WO2021023249A1

WO2021023249A1 - Generation of recommendation reason

Info

Publication number: WO2021023249A1
Application number: PCT/CN2020/107285
Authority: WO
Inventors: 王金刚; 王倩舒; 兰田; 陆源源; 张富峥; 王仲远
Original assignee: 北京三快在线科技有限公司
Priority date: 2019-08-06
Filing date: 2020-08-06
Publication date: 2021-02-11
Also published as: CN110532463A

Abstract

A recommendation reason generation apparatus (100) capable of generating a recommendation reason for personalized query. The recommendation reason generation apparatus (100) comprises: a query encoder (101) used for encoding the query information of a user into a query vector; a selective encoder (102) used for encoding the comment of the user into a comment vector selective hiding state according to the query vector; and a decoder (103) constituting an encoder-decoder structure together with the selective encoder (102) and used for decoding the comment vector selective hiding state into a recommendation reason. The selective encoder (102) comprises a comment encoding unit (1021) and a selective computing network (1022); the comment encoding unit (1021) is used for encoding the comment of the user into a comment vector intermediate hiding state; the selective computing network (1022) is used for performing computing on the query vector and the comment vector intermediate hiding state to generate the comment vector selective hiding state.

Description

Recommendation reason generation

Cross references to related applications

This patent application claims the priority of the Chinese patent application with application number 2019107204036 filed on August 6, 2019, and the full text of the application is incorporated herein by reference.

Technical field

The present disclosure relates to the field of recommendation reason generation, in particular to an artificial intelligence-based recommendation reason generation device and method, computer-readable storage media, and electronic equipment.

Background technique

The reason for recommendation is a natural language text displayed to the user on a search result page or a discovery page (such as a scene decision, a must-eat list, etc.) for highlight recommendation to assist the user in decision-making, which is usually a short sentence. For example, if you enter "steak" on the search page of a review website APP, multiple businesses related to "steak" will be listed on the search results page, and there is usually a short recommendation reason in the brief introduction of each business, for example, "Selected filet steak is very generous", "Selected sirloin wagyu steak is fresh and delicious, very tender". For another example, there are many must-eat restaurants recommended on the "must-eat list" of the review website APP, and the brief introduction of each recommended restaurant has a recommendation reason called "must-eat reason", for example, "Broiled pork is great , The service attitude is very good, very good", "Roast duck is well-deserved, not only the duck skin is crispy, but the duck meat is very sticky."

The reason for recommendation in the form of natural language can be seen as a high degree of concentration of real user reviews, explaining the recall results for users, digging out merchant characteristics, attracting users to click, and providing scene-oriented guidance to users, enhancing user experience and trust in the platform .

Summary of the invention

According to one aspect of the present disclosure, a recommendation reason generation device is provided, which includes: a query encoder for encoding user query information into a query vector; and a selective encoder for encoding user comments into a query vector based on the query vector The selective hidden state of the review vector; and a decoder, which forms an encoder-decoder structure with the selective encoder for decoding the selective hidden state of the review vector into a recommendation reason, wherein the selective encoder It includes a comment coding unit and a selective calculation network. The comment coding unit is used to code the user comment into a comment vector intermediate implicit state, and the selective calculation network is used to divide the query vector with the comment vector. The hidden state is operated to generate the selective hidden state of the comment vector.

According to another aspect of the present disclosure, a method for generating a recommendation reason is provided, including: encoding user query information as a query vector; encoding user comments as an intermediate hidden state of the comment vector; and combining the query vector with the comment vector The intermediate hidden state is operated to generate a selective hidden state of the review vector; and the selective hidden state of the review vector is decoded as a recommendation reason.

According to another aspect of the present disclosure, there is provided a computer-readable storage medium having a computer program stored thereon, and the program is executed by a processor to implement the above method.

According to another aspect of the present disclosure, there is provided an electronic device including a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor implements the above method when executed.

According to the above-mentioned aspects of the present disclosure, the user's query information is taken into consideration when the user comments are encoded, so it is possible to generate a personalized recommendation reason for the query, thereby improving user experience.

Description of the drawings

These and/or other aspects, features and advantages of the present disclosure will become clearer and easier to understand through the following description of the embodiments in conjunction with the accompanying drawings, in which:

Fig. 1 shows a schematic block diagram of a recommendation reason generating device according to an embodiment of the present disclosure;

FIG. 2 shows a schematic diagram of an operation example of a recommendation reason generating device according to an embodiment of the present disclosure;

FIG. 3 shows a schematic flowchart of a method for generating a recommendation reason according to an embodiment of the present disclosure;

Fig. 4 shows a schematic structural diagram of an electronic device for generating a recommendation reason according to an embodiment of the present disclosure.

detailed description

The present disclosure will be described in detail below with reference to exemplary embodiments of the present disclosure. However, the present disclosure is not limited to the embodiments described herein, and it can be implemented in many different forms. The described embodiments are only used to make the present disclosure thorough and complete, and fully convey the concept of the present disclosure to those skilled in the art. The features of the various described embodiments can be combined or replaced with each other, unless explicitly excluded or should be excluded according to the context.

Recommendation reasons in natural language can enhance user experience and trust in the platform. At present, a method for generating a recommendation reason is specifically: generating a one-sentence recommendation reason based on the high-quality user reviews of the POI (Point of Interest). The wording of the generated recommendation reason does not need to be strictly extracted from the original text, but is based on Artificial intelligence technology encodes the original text through an encoder (Encoder) and then generates it through a decoder (Decoder).

However, the generation method does not take into account the user's real-time intentions, and generates recommendation reasons that are irrelevant to the user's query, resulting in uneven generation results and reducing user experience. For example, if you search for "Take a baby to eat" in a gourmet app, the app will recommend a number of restaurants. However, the reason for the recommendation of these restaurants may not be related to "Take a baby to eat", but is only a general recommendation about the restaurant, such as "Brown" The meat tastes good" because the generation method does not consider the information queried by the user when generating the recommendation reason.

The present disclosure provides an artificial intelligence-based recommendation reason generating device and method, which can generate a personalized recommendation reason for query. The recommendation reason generation method proposed in the present disclosure is an improvement of the above-mentioned generative method, which can generate personalized recommendation reasons for recalled POIs according to different user query information, which is more intelligent and improves user experience. In the present disclosure, "query" includes any behavior of the user searching for POIs, so "query information" includes any information input by the user for searching for POIs, such as search information such as text or voice input by the user (for example, "steak ", "Take a baby to eat"), but also include user-selected themes (such as "ramen"), etc. In the present disclosure, a business of interest (POI) includes any type of business that can be searched, such as restaurants, swimming pools, movie theaters, etc.

In recent years, with the large-scale application of artificial intelligence deep learning methods in the field of natural language processing, the sequence-to-sequence (Seq2Seq) model based on the encoder-decoder structure has achieved good results in the field of text generation . Recommendation reason generation can be regarded as a generative text summary task, which is a kind of text generation task. The recommendation reason generation device and method of the embodiments of the present disclosure are implemented based on the sequence-to-sequence model of the encoder-decoder structure of artificial intelligence.

Fig. 1 shows a schematic block diagram of a recommendation reason generating device 100 according to an embodiment of the present disclosure. As shown in FIG. 1, the recommendation reason generating device 100 includes a query encoder 101, a selective encoder 102 and a decoder 103. The query encoder 101 is used to encode the user's query information into a query vector. The selective encoder 102 is configured to encode user comments into a comment vector selective hidden state based on the query vector. The decoder 103 and the selective encoder 102 form an encoder-decoder structure for decoding the selective hidden state of the comment vector into a recommendation reason. The selective encoder 102 includes a comment encoding unit 1021 and a selective calculation network 1022. The comment encoding unit 1021 is configured to encode the user comment into an intermediate hidden state of the comment vector. The selective calculation network 1022 is configured to perform operations on the query vector and the intermediate hidden state of the review vector to generate the selective hidden state of the review vector, so that the recommendation reason includes content related to the query information. In some embodiments, the decoder 103 may also perform decoding based on the query vector generated by the query encoder 101.

The recommendation reason generating device 100 of the embodiment of the present disclosure is implemented based on an encoder-decoder structure of artificial intelligence technology. In the encoder-decoder structure, the encoder encodes the source text into a vector, and the decoder decodes the vector generated by the encoder into the target text. In the embodiment of the present disclosure, a recommendation reason needs to be generated based on user comments. Therefore, the user comment is the source text in the encoder-decoder structure, and the recommendation reason is the target text. In the embodiment of the present disclosure, the selective encoder 102 is used when encoding user comments. When encoding user comments into a comment vector, it is not only based on the input user comment itself, but also based on the query vector generated by the query encoder 101 , Where the query vector corresponds to the user’s query information. The information may be filtered based on the query vector when the selective encoder 102 encodes the user comment, so that the encoded comment vector contains information related to the user's query. Therefore, the corresponding recommendation reason after decoding by the decoder 103 based on the comment vector can reflect the user's query information, that is, the generated recommendation reason is the query personalized recommendation reason. In some embodiments, the decoder 103 decodes based on the query vector, so that the decoded recommendation reason better reflects the user's query information. For example, the decoder 103 adopts an attention mechanism, and uses the query vector as an input parameter of the context vector generation function when generating a context vector based on the attention mechanism.

In the embodiments of the present disclosure, the query encoder 101 can use any existing or future-developed encoder technology in the field of artificial intelligence to encode the user's query information into a query vector, for example, using one-way or two-way long short-term memory (Long Short-term Memory). -Term Memory, LSTM) network, Word2Vec and other word vector tools. As mentioned above, the "query information" here means any information entered by the user to search for POIs. When the user uses non-text input such as voice, for example, the non-text information can be recognized as text information first, and then vector encoding is performed.

The structure of the selective encoder 102-decoder 103 according to the embodiments of the present disclosure may be based on any existing or future neural network encoder-decoder model developed in the field of artificial intelligence, for example, a pointer generator (Pointer-Generator) Network, Recurrent Neural Network-Long Short-Term Memory (RNN-LSTM) network, ConS2S (sequence to sequence framework based on Convolutional Neural Network), Transformer network, etc. The selective encoder 102 may include a comment coding unit 1021 and a selective calculation network 1022 to encode user comments into a selective hidden state of the comment vector filtered by the query vector, so that the reason for recommendation decoded by the decoder 103 contains the same Describe the content related to the query information. Specifically, the comment coding unit 1021 may adopt any existing or future general-purpose encoder technology in the field of artificial intelligence, for example, a one-way or two-way LSTM network, to encode user comments into the intermediate hidden state of the comment vector, and the intermediate hidden state is the same as The query vector is irrelevant. The selective calculation network 1022 calculates the query vector and the intermediate hidden state of the review vector, so that the generated selective hidden state of the review vector contains query vector information, so that the recommendation reason includes information related to the query information. content. The selective computing network 1022 may be a Multilayer Perceptron (MLP) network, for example. In the present disclosure, "selective hidden state" means that the hidden state contains query vector information, and the query vector information has a selective filtering effect on the content of user comments.

In the following, taking the Pointer-Generator network as an example, the recommendation reason generating device 100 of the present disclosure will be described in detail. FIG. 2 shows a schematic diagram of an operation example of the recommendation reason generation device 100 according to an embodiment of the present disclosure. As shown in FIG. 2, the query encoder 101 encodes the user's query information into a query vector q* through, for example, a two-way LSTM, to be input to the selective encoder 102 and the decoder 103. It should be noted that, in some embodiments, the query vector q* is only used in the selective encoder 102 and not in the decoder 103.

The selective encoder 102 uses the query vector q* as a filter parameter to divide the constituent words x ₁ , x ₂ , x ₃ , x ₄ , x ₅ , and x _{6 of the} user review (in the example of FIG. 2, the 6 constituent words are For example, the actual user comment can have more or less than 6 constituent words) coded as a comment vector. Selective hidden states h ₁ ', h ₂ ', h ₃ ', h ₄ ', h ₅ ', h ₆ '( collectively referred to as h _i ') for output to the decoder 103 for decoding. In other words, the encoder 102 may be based on selective query vector q * selectivity hidden state h _i '. Specifically, as shown in Fig. 2, the comment coding unit 1021 of the selective encoder 102 first generates the intermediate hidden states h ₁ , h ₂ , h ₃ , h ₄ , h ₅ , h ₆ (collectively referred to as h _i ), then, the selective calculation network 1022 of the selective encoder 102 generates a comment vector selection for output to the decoder 103 based on the intermediate hidden state h _{i of the} comment vector and the query vector q* Sexual recessive state h _i '. For example, the comment encoding unit 1021 may encode the input user comments through a conventional two-way LSTM to generate intermediate implicit words corresponding to each constituent word x ₁ , x ₂ , x ₃ , x ₄ , x ₅ , and x ₆ in the user comment sequence. The state h ₁ , h ₂ , h ₃ , h ₄ , h ₅ , h ₆ , and then, the selective computing network 1022 can combine the intermediate hidden states h ₁ , h ₂ , h ₃ , h ₄ , h ₅ , h ₆ and query The vector q* is operated to generate selective hidden states h ₁ ', h ₂ ', h ₃ ', h ₄ ', h ₅ ', h ₆ 'in consideration of the query vector q*. In the embodiment of the present disclosure, the selective computing network 1022 may adopt a multilayer perceptron (MLP), as shown in FIG. 2.

As an example, network 1022 may calculate the selectivity based on the intermediate hidden state and query vector h _i q * and selectivity weighting vector g _i, h _i and an intermediate hidden state with a selective vector elements g _i is multiplied by the selectivity The hidden state h _i ', for example, is shown in the following equations (1) and (2):

g _i =σ(W _h h _i +W _q q*+b) (1)

h _i '=g _i ⊙h _i (2)

Wherein, W _h, W _q h _i are weights and q *, b is the offset value, W _h, W _q and b are parameters can learn, σ represents a sigmoid activation function, ⊙ denotes element-multiplication.

In some embodiments, the comment coding unit 1021 encodes the user comments by LSTM and other methods to form an entire sequence vector h*, for example, it may be h ₁ or h ₆ in FIG. 2. The selective calculation network 1022 is generating selective implicit the user also reviews the state of the entire vector sequence h * calculates the intermediate hidden state h _i, for example, the user reviews the entire vector sequence h * and the arithmetic or weighted splicing intermediate hidden state h _i. As an example, in the case of the entire sequence vector h* based on user comments, the above formula (1) can be transformed into formula (3) or formula (4):

g _i =σ(W _h h _i +W _q q*+W _c h*+b) (3)

Or, g _i =σ(W _h [h _i ; h*]+W _q q*+b) (4)

Among them, g _i in formula (3) represents the weighted sum of h*, h _i and q*, and W _c represents the learnable weight of the entire sequence vector h*; g _i in formula (4) represents h* Spliced with h _i .

The selective hidden states h ₁ ′, h ₂ ′, h ₃ ′, h ₄ ′, h ₅ ′, and h ₆ ′ output by the selective computing network 1022 in the selective encoder 102 are used by the decoder 103 for decoding. The decoder 103 can adopt any conventional decoder with or without an attention mechanism. In the example shown in Figure 2, the decoder 103 uses an attention mechanism, and the decoder 103 is generating context vectors C1, C2, C3, C4 based on the attention mechanism (collectively referred to as Ct, in Figure 2 with 4 context vectors For example), the query vector q* is used as the input parameter of the context vector generation function. This function can use Multilayer perceptron (MLP). At this time, the context vector generation function generates the context vector using MLP, as shown in Figure 2. Shown. For example, decoder 103 may perform the calculation in the query vector q C _t * and Exposition selectively hidden state vector h _i and the weighting operation, each of the weights' weight to be learning parameters. In some embodiments, the input parameters of the context vector generation function based on the attention mechanism may also include the entire sequence vector h* of user comments. For example, the entire sequence vector h can be executed when calculating the context vector C _t based on the attention mechanism. *With the query vector q* and the comment vector selective hidden state h _i'weighted sum operation, or concatenate the entire sequence vector h* with the comment vector selective hidden state h _t '. According to an embodiment of the present disclosure, C _t is calculated unless otherwise stated conventional C _t calculation may be employed.

After the decoder 103 obtains the context vector C _t , it can generate target sequence hidden states S ₁ , S ₂ , S ₃ , S ₄ and output the constituent words y ₁ , y ₂ , y ₃ , and y _{4 of the} target sequence by conventional methods.

The decoder 103 of the embodiment of the present disclosure considers the query vector q* in the context vector based on the attention mechanism, and thus decodes the information related to the user query in the generated target text (that is, the reason for recommendation), so that the query personality can be better obtained Reason for recommendation.

The operation of the recommendation reason generation device in the embodiment of the present disclosure has been exemplarily described above in conjunction with FIG. For example, the comment encoding unit 1021 of the selective encoder 102 may be based on other encoder structures, such as word vector tools such as one-way LSTM or Word2Vec. The selective encoder 102 can output the selective hidden state of all the constituent words of the user comment to the decoder 103 for decoding, or it can only output one selective hidden state of the entire sequence of information representing the user comment, such as the one in Figure 2 h ₁ 'or h ₆ ', output to the decoder 103 for decoding.

The recommendation reason generating device of the embodiment of the present disclosure can be trained through POI user comments, user queries, and user clicked recommendation reasons. For example, the recommendation reason has accumulated a large number of user search and click logs after the review site went online in multiple scenarios. In an example, a certain number of four-tuples (POI, high-quality user reviews, query information, user click recommendation reasons) can be extracted from the user's click log as training data, and the above-mentioned recommendation reason generation device can be trained. For example, from these logs, about one million quaternions are selected to train the model. Table 1 below lists a quaternion for training.

Table 1

The trained model was tested. Table 2 below is a test example based on the recommended reasons generated by three different scenarios.

Table 2

In Table 2, the first solution is a conventional Pointer-Generator model that does not consider query information, and the recommended reason generated is "the shop is large with billiards"; the second solution is to consider query information only in the encoder Solution (also based on the Pointer-Generator model), the recommended reason generated is "The shop is large, there are billiards, and the sound effects are not bad"; the third solution is that both the encoder and the decoder consider the query information solution, so The recommended reason generated is "It is comfortable to lie down and watch a movie." From the results of these three schemes, it can be seen that the recommendation reason generated by the first scheme is very poorly related to the user’s query information "Private Cinema", which cannot meet the requirements of personalized query; the second scheme generates The recommendation reason contains "the sound effect is not bad", which is generally related to the query information "private theater", which can meet the personalized requirements of the query; and the recommendation reason generated by the third solution is "comfortable to lie down and watch a movie" and The query information "Private Cinema" has a strong relevance, which better meets the requirements of individual query.

In summary, the recommendation reason generation device of the embodiment of the present disclosure can generate a personalized query recommendation reason by considering the user's query information in the encoding process and optionally in the decoding process.

The recommendation reason generating apparatus in the embodiment of the present disclosure may be any device or system that can be used to generate recommendation reasons, for example, it may be a server, a workstation, a personal computer, a tablet circuit, a smart phone, a personal digital assistant, etc., or may be Cloud computing equipment, or devices in the above equipment. Although the operation of the recommendation reason generation device in the embodiment of the present disclosure involves the user's query information and user comments, this does not mean that the user directly enters the query information or user comments on the recommendation reason generation device, only that the recommendation reason generation device will use Query information or user comments, which can be sent to the recommendation reason generation device after receiving user input by other devices, and the transmitted query information or user comments are not necessarily in exactly the same form as the original input information of the user, but only the recommendation reason generation is required The device can obtain the content of the corresponding information.

FIG. 3 shows a schematic flowchart of a method 300 for generating a recommendation reason according to an embodiment of the present disclosure. The recommendation reason generation method 300 includes steps S301 to S304. As shown in FIG. 3, in step S301, the user's query information is encoded as a query vector; in step S302, the user comment is encoded as an intermediate hidden state of the comment vector; in step S303, the query vector and the comment vector are intermediate The hidden state is operated to generate the selective hidden state of the review vector; in step S304, the selective hidden state of the review vector is decoded as a recommendation reason.

In the exemplary embodiment of the present disclosure, the step S304 may adopt an attention mechanism in the process of decoding the selective hidden state of the comment vector into a recommendation reason, and when generating the context vector based on the attention mechanism, The query vector is used as the input parameter of the context vector generation function (the function adopts MLP, for example).

In the exemplary embodiment of the present disclosure, the step S303 calculating the intermediate hidden state of the query vector and the review vector to generate the selective hidden state of the review vector includes: based on the intermediate hidden state of the review vector and the hidden state of the review vector. The weighted sum of the query vector generates a selectivity vector; and the intermediate hidden state of the review vector and the selectivity vector are multiplied by elements to generate the selectivity hidden state of the review vector.

In the exemplary embodiment of the present disclosure, when the selective hidden state of the comment vector is generated, the entire sequence vector of the user comment and the intermediate hidden state of the comment vector are also calculated.

In an exemplary embodiment of the present disclosure, when generating a context vector based on the attention mechanism, the query vector and the comment vector are selectively hidden by performing a weighted sum operation.

The above description of the recommendation reason generation device also applies to the recommendation reason generation method. Similarly, the method for generating a recommendation reason according to an embodiment of the present disclosure can generate a personalized query recommendation reason by considering the user's query information in the encoding process and optionally in the decoding process.

The recommendation reason generation method according to the embodiment of the present disclosure can be executed by any electronic device or system, including but not limited to servers, workstations, personal computers, tablet circuits, smart phones, personal digital assistants, cloud computing devices, and so on.

As an example, an electronic device for performing the above method may include a memory, a processor, and a computer program stored in the memory and running on the processor. The processor implements the steps of the method according to any embodiment of the present disclosure when executed. FIG. 4 shows a schematic structural diagram of an electronic device 400 according to an exemplary embodiment of the present disclosure. The electronic device 400 includes a processor 401 and a memory 402. The processor 401 may execute corresponding processing according to a program stored or loaded in the memory 402, that is, it may execute the steps of the method according to any embodiment of the present disclosure. The processor 401 and the memory 402 may be connected to each other through a bus 403. According to an embodiment of the present disclosure, the electronic device 400 may further include an input/output (I/O) interface 404, and the interface 404 may also be connected to the bus 403.

In addition, an embodiment of the present disclosure also provides a computer-readable storage medium on which a computer program is stored, and when the program is executed by a processor, the steps in the method according to any embodiment of the present disclosure are implemented. The computer-readable storage medium according to an embodiment of the present disclosure may be one computer-readable storage medium or a combination of multiple computer-readable storage media.

It should be noted that the embodiments of the present disclosure can be implemented in hardware or dedicated circuits, software, firmware or any combination thereof. The processor in the embodiments of the present disclosure may be an integrated circuit chip or discrete hardware component with signal processing capabilities, for example, it may be a general-purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), or a ready-made programmable Programmable logic devices such as gate arrays (FPGAs), discrete gates or transistor logic devices, etc. The methods, steps, and logical block diagrams disclosed in the embodiments of the present application can be implemented or executed. The computer-readable storage medium or memory in the embodiment of the present disclosure may be a volatile memory or a non-volatile memory, or include both volatile and non-volatile memory. The non-volatile memory may be read only memory (ROM), programmable read only memory (PROM), erasable programmable read only memory (EPROM), electrically erasable programmable read only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access Access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous connection dynamic random access memory (SLDRAM), and direct memory bus random access memory (DR RAM).

Those skilled in the art should understand that the specific embodiments described above are only examples and not limitations. Various modifications, combinations, partial combinations and replacements can be made to the embodiments of the present disclosure according to design requirements and other factors, as long as they are in the appended claims. Within the scope of its equivalent, it belongs to the scope of rights to be protected by this disclosure.

Claims

A recommendation reason generating device includes:

The query encoder is used to encode the user's query information into a query vector;

A selective encoder for encoding user comments into a comment vector selective hidden state based on the query vector; and

A decoder, which forms an encoder-decoder structure with the selective encoder, and is used to decode the selective hidden state of the comment vector into a recommendation reason,

Wherein, the selective encoder includes a comment coding unit and a selective calculation network,

The comment coding unit is used for coding the user comment into an intermediate hidden state of the comment vector, and

The selective calculation network is used for calculating the query vector and the intermediate hidden state of the review vector to generate the selective hidden state of the review vector.
The device of claim 1, wherein

The decoder adopts an attention mechanism, and uses the query vector as an input parameter of the context vector generation function when generating a context vector based on the attention mechanism.
The device of claim 1, wherein

The selectivity calculation network generates a selectivity vector based on the weighted sum of the intermediate hidden state of the review vector and the query vector, and multiplies the intermediate hidden state of the review vector and the selectivity vector by element to generate the comment Vector selective hidden state.
The device of claim 1, wherein

When generating the selective hidden state of the comment vector, the selective computing network also performs an operation on the entire sequence vector of the user comment and the intermediate hidden state of the comment vector.
The device of claim 2, wherein

When generating the context vector based on the attention mechanism, the decoder performs a weighted sum operation on the query vector and the selective hidden state of the comment vector.
The device of claim 1, wherein

The recommendation reason generating device is trained through user comments of interested merchants, user queries, and user clicked recommendation reasons.
A method for generating recommended reasons based on the encoder-decoder structure includes:

Encode the user's query information into a query vector;

Encode user comments into the intermediate hidden state of the comment vector;

Computing the query vector and the intermediate hidden state of the review vector to generate the selective hidden state of the review vector; and

The selective hidden state of the comment vector is decoded as a recommendation reason.
The method according to claim 7, wherein

The attention mechanism is used to decode the selective hidden state of the comment vector into a recommendation reason, and the query vector is used as the input parameter of the context vector generation function when generating the context vector based on the attention mechanism.
8. The method according to claim 7, wherein said calculating said query vector and said review vector intermediate hidden state to generate said review vector selective hidden state comprises:

Generating a selectivity vector based on the weighted sum of the intermediate hidden state of the comment vector and the query vector; and

Multiplying the intermediate hidden state of the review vector and the selective vector by element to generate the selective hidden state of the review vector.
The method according to claim 7, wherein

When generating the selective hidden state of the comment vector, the entire sequence vector of the user comment and the intermediate hidden state of the comment vector are also calculated.
The method according to claim 8, wherein

When generating a context vector based on the attention mechanism, a weighted sum operation is performed on the query vector and the selective hidden state of the comment vector.
A computer-readable storage medium having a computer program stored thereon, wherein the program is executed by a processor to implement the method according to any one of claims 7-11.
An electronic device, comprising a memory, a processor, and a computer program stored on the memory and runnable on the processor, and the processor implements any one of claims 7-11 when executed Methods.