Multitask learning network and method based on multitask learning network

Abstract

To provide a multi-task learning network and a method based on the multi-task learning network.SOLUTION: The multi-task learning network includes a task network and a merging network. The task network includes a plurality of sub-task networks, and the plurality of sub-task networks generates corresponding feature vectors with respect to the same input, and the merging network merges the corresponding feature vectors and performs optimization processing on a feature vector resulting from merging to obtain a final output. Each of the sub-task networks includes a coding layer and a task layer, and the coding layer converts an input to a coded vector including dimensionality information and transmits the coded vector to the task layer, and the task layer performs processing on the coded vector to generate a feature vector.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present invention relates to Natural Language Processing (NLP), and in particular to multi-task learning networks and methods based on multi-task learning networks for document classification. [Background technology]

[0002] With the development of information technology and globalization, the need for document classification is increasing. Multi-task learning (MTL) technology has been applied to the field of document classification so far.

[0003] Prior art document classification methods based on multi-task learning exist, which use a main task for document classification and subtasks for non-document classification to achieve classification of target documents. For example, in paper 1, "Stochastic Tokenization with a Language Model for Neural Text Classification," presented by Tatsuya Hiraoka, Hiroyuki Shindo, and Yuji Matsumoto at the 57th Annual Meeting of the Association for Computational Linguistics in 2019, such a multi-task learning-based document classification method is employed.

[0004] In Literature 1, a main task for document classification and a subtask for sentence segmentation based on a language model are designed. The subtask randomly segments sentences in an input document, selects the optimal segmentation result using a pre-trained language model, and then inputs word information from the segmentation result into the main task to assist the main task in document classification. The document classification method proposed in Literature 1 has a problem in that the main task and the subtask are different and the subtask must be specially designed, making it difficult to design an effective subtask. Furthermore, the application scenario of the method in Literature 1 is limited to the classification of Japanese and Chinese documents, and cannot be applied to the classification of English documents, etc. Therefore, the method in Literature 1 lacks scalability.

[0005] Currently, there are other document classification methods based on multi-task learning, in which each task is a document classification task. For example, Pengfei Liu, Xipeng Qiu, and Xuanjing Huang adopted such a multi-task learning-based document classification method in their paper "Adversarial Multi-task Learning for Text Classification," presented at the 55th Annual Meeting of the Association for Computational Linguistics in 2017.

[0006] In Reference 2, each subtask is a document classification task, and although the structure of each subtask is the same, the input for each subtask is different. By training using a shared long-short-term memory (LSTM) structure between the subtasks, the accuracy of the classification results for each subtask can be improved. The document classification method proposed in Reference 2 requires different input resources, but in practical applications, other resources are often not available. Summary of the Invention [Problem to be solved by the invention]

[0007] The present invention aims to solve the above-mentioned problems in the prior art by providing a multi-task learning network and a method based on the multi-task learning network. Since the sub-task networks in the multi-task learning network disclosed in the present invention have the same configuration and the same inputs, there is no need to specially design the sub-task networks, nor is there a need to input different resources to train the sub-task networks. Furthermore, the multi-task learning network disclosed in the present invention has excellent scalability. [Means for solving the problem]

[0008] According to one aspect of the present invention, there is provided a multi-task learning network, which comprises: a task network including multiple sub-task networks, the multiple sub-task networks generating corresponding feature vectors for the same input; and The merger network includes a merger network that performs a splicing operation on corresponding feature vectors and an optimization process on the spliced ​​feature vectors to obtain a final output.

[0009] According to one aspect of the present invention, each sub-task network in the plurality of sub-task networks includes a coding layer and a task layer, where the coding layer converts the input into a coded vector including dimensionality information and sends the coded vector to the task layer, which processes the coded vector to generate a feature vector.

[0010] According to one aspect of the present invention, multiple subtask networks share the same coding layer.

[0011] According to one aspect of the present invention, the merger network performs a fully connected process and a softmax process on the concatenated feature vector to obtain the final output.

[0012] According to one aspect of the invention, the encoding layer converts the input into the encoded vector by a token insertion process and an RNN process.

[0013] According to another aspect of the present invention, there is provided a method based on a multi-task learning network, which includes: generating corresponding feature vectors for the same input using multiple sub-task networks; performing a splicing process on the corresponding feature vectors; performing an optimization process on the spliced ​​feature vectors; and outputting the optimal feature vector.

[0014] According to other aspects of the present invention, there are further provided corresponding computer program code, a computer readable storage medium and a computer program product. [Brief explanation of the drawings]

[0015] [Figure 1] FIG. 1 is a block diagram illustrating the configuration of a typical single-task document classification network based on a recurrent neural network (RNN) and a convolutional neural network (CNN). [Figure 2] FIG. 1 is a block diagram showing the configuration of a multi-task learning network according to an embodiment of the present invention. [Figure 3] FIG. 2 is a block diagram showing in detail the configuration of a multi-task learning network in an embodiment of the present invention. [Figure 4] FIG. 1 illustrates an example of converting a single-task document classification network into a multi-task learning network in an embodiment of the present invention. [Figure 5] FIG. 1 is a diagram illustrating an example of document classification using a multi-task learning network according to an embodiment of the present invention. [Figure 6] 1 is a flowchart of a method based on a multi-task learning network in an embodiment of the present invention. [Figure 7]FIG. 1 is a block diagram illustrating an exemplary configuration of a general-purpose personal computer in which methods and / or apparatus according to embodiments of the present invention may be implemented. DETAILED DESCRIPTION OF THE INVENTION

[0016] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings. However, these embodiments are merely illustrative and are not intended to limit the scope of the present invention.

[0017] 1 illustrates the main components of a conventional single-task document classification network 100 based on a recurrent neural network (RNN) and a convolutional neural network (CNN). The single-task document classification network 100 includes a token insertion layer 110, an RNN layer 120, one or more CNN layers 1301-130N (where N is a positive integer greater than or equal to 1), one or more pooling layers 1401-140N (where N is a positive integer greater than or equal to 1), a vector splicing layer 150, a Softmax layer 160, and a single-task output layer 170.

[0018] When using the single-task document classification network 100 to perform document classification, an input is first transmitted to the single-task document classification network 100. The input may include the number of documents, the character token sequence in each document, etc. For example, the input may be a two-dimensional array sing_task_input[batch_size][document_token], where batch_size indicates the number of input documents and document_token indicates the character token sequence in each document.

[0019] The token embedding layer 110 performs embedding on the input. For example, the token embedding layer performs word vectorization on the character token sequences of each input document and adds dimensionality information. For example, the output may be a three-dimensional array embedding[batch_size][document_token][embedding_dim], where batch_size indicates the number of input documents, document_token indicates the character token sequences in each document, and embedding_dim indicates the embedding dimensionality of each character in the document.

[0020] The RNN layer 120 performs RNN processing on the input of the token insertion layer 110 by encoding it, and then transmits the output to the CNN layers 1301-130N (where N is a positive integer greater than or equal to 1). For example, the output of the RNN layer 120 may be a three-dimensional array RNN[batch_size][document_token][hidden_dim], where batch_size indicates the number of input documents, document_token indicates the character token sequence in each document, and hidden_dim indicates the number of encoding dimensions of the RNN layer 120.

[0021] The CNN layers 1301-130N and pooling layers 1401-140N extract features from the output of the RNN layer 120 and output the resulting feature vectors to the vector splicing layer 150. The vector splicing layer 150 then splices the features extracted by the CNN layers 1301-130N and corresponding pooling layers 1401-140N (e.g., the splicing result is a two-dimensional array feat_vector[batch_size][feat_dim], where batch_size represents the number of input documents and feat_dim represents the dimension of the feature vector), and outputs the spliced ​​feature vectors to the Softmax layer 160. The Softmax layer 160 processes the spliced ​​feature vectors to obtain document classification results, which are then output to the single-task output layer 170. The single-task output layer 170 outputs the final classification results.

[0022] Unlike the typical single-task document classification network architecture based on recurrent neural networks (RNNs) and convolutional neural networks (CNNs) mentioned above, the multi-task learning network 200 disclosed in this invention includes a task network 210 and a merger network 220, as shown in FIG. 2 .

[0023] Figure 3 is a block diagram showing in detail the configuration of a multi-task learning network in an embodiment of the present invention. Figure 3 shows a multi-task learning network 300 including a task network 301 and a merger network 302. The left side of Figure 3 is a block diagram showing the configuration of a task network in the multi-task learning network in an embodiment of the present invention, and the right side of Figure 3 is a block diagram showing the configuration of a merger network in the multi-task learning network in an embodiment of the present invention.

[0024] Task network 301 includes one or more subtask networks 3001-300N (where N is a positive integer greater than or equal to 1). Each of the one or more subtask networks 3001-300N includes a task layer and an output layer. For example, as shown in Figure 3, subtask network 3001 includes task layer 3201 and output layer 3301, and subtask network 300N includes task layer 320N and output layer 330N. The one or more subtask networks 3001-300N share coding layer 310, and the outputs from coding layer 310 to task layers 3201-320N are the same.

[0025] Note that the configuration of the multi-task learning network shown in Figure 3 is merely an example and does not limit the present invention. In other embodiments of the present invention, the output layer of the sub-task network is not required. One or some or all of the sub-task networks in a task network may not include an output layer. Also, in other embodiments of the present invention, multiple sub-task networks may each have a coding layer, i.e., they may not share a coding layer.

[0026] When using the multi-task learning network 300 to perform document classification, an input is first sent to the multi-task learning network 300. The input may include the number of documents, the character token sequences in each document, etc. The encoding layer 310 converts the input into an encoded vector including dimensionality information through token insertion and RNN processing, and then sends the encoded vector to the task layers 3201-320N. For example, the input may be a two-dimensional array encode_input[batch_size][document_token], where batch_size indicates the number of input documents and document_token indicates the character token sequence in each document. The encoding layer 310 converts the input two-dimensional array into a three-dimensional array encode_proc[batch_size][document_token][hidden_dim] through token insertion (e.g., word insertion) and RNN processing, where batch_size represents the number of input documents, document_token represents the character token sequence in each document, and hidden_dim represents the number of encoding dimensions.

[0027] Each of the task layers 3201-320N generates a feature vector based on the received encoded vector through convolution and pooling, and outputs the generated feature vector to the corresponding output layer. For example, as shown in Figure 3, the task layer 3201 outputs the generated feature vector to the output layer 3301, and the task layer 320N outputs the generated feature vector to the output layer 330N. For example, the output from the task layer 3201 to the output layer 3301 may be a two-dimensional array task_feat_vector[batch_size][feat_dim], where batch_size represents the number of input documents and feat_dim represents the number of dimensions of the feature vector.

[0028] Each of the output layers 3301-330N performs a Softmax operation on the received feature vector to obtain the output of the subtasking network.

[0029] Multi-task learning network 300 can construct a loss function for each sub-task network 3001-300N based on the output of each sub-task network in task network 301. Multi-task learning network 300 then constructs a total loss function for the task network based on the loss functions of each sub-task network, and trains the task network based on the total loss function.

[0030] Note that constructing the loss function of the subtask network described above is not essential, and in other embodiments of the present invention, for example, in applications where the subtask network does not include an output layer, the process of constructing the loss function of the subtask network may be omitted.

[0031] Hereinafter, the configuration of the merge network in the multi-task learning network in the embodiment of the present invention will be described in detail with reference to the right side of FIG.

[0032] As shown in FIG. 3, merger network 302 concatenates the corresponding feature vectors output by the task layers of each of multiple subtask networks 3001-300N, and then performs optimization processing on the concatenated feature vectors to obtain the final output.

[0033] Specifically, merger network 302 includes a concatenation layer 340, a fully connected layer 350, and a final output layer 360. Concatenation layer 340 concatenates corresponding feature vectors output by the task layers of each of multiple subtask networks 3001-300N into a single vector.

[0034] The fully connected layer 350 performs a fully connected transformation on the spliced ​​feature vector and inputs it to the final output layer 360 of the merger network 302. For example, the output of the fully connected layer 350 may be a two-dimensional array FC[batch_size][feat_dim], where batch_size represents the number of input documents and feat_dim represents the number of dimensions of the feature vector.

[0035] The final output layer 360 of the merger network 302 performs a softmax process on the received inputs to normalize them, and then outputs the final output.

[0036] The multi-task learning network 300 can construct a loss function for the merger network based on the final output, and then train the merger network based on the loss function for the merger network.

[0037] Hereinafter, how to convert a single-task document classification network into a multi-task learning network in an embodiment of the present invention will be described in detail with reference to Figure 4. The left side of the arrow in Figure 4 is a block diagram of the configuration of the conventional single-task document classification network 100 shown in Figure 1, and the right side of the arrow in Figure 4 is a block diagram of the configuration of the multi-task learning network constructed in an embodiment of the present invention.

[0038] First, the single-task document classification network 100 is divided. Specifically, the single-task document classification network is divided into an encoding layer, a task layer, and an output layer. The encoding layer includes a token insertion layer 110 and an RNN layer 120. The task layer includes one or more CNN layers 1301-130N and one or more pooling layers 1401-140N. The output layer includes a vector splicing layer 150, a Softmax layer 160, and a single-task output layer 170.

[0039] Then, multiple subtask networks 4001-400N are constructed by reusing the task layer and output layer and sharing the coding layer, thereby constructing task network 401 in multi-task learning network 400.

[0040] Then, a merger network 402 is constructed in the multi-task learning network 400, which includes a splicing layer 440, a fully connected layer 450, and a final output layer 460. The configurations of the task network 401 and the merger network 402 in the multi-task learning network 400 in Figure 4 and the interactions between each layer are similar to those described above with reference to Figure 3, so a detailed description thereof will be omitted here.

[0041] It should be noted that the single-task learning network 100 to be converted shown in Figure 4 is merely an example and does not limit the present invention. For other networks that can be divided into an encoding layer, a task layer, and an output layer, a multi-task learning network can be constructed in the above manner.

[0042] An example of document classification using a multi-task learning network according to an embodiment of the present invention will be described below with reference to Fig. 5. Fig. 5 shows an example of classification of a document containing 100 words shown in Table 1 below. [Table 1]

[0043] As shown in FIG. 5, the multi-task learning network 500 includes a task network 501 and a merge network 502.

[0044] When classifying the 100-word document shown in Table 1 using the multi-task learning network 500, the input is the array input[batch_size][document_token], where batch_size=1 and document_token=100, where batch_size represents the number of input documents and document_token represents the character token sequence in the document.

[0045] The encoding layer 510 converts the input into an array encoding[batch_size][document_token][hidden_dim] through token insertion processing (e.g., word insertion processing) and RNN processing, where batch_size=1 (batch_size represents the number of input documents), document_token=100 (document_token represents the character token sequence in this document), and hidden_dim=128 (hidden_dim represents the number of encoding dimensions), and then transmits the array to task layers 5201 and 5202 respectively.

[0046] The task layers 5201 and 5202 generate corresponding feature vectors through convolution and pooling, respectively, and then transmit the generated feature vectors to the corresponding output layers 5301 and 5302 in the subtask network. The task layers 5201 and 5202 then transmit the generated feature vectors to the splicing layer 511 in the merger network. For example, the task layer 5201 generates a feature vector feat_A[batch_size][feat_dim], where batch_size=1 and feat_dim=32, and the task layer 5202 generates a feature vector feat_B[batch_size][feat_dim], where batch_size=1 and feat_dim=32. In this example, batch_size indicates the number of input documents, and feat_dim indicates the number of dimensions of the feature vector.

[0047] The output layer 5301 performs Softmax processing on the received feature vector feat_A[batch_size][feat_dim], obtaining the classification result (1, 117, 242) output by the subtasking network 5001. The output layer 5302 performs Softmax processing on the received feature vector feat_B[batch_size][feat_dim], obtaining the classification result (1, 121, 1212, 13) output by the subtasking network 5002. Note that there is a possibility that an incorrect classification result may be present in the output of the subtasking network. In Figure 5, the shaded areas indicate incorrect classification results in the output of the subtasking network.

[0048] The splicing layer 511 in the merger network 501 splices the feature vectors output by the task layers 5201 and 5201, and then transmits the spliced ​​feature vectors to the fully connected layer 521. For example, the splicing layer 511 splices A[batch_size][feat_dim] and B[batch_size][feat_dim], and then transmits the spliced ​​feature vectors to the fully connected layer 521.

[0049] The fully connected layer 521 performs a fully connected transformation on the spliced ​​feature vector and outputs it to the final output layer 531. For example, the fully connected layer 521 performs a fully connected transformation on the spliced ​​feature vector and outputs the fully connected transformed array FC[batch_size][feat_dim] (where batch_size=1, feat_dim=32, batch_size indicates the number of input documents, and feat_dim indicates the number of dimensions of the feature vector) to the final output layer 531.

[0050] The final output layer 531 performs Softmax processing on the received input to normalize it, and then outputs the final classification result (1, 1212, 13, 117). As can be seen, the multi-task learning network in this embodiment of the present invention can eliminate incorrect features by selecting appropriate features from the feature vectors of the two sub-tasks through learning.

[0051] The method based on the multi-task learning network in the embodiment of the present invention will now be described with reference to FIG.

[0052] As shown in Figure 6, the method based on the multi-task learning network according to the embodiment of the present invention starts in step S110. In step S110, corresponding feature vectors are generated by multiple sub-task networks in the multi-task learning network for the same input, and step S110 corresponds to the coding layer and task layer processes shown in Figures 3 to 5. Note that the coding layer and task layer processes shown in Figures 3 to 5 have already been described in detail, so a detailed description thereof will be omitted here.

[0053] Next, in step S120, the corresponding feature vectors are spliced, and step S120 corresponds to the splicing layer process shown in Figures 3 to 5. Note that the splicing layer process shown in Figures 3 to 5 has already been described in detail, so a detailed description thereof will be omitted here.

[0054] Next, in step S130, optimization processing is performed on the spliced ​​feature vector.

[0055] Then, in step S140, the optimum feature vector is output.

[0056] Steps S130-S140 correspond to the processing of the fully connected layer and the final output layer shown in Figures 3 to 5. Note that the processing of the fully connected layer and the final output layer shown in Figures 3 to 5 has already been explained in detail, so a detailed explanation thereof will be omitted here.

[0057] The process then ends.

[0058] Each component or unit in the above-described multi-task learning network can be configured in the form of software, firmware, hardware, or a combination thereof. The specific means or methods used for configuration are well known to those skilled in the art, and therefore will not be described in detail here. When implemented by software or firmware, the programs constituting the software are installed from a storage medium or a network onto a computer having a dedicated hardware configuration (e.g., a general-purpose computer 700 shown in FIG. 7), and the computer can perform various functions when various programs are installed.

[0059] FIG. 7 is a structural diagram of a hardware configuration (general-purpose computer) 700 that can implement the method and apparatus of an embodiment of the present invention.

[0060] The general-purpose computer 700 may be, for example, a computer system. Note that the general-purpose computer 700 is merely an example and does not limit the scope or functionality of the method and apparatus according to the present invention. Furthermore, the general-purpose computer 700 does not depend on any module, assembly, or combination thereof in the above-described method and apparatus.

[0061] 7, a central processing unit (CPU) 701 performs various processes based on programs stored in a ROM 702 or programs loaded from a storage unit 708 into a RAM 703. The RAM 703 can also store data required by the CPU 701 to perform various processes, depending on the needs. The CPU 701, ROM 702, and RAM 703 are connected to one another via a bus 704. An input / output interface 705 is also connected to the bus 704.

[0062] The input / output interface 705 is further connected to the following components: an input unit 706 including a keyboard, an output unit 707 including a display such as a liquid crystal display (LCD) and a speaker, a storage unit 708 including a hard disk, and a communication unit 709 including a network interface card such as a LAN card or a modem. The communication unit 709 performs communication processing via a network such as the Internet or a LAN. A drive 710 may be connected to the input / output interface 705 as needed. A removable medium 711, such as a semiconductor memory, can be inserted into the drive 710 as needed, allowing a computer program read from the medium to be installed in the storage unit 708.

[0063] The present invention also provides a program product including machine-readable instruction codes, which, when read and executed by a machine, can perform the methods of the above-described embodiments of the present invention. Accordingly, various storage media for carrying such program products, such as magnetic disks (including floppy disks (registered trademark)), optical disks (including CD-ROMs and DVDs), magneto-optical disks (including MDs (registered trademark)), and semiconductor storage devices, are also included in the present invention.

[0064] The storage medium may include, for example, a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory device, etc., but is not limited to these.

[0065] Furthermore, each operation (process) in the above-described method can also be realized in the form of a computer-executable program stored in various machine-readable storage media.

[0066] Furthermore, the above-mentioned embodiments are further disclosed as supplementary notes as follows.

[0067] (Appendix 1) A multi-task learning network, a task network including a plurality of sub-task networks, the plurality of sub-task networks generating corresponding feature vectors for the same input; and A multi-task learning network including a merging network that performs a concatenation of the corresponding feature vectors and performs an optimization process on the concatenated feature vectors to obtain a final output.

[0068] (Appendix 2) 2. The multi-task learning network of claim 1, a task layer that processes the input and outputs a feature vector; and a coding layer that converts the input into a coded vector that includes dimensionality information. The task layer then processes the coded vector and outputs a feature vector.

[0069] (Appendix 3) 3. The multi-task learning network of claim 2, A multi-task learning network, wherein the multiple sub-task networks share the same coding layer.

[0070] (Appendix 4) 2. The multi-task learning network of claim 1, The merger network performs a fully connected process and a softmax process on the combined feature vector to obtain the final output.

[0071] (Appendix 5) 3. The multi-task learning network of claim 2, The encoding layer converts the input into the encoded vector through a token insertion process and an RNN process.

[0072] (Appendix 6) 3. The multi-task learning network of claim 2, The task layer generates the feature vector based on the encoded vector through convolution and pooling processes.

[0073] (Appendix 7) 3. The multi-task learning network of claim 2, a multi-task learning network, wherein each sub-task network of the plurality of sub-task networks further includes an output layer, the output layer receiving the feature vectors from the task layer and performing a Softmax operation to obtain an output of the sub-task network.

[0074] (Appendix 8) 8. The multi-task learning network of claim 7, constructing a loss function for each subtask network based on the outputs of the subtask networks, constructing a total loss function for the task network based on the loss functions for each subtask network, and training the task network based on the total loss function.

[0075] (Appendix 9) 5. The multi-task learning network of claim 4, A multi-task learning network that constructs a loss function for the merger network based on the final output, and trains the merger network based on the loss function for the merger network.

[0076] (Appendix 10) 2. The multi-task learning network of claim 1, The multi-task learning network is used for natural language processing.

[0077] (Appendix 11) A method based on a multi-task learning network, comprising: generating corresponding feature vectors by multiple subtask networks in the multi-task learning network for the same input; performing stitching on the corresponding feature vectors; Optimize the stitched feature vector; and outputting the optimal feature vector.

[0078] (Appendix 12) 12. The method of claim 11, 1. A method according to claim 1, wherein each sub-task network of a plurality of sub-task networks includes a coding layer and a task layer, the coding layer converting the input into a coded vector including dimensionality information, and the task layer operating on the coded vector to generate a feature vector.

[0079] (Appendix 13) 12. The method of claim 11, The method, wherein the multiple subtask networks share the same coding layer.

[0080] (Appendix 14) 12. The method of claim 11, A method for obtaining the final output by performing a full connection process and a softmax process on the combined feature vector.

[0081] (Appendix 15) 13. The method of claim 12, The method converts the input into the coded vector through a token insertion process and an RNN process.

[0082] (Appendix 16) 13. The method of claim 12, generating the feature vector based on the encoded vector by convolution and pooling processes.

[0083] (Appendix 17) 13. The method of claim 12, A method comprising: performing a Softmax operation on feature vectors generated by each sub-tasking network of a plurality of sub-tasking networks to obtain an output of the sub-tasking network.

[0084] (Appendix 18) 18. The method of claim 17, constructing a loss function for each subtask network based on the outputs of the subtask networks, and constructing an overall loss function based on the loss functions of each subtask network.

[0085] (Appendix 19) 12. The method of claim 11, A method based on multi-task learning networks is used for natural language processing.

[0086] (Appendix 20) 1. A computer-readable storage medium, comprising: The computer-readable storage medium stores a program, and the program can be executed by a processor to perform the following operations: Generate corresponding feature vectors for the same input; performing stitching on the corresponding feature vectors; Optimize the stitched feature vector; and Output the optimal feature vector A computer-readable storage medium.

[0087] Although the preferred embodiment of the present invention has been described above, the present invention is not limited to this embodiment, and any modification to the present invention falls within the technical scope of the present invention as long as it does not depart from the spirit of the present invention.

Claims

1. A multi-task learning network including a task network and a merger network, the task network includes a plurality of sub-task networks, each of which generates a corresponding feature vector for a same input; The merger network combines the corresponding feature vectors into one feature vector, and performs full connection processing and normalization processing on the combined feature vector to obtain a final output.

2. 2. The multi-task learning network of claim 1, a task layer that processes the input and outputs a feature vector; and a coding layer that converts the input into a coded vector that includes dimensionality information. The task layer then processes the coded vector and outputs a feature vector.

3. 3. A multi-task learning network as claimed in claim 2, comprising: A multi-task learning network, wherein the multiple sub-task networks share the same coding layer.

4. 3. A multi-task learning network as claimed in claim 2, comprising: The encoding layer converts the input to the encoded vector through a token insertion process and an RNN process.

5. 3. A multi-task learning network as claimed in claim 2, comprising: The task layer generates the feature vector based on the encoded vector through convolution and pooling processes.

6. 3. A multi-task learning network as claimed in claim 2, comprising: a multi-task learning network, wherein each sub-task network of the plurality of sub-task networks further includes an output layer, the output layer receiving feature vectors from the task layer and performing Softmax processing on the feature vectors to obtain an output of the sub-task network.

7. 7. A multi-task learning network as claimed in claim 6, comprising: constructing a loss function for each subtask network based on the outputs of the subtask networks, constructing a total loss function for the task network based on the loss functions for each subtask network, and training the task network based on the total loss function.

8. 2. The multi-task learning network of claim 1, A multi-task learning network that constructs a loss function for the merger network based on the final output and trains the merger network based on the loss function for the merger network.

9. A method performed by a multi-task learning network, comprising: generating corresponding feature vectors by a plurality of sub-task networks in the multi-task learning network for the same input; Concatenating the corresponding feature vectors into a single feature vector; Performing full connection processing and normalization processing on the stitched feature vector; and outputting the optimal feature vector.

Images