Multi-task based label generation method, device, electronic device and medium

Abstract

The present invention discloses a label generation method, device, electronic device and medium based on multi-task, the method includes processing semantic data through a preset semantic intent label mining model to generate semantic intent labels; and calling a preset semantic intent recognition engine and a semantic fine-grained emotion judgment engine respectively to perform semantic intent recognition and semantic fine-grained emotion judgment based on multi-task learning, wherein the semantic intent recognition engine associates semantics with semantic intent labels through a preset semantic intent recognition model, and the semantic fine-grained emotion judgment engine judges the semantic emotion labels of semantics through a preset semantic fine-grained emotion judgment model, thereby clustering semantics under semantic intent labels according to the semantic emotion labels. Therefore, it is possible to improve the decision-making efficiency of users, enhance the user experience, and improve the conversion effect of traffic.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a label generation method, device, electronic device and medium based on multi-task. Background Art

[0002] Most of the existing semantic search applications or search operations performed in the applications only provide a search box for users to perform semantic searches of interest.

[0003] In the process of implementing the present invention, the inventors found that there are at least the following semantics in the prior art:

[0004] On the one hand, the search is a semantic search in the way of full string matching. Due to the diversity of user expressions, most users cannot find satisfactory results by inputting in their own habitual input methods. On the other hand, the semantic data is relatively messy and has not been clustered. As a result, the matching degree of user search results is low and the user's search experience is reduced. Summary of the invention

[0005] In view of this, the embodiments of the present invention provide a label generation method, device, electronic device and medium based on multi-task, which can mine the semantic dimensions of users' concerns based on user semantic data to establish semantic intent labels, and perform semantic intent recognition and semantic fine-grained sentiment judgment based on multi-task learning, associate semantics with semantic intent labels and provide key dimensions of semantic sorting for semantic clustering, thereby improving the execution efficiency of the entire solution. This can improve the user's decision-making efficiency, enhance the user's usage experience, and improve the conversion effect of traffic, such as clicks, dwell time, etc.

[0006] To achieve the above object, according to one aspect of an embodiment of the present invention, a multi-task based label generation method is provided, which is characterized by comprising:

[0007] Processing the semantic data through a preset semantic intent label mining model to generate semantic intent labels; and

[0008] The preset semantic intent recognition engine and semantic fine-grained sentiment judgment engine are called respectively to perform semantic intent recognition and semantic fine-grained sentiment judgment based on multi-task learning.

[0009] Among them, the semantic intent recognition engine associates the semantics with the semantic intent label through a preset semantic intent recognition model, and the semantic fine-grained sentiment judgment engine judges the semantic sentiment label of the semantics through a preset semantic fine-grained sentiment judgment model, thereby clustering the semantics under the semantic intent label according to the semantic sentiment label to generate a semantic clustering label.

[0010] Furthermore, the semantic intent recognition engine and the semantic fine-grained sentiment judgment engine share a two-layer bidirectional long short-term memory network for feature extraction, and

[0011] The double-layer bidirectional long short-term memory network inputs the embedding representation of the semantics, and the hidden layer of the double-layer bidirectional long short-term memory network outputs the sentence feature vector of the semantics.

[0012] Furthermore, the semantic intent recognition model includes using a fully connected layer to determine a semantic intent label vector representation based on the sentence feature vector, and identifying the semantic intent label of the semantics through a softmax activation function based on the semantic intent label vector representation.

[0013] Furthermore, the semantic fine-grained sentiment judgment model includes:

[0014] Through a layer of convolutional neural network, semantic representation features S_1 and S_2 are obtained based on the sentence feature vector;

[0015] Input the representation feature S_1 into the hyperbolic tangent activation function tanh to obtain the output vector tanh(S_1) of the representation feature S_1 as the first output vector, add the representation feature S_2 and the semantic intent label vector representation A into the rectified linear unit activation function relu to obtain the output vector relu(S_2+A) of the sum of the representation feature S_2 and the semantic intent label vector representation A as the second output vector, and perform element-wise product of the first output vector and the second output vector to obtain a final feature vector; and

[0016] After the final feature vector is max-pooled and passes through a fully connected layer, a softmax activation function is used to obtain the final classification result as the semantic sentiment label.

[0017] Further, when semantic clustering is performed according to the semantic intent label, the semantic sentiment label is a ranking factor when the semantics are clustered under the semantic intent label.

[0018] Furthermore, the semantic intent label mining model includes:

[0019] Based on the semantic data, semantic preprocessing is performed to obtain words;

[0020] Aggregating the words into phrases to generate candidate intent label words;

[0021] Performing low-frequency filtering and fluency detection on the candidate intent label words for screening; and

[0022] The screened candidate intent label words are aggregated, and the aggregated results are manually summarized to obtain a central phrase representing the class family as the semantic intent label.

[0023] Furthermore, the loss of the multi-task learning is equal to the sum of the cross entropy loss of the semantic intent recognition and the cross entropy loss of the semantic fine-grained sentiment judgment.

[0024] According to another aspect of the present invention, there is provided a multi-task based label generation device, comprising:

[0025] A semantic intent label mining module, the semantic intent label mining module is used to generate semantic intent labels based on semantic data; and

[0026] A semantic intent recognition and semantic fine-grained sentiment judgment module based on multi-task learning, which is used to perform semantic intent recognition and semantic fine-grained sentiment judgment based on multi-task learning, associate semantics with the semantic intent label using the semantic intent recognition, determine the semantic sentiment label of the semantics using the semantic fine-grained sentiment judgment, and cluster the semantics under the semantic intent label according to the semantic sentiment label to generate a semantic cluster label.

[0027] Furthermore, the semantic intent recognition and semantic fine-grained sentiment judgment module based on multi-task learning includes a multi-task learning unit, which shares a double-layer bidirectional long short-term memory network for feature extraction, and the double-layer bidirectional long short-term memory network inputs the embedded representation of semantics, and the hidden layer of the double-layer bidirectional long short-term memory network outputs the sentence feature vector of the semantics.

[0028] Furthermore, the semantic intent recognition and semantic fine-grained sentiment judgment module based on multi-task learning includes a semantic intent recognition unit, which uses a fully connected layer to determine a semantic intent label vector representation based on the sentence feature vector, and identifies the semantic intent label of the semantics through a softmax activation function based on the semantic intent label vector representation.

[0029] Furthermore, the semantic intent recognition and semantic fine-grained sentiment judgment module based on multi-task learning includes a semantic fine-grained sentiment judgment unit, which performs the following processing:

[0030] Through a layer of convolutional neural network, semantic representation features S_1 and S_2 are obtained based on the sentence feature vector;

[0031] Input the representation feature S_1 into the hyperbolic tangent activation function tanh to obtain the output vector tanh(S_1) of the representation feature S_1 as the first output vector, add the representation feature S_2 and the semantic intent label vector representation A into the rectified linear unit activation function relu to obtain the output vector relu(S_2+A) of the sum of the representation feature S_2 and the semantic intent label vector representation A as the second output vector, and perform element-wise product of the first output vector and the second output vector to obtain a final feature vector; and

[0032] After the final feature vector is max-pooled and passes through a fully connected layer, a softmax activation function is used to obtain the final classification result as the semantic sentiment label.

[0033] Further, when semantic clustering is performed according to the semantic intent label, the semantic sentiment label is a ranking factor when the semantics are clustered under the semantic intent label.

[0034] Furthermore, the semantic intent label mining module performs the following processing:

[0035] Based on the semantic data, semantic preprocessing is performed to obtain words;

[0036] Aggregating the words into phrases to generate candidate intent label words;

[0037] Performing low-frequency filtering and fluency detection on the candidate intent label words for screening; and

[0038] The screened candidate intent label words are aggregated, and the aggregated results are manually summarized to obtain a central phrase representing the class family as the semantic intent label.

[0039] Furthermore, the loss of the multi-task learning performed by the multi-task learning unit is equal to the sum of the cross-entropy loss of the semantic intent recognition and the cross-entropy loss of the semantic fine-grained sentiment judgment.

[0040] According to another aspect of the present invention, there is provided an electronic device for generating a label based on multi-task, comprising:

[0041] one or more processors;

[0042] a storage device for storing one or more programs,

[0043] When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in one aspect above.

[0044] According to another aspect of the present invention, there is provided a computer-readable medium having a computer program stored thereon, wherein when the program is executed by a processor, the method described in one of the above aspects is implemented.

[0045] One embodiment of the above invention has the following advantages or beneficial effects: the present invention mines the semantic dimensions that users care about based on user semantic data to establish semantic intent labels, and performs semantic intent recognition and semantic fine-grained sentiment judgment based on multi-task learning, associates semantics with semantic intent labels, and provides key dimensions of semantic sorting for semantic clustering, thereby improving the execution efficiency of the entire solution. This can improve the user's decision-making efficiency, enhance the user's experience, and improve the conversion effect of traffic, such as clicks, dwell time, etc.

[0046] The further effects of the above-mentioned non-conventional optional manner will be described below in conjunction with specific implementation examples. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The accompanying drawings are used to better understand the present invention and do not constitute an improper limitation of the present invention.

[0048] Figure 1 is a schematic diagram of the main process of the multi-task based label generation method according to an embodiment of the present invention;

[0049] Figure 2 is a flowchart of each sub-step of the semantic intent label mining step of the multi-task based label generation method according to an embodiment of the present invention;

[0050] Figure 3 It is an example of a process of manually screening and summarizing semantic intent labels mined by a multi-task based label generation method according to an embodiment of the present invention;

[0051] Figure 4 is a schematic diagram illustrating a detailed process of semantic intent recognition and semantic fine-grained emotion judgment steps based on multi-task learning of a multi-task based label generation method according to an embodiment of the present invention;

[0052] Figure 5 is a schematic diagram illustrating a calculation process of semantic fine-grained sentiment judgment based on a multi-task label generation method according to an embodiment of the present invention;

[0053] Figure 6 is a schematic diagram of main modules of a multi-task based label generation device according to an embodiment of the present invention;

[0054] Figure 7 is a schematic diagram of main components of a semantic intent tag mining module of a multi-task based tag generation device according to an embodiment of the present invention;

[0055] Figure 8 Schematic diagram of main components of a semantic intent recognition and semantic fine-grained emotion judgment module based on multi-task learning of a multi-task label generation device according to an embodiment of the present invention;

[0056] Fig. 9 is a schematic diagram of an example of performing semantic clustering using a multi-task based label generation method and apparatus according to an embodiment of the invention;

[0057] Fig.10 An example of an interface before applying the multi-task based label generation method of the present invention is shown;

[0058] Figure 11(a) and 11(b) An example of an interface after applying the multi-task based label generation method of the present invention is shown;

[0059] Fig.12 is an exemplary system architecture diagram to which embodiments of the present invention may be applied;

[0060] Fig.13 It is a schematic diagram of the structure of a computer system of a terminal device or a server suitable for implementing an embodiment of the present invention. DETAILED DESCRIPTION

[0061] The following is a description of exemplary embodiments of the present invention in conjunction with the accompanying drawings, including various details of the embodiments of the present invention to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present invention. Similarly, for clarity and conciseness, the description of well-known functions and structures is omitted in the following description.

[0062] Figure 1 is a schematic diagram of the main process of the label generation method based on multi-task according to an embodiment of the present invention, such as Figure 1 As shown, the multi-task based label generation method according to an embodiment of the present invention includes a semantic intent label mining step S101 and a semantic intent recognition and semantic fine-grained emotion judgment step S102 based on multi-task learning.

[0063] The following will refer to the attached Figure 1-4 The above steps are described in detail. Figure 2 is a flowchart of each sub-step of the semantic intent label mining step of the multi-task based label generation method according to an embodiment of the present invention; Figure 3is a schematic diagram illustrating a detailed process of semantic intent recognition and semantic fine-grained emotion judgment steps based on multi-task learning of a multi-task based label generation method according to an embodiment of the present invention; Figure 4 It is a schematic diagram illustrating the calculation process of semantic fine-grained sentiment judgment based on the multi-task label generation method according to an embodiment of the present invention.

[0064] Step S101: semantic intent label mining step

[0065] The semantic data is processed by a preset semantic intent label mining model to generate semantic intent labels. Specifically, the semantic intent label mining model includes the following processing steps: Steps: S1011, preprocessing step; S1012, candidate intent label word generation step; S1013, screening step; and S1014, aggregation and summary step. The above sub-steps are described in detail below.

[0066] S1011, pre-processing step

[0067] Based on the semantic data, the semantics are segmented and the body words are removed to obtain the words.

[0068] S1012, candidate intent label word generation step

[0069] The obtained words are aggregated into phrases to generate candidate intent label words.

[0070] Specifically, the present invention generates candidate intent label words based on word2phrase (an algorithm). The idea of ​​word2phrase is to aggregate words into phrases based on mutual information. For words that often appear together, the judgment of whether two words form a phrase is completed based on the following formula:

[0071]

[0072] Among them, score represents the calculated score, wi, wj represent the word vectors obtained after preprocessing by word2phrase, and δ is a preset parameter (for example, it can be set to 5). Input the vectors of two words, if the calculated score is greater than the preset threshold, then the two words are considered to be "together".

[0073] In order to take into account longer phrases, 2-4 words are used as training data, and the threshold is lowered successively. Thus, candidate intent label words can be obtained.

[0074] S1013, screening step

[0075] Perform word frequency statistics on candidate intent label words, filter low-frequency label words, and perform phrase fluency detection, for example, based on the N-gram language model (this language model is the product of the previous word2phrase) to filter out candidate intent label words with fluency lower than a predetermined value.

[0076] S1014, Aggregation and Summarization Steps

[0077] The candidate intent label words after screening are aggregated, and the number of aggregations is set according to the actual situation of the category (generally set to 100-200). The results of the aggregation are manually screened and summarized to obtain the central phrase representing the category family as the semantic intent label.

[0078] For example, Figure 3 As shown, the candidate intent label words such as easy to wear, comfortable to wear, and comfortable to wear are aggregated and manually screened and summarized into the central phrase of "wearing comfort" as the semantic intent label.

[0079] By summarizing the filtered central phrases, we can obtain the core intent phrases sorted out under the mobile phone category, such as shown in the following Table 1, that is, the semantic intent labels (split according to the two-level intent hierarchy).

[0080] Table 1: Two-level intent hierarchy (mobile phone category)

[0081]

[0082]

[0083] Step S102: semantic intent recognition and semantic fine-grained sentiment judgment based on multi-task learning

[0084] The present invention models semantic intent recognition and semantic fine-grained sentiment judgment in a multi-task learning manner, that is, semantic intent recognition and semantic fine-grained sentiment judgment are performed based on multi-task learning, wherein semantic intent recognition realizes the association of semantics with the semantic intent label obtained in step S101, and semantic fine-grained sentiment judgment determines the semantic sentiment label of the semantics. In step S102, the preset semantic intent recognition engine and semantic fine-grained sentiment judgment engine are respectively called to perform semantic intent recognition and semantic fine-grained sentiment judgment based on multi-task learning, wherein the semantic intent recognition engine associates the semantics with the semantic intent label through a preset semantic intent recognition model, and the semantic fine-grained sentiment judgment engine determines the semantic sentiment label of the semantics through a preset semantic fine-grained sentiment judgment model, thereby clustering the semantics under the semantic intent label according to the semantic sentiment label to generate semantic cluster labels in which the semantics (e.g., text, question, evaluation, etc.) are sorted under the semantic intent label. Refer to the attached figure below. Figure 4 and5 The above multi-task learning process is described in detail, which includes the following sub-steps:

[0085] Step S1021, semantic processing step

[0086] The word embedding vector is used to process the input semantics to obtain the semantic embedding representation.

[0087] Step S1022, multi-task learning step

[0088] A BiLstm-based multi-task shared encoder is used for multi-task learning. The semantic intent recognition engine and the semantic fine-grained sentiment judgment engine share a double-layer BiLstm for feature extraction, so that when a semantic embedding representation is input, the hidden layer outputs a semantic sentence feature vector.

[0089] Step S1023, semantic intent recognition step

[0090] In the semantic intent recognition step, the semantic intent recognition model is used to perform semantic recognition. Specifically, the semantic intent label vector representation A is determined based on the output sentence feature vector using the fully connected layer, and the semantic intent label of the semantics is recognized based on the semantic intent label vector representation A through the softmax activation function.

[0091] Step S1024: semantic fine-grained sentiment judgment step

[0092] When performing semantic clustering according to semantic intent labels, it is necessary to sort the semantics under this label (semantic intent label). After exploring the data effects, it is found that semantic sentiment is the key feature dimension that affects online conversion. Therefore, when performing semantic sorting, semantic sentiment is mainly used as the main sorting factor.

[0093] In order to achieve refined analysis of semantic sentiment, the present invention uses the Aspect-based fine-grained sentiment analysis solution ABSA (Aspect based sentiment analysis) to achieve Aspect-based sentiment analysis. The main task of ABSA is to determine the sentiment polarity of the scene in which the sentence is located in a certain aspect. Since ABSA generally has two inputs: the text features of the sentence and the aspect words, the information of the entity (entity) or aspect (aspect) word is generally used to perform attention (attention mechanism) on the feature text to extract features with a high correlation with the entity and aspect words for sentiment analysis. Attention-based LSTM (Long Short-Term Memory Network) and CNN (Convolutional Neural Network) are frequently used algorithms, but since the computational efficiency of attention and LSTM is relatively low, in order to make full use of the information of entity and aspect words, CNN based on gate control mechanism (Gate-mechanisms) is also an important algorithm for solving such semantics.

[0094] Therefore, the present invention is based on the deep model GCAE (Gated Convolutional Networks Aspcet Embedding) under ABA to provide a key dimension of semantic sorting when performing semantic clustering according to semantic intent labels.

[0095] In the semantic fine-grained sentiment judgment step of the embodiment of the present invention, the semantic fine-grained sentiment judgment model is used to perform semantic fine-grained sentiment judgment, specifically, as follows: Figure 5 As shown, including:

[0096] Using a convolutional neural network (CNN), the representation features S_1 and S_2 of the sentence are obtained based on the sentence feature vector output in step S1022;

[0097] The feature representation S_1 is fed into a hyperbolic tangent activation function tanh to obtain an output vector tanh(S_1) representing the feature S_1 as a first output vector, and the feature representation S_2 and the semantic intent label vector representation A are added and fed into a rectified linear unit activation function relu to obtain an output vector relu(S_2+A) representing the sum of the feature representation S_2 and the semantic intent label vector representation A as a second output vector, and the first output vector tanh(S_1) and the second output vector relu(S_2+A) are element-wise-producted to obtain a final feature vector, thereby performing gate mechanism control; and

[0098] After the final feature vector is max pooled and passed through a fully connected layer, a softmax activation function is used to obtain the final classification result as the semantic sentiment label. When semantic clustering is performed based on the semantic intent label, the semantic sentiment label is the main ranking factor when the semantics are sorted (semantic clustering).

[0099] The above calculation process gets rid of time-consuming and memory-consuming network structures such as RNN (recurrent neural network) and attention (attention mechanism). At the same time, the gate mechanism makes good use of the information of aspect words, which increases the accuracy of the model.

[0100] In the above-mentioned step S102 of semantic intent recognition and semantic fine-grained sentiment judgment based on multi-task learning, a sentence feature vector is obtained by sharing a double-layer Bilstm. On the one hand, semantic intent recognition is performed based on the sentence feature vector through a fully connected layer, and on the other hand, the sentence feature vector can also be used as an input parameter for fine-grained sentiment judgment, which greatly saves the model network storage space and improves the execution efficiency of the entire process.

[0101] It should be pointed out that the Loss (loss function) construction of the above multi-task learning is mainly composed of two tasks in this paper: semantic intent recognition and semantic fine-grained sentiment judgment, so the loss of multi-task learning is also composed of two parts: Loss = Loss_senti + Loss_tag, where: Loss_senti is the Entropy Loss (cross quotient loss) of sentiment recognition of semantic intent tags, and Loss_tag is the Entropy Loss (cross quotient loss) of semantic intent tag recognition. The Loss calculation of multi-task recognition of the present invention does not adopt the weighted calculation method, that is, it is assumed that the importance of the two tasks is the same. After the Loss calculation is completed, the error backpropagation of the entire network structure and the parameter gradient update are performed.

[0102] The above describes in detail the multi-task label generation method of the present invention. The multi-task label generation method of the embodiment of the present invention can mine the semantic dimensions that users care about based on user semantic data to establish semantic intent labels, and perform semantic intent recognition and semantic fine-grained sentiment judgment based on multi-task learning, associate semantics with semantic intent labels, and provide key dimensions of semantic sorting for semantic clustering, thereby improving the execution efficiency of the entire solution. This can improve the user's decision-making efficiency, enhance the user's usage experience, and improve the conversion effect of traffic, such as clicks, dwell time, etc.

[0103] The above describes an embodiment of a label generation method based on multiple tasks according to the present invention. In another aspect of the present invention, a label generation device based on multiple tasks is also provided.

[0104] The following describes a multi-task based label generation device 200 according to an embodiment of the present invention. As shown in Figure 6, the multi-task based label generation device 200 includes a semantic intent label mining module 201 and a semantic intent recognition and semantic fine-grained sentiment judgment module 202 based on multi-task learning.

[0105] The following will refer to the attached Figure 6-8 The above steps are described in detail. Figure 6 is a schematic diagram of main modules of a multi-task based label generation device according to an embodiment of the present invention; Figure 7 is a schematic diagram of main components of a semantic intent tag mining module of a multi-task based tag generation device according to an embodiment of the present invention; Figure 8 It is a schematic diagram of the main components of the semantic intent recognition and semantic fine-grained sentiment judgment modules based on multi-task learning of the multi-task based label generation device according to an embodiment of the present invention.

[0106] Semantic intent label mining module 201

[0107] The semantic intent label mining module 201 is used to generate semantic intent labels based on semantic data. Specifically, the semantic intent label mining module 201 includes the following units: a preprocessing unit 2011, a candidate intent label word generation unit 2012, a screening unit 2013, and an aggregation and summary unit 2014. The above-mentioned units are described in detail below.

[0108] Preprocessing Unit 2011

[0109] The preprocessing unit 2011 is used to perform preprocessing such as word segmentation and removal of body words on the semantic data to obtain words.

[0110] Candidate intent label word generation unit 2012

[0111] The candidate intent label word generating unit 2012 is used to aggregate the obtained words into phrases to generate candidate intent label words.

[0112] Specifically, the candidate intent label word generation unit 2012 generates candidate intent label words based on word2phrase. The idea of ​​word2phrase is to aggregate words into phrases based on mutual information. For words that often appear together, the judgment of whether two words form a phrase is completed based on the following formula:

[0113]

[0114] The vectors of two words are input. If the calculated score is greater than a preset threshold, the candidate intention label word generation unit 2012 determines that the two words are "together".

[0115] In order to take longer phrases into consideration, the candidate intention label word generation unit 2012 uses 2-4 words as training data and sequentially lowers the threshold, thereby obtaining candidate intention label words.

[0116] Screening Unit 2013

[0117] The screening unit 2013 is used to perform word frequency statistics on candidate intent label words, filter low-frequency label words, and perform phrase fluency detection. For example, the screening unit 2013 is based on the N-gram language model (this language model is the product of the previous word2phrase) to filter out candidate intent label words whose fluency is lower than a predetermined value.

[0118] Aggregation and Summarization Unit 2014

[0119] The aggregation and summarization unit 2014 is used to aggregate the candidate intent label words after screening. The number of aggregations is set according to the actual situation of the category (general setting: 100-200), and the results after aggregation are manually screened and summarized to obtain the central phrase representing the category family as the semantic intent label.

[0120] For example, Figure 3 As shown, the aggregation and summarization unit 2014 aggregates the candidate intent label words such as easy to wear, comfortable to wear, and comfortable to wear, and manually selects and summarizes them into the central phrase of "wearing comfort" as the semantic intent label.

[0121] The aggregation and summarization unit 2014 summarizes the filtered central phrases to obtain, for example, the core intent phrases sorted out under the mobile phone category as shown in Table 1 above, that is, semantic intent labels.

[0122] Semantic intent recognition and semantic fine-grained sentiment judgment module based on multi-task learning202

[0123] The present invention models semantic intent recognition and semantic fine-grained sentiment judgment in a multi-task learning manner, that is, semantic intent recognition and semantic fine-grained sentiment judgment are performed based on multi-task learning, wherein semantic intent recognition realizes the association of semantics with the semantic intent labels obtained by the semantic intent label mining module 201, and semantic fine-grained sentiment judgment determines the semantic sentiment labels of semantics, thereby clustering the semantics under the semantic intent labels according to the semantic sentiment labels to generate the semantic clustering labels. Figure 7 and 8The semantic intent recognition and semantic fine-grained sentiment judgment module 202 based on multi-task learning is described in detail, and includes the following units:

[0124] Semantic Processing Unit 2021

[0125] The semantic processing unit 2021 processes the input semantics using a word embedding vector to obtain an embedded representation of the semantics.

[0126] Multi-Task Learning Unit 2022

[0127] The multi-task learning unit 2022 uses a BiLstm-based multi-task shared encoder (Encoder) to perform multi-task learning, so that semantic intent recognition and semantic fine-grained sentiment judgment share a double-layer BiLstm for feature extraction, so that when the semantic embedding representation is input, the hidden layer outputs the semantic sentence feature vector.

[0128] Semantic Intent Recognition Unit 2023

[0129] The semantic intent recognition unit 2023 uses the fully connected layer to determine the semantic intent label vector representation A based on the output sentence feature vector, and identifies the semantic intent label of the semantics based on the semantic intent label vector representation A through the softmax activation function.

[0130] Semantic fine-grained sentiment judgment unit 2024

[0131] When performing semantic clustering according to semantic intent labels, it is necessary to sort the semantics under this label (semantic intent label). After exploring the data effects, it is found that semantic sentiment is the key feature dimension that affects online conversion. Therefore, when performing semantic sorting, semantic sentiment is mainly used as the main sorting factor.

[0132] In order to achieve refined analysis of semantic sentiment, the present invention uses the Aspect-based fine-grained sentiment analysis solution ABSA (Aspect based sentiment analysis) to achieve Aspect-based sentiment analysis. The main task of ABSA is to determine the sentiment polarity of the scene in which the sentence is located in a certain aspect. Since ABSA generally has two inputs: the text features of the sentence and the aspect words, the information of the entity (entity) or aspect (aspect) word is generally used to perform attention (attention mechanism) on the feature text to extract features with a high correlation with the entity and aspect words for sentiment analysis. Attention-based LSTM (Long Short-Term Memory Network) and CNN (Convolutional Neural Network) are frequently used algorithms, but since the computational efficiency of attention and LSTM is relatively low, in order to make full use of the information of entity and aspect words, CNN based on gate control mechanism (Gate-mechanisms) is also an important algorithm for solving such semantics.

[0133] Therefore, the semantic fine-grained sentiment judgment unit 2024 provides the key dimension of semantic sorting, i.e., sorting factor, when performing semantic clustering according to semantic intent labels based on the deep model GCAE (Gated Convolutional Networks Aspcet Embedding) under ABSA. Figure 5 As shown, the specific calculation process of the semantic fine-grained sentiment judgment unit 2024 is as follows:

[0134] Using a convolutional neural network (CNN), the representation features S_1 and S_2 of the sentence are obtained based on the sentence feature vector output in step S1022;

[0135] The feature representation S_1 is fed into a hyperbolic tangent activation function tanh to obtain an output vector tanh(S_1) representing the feature S_1 as a first output vector, and the feature representation S_2 and the semantic intent label vector representation A are added and fed into a rectified linear unit activation function relu to obtain an output vector relu(S_2+A) representing the sum of the feature representation S_2 and the semantic intent label vector representation A as a second output vector, and the first output vector tanh(S_1) and the second output vector relu(S_2+A) are element-wise-producted to obtain a final feature vector, thereby performing gate mechanism control; and

[0136] After the final feature vector is max pooled and passed through a fully connected layer, a softmax activation function is used to obtain the final classification result as the semantic sentiment label. When semantic clustering is performed based on the semantic intent label, the semantic sentiment label is the main ranking factor when the semantics are sorted.

[0137] The above calculation process gets rid of time-consuming and memory-consuming network structures such as RNN (recurrent neural network) and attention (attention mechanism). At the same time, the gate mechanism makes good use of the information of aspect words, which increases the accuracy of the model.

[0138] The above-mentioned semantic intent recognition and semantic fine-grained sentiment judgment module 202 based on multi-task learning obtains a sentence feature vector by sharing a two-layer Bilstm. On the one hand, semantic intent recognition is performed based on the sentence feature vector through a fully connected layer. On the other hand, the sentence feature vector can also be used as an input parameter for fine-grained sentiment judgment, which greatly saves the model network storage space and improves the execution efficiency of the entire process.

[0139] The above describes in detail the multi-task label generation device of the present invention. The multi-task label generation device of the embodiment of the present invention can mine the semantic dimensions that users care about based on user semantic data to establish semantic intent labels, and perform semantic intent recognition and semantic fine-grained sentiment judgment based on multi-task learning, associate semantics with semantic intent labels, and provide key dimensions of semantic sorting for semantic clustering, thereby improving the execution efficiency of the entire solution. This can improve the user's decision-making efficiency, enhance the user's usage experience, and improve the conversion effect of traffic, such as clicks, dwell time, etc.

[0140] Reference below Figure 9-1 1. Describe the specific implementation of semantic intent recognition and semantic fine-grained sentiment judgment based on multi-task learning in the multi-task based label generation method and device according to the present invention.

[0141] like Fig. 9 As shown, enter the text "Supor's soup pot is of good quality, but the black spots inside the pot are difficult to clean after cooking porridge. Has anyone noticed this situation?"

[0142] At this time, after processing in step S102 of the present invention or module 202 of the present invention, the semantic intention labels "product quality" and "difficulty of cleaning" are finally generated, and the corresponding semantic emotion labels "3" and "1" are generated.

[0143] Therefore, the above input text "Supor's soup pot is of good quality, but the black spots in the pot are difficult to wash after cooking porridge. Has anyone noticed this situation?" can be clustered under the "product quality" and "difficulty of cleaning" labels, and the sentiment labels "3" and "1" are used as the sorting basis for the input text under the "product quality" and "difficulty of cleaning" labels.

[0144] Further, refer to Fig.10 As shown in FIG. 11 , the effect of applying the multi-task-based label generation method and device of the present invention can be intuitively displayed.

[0145] like Fig.10 As shown, when multi-task based label generation is not performed, most users may not be able to find satisfactory results when inputting according to their own accustomed input methods, and the semantic data is relatively messy.

[0146] In contrast, the specific embodiment of the label generation method and device based on multi-task according to the present invention is shown in FIG11, which performs clustering optimization processing on the question and answer data, sorts and clusters the questions under the question intention label based on the question sentiment label, and obtains the semantic clustering result shown in FIG11. In this case, the conversion rate of indicators such as the user's click-through rate and the user's stay time is greatly improved, which improves the user's shopping experience and usage experience to a certain extent, and can not only mine the semantic dimensions that the user cares about, but also provide high-quality positive semantics as a reference according to the user's intention, thereby facilitating the user's decision-making.

[0147] Fig.12 An exemplary system architecture 1200 is shown to which a multi-task based label generation method or a multi-task based label generation apparatus according to an embodiment of the present invention can be applied.

[0148] like Fig.12 As shown, the system architecture 1200 may include terminal devices 1201, 1202, 1203, a network 1204, and a server 1205. The network 1204 is used to provide a medium for communication links between the terminal devices 1201, 1202, 1203 and the server 1205. The network 1204 may include various connection types, such as wired, wireless communication links, or optical fiber cables, etc.

[0149] Users can use terminal devices 1201, 1202, 1203 to interact with server 1205 through network 1204 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 1201, 1202, 203, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social platform software, etc. (only as examples).

[0150] The terminal devices 1201 , 1202 , and 1203 may be various electronic devices having a display screen and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, and desktop computers, etc.

[0151] The server 1205 may be a server that provides various services, such as a backend management server (only an example) that provides support for shopping websites browsed by users using the terminal devices 1201, 1202, and 1203. The backend management server may analyze and process the received data such as product information query requests, and feed back the processing results to the terminal device.

[0152] It should be noted that the multi-task based label generation method provided in the embodiment of the present invention is generally executed by the server 1205 , and accordingly, the multi-task based label generation device is generally set in the server 1205 .

[0153] It should be understood that Fig.12 The number of terminal devices, networks and servers in the embodiment is only for illustration. Any number of terminal devices, networks and servers may be provided according to implementation requirements.

[0154] Reference below Fig.13 , which shows a schematic diagram of the structure of a computer system 1300 of a terminal device suitable for implementing an embodiment of the present invention. Fig.13 The terminal device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.

[0155] like Fig.13 As shown, the computer system 1300 includes a central processing unit (CPU) 1301, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1302 or a program loaded from a storage part 1308 into a random access memory (RAM) 1303. In the RAM 1303, various programs and data required for the operation of the system 1300 are also stored. The CPU 1301, the ROM 1302, and the RAM 1303 are connected to each other via a bus 1304. An input / output (I / O) interface 1305 is also connected to the bus 1304.

[0156] The following components are connected to the I / O interface 1305: an input section 1306 including a keyboard, a mouse, etc.; an output section 1307 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 1308 including a hard disk, etc.; and a communication section 1309 including a network interface card such as a LAN card, a modem, etc. The communication section 1309 performs communication processing via a network such as the Internet. A drive 1310 is also connected to the I / O interface 1305 as needed. A removable medium 1311, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 1310 as needed, so that a computer program read therefrom is installed into the storage section 1308 as needed.

[0157] In particular, according to the embodiments disclosed in the present invention, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present invention include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from the network through the communication part 1309, and / or installed from the removable medium 1311. When the computer program is executed by the central processing unit (CPU) 1301, the above-mentioned functions defined in the system of the present invention are executed.

[0158] It should be noted that the computer-readable medium shown in the present invention may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present invention, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device. In the present invention, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries a computer-readable program code. This propagated data signal may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, which may send, propagate or transmit a program for use by or in conjunction with an instruction execution system, apparatus or device. The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0159] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present invention. In this regard, each box in the flow chart or block diagram can represent a module, a program segment, or a part of a code, and the above-mentioned module, program segment, or a part of a code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flow chart, and the combination of the boxes in the block diagram or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0160] The modules and units involved in the embodiments of the present invention may be implemented in software or hardware. The modules and units described may also be arranged in a processor. For example, they may be described as follows: a processor includes a semantic intent label mining module and a semantic intent recognition and semantic fine-grained sentiment judgment module based on multi-task learning. Among them, the names of these modules do not constitute a limitation on the module itself in some cases. For example, the semantic intent recognition and semantic fine-grained sentiment judgment module based on multi-task learning may also be described as a "multi-task learning module".

[0161] As another aspect, the present invention further provides a computer-readable medium, which may be included in the device described in the above embodiment; or may exist independently without being assembled into the device. The above computer-readable medium carries one or more programs, and when the above one or more programs are executed by a device, the device includes:

[0162] Processing the semantic data through a preset semantic intent label mining model to generate semantic intent labels; and

[0163] The preset semantic intent recognition engine and semantic fine-grained sentiment judgment engine are called respectively to perform semantic intent recognition and semantic fine-grained sentiment judgment based on multi-task learning.

[0164] Among them, the semantic intent recognition engine associates the semantics with the semantic intent label through a preset semantic intent recognition model, and the semantic fine-grained sentiment judgment engine judges the semantic sentiment label of the semantics through a preset semantic fine-grained sentiment judgment model, thereby clustering the semantics under the semantic intent label according to the semantic sentiment label.

[0165] According to the technical solution of the embodiment of the present invention, it is possible to mine the semantic dimensions that users care about based on user semantic data to establish semantic intent labels, and to perform semantic intent recognition and semantic fine-grained sentiment judgment based on multi-task learning, associate semantics with semantic intent labels, and provide key dimensions of semantic sorting for semantic clustering, thereby improving the execution efficiency of the entire solution. This can improve the user's decision-making efficiency, enhance the user's usage experience, and improve the conversion effect of traffic, such as clicks, dwell time, etc.

[0166] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may occur depending on design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A multi-task based label generation method, It is characterized in that include: Processing semantic data through a preset semantic intent label mining model to generate semantic intent labels; as well as The preset semantic intent recognition engine and semantic fine-grained sentiment judgment engine are called respectively to perform semantic intent recognition and semantic fine-grained sentiment judgment based on multi-task learning. The semantic intent recognition engine associates the semantics with the semantic intent label through a preset semantic intent recognition model, and the semantic fine-grained emotion judgment engine judges the semantic emotion label of the semantics through a preset semantic fine-grained emotion judgment model, thereby clustering the semantics under the semantic intent label according to the semantic emotion label to generate a semantic cluster label; Among them, the process of multi-task learning includes: Use word embedding vectors to process the input semantics to obtain semantic embedding representation; The semantic intent recognition engine and the semantic fine-grained sentiment judgment engine share a double-layer bidirectional long short-term memory network for feature extraction, and the double-layer bidirectional long short-term memory network inputs the embedding representation of the semantics, and the hidden layer of the double-layer bidirectional long short-term memory network outputs the sentence feature vector of the semantics; The semantic intent recognition model includes determining a semantic intent label vector representation A based on the sentence feature vector using a fully connected layer, and identifying the semantic intent label of the semantics through a softmax activation function based on the semantic intent label vector representation A; Among them, the semantic fine-grained sentiment judgment model includes: obtaining semantic representation features S_1 and S_2 based on the sentence feature vector through a layer of convolutional neural network; sending the representation feature S_1 into the hyperbolic tangent activation function tanh to obtain the output vector tanh(S_1) of the representation feature S_1 as the first output vector, adding the representation feature S_2 and the semantic intent label vector representation A and sending them into the rectified linear unit activation function relu to obtain the output vector relu(S_2+A) of the sum of the representation feature S_2 and the semantic intent label vector representation A as the second output vector, and performing element-wise product of the first output vector and the second output vector to obtain the final feature vector; and, after maximum pooling the final feature vector and passing through the fully connected layer, using the softmax activation function to obtain the final classification result as the semantic sentiment label.

2. The method according to claim 1, It is characterized in that in, When semantic clustering is performed according to the semantic intent label, the semantic sentiment label is a ranking factor when the semantics are clustered under the semantic intent label.

3. The method according to any one of claims 1 to 2, It is characterized in that in, The semantic intent label mining model includes: Based on the semantic data, semantic preprocessing is performed to obtain words; Aggregating the words into phrases to generate candidate intent label words; Performing low-frequency filtering and fluency detection on the candidate intent label words for screening; and The screened candidate intent label words are aggregated, and the aggregated results are manually summarized to obtain a central phrase representing the class family as the semantic intent label.

4. The method according to any one of claims 1 to 2, It is characterized in that in, The loss of the multi-task learning is equal to the sum of the cross entropy loss of the semantic intent recognition and the cross entropy loss of the semantic fine-grained sentiment judgment.

5. A label generation device based on multi-task, It is characterized in that include: A semantic intent label mining module, which is used to generate semantic intent labels based on semantic data; as well as A semantic intent recognition and semantic fine-grained sentiment judgment module based on multi-task learning, wherein the semantic intent recognition and semantic fine-grained sentiment judgment module is used to perform semantic intent recognition and semantic fine-grained sentiment judgment based on multi-task learning, associate semantics with the semantic intent label using the semantic intent recognition, determine the semantic sentiment label of the semantics using the semantic fine-grained sentiment judgment, and cluster the semantics under the semantic intent label according to the semantic sentiment label to generate a semantic cluster label; The semantic intent recognition and semantic fine-grained sentiment judgment module based on multi-task learning includes the following units: The semantic processing unit processes the input semantics using the word embedding vector to obtain the semantic embedding representation; A multi-task learning unit, wherein the semantic intent recognition engine and the semantic fine-grained sentiment judgment engine share a double-layer bidirectional long short-term memory network for feature extraction, and the double-layer bidirectional long short-term memory network inputs the embedding representation of the semantics, and the hidden layer of the double-layer bidirectional long short-term memory network outputs the sentence feature vector of the semantics; The semantic intent recognition unit determines the semantic intent label vector representation A based on the output sentence feature vector using the fully connected layer, and recognizes the semantic intent label of the semantics through the softmax activation function based on the semantic intent label vector representation A; The semantic fine-grained sentiment judgment unit obtains semantic representation features S_1 and S_2 based on the sentence feature vector through a layer of convolutional neural network; the representation feature S_1 is sent to the hyperbolic tangent activation function tanh to obtain the output vector tanh(S_1) of the representation feature S_1 as the first output vector, the representation feature S_2 and the semantic intent label vector representation A are added and sent to the rectified linear unit activation function relu to obtain the output vector relu(S_2+A) of the sum of the representation feature S_2 and the semantic intent label vector representation A as the second output vector, and the first output vector and the second output vector are element-wise multiplied to obtain the final feature vector; and, after maximum pooling the final feature vector and passing through the fully connected layer, the softmax activation function is used to obtain the final classification result as the semantic sentiment label.

6. A multi-task based label generation electronic device, It is characterized in that include: one or more processors; a storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 4.

7. A computer readable medium having a computer program stored thereon, It is characterized in that When the program is executed by a processor, the method according to any one of claims 1 to 4 is implemented.

Images