Method and device for processing dialogue information, equipment, and storage medium

Abstract

The one or more embodiments of the specification provide a processing method and device of dialogue information, equipment and storage medium; the method can include: obtaining user data, the user data includes conversation content input by a user in at least one round of dialogue and auxiliary reference information related to the user distinguished from the conversation content;Respectively extracting features of the conversation content and the auxiliary reference information to obtain corresponding feature vectors, and splicing the feature vectors to obtain a representation vector;The representation vector is input into a preset machine learning model, and the output information of the machine learning model is used to predict the response result of the user to a preset interaction request;If the predicted response result is response success, trigger the process for the preset interaction request.

Description

Technical Field

[0001] This specification relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device, and storage medium for processing dialogue information. Background Technology

[0002] In related technologies, machine learning techniques utilize algorithms to learn from existing data and make judgments and decisions about real-world situations. Machine learning techniques include supervised learning, unsupervised learning, semi-supervised learning, reinforcement learning, and so on.

[0003] In the training process of supervised learning, the input sample data is called the "training set". The sample data in the training set has a clear identifier or result (i.e., sample label). When using supervised learning algorithms to build machine learning models, the supervised learning algorithm establishes a learning process, compares the prediction results with the actual results of the "training set", and continuously adjusts the machine learning model until the model's prediction results reach a expected accuracy. Summary of the Invention

[0004] In view of the above, one or more embodiments of this specification provide a method, apparatus, device, and storage medium for processing dialogue information.

[0005] To achieve the above objectives, one or more embodiments of this specification provide the following technical solutions:

[0006] According to a first aspect of one or more embodiments of this specification, a method for processing dialogue information is provided, comprising:

[0007] Acquire user data, which includes conversation content entered by the user in at least one round of dialogue and auxiliary reference information related to the user that is different from the conversation content;

[0008] Feature vectors are obtained by extracting features from the conversation content and the auxiliary reference information respectively, and the feature vectors are concatenated to obtain a representation vector;

[0009] The representation vector is input into a preset machine learning model, and the output information of the machine learning model is used to predict the user's response to the preset interaction request.

[0010] If the predicted response is successful, the process for the preset interaction request is triggered. According to a second aspect of one or more embodiments of this specification, a method for training a machine learning model based on dialogue information is proposed, comprising:

[0011] Acquire user sample data and the user's response to a preset interaction request. The user sample data includes the conversation content entered by the user in at least one round of dialogue and auxiliary reference information related to the user that is different from the conversation content.

[0012] Feature vectors are obtained by extracting features from the conversation content and the auxiliary reference information respectively, and the feature vectors are concatenated to obtain a representation vector;

[0013] The machine learning model is trained using the representation vector and the response result.

[0014] According to a third aspect of one or more embodiments of this specification, a dialogue information processing apparatus is provided, comprising:

[0015] The acquisition unit acquires user data, which includes conversation content input by the user in at least one round of dialogue and auxiliary reference information related to the user that is different from the conversation content.

[0016] The extraction unit performs feature extraction on the session content and the auxiliary reference information respectively to obtain corresponding feature vectors, and concatenates the feature vectors to obtain a representation vector;

[0017] The input unit inputs the representation vector into a preset machine learning model and uses the output information of the machine learning model to predict the user's response to the preset interaction request.

[0018] If the predicted response result is a successful response, the triggering unit will trigger the process for the preset interaction request.

[0019] According to a fourth aspect of one or more embodiments of this specification, a machine learning model training apparatus based on dialogue information is provided, comprising:

[0020] The acquisition unit acquires user sample data of the user and the response results of the user to a preset interaction request. The user sample data includes the conversation content entered by the user in at least one round of dialogue and auxiliary reference information related to the user that is different from the conversation content.

[0021] The extraction unit performs feature extraction on the session content and the auxiliary reference information respectively to obtain corresponding feature vectors, and concatenates the feature vectors to obtain a representation vector;

[0022] The training unit trains a machine learning model using the representation vector and the response result.

[0023] According to a fifth aspect of one or more embodiments of this specification, an electronic device is provided, comprising:

[0024] processor;

[0025] Memory used to store processor-executable instructions;

[0026] The processor executes the executable instructions to implement the method as described in any of the above embodiments.

[0027] According to a sixth aspect of one or more embodiments of this specification, a computer-readable storage medium is provided that stores computer instructions thereon, which, when executed by a processor, implement the steps of the method as described in any of the above embodiments.

[0028] As can be seen from the above embodiments, for users engaged in dialogue, the technical solution provided in this specification improves the training set of the machine learning model used to predict whether the user will respond to an interaction request. During the training of this machine learning model, the training set integrates the conversational content and auxiliary reference information related to the user, that is, it extracts the textual features of the conversational content and other non-textual features related to the user, thereby improving the prediction accuracy of the trained machine learning model.

[0029] Correspondingly, when using the trained machine learning model to predict the user's current interaction, features from the user's current conversation content and auxiliary reference information are extracted and input into the model for prediction. Based on the model's output, the model can predict the user's response to a preset interaction request. If the predicted response is successful, the process for the preset interaction request is triggered. Through this prediction process, the user's need for a preset interaction request can be identified promptly, and the relevant process can be proactively triggered in a timely manner. This improves the communication efficiency of the conversation, simplifies the user's actions in triggering the process, and enhances the user experience. Attached Figure Description

[0030] Figure 1 This is a flowchart of an exemplary embodiment of a machine learning model training method based on dialogue information.

[0031] Figure 2 This is a flowchart of a method for processing dialogue information provided in an exemplary embodiment.

[0032] Figure 3 This is a schematic diagram of the architecture of a session system provided in an exemplary embodiment.

[0033] Figure 4 This is a flowchart of another method for training a machine learning model based on dialogue information, provided in an exemplary embodiment.

[0034] Figure 5This is a schematic diagram of a machine learning model provided in an exemplary embodiment.

[0035] Figure 6 This is a flowchart of another method for processing dialogue information provided in an exemplary embodiment.

[0036] Figure 7 This is a schematic diagram of a terminal device provided in an exemplary embodiment.

[0037] Figure 8 This is a block diagram of a machine learning model training apparatus based on dialogue information, provided in an exemplary embodiment.

[0038] Figure 9 This is a schematic diagram of the structure of another device provided in an exemplary embodiment.

[0039] Figure 10 This is a block diagram of a dialogue information processing apparatus provided in an exemplary embodiment. Detailed Implementation

[0040] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with one or more embodiments of this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of one or more embodiments of this specification as detailed in the appended claims.

[0041] It should be noted that the steps of the corresponding methods are not necessarily performed in the order shown and described in this specification in other embodiments. In some other embodiments, the methods may include more or fewer steps than described in this specification. Furthermore, a single step described in this specification may be broken down into multiple steps in other embodiments; and multiple steps described in this specification may be combined into a single step in other embodiments.

[0042] In related technologies, business platforms can provide online customer service to users through a conversational system, allowing users to engage in human-computer dialogue with an online chatbot using their client applications. For example, users can ask questions, handle business transactions, and inquire about information through the online chatbot.

[0043] This specification aims to provide a training scheme for a machine learning model based on dialogue information and a processing scheme for dialogue information based on the model. By improving the training set of the machine learning model, the prediction accuracy of the trained model can be improved. In this way, the model can be used to promptly discover the user's needs for preset interaction requests in the current dialogue and proactively trigger the relevant process, thereby improving the communication efficiency of the user's dialogue and simplifying the user's operation of triggering the process.

[0044] Please see Figure 1 , Figure 1 This is a flowchart illustrating an exemplary embodiment of a machine learning model training method based on dialogue information. Figure 1 As shown, this method, when applied to any electronic device that can be used to train a machine learning model, may include the following steps:

[0045] Step 102: Obtain user sample data and the user's response to the preset interaction request. The user sample data includes the conversation content entered by the user in at least one round of dialogue and auxiliary reference information related to the user that is different from the conversation content.

[0046] In this embodiment, a conversational system providing online customer service can establish a dialogue with a user's client. The user inputs conversation content through the client, and the online chatbot of the conversational system responds, thus realizing the dialogue. The conversational system can store the conversation content input by each user in initiated conversations (i.e., historical conversation content) through system logs, and can also store auxiliary reference information related to the user during the conversation. Based on the system logs configured in the conversational system, when it is necessary to obtain user data (at least one) to train the model, the conversation content and corresponding auxiliary reference information stored in the system logs of the conversational system used to establish the dialogue can be obtained.

[0047] Each user input and corresponding response can be considered as one round of dialogue. Taking a conversational system provided by a network operator as an example, user A engages in three rounds of human-computer dialogue with the online robot of the conversational system, as shown in Table 1:

[0048]

[0049]

[0050] Table 1

[0051] In other words, a question-and-answer session is counted as one round of human-computer dialogue. In each round of human-computer dialogue, the conversation content input by the user is used as part of the user sample data for training.

[0052] It should be noted that the above-mentioned scenario of network operators providing online customer service is merely an example, and the dialogue scenario in this specification can also be applied to any other type of business platform. For example, e-commerce, government affairs, finance, education, entertainment, health, tourism, and other types of business platforms can all provide online customer service to users.

[0053] Taking e-commerce as an example, e-commerce platforms' chat systems provide online customer service. Users can interact with online chatbots through their clients, asking questions about promotional activities, product information, and user rights. In this scenario, the preset interaction request could be a product purchase request, inquiring whether the user needs to buy the corresponding product. A machine learning model predicts whether the user will respond to this purchase request, i.e., whether the user will buy the product recommended in the request. If the product meets the user's needs, the user can respond to the purchase request through the client, thus completing the purchase. Through this predictive process, the potential product needs of the user in the current conversation can be identified in a timely manner, and the relevant purchase process can be proactively triggered. On the one hand, this helps merchants better recommend products to users; on the other hand, it improves the communication efficiency of user conversations, simplifies the product purchase process, and enhances the user experience.

[0054] As for auxiliary reference information, it may include at least one of the following: user profile, user location information, dialogue time information, and the user's historical response results to preset interaction requests.

[0055] Since the machine learning model in this manual is a supervised machine learning model, the user sample data needs to be labeled when training the model using the user sample data as the training set. The machine learning model in this manual is used to predict whether the user currently in a conversation will respond to a preset interaction request; therefore, the label for the user sample data is the user's response to the preset interaction request. The response result includes successful and failed responses. For example, a successful response is represented by the value "1", and a failed response is represented by the value "0".

[0056] In this embodiment, the function of the preset interaction request can be flexibly set according to actual needs. For example, in order to promptly identify potential business processing needs of users, the preset interaction request can be a business processing request, used to inquire whether the user wants to process the corresponding business. If the business meets the user's needs, the user can respond to the business processing request through the client to complete the business processing. As another example, the preset interaction request can be a product purchase request, used to inquire whether the user needs to purchase the corresponding product. If the product meets the user's needs, the user can respond to the product purchase request through the client to complete the purchase. Yet another example is a task reminder request, used to inquire whether the user needs to set a reminder for a pending task. If the user confirms that a reminder needs to be set for the pending task, they can respond to the task reminder request through the client to complete the reminder service settings. Of course, the preset interaction request can also be any other form of request, and this specification does not limit it.

[0057] Given that the aforementioned preset interactive request is a business processing request, when outputting the business processing request to the user's client, it can be output through a human customer service representative or by continuing to use a human-computer dialogue method.

[0058] Specifically, the human-machine interface of the dialogue can be switched to a human customer service interface, allowing the service request to be output through the human interface. For example, after switching to a human customer service interface, the customer service representative will converse with the user, and upon confirming that the user needs to handle the relevant service, will push the service request to the user's client through the conversation system for the user to respond to. Of course, it is also possible to output the service request directly through the human-machine interface of the dialogue without switching to a human interface, allowing the user to respond.

[0059] Step 104: Extract features from the conversation content and the auxiliary reference information to obtain corresponding feature vectors, and concatenate the feature vectors to obtain a representation vector.

[0060] During a dialogue, there are usually connections between each round, which often reflect the user's underlying needs, i.e., features that characterize the user's response to pre-defined interaction requests. Therefore, feature extraction can be performed by simultaneously selecting the conversation content entered by the user in a particular round of dialogue and the conversation content entered in several previous rounds of dialogue (i.e., historical rounds of dialogue) for subsequent training of a machine learning model, thereby improving the prediction accuracy of the trained machine learning model.

[0061] Specifically, the conversation content may include the conversation content entered by the user in any round of dialogue and the historical conversation content entered in several previous rounds of dialogue corresponding to that round. In this case, a first feature vector corresponding to the conversation content in that round of dialogue can be extracted, and the historical conversation content can be converted into a word sequence, and a second feature vector corresponding to the word sequence can be extracted. Then, the first feature vector and the second feature vector are concatenated to obtain the feature vector corresponding to the aforementioned conversation content.

[0062] For any given round of dialogue, CNN (Convolutional Neural Networks) can be used for feature extraction. Specifically, the conversation content in any given round of dialogue can first be vectorized to obtain a first text matrix, that is, the conversation content can be converted into the form of embedding vectors. Then, multiple convolutional kernels are used to convolve the first text matrix to obtain the corresponding convolutional feature vectors. Finally, the convolutional feature vectors corresponding to each convolutional kernel are concatenated to obtain the first feature vector.

[0063] For word sequences corresponding to historical conversation content, in one case, the same CNN-based feature extraction method described above can be used. Specifically, the word sequence is vectorized to obtain a second text matrix, that is, each word in the word sequence is converted into an embedding form, and multiple convolutional kernels are used to convolve the second text matrix to obtain corresponding convolutional feature vectors. The convolutional feature vectors corresponding to each kernel are then concatenated to obtain a first sequence feature vector corresponding to the word sequence. This first sequence feature vector can then be used as the second feature vector corresponding to the word sequence.

[0064] In another scenario, to better extract the spatial information between words in a word sequence and thus capture the contribution of each word to the overall historical conversation content, an attention mechanism is introduced. The attention mechanism can capture the dependencies and importance between words in the text, thereby better representing the text's features. Therefore, the attention mechanism can be used to determine the attention weight vector (also called attention probability) of the word vectors corresponding to each word in the word sequence. Based on the word vectors and attention weight vectors corresponding to each word in the word sequence, a second sequence feature vector corresponding to the word sequence is obtained. This second sequence feature vector can then be used as the second feature vector corresponding to the word sequence.

[0065] In another case, the first sequence feature vector and the second sequence feature vector can be concatenated to obtain the concatenated vector as the second feature vector corresponding to the word sequence.

[0066] In this embodiment, to further improve the prediction accuracy of the trained model, the user sample data may also include intent recognition results for the conversation content. For example, the conversation system can use semantic recognition algorithms in related technologies to identify user intent based on the conversation content input by the user.

[0067] For auxiliary reference information and intent recognition results, these data can also be converted into embedding form first, and then concatenated with the feature vectors corresponding to the aforementioned conversation content to obtain the representation vector corresponding to the user sample data. For example, following the above method of using convolutional neural networks for feature extraction, the feature vectors corresponding to each content (conversation content input in any round of dialogue, historical conversation content, auxiliary reference information, intent recognition results, etc.) can be concatenated in the fully connected layer of the convolutional neural network to obtain the representation vector corresponding to the user sample data.

[0068] Step 106: Train a machine learning model using the representation vector and the response result.

[0069] In this embodiment, any supervised learning algorithm can be used to train the representation vector and response results to obtain a machine learning model, and this specification does not impose any limitations on this. Examples include neural networks, decision trees, support vector machines, Bayesian classifiers, etc.

[0070] In response to the above Figure 1 The illustrated training model example is accompanied by corresponding dialogue information processing solutions provided in this specification. Please refer to... Figure 2 , Figure 2 This is a flowchart illustrating a method for processing dialogue information as provided in an exemplary embodiment. Figure 2 As shown, this method, applied to a session system, may include the following steps:

[0071] Step 202: Obtain user data, which includes conversation content entered by the user in at least one round of dialogue and auxiliary reference information related to the user that is different from the conversation content.

[0072] In this embodiment, based on the above Figure 1 The illustrated machine learning model can be configured in the conversation system. After establishing a dialogue with the user, the system can acquire the user's input conversation content and auxiliary reference information different from the conversation content in real time and input them into the machine learning model. Based on the output of the machine learning model, the system can predict the user's response to a preset interaction request. Specifically, if the predicted response (including success and failure) is successful, the process for the preset interaction request can be triggered, allowing the user to respond to the preset interaction request.

[0073] Step 204: Extract features from the conversation content and the auxiliary reference information to obtain corresponding feature vectors, and concatenate the feature vectors to obtain a representation vector.

[0074] As mentioned earlier, the conversation content may include the current conversation content entered by the user in the current round of dialogue and the historical conversation content entered in several previous rounds of dialogue. Therefore, a first feature vector corresponding to the current conversation content can be extracted first, and the historical conversation content can be converted into a word sequence, with a second feature vector corresponding to the word sequence extracted. Then, the first feature vector and the second feature vector are concatenated to obtain the feature vector corresponding to the conversation content.

[0075] As mentioned above, the current session content can be vectorized to obtain a first text matrix. Then, multiple convolution kernels are used to convolve the first text matrix to obtain corresponding convolution feature vectors. Finally, the convolution feature vectors corresponding to each convolution kernel are concatenated to obtain the first feature vector.

[0076] As mentioned above, the word sequence can be vectorized to obtain a second text matrix, and multiple convolution kernels can be used to convolve the second text matrix to obtain corresponding convolutional feature vectors. The convolutional feature vectors corresponding to each convolution kernel are then concatenated to obtain a first sequence feature vector corresponding to the word sequence. Next, an attention mechanism is used to determine the attention weight vectors of the word vectors corresponding to each word in the word sequence. Based on the word vectors and attention weight vectors corresponding to each word in the word sequence, a second sequence feature vector corresponding to the word sequence is obtained. Therefore, either the first sequence feature vector or the second sequence feature vector can be used as the second feature vector; or, the first sequence feature vector and the second sequence feature vector can be concatenated to obtain the second feature vector.

[0077] Step 206: Input the representation vector into a preset machine learning model, and use the output information of the machine learning model to predict the user's response to the preset interaction request.

[0078] Step 208: If the predicted response result is a successful response, trigger the process for the preset interaction request.

[0079] As mentioned above, the preset interaction request may include a business processing request. If the predicted response result is a successful response, the human-computer interface of the dialogue can be switched to the human interface of the human customer service dialogue, and the business processing request is output by the human interface.

[0080] Furthermore, regarding the switching to human customer service, when annotating the response results corresponding to the user sample data, the service type of the final transaction handled by the user can be indicated. Accordingly, the session system configures a human interface for each service type to output the corresponding service request. Therefore, if the predicted response result for a service request of any service type is successful, the human-machine interface can be switched to the human interface corresponding to that service type. In this way, each human interface is responsible for outputting the service request of its corresponding service type, allowing each customer service representative to be familiar with the corresponding service type, thus simplifying and refining their workload.

[0081] As mentioned above, the preset interaction request includes a business processing request. Therefore, if the predicted response to the business processing request is successful, the business processing request can be output through the human-computer interface of the human-computer dialogue for the user to respond to.

[0082] As mentioned above, the auxiliary reference information includes at least one of the following: the user's user profile, the user's location information, the time information of the dialogue, and the user's historical response results to preset interaction requests.

[0083] As previously mentioned, the user data also includes intent recognition results for the session content.

[0084] It should be noted that the above Figure 2 The contents of each part can be referenced above. Figure 1 The corresponding parts of the illustrated embodiment process user data in a similar manner to the above-described processing method for user sample data, and will not be repeated here.

[0085] To facilitate understanding, the following example uses a preset interactive request as a business processing request, combined with... Figure 3-6 The training process of the machine learning model and the process of using the machine learning model to predict whether a user will respond to a service request are described in detail.

[0086] Figure 3 This is a schematic diagram of the architecture of a session system provided in an exemplary embodiment. Figure 3 As shown, the system may include servers 31-34 and network 35. Servers 31-34 may be physical servers containing independent hosts or virtual servers hosted in a host cluster.

[0087] Server 31 stores the conversation content and intent recognition results for the human-computer dialogue, while server 32 stores the user's auxiliary reference information. Of course, the conversation content, intent recognition results, and user's auxiliary reference information can be stored on the same server; this specification does not impose this limitation. Server 33 retrieves relevant content from servers 31 and 32 as user sample data to train a machine learning model, and provides the trained machine learning model to server 34. Server 34 engages in human-computer dialogue with the user and uses the machine learning model to predict whether the user will respond to a business processing request. Simultaneously, server 34 can provide the conversation content and intent recognition results for the conversation content to server 31 for storage. It should be noted that server 33 can retrieve relevant content from servers 31 and 32 at preset intervals to train and update the machine learning model.

[0088] The network 35, which facilitates interaction between servers 31-34, can include various types of wired or wireless networks. For example, network 35 could include the Public Switched Telephone Network (PSTN) and the Internet.

[0089] Please see Figure 4 , Figure 4 This is a flowchart illustrating another method for training a machine learning model based on dialogue information, as provided in an exemplary embodiment. Figure 4 As shown, this method, when applied to any electronic device that can be used to train a machine learning model, may include the following steps:

[0090] Step 402: Extract the feature vector E1 of the current speech.

[0091] In this embodiment, user sample data can be obtained. This user sample data includes conversation content input by the user in at least one round of human-computer dialogue and external information. The external information includes auxiliary reference information for the conversation content and intent recognition results for the conversation content. For the conversation content in any round of human-computer dialogue, the conversation content can be understood as the current dialogue in this round of human-computer dialogue, and the human-computer dialogues before this round of human-computer dialogue can be understood as the dialogues in previous rounds.

[0092] CNNs can be used to extract features from the current speech. For example... Figure 5As shown, a CNN can include convolutional layers, pooling layers, and fully connected layers. Convolutional layers are used to compute feature vectors from the input data. Pooling layers, typically placed after convolutional layers, reduce the dimensionality of the feature vectors to simplify network computation and reduce the output feature vectors of the convolutional layers, thus preventing overfitting. Fully connected layers map the learned feature vectors to the label space of the samples, for example, transforming the two-dimensional feature vectors output by the pooling layers into one-dimensional vectors.

[0093] Assuming the current utterance is text of length L, after vectorization, the resulting embedding matrix is ​​W = [w1, w2, ..., w...]. L The process involves using multiple convolutional kernels of varying widths to perform convolutions along the direction of the current utterance. The number and width of the convolutional kernels can be empirical values; for example, five kernels could be used, each with widths of 2d, 3d, 4d, 5d, and 6d, where d represents the dimension of a word vector. Of course, any other number and width of convolutional kernels can be used, and this specification does not impose any restrictions on this.

[0094] After obtaining the feature matrix X through convolution kernel, the feature matrix X can be padded appropriately. Then, the feature matrix X can be reduced to a specific feature dimension through pooling operation. Average pooling and max pooling are performed on the feature vector at each position to obtain the corresponding feature vector. Finally, these feature vectors are concatenated to obtain the embedding vector E1 of the current speech.

[0095] Step 404: Extract the feature vector E2 of several rounds of historical dialogue.

[0096] Several rounds of dialogue can be concatenated into a word sequence in chronological order. Then, feature extraction is performed on this text using the method described in step 403 above to obtain E. a .

[0097] Furthermore, the dependencies and importance between words in a word sequence can be obtained through attention mechanisms, thereby better representing the features of the word sequence. Therefore, attention mechanisms can be used to determine the attention weight vector (also called attention probability) of the word vector corresponding to each word in the word sequence, and based on the word vector and attention weight vector corresponding to each word in the word sequence, the feature vector E corresponding to the word sequence can be obtained. b The eigenvector E b Characterizing word sequences from a categorical feature perspective. Therefore, vector E can be... a and E b The feature vector E2 is obtained by concatenating the features.

[0098] The following explanation uses LEAM (Label-Embedding Attentive Model), which incorporates an attention mechanism, as an example. LEAM treats text classification as a joint label-word embedding problem: each label and word vector are embedded in the same space. It proposes an attention framework to measure the compatibility between text sequences and label embeddings. The attention is learned on a set of labeled training samples to ensure that, given a text sequence, relevant words have higher weights than irrelevant words. This method maintains the interpretability of word embeddings and has the ability to utilize information beyond the input text sequence.

[0099] The specific implementation process of LEAM is as follows:

[0100] 1. Joint embedding of words and tags

[0101] Words and embeddings are grouped into the same space, and the word embedding matrix of a sentence is initialized using pre-trained word vectors. For descriptive class labels (class labels are words), the pre-trained word embeddings corresponding to each label are used as the initialization for the label embeddings. For datasets without representative class descriptions (e.g., labels are numbers), the label embeddings can be initialized with random samples drawn from a standard Gaussian distribution.

[0102] 2. Weight the word embedding matrix using tag information.

[0103] First calculate Where G^ is a normalized matrix of size K×L, and each element is the product of the l2 norms of the c-th label embedding and the L-th word embedding: g^ kl =||c k ||||v l Then, the weight vector β is calculated using G, as follows:

[0104] u l =ReLU(G l-r:l+r W1+b1)

[0105] Where W1∈R 2r+1 b1∈R K All of these are parameters that need to be learned. Gl-r:l+r represents a segment centered on word l, and the obtained u l ∈R K .

[0106] m l =max-pooling(u l )

[0107] Where m is a vector of length L, and the attention score of the text sequence is:

[0108] β = SoftMax(m)

[0109] Text representation is obtained through label-based attention score weighted word embeddings:

[0110] z = ∑ l β l u l

[0111] Step 406: Extract the feature vector E3 of external information.

[0112] Step 408: Concatenate to obtain the representation vector E.

[0113] External information may include the intent recognition results for the session content and the user's auxiliary reference information, as mentioned above. Similarly, the external information can first be converted into an embedding form to obtain a feature vector E3, and then the feature vectors E1-E3 can be concatenated to finally obtain the representation vector E corresponding to the user sample data. For example, it can be done as described above. Figure 5 The fully connected layer of the convolutional neural network shown concatenates the feature vectors E1-E3.

[0114] Step 410: Train the machine learning model.

[0115] For example, such as Figure 5 As shown, the representation vector E can be input into an LSTM (Long Short-Term Memory) network, and the classification vector generated by the LSTM can represent the user's response to the business processing request.

[0116] Please see Figure 6 , Figure 6 This is a flowchart illustrating another method for processing dialogue information provided in an exemplary embodiment. For example... Figure 6 As shown, the method may include the following steps:

[0117] Step 602: Extract the feature vector F1 of the current speech.

[0118] Step 604: Extract the feature vector F2 of several rounds of historical dialogue.

[0119] Step 606: Extract the feature vector F3 of external information.

[0120] Step 608: Concatenate to obtain the representation vector F.

[0121] For the specific implementation details of steps 602-608 above, please refer to steps 402-408 above, which will not be repeated here.

[0122] Step 610: Input the representation vector F into the machine learning model.

[0123] Step 612: If the output result is a successful response, proceed to step 614; otherwise, return to step 602.

[0124] Step 614: Switch to human customer service.

[0125] Step 616: Output the business processing request.

[0126] When the predicted response is successful, the human-computer dialogue interface can be switched to a human customer service dialogue interface, allowing the service request to be output through the human interface. For example, after switching to a human customer service dialogue, the customer service representative will converse with the user. Once it is confirmed that the user needs to handle the corresponding service, the service request will be pushed to the user's client through the conversation system (e.g., via a link) for the user to respond to.

[0127] Corresponding to the above method embodiments, this specification also provides an embodiment of a machine learning model training device based on human-computer dialogue information.

[0128] Figure 7 This is a schematic structural diagram of a device provided in an exemplary embodiment. Please refer to... Figure 7 At the hardware level, the device includes a processor 702, an internal bus 704, a network interface 706, memory 708, and non-volatile memory 710, and may also include other hardware required for business operations. The processor 702 reads the corresponding computer program from the non-volatile memory 710 into the memory 708 and then runs it, forming a machine learning model training device based on dialogue information at the logical level. Of course, besides the software implementation, one or more embodiments in this specification do not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution subject of the following processing flow is not limited to individual logic units, but can also be hardware or logic devices.

[0129] Please refer to Figure 8 In one software implementation, the training device may include:

[0130] The acquisition unit 81 acquires user sample data of the user and the response result of the user to the preset interaction request. The user sample data includes the conversation content entered by the user in at least one round of dialogue and auxiliary reference information related to the user that is different from the conversation content.

[0131] Extraction unit 82 performs feature extraction on the session content and the auxiliary reference information respectively to obtain corresponding feature vectors, and concatenates the feature vectors to obtain a representation vector;

[0132] Training unit 83 trains a machine learning model using the representation vector and the response result.

[0133] Optionally, the acquisition unit 81 is specifically used for:

[0134] Obtain the session content and auxiliary reference information stored in the system log of the session system used to establish the dialogue.

[0135] Optionally, the conversation content includes the conversation content entered by the user in any round of dialogue and the historical conversation content entered in several historical rounds of dialogue corresponding to that round of dialogue; the extraction unit 82 is specifically used for:

[0136] Extract the first feature vector corresponding to the conversation content in any of the rounds of dialogue;

[0137] The historical conversation content is converted into a word sequence, and a second feature vector corresponding to the word sequence is extracted;

[0138] The first feature vector and the second feature vector are concatenated to obtain a feature vector corresponding to the session content.

[0139] Optionally, the extraction unit 82 is further configured to: vectorize the conversation content in any round of dialogue to obtain a first text matrix;

[0140] The first text matrix is ​​convolved using multiple convolution kernels to obtain the corresponding convolutional feature vectors.

[0141] The first feature vector is obtained by concatenating the convolutional feature vectors corresponding to each convolutional kernel.

[0142] Optionally, the extraction unit 82 is further configured to: vectorize the word sequence to obtain a second text matrix, and perform convolution processing on the second text matrix using multiple convolution kernels to obtain corresponding convolution feature vectors, and concatenate the convolution feature vectors corresponding to each convolution kernel to obtain a first sequence feature vector corresponding to the word sequence;

[0143] An attention mechanism is used to determine the attention weight vector of the word vector corresponding to each word in the word sequence, and a second sequence feature vector corresponding to the word sequence is obtained based on the word vector and attention weight vector corresponding to each word in the word sequence.

[0144] The first sequence feature vector or the second sequence feature vector is used as the second feature vector; or, the first sequence feature vector and the second sequence feature vector are concatenated to obtain the second feature vector.

[0145] Optionally, the preset interaction request includes a business processing request, which is output by the human interface after the human-computer interface of the dialogue is switched to the human interface of the human customer service dialogue.

[0146] Optionally, the preset interaction request includes a business processing request; the business processing request is output by the human-machine interface of the dialogue.

[0147] Optionally, the auxiliary reference information includes at least one of the following:

[0148] The user's profile, the user's location information, the time information of the conversation, and the user's historical response results to preset interaction requests.

[0149] Optionally, the user sample data may also include intent recognition results for the session content.

[0150] Corresponding to the above method embodiments, this specification also provides an embodiment of a dialogue information processing apparatus.

[0151] Figure 9 This is a schematic structural diagram of a device provided in an exemplary embodiment. Please refer to... Figure 9 At the hardware level, the device includes a processor 902, an internal bus 904, a network interface 906, memory 908, and non-volatile memory 910, and may also include other hardware required for business operations. The processor 902 reads the corresponding computer program from the non-volatile memory 910 into memory 908 and then runs it, forming a processing device for dialogue information at the logical level. Of course, in addition to the software implementation, one or more embodiments in this specification do not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0152] Please refer to Figure 10 In one software implementation, the device for processing the dialogue information may include:

[0153] Acquisition unit 1001 acquires user data, the user data including conversation content input by the user in at least one round of dialogue and auxiliary reference information related to the user that is different from the conversation content;

[0154] The extraction unit 1002 performs feature extraction on the session content and the auxiliary reference information respectively to obtain corresponding feature vectors, and concatenates the feature vectors to obtain a representation vector;

[0155] The input unit 1003 inputs the representation vector into a preset machine learning model and uses the output information of the machine learning model to predict the user's response to the preset interaction request.

[0156] If the predicted response result is a successful response, the triggering unit 1004 triggers the process for the preset interaction request.

[0157] Optionally, the conversation content includes the current conversation content entered by the user in the current round of dialogue and the historical conversation content entered in several previous rounds of dialogue; the extraction unit 1002 is specifically used for:

[0158] Extract the first feature vector corresponding to the current session content;

[0159] The historical conversation content is converted into a word sequence, and a second feature vector corresponding to the word sequence is extracted;

[0160] The first feature vector and the second feature vector are concatenated to obtain a feature vector corresponding to the session content.

[0161] Optionally, the extraction unit 1002 is further configured to:

[0162] The current session content is vectorized to obtain the first text matrix;

[0163] The first text matrix is ​​convolved using multiple convolution kernels to obtain the corresponding convolutional feature vectors.

[0164] The first feature vector is obtained by concatenating the convolutional feature vectors corresponding to each convolutional kernel.

[0165] Optionally, the extraction unit 1002 is further configured to:

[0166] The word sequence is vectorized to obtain a second text matrix, and multiple convolution kernels are used to convolve the second text matrix to obtain corresponding convolution feature vectors. The convolution feature vectors corresponding to each convolution kernel are concatenated to obtain a first sequence feature vector corresponding to the word sequence.

[0167] An attention mechanism is used to determine the attention weight vector of the word vector corresponding to each word in the word sequence, and a second sequence feature vector corresponding to the word sequence is obtained based on the word vector and attention weight vector corresponding to each word in the word sequence.

[0168] The first sequence feature vector or the second sequence feature vector is used as the second feature vector; or, the first sequence feature vector and the second sequence feature vector are concatenated to obtain the second feature vector.

[0169] Optionally, the preset interaction request includes a business processing request; the triggering unit 1004 is specifically used for:

[0170] If the predicted response is successful, the human-machine interface of the dialogue is switched to the human interface for human customer service, and the business processing request is output by the human interface.

[0171] Optionally, the service processing request includes multiple service types, and a manual interface is configured for each service type to output the corresponding service processing request; the triggering unit 1004 is further used for:

[0172] If the predicted response to a business processing request for any business type is successful, the human-machine interface is switched to the human interface corresponding to that business type.

[0173] Optionally, the preset interaction request includes a business processing request; the triggering unit 1004 is specifically used for:

[0174] If the predicted response to the service request is a successful response, the service request is output through the human-computer interface of the human-computer dialogue for the user to respond to.

[0175] Optionally, the auxiliary reference information includes at least one of the following:

[0176] The user's profile, the user's location information, the time information of the conversation, and the user's historical response results to preset interaction requests.

[0177] Optionally, the user data may also include intent recognition results for the session content.

[0178] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer, which can take the form of a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email sending and receiving device, game console, tablet computer, wearable device, or any combination of these devices.

[0179] In a typical configuration, a computer includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0180] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0181] Computer-readable media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage, quantum memory, graphene-based storage media or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0182] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0183] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0184] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of one or more embodiments of this specification. The singular forms “a,” “described,” and “the” used in one or more embodiments of this specification and in the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0185] It should be understood that although the terms first, second, third, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first information may also be referred to as second information without departing from the scope of one or more embodiments of this specification, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "in response to a determination," or "when," or "in the event of a determination."

[0186] The above description is merely a preferred embodiment of one or more embodiments of this specification and is not intended to limit the scope of one or more embodiments of this specification. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments of this specification should be included within the protection scope of one or more embodiments of this specification.

Claims

1. A method of processing dialogue information, characterized by, The method is applied to a conversation system including an online robot and a machine learning model for human-computer dialogue, and the method comprises: obtaining user data in real time in the process of human-computer dialogue between the online robot and a client of a user, the user data including conversation content input by the user in at least one round of dialogue and auxiliary reference information related to the user and different from the conversation content; the auxiliary reference information including at least one of a user portrait of the user, location information of the user, time information of human-computer dialogue, and a historical response result of the user to a preset interaction request; performing feature extraction on the conversation content and the auxiliary reference information respectively to obtain corresponding feature vectors, and splicing the feature vectors to obtain a representation vector; inputting the representation vector into the machine learning model, and predicting a response result of the user to a preset interaction request by using output information of the machine learning model; the preset interaction request including a service handling request of a preset service provided by a service party for the user; a label of a user sample data in a training set of the machine learning model including the response result of the user to the preset interaction request; the response result representing whether the user has a potential service handling demand in the current conversation; if the predicted response result is a successful response, automatically triggering a process for the preset interaction request in the current human-computer dialogue; the process including outputting the service handling request for the user to complete service handling on the client.

2. The method of claim 1, wherein, The conversation content includes current conversation content input by the user in the current round of dialogue and historical conversation content input by the user in historical rounds of dialogue; performing feature extraction on the conversation content to obtain a corresponding feature vector includes: extracting a first feature vector corresponding to the current conversation content; converting the historical conversation content into a word sequence, and extracting a second feature vector corresponding to the word sequence; splicing the first feature vector and the second feature vector to obtain a feature vector corresponding to the conversation content.

3. The method of claim 2, wherein, The extracting of the first feature vector corresponding to the current conversation content includes: vectorizing the current conversation content to obtain a first text matrix; performing convolution processing on the first text matrix by using multiple convolution kernels respectively to obtain corresponding convolution feature vectors; splicing the convolution feature vectors corresponding to the respective convolution kernels to obtain the first feature vector.

4. The method of claim 2, wherein, The extracting of the second feature vector corresponding to the word sequence includes: vectorizing the word sequence to obtain a second text matrix, and performing convolution processing on the second text matrix by using multiple convolution kernels respectively to obtain corresponding convolution feature vectors, and splicing the convolution feature vectors corresponding to the respective convolution kernels to obtain a first sequence feature vector corresponding to the word sequence; determining an attention weight vector of a word vector corresponding to each word in the word sequence by using an attention mechanism, and obtaining a second sequence feature vector corresponding to the word sequence according to the word vector corresponding to each word in the word sequence and the attention weight vector respectively; The first sequence feature vector or the second sequence feature vector is taken as the second feature vector, or the first sequence feature vector and the second sequence feature vector are spliced to obtain the second feature vector.

5. The method of claim 1, wherein, The method further includes the following steps: If the predicted response result is a response success, the man-machine interface of the current man-machine conversation is automatically switched to a manual interface of a manual customer service conversation, and the service handling request is output by the manual interface.

6. The method of claim 5, wherein, The service handling request includes multiple business types, and a manual interface for outputting a service handling request of a corresponding business type is configured for each business type; if the predicted response result is a response success, the man-machine interface of the current man-machine conversation is automatically switched to a manual interface of a manual customer service conversation, and the service handling request is output by the manual interface. If the predicted response result of the service handling request of any business type is a response success, the man-machine interface is automatically switched to a manual interface corresponding to the any business type.

7. The method of claim 1, wherein, The method further includes the following steps: If the predicted response result of the service handling request is a response success, the service handling request is automatically output by the man-machine interface of the man-machine conversation for response by the user.

8. The method of claim 1, wherein, The user data further includes an intent recognition result for the conversation content. 9.A method for training a machine learning model based on dialogue information, the method comprising: The method further includes the following steps: Obtaining user sample data of a user and a response result of the user to a preset interactive request, the user sample data including conversation content input by the user in at least one round of conversation and auxiliary reference information related to the user and different from the conversation content; the auxiliary reference information including at least one of the following: a user portrait of the user, location information of the user, time information of a man-machine conversation, and a historical response result of the user to a preset interactive request; The preset interactive request includes a service handling request of a preset service provided by a service party for the user; The label of the user sample data includes a response result of the user to a preset interactive request; Feature extraction is performed on the conversation content and the auxiliary reference information to obtain corresponding feature vectors, and the feature vectors are spliced to obtain a representation vector; The representation vector and the response result are used to train a machine learning model; wherein the trained machine learning model is used to predict a response result of a user to a preset interactive request in the method for processing conversation information according to any one of claims 1 to 8.

10. The method of claim 9, wherein, The method further includes the following steps: The conversation content and the auxiliary reference information stored in a system log of a conversation system used to establish a conversation are obtained.

11. The method of claim 9, wherein, The conversation content includes conversation content input by the user in any round of dialogue and historical conversation content input in a number of historical rounds of dialogue corresponding to the any round of dialogue; feature extraction is performed on the conversation content to obtain a corresponding feature vector, including: extracting a first feature vector corresponding to the conversation content in the any round of dialogue; converting the historical conversation content into a word sequence and extracting a second feature vector corresponding to the word sequence; splicing the first feature vector and the second feature vector to obtain a feature vector corresponding to the conversation content.

12. The method of claim 11, wherein, The extracting of the first feature vector corresponding to the conversation content in the any round of dialogue includes: vectorizing the conversation content in the any round of dialogue to obtain a first text matrix; performing convolution processing on the first text matrix by using a plurality of convolution kernels respectively to obtain corresponding convolution feature vectors; splicing the convolution feature vectors corresponding to each convolution kernel to obtain the first feature vector.

13. The method of claim 11, wherein, The extracting of the second feature vector corresponding to the word sequence includes: vectorizing the word sequence to obtain a second text matrix, performing convolution processing on the second text matrix by using a plurality of convolution kernels respectively to obtain corresponding convolution feature vectors, and splicing the convolution feature vectors corresponding to each convolution kernel to obtain a first sequence feature vector corresponding to the word sequence; determining an attention weight vector of a word vector corresponding to each word in the word sequence by using an attention mechanism, and obtaining a second sequence feature vector corresponding to the word sequence according to the word vector corresponding to each word in the word sequence and the attention weight vector; taking the first sequence feature vector or the second sequence feature vector as the second feature vector; or splicing the first sequence feature vector and the second sequence feature vector to obtain the second feature vector.

14. The method of claim 9, wherein, The service handling request is output by the artificial interface after the man-machine interface of the dialogue is switched to the artificial interface of the artificial customer service dialogue.

15. The method of claim 9, wherein, The preset interaction request includes a service handling request; the service handling request is output by the man-machine interface of the dialogue.

16. The method of claim 9, wherein, The user sample data further includes an intent recognition result for the conversation content.

17. A processing apparatus of dialogue information, characterized by, The device is applied to a conversation system including an online robot for man-machine dialogue and a machine learning model, including: An acquisition unit acquires user data in real time in the process of man-machine dialogue between the online robot and a client of a user, the user data including conversation content input by the user in at least one round of dialogue and auxiliary reference information related to the user and different from the conversation content; the auxiliary reference information includes at least one of the following: a user portrait of the user, location information of the user, time information of man-machine dialogue, and a historical response result of the user to a preset interaction request; A feature extraction unit extracts a feature vector corresponding to the conversation content and the auxiliary reference information respectively, and splices the feature vectors to obtain a representation vector. The input unit inputs the representation vector into the machine learning model, and predicts a response result of the user to a preset interaction request by using output information of the machine learning model; the response result represents whether the user has a potential service handling demand in a current session; the preset interaction request includes a service handling request of a preset service provided by a service party for the user; and a label of a user sample data in a training set of the machine learning model includes the response result of the user to the preset interaction request. The triggering unit automatically triggers a flow for the preset interaction request in the current human-computer conversation if the predicted response result is a response success; and the flow includes outputting the service handling request for the user to complete service handling on the client.

18. The apparatus of claim 17, wherein, The session content includes current session content input by the user in a current round of conversation and historical session content input in historical rounds of conversation; and the extraction unit is specifically configured to: extract a first feature vector corresponding to the current session content; convert the historical session content into a word sequence, and extract a second feature vector corresponding to the word sequence; splice the first feature vector and the second feature vector to obtain a feature vector corresponding to the session content.

19. The apparatus of claim 18, wherein, The extraction unit is further configured to: vectorize the current session content to obtain a first text matrix; perform convolution processing on the first text matrix by using a plurality of convolution kernels respectively to obtain corresponding convolution feature vectors; splice the convolution feature vectors corresponding to the respective convolution kernels to obtain the first feature vector.

20. The apparatus of claim 18, wherein, The extraction unit is further configured to: vectorize the word sequence to obtain a second text matrix, and perform convolution processing on the second text matrix by using a plurality of convolution kernels respectively to obtain corresponding convolution feature vectors, and splice the convolution feature vectors corresponding to the respective convolution kernels to obtain a first sequence feature vector corresponding to the word sequence; determine an attention weight vector of a word vector corresponding to each word in the word sequence by using an attention mechanism, and obtain a second sequence feature vector corresponding to the word sequence according to the word vector corresponding to each word in the word sequence and the attention weight vector; use the first sequence feature vector or the second sequence feature vector as the second feature vector, or splice the first sequence feature vector and the second sequence feature vector to obtain the second feature vector.

21. The apparatus of claim 17, wherein, The triggering unit is specifically configured to: if the predicted response result is a response success, automatically switch a human-computer interface of the conversation to an artificial interface of an artificial customer service conversation in the current human-computer conversation, and the service handling request is output by the artificial interface.

22. The apparatus of claim 21, wherein, The service handling request includes a plurality of business types, and an artificial interface for outputting a service handling request of a corresponding business type is configured for each business type; and the triggering unit is further configured to: if the predicted response result of the service handling request of any business type is a response success, automatically switch the human-computer interface to an artificial interface corresponding to the any business type in the current human-computer conversation.

23. The apparatus of claim 17, wherein, The triggering unit is specifically configured to: if the predicted response result for the service handling request is a response success, automatically outputting the service handling request in the current human-computer conversation through the human-computer interface of the human-computer conversation for response by the user.

24. The apparatus of claim 17, wherein, The user data further includes an intent recognition result for the conversation content.

25. A machine learning model training device based on dialogue information, characterized in that, Comprise: An acquisition unit acquires user sample data of a user and a response result of the user to a preset interaction request, the user sample data including conversation content input by the user in at least one round of conversation and auxiliary reference information related to the user and different from the conversation content; the auxiliary reference information includes at least one of a user portrait of the user, location information of the user, time information of the human-computer conversation, and a historical response result of the user to a preset interaction request; The preset interaction request includes a service handling request of a preset service provided by a service party for the user; The label of the user sample data includes a response result of the user to a preset interaction request; An extraction unit respectively extracts features from the conversation content and the auxiliary reference information to obtain corresponding feature vectors, and splices the feature vectors to obtain a representation vector; A training unit trains a machine learning model using the representation vector and the response result; wherein the trained machine learning model is used in the conversation information processing method of any one of claims 1 to 8 to predict a response result of the user to a preset interaction request.

26. The apparatus of claim 25, wherein, The acquisition unit is specifically configured to: Acquire the conversation content and the auxiliary reference information stored in a system log of a conversation system used to establish a conversation.

27. The apparatus of claim 25, wherein, The conversation content includes conversation content input by the user in any round of conversation and historical conversation content input in a number of historical rounds of conversation corresponding to the any round of conversation; the extraction unit is specifically configured to: Extract a first feature vector corresponding to the conversation content in the any round of conversation; Convert the historical conversation content into a word sequence and extract a second feature vector corresponding to the word sequence; Splice the first feature vector and the second feature vector to obtain a feature vector corresponding to the conversation content.

28. The apparatus of claim 27, wherein, The extraction unit is further configured to: Vectorize the conversation content in the any round of conversation to obtain a first text matrix; Respectively use a plurality of convolution kernels to perform convolution processing on the first text matrix to obtain corresponding convolution feature vectors; Splice the convolution feature vectors corresponding to each convolution kernel to obtain the first feature vector.

29. The apparatus of claim 27, wherein, The extraction unit is further configured to: Vectorize the word sequence to obtain a second text matrix, and respectively use a plurality of convolution kernels to perform convolution processing on the second text matrix to obtain corresponding convolution feature vectors, and splice the convolution feature vectors corresponding to each convolution kernel to obtain a first sequence feature vector corresponding to the word sequence; Determine an attention weight vector of a word vector corresponding to each word in the word sequence using an attention mechanism, and obtain a second sequence feature vector corresponding to the word sequence according to the word vector corresponding to each word in the word sequence and the attention weight vector; The first sequence feature vector or the second sequence feature vector is taken as the second feature vector; or the first sequence feature vector and the second sequence feature vector are spliced to obtain the second feature vector.

30. The apparatus of claim 28, wherein, The service handling request is output by the artificial interface after the dialogue man-machine interface is switched to the artificial interface of the artificial customer service dialogue.

31. The apparatus of claim 28, wherein, The service handling request is output by the dialogue man-machine interface.

32. The apparatus of claim 28, wherein, The user sample data further comprises an intent recognition result for the conversation content.

33. An electronic device, comprising: Comprise: a processor; a memory for storing processor-executable instructions; wherein the processor implements the method of any one of claims 1-16 by running the executable instructions.

34. A computer-readable storage medium, characterized in that, A computer program product, comprising a computer readable medium having stored thereon computer instructions that, when executed by a processor, implement the steps of the method of any one of claims 1-16.

Images