Generation of point of interest copy

Abstract

Provided in the present application are a method and device for generating a point of interest (POI) copy. According to one example of the method, after acquiring POI information to be used for generating a copy, subject matter orientation prediction may be carried out for the POI information so as to determine subject matter words of the POI information, and a POI copy to which the POI information pertains is generated according to the subject matter words.

Description

GENERATION OF POINT OF INTEREST COPY Technical Field

[02] The present application relates to the generation of point of interest copy. Background Art

[03] Generating and displaying a Point of Interest (POI) copy to convey the key information of POI is an effective means to improve user click rate and purchase rate. It has been widely used in product display or trading platforms. However, manually creating a POI copy not only is costly, but also may have poor real-time performance when the number of POIs is greatly increased. In addition, the automatic extraction of the POI copy from the relevant description information of the POI can improve the extraction efficiency of the copy. However, since the relevant description text can only be extracted from the relevant description information of the POI, the obtained copy may not be accurate for the representation of POI. Summary of the Invention

[04] The embodiments of the present application provide a method for generating a POI copy to overcome the foregoing problems at least to a certain degree.

[05] In a first aspect, an embodiment of the present application provides a method for generating a point of interest (POI) copy, the method comprises: obtaining POI information; performing subject matter orientation prediction for the POI information to determine a subject matter word of the POI information; generating a copy of a POI to which the POI information pertains based on the subject matter word.

[06] In a second aspect, an embodiment of the present application provides a device for generating a point of interest (POI) copy, the device comprises: a POI information obtaining module, which is used for obtaining POI information; a subject matter word determining module, which is used for performing subject matter orientation prediction for the POI information to determine a subject matter word of the POI information; a copy generating module, which is used for generating a copy of a POI to which the POI information pertains based on the subject matter word.

[07] In a third aspect, an embodiment of the present application provides an electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that when the processor executes the computer program, the method for generating a point of interest (POI) copy according to the embodiments of the present application is implemented.

[08] In a fourth aspect, an embodiment of the present application provides a computer readable storage medium, having a computer program stored thereon, characterized in that when the computer program is executed by a processor, the steps of the method for generating a point of interest (POI) copy according to the embodiments of the present application are implemented.

[09] With the method and device for generating a POI copy disclosed in the embodiments of the present application, after obtaining the POI information, subject matter orientation prediction may be performed for the POI information to determine the subject matter word of the POI information, and then based on the subject matter word a copy of POI to which the POI information pertains is generated, which can effectively improve the expression ability and accuracy of the obtained copy for POI. With the subjective orientation prediction of the POI information and then using the predicted subject matter word as the input text of the copy to be generated, the data range of the input text to be used to generate the copy can be effectively reduced, and the input text is thus more targeted. This can effectively improve the accuracy and expressiveness of the generated copy. In addition, by means of automatically extracting the subject matter word, the efficiency of copy generation can also be effectively improved. Brief Description of the Drawings

[10] In order to describe the technical solutions of the embodiments of the present application in a clear way, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings for the following description are only some embodiments of the present application. For a person of ordinary skill in the art, other drawings can be obtained based on these drawings without inventive skills involved.

[11] FIG. 1 is a flowchart of a method for generating a POI copy according to an embodiment of the present application.

[12] FIG. 2 is a flowchart of a method for generating a POI copy according to another embodiment of the present application.

[13] FIG. 3 is a schematic structural diagram of a deep convolutional neural network model according to an embodiment of the present application.

[14] FIG. 4 is a schematic structural diagram of a POI copy generating device according to an embodiment of the present application.

[15] FIG. 5 is a schematic structural diagram of a POI copy generating device according to another embodiment of the present application.

[16] FIG. 6 is a schematic structural diagram of a POI copy generating device according to yet another embodiment of the present application. Description of the Embodiments

[17] The technical solutions of the embodiments of the present application will be clearly and completely described with reference to the accompanying drawings of the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all of the embodiments. Based on the embodiments in the present application, many other embodiments may be obtained by a person of ordinary skill in the art without inventive skills, which shall fall within the scope of protection of the present application.

[18] The present application discloses a method for generating a POI copy, as shown in FIG. 1, the method may include steps 100 to 120.

[19] Step 100 includes obtaining POI information.

[20] Point of interest (POI) may have different meanings on different platforms in practice, for example, it can be a tourist attraction, a dish, a product, and the like. The POI information may take various forms, such as in an image form and a text form. Different forms of POI information can show different dimensions of POI information. For example, the POI information in an image form can be used to display a picture of a POI, such as a home picture; the POI information in a text form can be used to display information such as the name, description, and comments of the POI.

[21] POI information is the source data used to generate the copy. For example, in order to generate a copy of a dish, the source data needed for generating a copy thereof may include an image of the dish, the name and the taste of the dish, as well as the description of the ingredients of the dish, and user reviews.

[22] Step 110 includes performing subject matter orientation prediction for the POI information to determine a subject matter word of the POI information.

[23] First, the POI information in a text form may be identified as a real number vector through a Bag of Words model to extract a feature of the POI information in a text form, and then a feature of the POI information in an image form is further extracted through a feature extraction method; next, the features of the POI information in different forms can be fused to obtain a comprehensive feature to represent the POI information; and finally, a pre-trained neural network model, such as an SVM classification model, is used to determine the subject matter word corresponding to the comprehensive feature of the obtained POI information as the subject matter word of the POI.

[24] In the embodiments of the present application, a neural network model may be trained in advance, and then POI information, such as any one or more of an image, a name, a comment, a description and the like may be input into the neural network model, such that the feature extraction and the determination of a subject matter word of the POI information may be performed through the neural network model. Thus, the subject matter orientation prediction is completed and the obtained subject matter word of the POI information is then output.

[25] Step 120 includes generating a copy of a POI to which the POI information pertains based on the subject matter word.

[26] A method similar to machine translation may be utilized to generate a copy of the POI information based on the subject matter word obtained. A deep neural network model may be pre-trained through a deep learning method, such as an encoding-decoding based deep neural network model. The encoding-decoding based deep neural network model may be employed to use the subject matter word output by the deep convolutional neural network model as its input to generate a copy corresponding to the POI information input to the deep convolutional neural network model.

[27] A neural network model based on encoding-decoding can be used herein, such as a seq2seq (sequence to sequence) deep neural network model based on the attention mechanism. During model training, a subject matter word sequence that may appear can be used as the input, and then the corresponding copy is used as the output so as to train the seq2seq deep neural network model. For example, the format of the training data for training a deep neural network model based on encoding-decoding may be [input: subject matter word sequence; output: copy], in which the subject matter word can be a text.

[28] The seq2seq deep neural network model has made great progress in the field of machine translation and natural language processing. Based on the Bayesian conditional probability formula, seq2seq consists of an encoder and a decoder. The encoder can encode the source sequence A into a fixed-length feature vector. This vector is transmitted to the decoder to obtain the target sequence B as an input. Among them, the attention mechanism used is mainly divided into two types, one is the additive attention mechanism proposed by Bahdanau, and the other is the multiplicative attention mechanism proposed by Luong. The characteristics of the seq2seq model based on the attention mechanism include: the encoding end uses a bidirectional recurrent neural network (or a variant thereof) to encode the input sequence, which can well model the entire context. The introduction of the attention mechanism makes the alignment information between the words in the target sequence and the words in the input sequence can be well used. Due to the time-dependent characteristics of the model (that is, subsequent outputs will be related to all previous inputs), the model can be learned using a back-propagation algorithm over time. For a specific implementation of training a seq2seq deep neural network model, refer to related technologies familiar to those skilled in the art, and details are not described herein again.

[29] After determining the subject matter word of the POI information, the subject matter word can be stitched into a text sequence and input to a deep neural network model based on encoding-decoding, so that output the corresponding copy.

[30] With the method for generating a POI copy disclosed in the embodiments of the present application, after obtaining the POI information, subject matter orientation prediction may be performed for the POI information to determine the subject matter word of the POI information, and then based on the subject matter word a copy of POI to which the POI information pertains is generated, which can effectively improve the expression ability and accuracy of the obtained copy for POI. With the subjective orientation prediction of the POI information and then using the predicted subject matter word as the input text of the copy to be generated, the data range of the input text to be used to generate the copy can be effectively reduced. This can effectively improve the accuracy and expressiveness of the generated copy. In addition, by means of automatically extracting the subject matter word, the efficiency of copy generation can also be effectively improved.

[31] As shown in FIG. 2, the method for generating a POI copy disclosed in another specific embodiment of the present application may include steps 200 to 230.

[32] Step 200 includes obtaining the POI information to be used for generating a copy.

[33] For a specific implementation manner of obtaining the POI information to be used for generating the copy, please refer to the foregoing description, and details will not be described herein again.

[34] After obtaining the POI information to be used for generating the copy, subject matter orientation prediction may be performed on the POI information to determine the subject matter word of the POI information. Then, a copy of POI to which the POI information pertains may be generated based on the subject matter word. Different neural network models can be used for subject matter word extraction and copy generation, respectively. For example, performing subject matter orientation prediction on the POI information to determine the subject matter word of the POI information may include: subject matter orientation prediction of the POI information through a first neural network model, to determine the subject matter word of the POI information. Generating a copy of POI to which the POI information pertains based on the subject matter word may include: using a second neural network model to generate a copy of POI to which the POI information pertains based on the subject matter word.

[35] Step 210 includes performing subject matter orientation prediction for the POI information by using a first neural network model so as to determine a subject matter word of the POI information.

[36] In this embodiment of the present application, firstly a first neural network model may be trained.

[37] In a case where two subject matters (topics) of one POI need to be determined at the same time, for example, is the business open all day and does the business have a nice environment, for each subject matter, it would be a binary classification problem with a yes or no determination; using the "is it open all day" as an example, the reply would be either yes, it is open all day, or no, it is not open all day. In a case where a plurality of subject matters (topics) of one POI need to be determined, solve multiple binary classification problems need to be solved. Thus, in order to simultaneously solve multiple binary classification problems in one model so as to predict different subject matters of a POI at the same time, a multi-objective classification model capable of solving multi-objective classification problems may be needed. In the embodiment of the present application, the first neural network model for subject matter orientation prediction may be a multi-objective classification model based on a deep convolutional neural network.

[38] The POI information can be used as an input, the subject matter word sequence corresponding to the POI information can be used as an output, and a multi-objective classification model based on a convolutional neural network is trained based on the training data formed by the input and output. The format of the training data can be [input: picture, text] sequence; output: subject matter word sequence]. Among them, the POI information may include pictures and text sequences. When training deep convolutional neural network models with training data, a back-propagation algorithm can be used.

[39] The deep convolutional neural network model may include at least one data processing channel, such as a text processing channel or an image processing channel. As shown in FIG. 3, the data processing channel may include a text processing channel 310 and an image processing channel 320. Each data processing channel is an independent convolution module for processing different forms of POI information. For example, the text processing channel 310 may be used to process the POI information in a text form, such as a name, a description, a comment, and the like. The image processing channel 320 may be used to process the POI information in an image form, such as an image of the POI. The two data processing channels can be trained separately, for example, the image processing channel 320 may be trained with an image part of an input, and the text processing channel 310 may be trained with a text sequence part of the input. A deep convolutional neural network model training method that is well known to a person skilled in the art may be used herein to train the deep convolutional neural network model, which will not be repeated herein.

[40] When extracting a copy of a POI, the POI information of the POI, such as any one or more of the first picture, name, comment, description, etc., can be used as the input of the deep convolutional neural network model, for example, the format of [input: picture, text sequence] is input to the deep convolutional neural network model, and the deep convolutional neural network model performs subject matter orientation prediction on the POI information, and outputs the POI information corresponding to at least one subject matter word.

[41] Performing subject matter orientation prediction on the POI information through a first neural network model to determine the subject matter word of the POI information may include: determining characteristics of different forms of POI information separately; fusing features of the different forms of POI information to determine a comprehensive feature of the POI information; and mapping the comprehensive feature into a subject matter word of the POI information through the first neural network model. The different forms of POI information may include: the POI information in a text form and / or the POI information in an image form.

[42] For example, the subject matter orientation prediction of the corresponding form of POI information can be performed through each data processing channel in the deep convolutional neural network model (that is, the first neural network model), and the characteristics of each form of POI information can be extracted separately. Taking the POI information of a dish including the picture, name, and description of the dish as an example, the deep convolutional neural network model can assign the picture of the dish to the image processing channel for processing; distribute the name and description of the dish to the text processing channel for processing.

[43] The image processing channel can use a two-dimensional convolutional and maximum pooling (Maximum Pooling) alternately stacked network structure, such as a vgg16 convolutional neural network and a pre-trained model of the ImageNet dataset. After the image processing channel processes the input dish picture, a real number vector V1 can be obtained to represent the features of the input dish picture, that is, the features of the POI information in an image form. In convolutional neural networks, the role of the convolution layer is to extract the features of a local area, and each filter is equivalent to a feature extractor. When multiple features need to be extracted, multiple filters (also known as convolution kernels) can be used simultaneously. The output obtained by each feature extractor, that is, vector, matrix, tensor, etc., can be called a feature map. Moreover, the same filter can use the same set of weights on the entire input data, which is another important feature of the convolution layer, "weight sharing." This feature can significantly reduce the number of connections in the model. Although the convolution layer can significantly reduce the number of connections, the number of neurons in each feature map obtained after the convolution operation has not decreased significantly. In this way, if a classifier is followed, the input dimension of the classifier is still high, and it is easy to appear overfitting. To solve this problem, the deep convolutional neural network model of this application adds a pooling operation, which is, subsampling, after the convolution layer to form a subsampling layer. The sub-sampling layer can greatly reduce the dimensionality of the features, thereby avoiding overfitting.

[44] The text processing channel can use a layer of one-dimensional convolution and a layer of maximum pooling. After an input dish name and the description text thereof are processed, another real number vector V2 can be obtained to represent the input dish name and the feature(s) of the text, that is, the features of the POI information in a text form.

[45] Then, the features of each form of POI information can be fused to determine the comprehensive feature of the POI information. Taking the image feature V1 obtained by the image processing channel as a three-dimensional feature vector [1,2,3], and the text feature V2 obtained by the text processing channel as a two-dimensional feature vector [4,5] as an example, the features [1,2,3] of the POI information in an image form and the features [4,5] of the POI information in a text form are merged to form the comprehensive feature of the POI information <semantics>[1,2,3,4,5]<annotation encoding="application / x-tex">[1,2,3,4,5]< / annotation>< / semantics>.

[46] Finally, the comprehensive feature is mapped to the subject matter word of the POI information through a fully connected layer of a deep convolutional neural network model. As shown in FIG. 3, the deep convolutional neural network model has a fully connected layer 330 after the convolution layers 311 and 321 and the pooling layers 312 and 322. The deep . convolutional neural network model may include a multi-objective classifier based on a deep convolutional neural network. The fully connected layers (FCs) function as a "classifier" in a convolutional neural network. The operation of the convolution layer and pooling layer is to map the original data to a hidden layer feature space; and then the fully connected layer maps the "distributed features" of the hidden layer feature space to a sample tag space, that is, an output space.

[47] The fully connected layer 330 can predict a subject matter word list of the POI information according to the comprehensive feature of the input POI information. As the fully connected layer can implement a multi-objective classifier, each node in this layer is an independent binary classifier, which can be activated by sigmoid. In addition, different nodes are independent from each other. Accordingly, the final loss function of the entire network can be a sigmoid-based cross entropy. Based on the foregoing, a back-propagation algorithm may be used to learn the entire network. The fully connected layer can predict the subject matter word list corresponding to the POI information according to the comprehensive feature of the POI information. For example, the subject matter word list of the POI information of a dish can be predicted according to the feature <semantics>[1,2,3,4,5]<annotation encoding="application / x-tex">[1, 2, 3, 4, 5]< / annotation>< / semantics>.

[48] Each piece of POI information can correspond to a plurality of subject matter words, and plurality of subject matter words can be output in the form of a subject matter word sequence.

[49] Step 220 includes generating a copy of a POI to which the POI information pertains through a second neural network model based on the subject matter word.

[50] The second neural network model may be a deep neural network model based on encoding and decoding.

[51] The process of generating a copy of a POI to which the POI information pertains based on the subject matter word has been described previously, and thus will not be repeated again herein.

[52] Optionally, as shown in FIG. 2, before step 220 of generating a copy of a POI to which the POI information pertains through a second neural network model based on the subject matter word, the method may further include: step 230 of determining a user-defined subject matter word.

[53] After analyzing user information, a subject matter word related to the user's interests or preferences may be extracted as a user-defined subject matter word. Alternatively, according to the actual needs, a word describing the subject matter may be used as a user-defined subject matter word.

[54] When a user-defined subject matter word is input by the system, the step of generating a copy of a POI to which the POI information pertains based on the subject matter word comprises: generating the copy of the POI to which the POI information pertains based on the subject matter word obtained from the prediction according to the POI information and the user-defined subject matter word. That is to say, the subject matter word predicted by the first neural network model and the user-defined subject matter word can be simultaneously input into the second neural network model to generate a copy of POI to which the POI information pertains. For example, in the case where the subject matter word determined by subject matter orientation prediction of the POI information through a deep convolutional neural network model is "clean environment, spicy food," while the user-defined subject matter word is "spicy food," the subject matter word determined by the subject matter orientation prediction of the POI information through a deep convolutional neural network model and the user-defined subject matter word may be combined to obtain "clean environment, spicy food" and "spicy food" as the input for a deep neural network model based on encoding-decoding, so as to generate a copy. In this way, the subject matter regarding food preference is strengthened, and the final copy will be more in line with such a user preference.

[55] For the method for generating a POI copy disclosed in the embodiment of the present application, after obtaining the POI information, subject matter orientation prediction of the POI information may be performed through a first neural network model to determine the subject matter word of the POI information, and then the copy of the POI to which the POI information pertains may be generated based on the subject matter word obtained from the prediction according to the POI information and the user-defined subject matter word. This can effectively improve the expression ability and accuracy of the obtained copy for POI.

[56] Further, new subject matter words can be flexibly supplemented according to the needs of a user, which can make the subject matter words to be used to generate the copy more in line with the user preference, which can further increase the click rate of POI. By setting multiple data processing channels in the deep convolutional neural network model, each data processing channel corresponds to an independent convolution module, which can process multiple forms of POI information. This can further improve the accuracy and expression ability of the generated copy.

[57] The subject matter word sequence of each POI can be directly used as the input of the encoding-decoding neural network to lead to the generation of the final copy, and it can also provide an intuitive source explanation for the generated copy. In addition, when it is necessary to migrate the style of the final output copy based on the encoding-decoding deep neural network model, it only needs to retrain a neural network model based on encoding-decoding. This makes the copy style migration easy, and thus multiple styles can be flexibly switched therebetween as needed.

[58] As shown in FIG. 4, the POI copy generating device disclosed in an embodiment of the present application may include: a POI information obtaining module 410, which is used for obtaining POI information; a subject matter word determining module 420, which is used for performing subject matter orientation prediction for the POI information to determine a subject matter word of the POI information; and a copy generating module 430, which is used for generating a copy of a POI to which the POI information pertains based on the subject matter word.

[59] The subject matter word determining module 420 may be further configured as performing the subject matter orientation prediction for the POI information through a first neural network model to determine the subject matter word of the POI information. The first neural network model may be a deep convolutional neural network model, such as a multi-objective classifier based on a deep convolutional neural network. In addition, the first neural network model may include multiple data processing channels, such as a text processing channel and / or an image processing channel. The text processing channel may adopt a network structure including a layer of one-dimensional convolution and a layer of maximum pooling. The image processing channel may adopt a network structure including two-dimensional convolution and maximum pooling are alternately stacked.

[60] The copy generating module 430 may be further configured as generating the copy of the POI to which the POI information pertains through a second neural network model based on the subject matter word. In this case, the second neural network model may be a deep neural network model based on encoding and decoding.

[61] As shown in FIG. 5, the subject matter word determining module 420 may include: a partial feature extracting unit 4201, which is configured to determine the features of POI information in different forms respectively; a comprehensive feature determining unit 4202, which is configured to fuse the features of the different forms of POI information to determine a comprehensive feature of the POI information; and a subject matter word determining unit 4203, which is configured to map the comprehensive feature into a subject matter word of the POI information through the first neural network model. In this case, the different forms of POI information may include: the POI information in a text form and / or the POI information in an image form.

[62] With the device for generating a POI copy disclosed in the embodiments of the present application, after obtaining the POI information, subject matter orientation prediction may be performed for the POI information to determine the subject matter word of the POI information, and then based on the subject matter word a copy of POI to which the POI information pertains is generated, which can effectively improve the expression ability and accuracy of the obtained copy for POI. With the subjective orientation prediction of the POI information and then using the predicted subject matter word as the input text of the copy to be generated, the data range of the input text to be used to generate the copy can be effectively reduced, and the input text is thus more targeted. This can effectively improve the accuracy and expressiveness of the generated copy. In addition, by means of automatically extracting the subject matter word, the efficiency of copy generation can also be effectively improved.

[63] By means of setting multiple data processing channels in the deep convolutional neural network model, each data processing channel may correspond to an independent convolution module, which can process multiple forms of POI information. This can further improve the accuracy and expressiveness of the generated copy.

[64] The subject matter word sequence of each POI obtained from the prediction can be directly used as the input of the encoding-decoding neural network to lead to the generation of the final copy, and it can also provide an intuitive source explanation for the generated copy. In addition, when it is necessary to migrate the style of the final output copy based on the encoding-decoding deep neural network model, it only needs to retrain a neural network model based on encoding- decoding. This makes the copy style migration easy, and thus multiple styles can be flexibly switched therebetween as needed.

[65] As shown in FIG. 6, based on the POI copy generating device shown in FIG. 5, the POI copy generating device disclosed in another embodiment of the present application may further include a user-defined subject matter word determining module 440, which is used for determining a user-defined subject matter word. In such a case, the copy generating module 430 is configured as: generating the copy of the POI to which the POI information pertains based on the subject matter word determined by the subject matter word determining module 420 and the user-defined subject matter word determined by the user-defined subject matter word determining module 440.

[66] The POI copy generating device disclosed in the embodiments of the present application can flexibly supplement subject matter words according to user needs, which can make the subject matter words to be used for generating a copy more consistent with the user preference, thereby further improving the click rate of POI.

[67] Accordingly, the present application also discloses an electronic device including a memory, a processor, and a computer program stored on the memory and executable on the processor, in which by means of executing the computer program by the processor, the method for generating a POI copy as described above can be implemented. The electronic device may be a mobile terminal, a smart phone, a navigator, a personal digital assistant, a tablet computer, or the like.

[68] The present application also discloses a computer readable storage medium on which a computer program is stored. In this case, when the program is executed by the processor, the steps of the method for generating a POI copy as described above can be implemented.

[69] Each embodiment in this disclosure is described in a progressive manner. Each embodiment focuses on its differences from other embodiments. The same or similar parts between various embodiments may refer to each other. For the device embodiment, since it is basically similar to the method embodiment, the description thereof is relatively simple; for related parts, please refer to the description of the method embodiments.

[70] The method and device for generating a POI copy provided in the present application have been described in detail as above. Specific examples are used in this description to explain the principle and implementation of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core ideas; at the same time, for a person of ordinary skill in the art, based on the basic idea of the present application, there may be changes in certain specific implementations and applications. In summary, the content of the present disclosure should not be construed as a limitation on the present application.

[71] Through the description for the above embodiments, a person skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary universal hardware platform, and of course, it may be implemented by hardware. Based on this understanding, the above-mentioned technical solutions, especially the essential parts thereof or the parts that contribute to the existing technology, can be embodied in the form of a software product. The computer software product can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disc, and the like. Several instructions are included to cause a computer device (which may be a personal computer, a server, a network device, or the like) to perform the methods as described in various embodiments or certain parts of the embodiments. 4 • ·

Claims

1. A device comprising: a point of interest information obtaining module configured to obtain information about a point of interest; a subject matter word determining module configured to perform, through a first neural network model, subject matter orientation prediction for the information about the point of interest to determine a subject matter word of the information about the point of interest by: determining features of the information about a point of interest in different forms; fusing the features of the information about the point of interest in different forms to determine a comprehensive feature of the information about the point of interest; and mapping the comprehensive feature into the subject matter word of the information about the point of interest through the first neural network model; and a copy generating module configured to generate, through a second neural network model, copy based on the subject matter word of the information about the point of interest and user-defined subject matter word.

2. The device of claim 1, wherein the information about the point of interest includes various forms to display dimensions of the information about the point of interest.

3. The device of claim 1, wherein the subject matter word determining module further includes a partial feature extracting unit configured to determine features of the information about the point of interest in different forms.

4. The device of any one of claims 1 to 3, wherein the subject matter word determining module further includes a comprehensive feature determining unit configured to fuse the features of the information about the point of interest in different forms to determine a comprehensive feature of the information about the point of interest.

5. The device of any one of claims 1 to 4, wherein the subject matter word determining module further includes a subject matter word determining unit configured to map the comprehensive feature into the subject matter word of the information about the point of interest through the first neural network model.

6. The device of any one of claims 3 to 5, wherein the information about the point of interest in different forms at least includes the information about the point of interest in a text form.

7. The device of claim 6, wherein the information about the point of interest in a text form includes a name of the point of interest.

8. The device of any one of claims 6 to 7, wherein the information about the point of interest in a text form includes a description of the point of interest.

9. The device of any one of claims 6 to 8, wherein the information about the point of interest in the text form includes a comment of the point of interest.

10. The device of any one of claims 6 to 9, wherein the information about the point of interest in the text form is identified as a real number vector through a Bag of Words model to extract a feature of the information about the point of interest in a text form.

11. The device of any one of claims 2 to 5, wherein the information about the point of interest in different forms at least includes the information about the point of interest in an image form.

12. The device of claim 11, wherein the information about the point of interest in the image form includes a homepage image of the point of interest.

13. The device of any one of claims 11 to 12, wherein the information about the point of interest in the image form is to extract a feature of the information about the point of interest in an image form through a feature extraction method.

14. The device of any one of claims 1 to 13, wherein the device further includes a user-defined subject matter word determining module configured to determine a subject matter word defined by a user.

15. The device of any one of claims 1 to 14, wherein the copy generating module is further configured to generate the copy of the information about the point of interest to which the information about the point of interest pertains based on the subject matter word obtained from the subject matter orientation prediction according to the information about the point of interest and a subject matter word defined by a user.

16. The device of any one of claims 1 to 15, wherein the information about the point of interest is source data for generating the copy.

17. The device of claim 1, wherein the first neural network model includes a deep convolutional neural network model.

18. The device of claim 1, wherein the second neural network model includes an encoding- decoding based deep neural network model.

19. The device of any one of claims 1 to 18, wherein the subject matter word is stitched into a text sequence for inputting to an encoding-decoding based deep neural network model.

20. An electronic device, wherein the electronic device includes: a processor configured to: obtain information about a point of interest; perform, through a first neural network model, subject matter orientation prediction for the information about the point of interest to determine a subject matter word of the information about the point of interest, by: determining features of the information about a point of interest in different forms; fusing the features of the information about the point of interest in different forms to determine a comprehensive feature of the information about the point of interest; and mapping the comprehensive feature into the subject matter word of the information about the point of interest through the first neural network model; and generate, through a second neural network model, copy based on the subject matter word of the information about the point of interest and user-defined subject matter word.

21. The electronic device of claim 20, wherein the information about the point of interest in different forms at least includes the information about the point of interest in a text form.

22. The electronic device of claim 21, wherein the information about the point of interest in a text form includes a name of the point of interest.

23. The electronic device of any one of claims 21 to 22, wherein the information about the point of interest in the text form includes a description of the point of interest.

24. The electronic device of any one of claims 21 to 23, wherein the information about the point of interest in the text form includes a comment of the point of interest.

25. The electronic device of any one of claims 21 to 24, the information about the point of interest in the text form is identified as a real number vector through a Bag of Words model to extract a feature of the information about the point of interest in a text form.

26. The electronic device of any one of claims 20, wherein the information about the point of interest in different forms at least includes the information about the point of interest in an image form.

27. The electronic device of claim 26, wherein the information about the point of interest in the image form includes a homepage image of the point of interest.

28. The electronic device of any one of claims 26 to 27, wherein the information about the point of interest in the image form is to extract a feature of the information about the point of interest in an image form through a feature extraction method.

29. The electronic device of any one of claims 20 to 28, wherein the processor is further configured to determine a subject matter word defined by a user.

30. The electronic device of any one of claims 20 to 29, wherein the processor is further configured to generate the copy of the information about the point of interest to which the information about the point of interest pertains based on the subject matter word obtained from the subject matter orientation prediction according to the information about the point of interest and a subject matter word defined by a user.

31. The electronic device of any one of claims 20 to 30, wherein the information about the point of interest is source data for generating the copy.

32. The electronic device of claim 20, wherein the first neural network model includes a deep convolutional neural network model.

33. The electronic device of claim 20, wherein the second neural network model includes an encoding-decoding based deep neural network model.

34. The electronic device of any one of claims 20 to 33, wherein the subject matter word is stitched into a text sequence for inputting to an encoding-decoding based deep neural network model.

35. A machine-readable storage medium storing machine-executable instructions, wherein the machine-executable instructions include: obtaining information about a point of interest; performing, through a first neural network, subject matter orientation prediction for the information about the point of interest to determine a subject matter word of the information about the point of interest, by: determining features of the information about a point of interest in different forms; fusing the features of the information about the point of interest in different forms to determine a comprehensive feature of the information about the point of interest; and mapping the comprehensive feature into the subject matter word of the information about the point of interest through the first neural network model; and generating, through a second neural network model, copy based on the subject matter word of the information about the point of interest and user-defined subject matter word.

36. The machine-readable storage medium of claim 35, wherein the information about the point of interest in different forms at least includes the information about the point of interest in a text form.

37. The machine-readable storage medium of claim 36, wherein the information about the point of interest in the text form includes a name of the point of interest.

38. The machine-readable storage medium of any one of claims 36 to 37, wherein the information about the point of interest in the text form includes a description of the point of interest.

39. The machine-readable storage medium of any one of claims 36 to 38, wherein the information about the point of interest in the text form includes a comment of the point of interest.

40. The machine-readable storage medium of any one of claims 36 to 39, the information about the point of interest in a text form is identified as a real number vector through a Bag of Words model to extract a feature of the information about the point of interest in a text form.

41. The machine-readable storage medium of claim 35, wherein the information about the point of interest in different forms at least includes the information about the point of interest in an image form.

42. The machine-readable storage medium of claim 41, wherein the information about the point of interest in the image form includes a homepage image of the point of interest.

43. The machine-readable storage medium of any one of claims 41 to 42, wherein the information about the point of interest in the image form is to extract a feature of the information about the point of interest in an image form through a feature extraction method.

44. The machine-readable storage medium of any one of claims 35 to 43, wherein the machine- executable instructions further include determining a subject matter word defined by a user.

45. The machine-readable storage medium of claim 44, wherein the machine-executable instructions further include generating the copy of the information about the point of interest to which the information about the point of interest pertains based on the subject matter word obtained from the subject matter orientation prediction according to the information about the point of interest and the subject matter word defined by the user.

46. The machine-readable storage medium of any one of claims 35 to 45, wherein the information about the point of interest is source data for generating the copy.

47. The machine-readable storage medium of claim 35, wherein the first neural network model includes a deep convolutional neural network model.

48. The machine-readable storage medium of claim 35, wherein the second neural network model includes an encoding-decoding based deep neural network model.

49. The machine-readable storage medium of any one of claims 35 to 48, wherein the subject matter word is stitched into a text sequence for inputting to an encoding-decoding based deep neural network model.

50. A method for a point of interest copy generation, including: obtaining information about a point of interest; performing, through a first neural network, subject matter orientation prediction for the information about the point of interest to determine a subject matter word of the information about the point of interest, by: determining features of the information about a point of interest in different forms; fusing the features of the information about the point of interest in different forms to determine a comprehensive feature of the information about the point of interest; and mapping the comprehensive feature into the subject matter word of the information about the point of interest through the first neural network model; and generating, through a second neural network model, copy based on the subject matter word of the information about the point of interest and user-defined subject matter word.

51. The method of claim 50, wherein the information about the point of interest in different forms at least includes the information about the point of interest in a text form.

52. The method of claim 51, wherein the information about the point of interest in a text form includes a name of the point of interest.

53. The method of any one of claims 51 to 52, wherein the information about the point of interest in the text form includes a description of the point of interest.

54. The method of any one of claims 51 to 53, wherein the information about the point of interest in the text form includes a comment of the point of interest.

55. The method of any one of claims 50 to 54, the information about the point of interest in the text form is identified as a real number vector through a Bag of Words model to extract a feature of the information about the point of interest in a text form.

56. The method of claim 50, wherein the information about the point of interest in different forms at least includes the information about the point of interest in an image form.

57. The method of claim 56, wherein the information about the point of interest in the image form includes a homepage image of the point of interest.

58. The method of any one of claims 56 to 57, wherein the information about the point of interest in the image form is to extract a feature of the information about the point of interest in an image form through a feature extraction method.

59. The method of any one of claims 50 to 58, wherein the method further includes determining a subject matter word defined by a user.

60. The method of claim 59, wherein the method further includes generating the copy of the information about the point of interest to which the information about the point of interest pertains based on the subject matter word obtained from the subject matter orientation prediction according to the information about the point of interest and the subject matter word defined by the user.

61. The method of any one of claims 50 to 60, wherein the information about the point of interest is source data for generating the copy.

62. The method of claim 50, wherein the first neural network model includes a deep convolutional neural network model.

63. The method of claim 50, wherein the second neural network model includes an encoding- decoding based deep neural network model.

64. The method of any one of claims 50 to 63, wherein the subject matter word is stitched into a text sequence for inputting to an encoding-decoding based deep neural network model.