not repeated here. Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working process of the above-described devices and modules can refer to the corresponding process description in the foregoing method embodiments, and details are not repeated here.

It should be pointed out that, according to implementation requirements, each component/step described in the embodiment of the present invention can be divided into more components/steps, and two or more components/steps or partial operations of components/steps can also be combined into New components/steps to achieve the purpose of the embodiments of the present invention.

The above method according to the embodiment of the present invention can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as CD ROM, RAM, floppy disk, hard disk, or magneto-optical disk), or implemented by Computer code downloaded from a network that is originally stored on a remote recording medium or a non-transitory machine-readable medium and will be stored on a local recording medium so that the methods described herein can be stored on a computer code using a general-purpose computer, a dedicated processor, or a programmable Such software processing on a recording medium of dedicated hardware such as ASIC or FPGA. It will be appreciated that a computer, processor, microprocessor controller, or programmable hardware includes memory components (e.g., RAM, ROM, flash memory, etc.) that can store or receive software or computer code that, when When accessed and executed by a processor or hardware, implements the methods described herein. Furthermore, when a general purpose computer accesses code for implementing the methods shown herein, execution of the code transforms the general purpose computer into a special purpose computer for performing the methods shown herein.

Those skilled in the art can appreciate that the units and method steps of the examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed by hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art may use different methods to implement the described functions for each specific application, but such implementation should not be regarded as exceeding the scope of the embodiments of the present invention.

The above embodiments are only used to illustrate the embodiments of the present invention, rather than to limit the embodiments of the present invention. Those of ordinary skill in the relevant technical field can also make various Changes and modifications, so all equivalent technical solutions also belong to the category of the embodiments of the present invention, and the patent protection scope of the embodiments of the present invention should be defined by the claims.

Claims (14)

1. A conversation recommendation method comprising:

determining a plurality of recommendation elements in the current dialogue statement;

querying the incidence relation among the plurality of recommended elements in a pre-constructed recommended object map, wherein the entity of the recommended object map indicates the recommended elements, and the entity relation of the recommended object map indicates the incidence relation among different recommended elements;

constructing a preference representation of a current topic of the current dialog statement based on at least the contextual semantic association between the plurality of recommended elements and their associations;

predicting a next topic for the current topic based on the preferred representation of the current topic;

recommending the recommended object matched with the next theme.

2. The method of claim 1, wherein the recommending the recommended object that matches the next topic comprises:

and recommending the recommended object to which the object tag belongs when the next theme is matched with the object tag.

3. The method of claim 2, wherein the recommending the recommended object to which the object tag belongs comprises:

determining a plurality of candidate recommended objects with the object tags;

ranking the plurality of candidate recommended objects based on similarity of the plurality of candidate recommended objects to the preference representation of the current topic;

and selecting a recommended object from the multiple candidate recommended objects for recommendation based on the sequence of the multiple candidate recommended objects, wherein the sequence ordinal of the recommended object in the multiple candidate recommended objects is smaller than a preset ordinal.

4. The method of claim 2, wherein the method further comprises:

and when the next theme is not matched with the object label, generating a reply sentence of the current conversation sentence based on the preference representation of the current theme and the next theme.

5. The method of claim 1, wherein said constructing a preferred representation of a current topic of the current conversational sentence based at least on contextual semantic associations between the plurality of recommended elements and their associations comprises:

generating a plurality of initial vector representations of the plurality of recommended elements, respectively;

performing context semantic processing on the plurality of initial vector representations based on the association relation of the plurality of recommended elements to obtain a first context semantic representation of the plurality of recommended elements;

constructing a preference representation of a current topic of the current conversational sentence based on at least the first contextual semantic representation.

6. The method of claim 5, wherein the context semantic processing the initial vector representations based on the association of the recommended elements to obtain a first context semantic representation of the recommended elements comprises:

constructing an initial matrix representation of the plurality of initial vector representations;

constructing a relation matrix and a degree matrix among the plurality of initial vectors;

and coding and inputting the initial matrix representation, the relation matrix and the degree matrix into a graph convolution network to obtain a first context semantic representation of the plurality of recommended elements, wherein the graph convolution network is obtained by pre-training the initial matrix representation, the relation matrix, the degree matrix and classification labels of graph training data.

7. The method of claim 5, wherein the method further comprises:

generating respective second contextual semantic representations of at least one of historical conversation sentences and historical topics thereof, and conversational user preferences of the current conversation sentence;

the constructing of a preference representation of a current topic of the current conversational sentence based on at least the first contextual semantic representation comprises:

and inputting the first context semantic representation and each second context semantic representation into a pre-trained self-attention layer to obtain preference representation of the current topic of the current dialogue statement, wherein the self-attention layer is used for performing context representation based on the dialogue topic on the first context semantic representation and each second context semantic representation.

8. The method of claim 7, wherein the output of the self-attention layer is connected with the input of a theme guide layer, the self-attention layer and the theme guide layer form a guide theme guide model, and the theme guide model is obtained by training theme guide relations of different dialog sentences;

the predicting a next topic for the current topic based on the preferred representation of the current topic comprises:

and inputting the preference representation of the current theme into a theme guide layer to obtain a next theme of the current theme.

9. The method of claim 7, wherein the method further comprises:

and generating a reply sentence of the current dialog sentence based on the preference representation of the current theme and the recommended object.

10. A model training method, comprising:

obtaining a conversation statement sample and a guide theme of the conversation statement sample;

determining a plurality of recommendation elements in the dialog sentence sample;

querying the incidence relation among the plurality of recommended elements in a pre-constructed recommended object map, wherein the entity of the recommended object map indicates the recommended elements, and the entity relation of the recommended object map indicates the incidence relation among different recommended elements;

and training a theme guide model at least based on the context semantic association among the plurality of recommended elements and the association relation thereof as input and the guide theme as a supervision condition.

11. A conversation recommendation apparatus comprising:

the determining module is used for determining a plurality of recommended elements in the current conversation sentence;

the query module is used for querying the incidence relation among the plurality of recommended elements in a pre-constructed recommended object map, wherein the entity of the recommended object map indicates the recommended elements, and the entity relation of the recommended object map indicates the incidence relation among different recommended elements;

the building module is used for building preference expression of the current theme of the current dialogue statement at least based on the context semantic association among the plurality of recommended elements and the association relation thereof;

a prediction module to predict a next topic for the current topic based on the preferred representation of the current topic;

and the recommending module recommends the recommending object matched with the next theme.

12. A model training apparatus comprising:

the acquisition module acquires a conversation statement sample and a guide theme of the conversation statement sample;

the determining module is used for determining a plurality of recommended elements in the dialogue statement sample;

the query module is used for querying the incidence relation among the plurality of recommended elements in a pre-constructed recommended object map, wherein the entity of the recommended object map indicates the recommended elements, and the entity relation of the recommended object map indicates the incidence relation among different recommended elements;

and the training module is used for training the theme guide model at least based on the context semantic association among the plurality of recommended elements and the association relations thereof as input and the guide theme as a supervision condition.

13. An electronic device, comprising: the system comprises a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface complete mutual communication through the communication bus; the memory is used for storing at least one executable instruction, and the executable instruction causes the processor to execute the corresponding operation of the method according to any one of claims 1-10.

14. A computer storage medium having stored thereon a computer program which, when executed by a processor, carries out the method of any one of claims 1-10.

Images