Artificial intelligence-based automatic reply conversation method, device, equipment and medium

Abstract

The application provides an automatic reply dialogue method and device based on artificial intelligence, an electronic device and a storage medium. The automatic reply dialogue method based on artificial intelligence comprises the following steps: preprocessing collected dialogue text to obtain a dialogue data set; obtaining a reply candidate pool of each input text in the dialogue data set; labeling the reply text to obtain a dialogue training data set; performing multi-task learning based on the dialogue training data set and a preset neural network model to obtain a scene reply evaluation model; collecting dialogue text in real time to obtain a target reply candidate pool of a target input text; scoring the reply text in the target reply candidate pool based on the scene reply evaluation model to obtain a target reply score value; and replying to the target input text based on the target reply score value. The application can comprehensively consider whether the reply text is reasonable from multiple angles to screen out appropriate reply texts for more accurate reply.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an automatic response dialogue method, device, electronic device and storage medium based on artificial intelligence. Background Technology

[0002] Natural language processing has always been a popular approach in the field of artificial intelligence, with dialogue capabilities being considered one of the ultimate challenges in solving general artificial intelligence. As neural network technology matures, open-domain dialogue systems are facing increasingly higher demands: the ability to engage in meaningful dialogue within an open domain, rather than simply engaging in idle chatter.

[0003] With the emergence of more and more large-scale models and the application of knowledge fusion technologies, the responses generated by neural dialogue generation models are becoming increasingly natural and fluent. However, in current end-to-end dialogue models, many low-quality responses still frequently occur. For example, in intelligent customer service systems, there are often contextual logic conflicts, responses that are inconsistent with the background, or overly tedious dialogues due to generic templates. Some responses may even be biased or sensitive. These low-quality responses prevent these large-scale models or systems from accurately responding to user input text, thus severely limiting their practical application value. Summary of the Invention

[0004] In view of the above, it is necessary to propose an AI-based automatic response dialogue method, device, electronic device, and storage medium to solve the technical problem of how to accurately respond to user input text.

[0005] This application provides an artificial intelligence-based automatic response dialogue method, the method comprising:

[0006] The collected dialogue text is preprocessed to obtain a dialogue dataset, which includes multi-turn dialogue text, each turn of dialogue text including an input text and a response text;

[0007] Based on the dialogue generation model, multiple response texts for each input text in the dialogue dataset are obtained to obtain a response candidate pool for each input text;

[0008] The dialogue training dataset is obtained by annotating the response texts in all response candidate pools.

[0009] Based on the dialogue training dataset and the preset neural network model, multi-task learning is performed to obtain a scenario response evaluation model;

[0010] Real-time acquisition of dialogue text, and acquisition of a target response candidate pool of target input text in the dialogue text through a dialogue generation model;

[0011] The target response score value is obtained by scoring the response texts in the target response candidate pool based on the scenario response evaluation model.

[0012] Based on the target response score, response texts that match the target input text are selected to respond to the target input text.

[0013] In some embodiments, the preprocessing of the acquired dialogue text to obtain the dialogue dataset includes:

[0014] The dialogue text dataset is obtained by crawling dialogue text according to a preset method.

[0015] The dialogue text dataset is segmented into words to obtain a dialogue word dataset;

[0016] Statistical analysis of word frequencies in the aforementioned dialogue word dataset;

[0017] The dialogue dataset is obtained by filtering out low-frequency words from the dialogue word dataset.

[0018] In some embodiments, the step of annotating the response texts in all response candidate pools to obtain the dialogue training dataset includes:

[0019] Based on the dialogue generation model, obtain the response candidate pool corresponding to the user's input text in each round of dialogue in the dialogue dataset;

[0020] The reply text in the reply candidate pool is labeled according to the preset dialogue discrimination labels to obtain a reply text label set. The dialogue discrimination labels include global discrimination labels, local discrimination labels, security discrimination labels and scene discrimination labels.

[0021] The response candidate pool is traversed to obtain the response text label set generated in each round of dialogue, and all the input texts in the dialogue dataset and the response text label set corresponding to each input text are used as the dialogue training dataset.

[0022] In some embodiments, the step of obtaining a scenario response evaluation model through multi-task learning based on the dialogue training dataset and a preset neural network model includes:

[0023] The preset neural network model is trained based on the target discrimination labels in the dialogue training dataset so that the neural network model can obtain the target discrimination score value of the response text, wherein the target discrimination label is any one of the discrimination labels;

[0024] The global discrimination label, local discrimination label, security discrimination label and scene discrimination label are traversed to enable the neural network model to perform multi-task learning, and after multi-task learning, the global discrimination score, local discrimination score, security discrimination score and scene discrimination score of the response text can be obtained simultaneously.

[0025] A neural network model capable of simultaneously acquiring the global discrimination score, local discrimination score, security discrimination score, and scene discrimination score of the response text is used as the scene response evaluation model.

[0026] In some embodiments, training a preset neural network model based on target discrimination labels in the dialogue training dataset to enable the neural network model to obtain the target discrimination score value of the response text includes:

[0027] Based on a pre-defined neural network model, the dialogue data labeled with target discrimination tags is converted into dialogue word vectors;

[0028] Construct a dialogue word weight matrix based on the dialogue word vectors;

[0029] Based on the dialogue word weight matrix, the dialogue data labeled with target discrimination tags is transformed into dialogue sentence vectors;

[0030] The response text vector is constructed based on the dialogue statement vector;

[0031] The response text vector is classified based on the classifier in the neural network model to obtain the output target discrimination score value of the response text vector;

[0032] The classification loss between the output target discrimination score and the corresponding target discrimination label is calculated based on the cross-entropy loss function.

[0033] The neural network model is iteratively trained based on the classification loss so that the neural network model can obtain the target discrimination score value of the response text.

[0034] In some embodiments, the step of scoring the response texts in the target response candidate pool based on the scenario response evaluation model to obtain a target response score value includes:

[0035] Based on the scenario response evaluation model, the global discrimination score, local discrimination score, security discrimination score, and scenario discrimination score of each response text in the target response candidate pool are calculated;

[0036] The global discrimination score, local discrimination score, security discrimination score, and scene discrimination score of the response text are weighted according to the scene in which the target response candidate pool is located.

[0037] The target response score is obtained by weighted summing of the global discrimination score, local discrimination score, security discrimination score, scenario discrimination score, and the weights corresponding to each discrimination score of the response text. The target response score corresponds one-to-one with the response text.

[0038] In some embodiments, the step of filtering response texts that match the target input text based on the target response score to respond to the target input text includes:

[0039] The target response score values ​​of each response text in the target response candidate pool are sorted in descending order;

[0040] Select the response text corresponding to the target response score that ranks first as the response text that matches the target input text;

[0041] Reply to the target input text based on the response text that matches the target input text.

[0042] This application also provides an artificial intelligence-based automatic response dialogue device, which includes a preprocessing module, an acquisition module, an annotation module, a learning module, a collection module, a scoring module, and a filtering module.

[0043] The preprocessing module is used to preprocess the collected dialogue text to obtain a dialogue dataset, which includes multi-turn dialogue text, each turn of dialogue text including an input text and a response text.

[0044] The obtaining module is used to obtain multiple response texts for each input text in the dialogue dataset based on the dialogue generation model to obtain a response candidate pool for each input text;

[0045] The annotation module is used to annotate the response texts in all response candidate pools to obtain the dialogue training dataset;

[0046] The learning module is used to perform multi-task learning based on the dialogue training dataset and a preset neural network model to obtain a scenario response evaluation model;

[0047] The acquisition module is used to acquire dialogue text in real time and obtain the target response candidate pool of the target input text in the dialogue text through the dialogue generation model;

[0048] The scoring module is used to score the response texts in the target response candidate pool based on the scenario response evaluation model to obtain the target response score value;

[0049] The filtering module is used to filter out response texts that match the target input text based on the target response score, so as to respond to the target input text.

[0050] This application embodiment also provides an electronic device, the electronic device comprising:

[0051] Memory, storing at least one instruction;

[0052] The processor executes the instructions stored in the memory to implement the AI-based automatic response dialogue method.

[0053] This application also provides a computer-readable storage medium storing at least one instruction, which is executed by a processor in an electronic device to implement the AI-based automatic response dialogue method.

[0054] This application uses a multi-task learning approach to train a neural network model on the collected dialogue dataset to obtain a scenario response evaluation model. This model can comprehensively consider the rationality of the response text from multiple perspectives, thereby selecting response texts that match the user's input text and effectively improving the accuracy of automatic dialogue responses. Attached Figure Description

[0055] Figure 1 This is a flowchart of a preferred embodiment of the AI-based automatic response dialogue method involved in this application.

[0056] Figure 2 This is a functional block diagram of a preferred embodiment of the AI-based automatic response dialogue device involved in this application.

[0057] Figure 3 This is a schematic diagram of the structure of an electronic device that is a preferred embodiment of the AI-based automatic response dialogue method involved in this application. Detailed Implementation

[0058] To better understand the purpose, features, and advantages of this application, a detailed description of the application is provided below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of this application can be combined with each other. Numerous specific details are set forth in the following description to provide a thorough understanding of this application; the described embodiments are only a part of the embodiments of this application, and not all of them.

[0059] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "a plurality of" means two or more, unless otherwise explicitly specified.

[0060] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0061] This application provides an automatic response dialogue method based on artificial intelligence, which can be applied to one or more electronic devices. An electronic device is a device that can automatically perform numerical calculations and / or information processing according to pre-set or stored instructions. Its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0062] Electronic devices can be any electronic product that allows human-computer interaction with a customer, such as personal computers, tablets, smartphones, personal digital assistants (PDAs), game consoles, interactive network television (IPTV), smart wearable devices, etc.

[0063] Electronic devices may also include network devices and / or client devices. The network devices include, but are not limited to, a single network server, a server group consisting of multiple network servers, or a cloud based on cloud computing consisting of a large number of hosts or network servers.

[0064] The networks in which electronic devices are located include, but are not limited to, the Internet, wide area networks, metropolitan area networks, local area networks, and virtual private networks (VPNs).

[0065] like Figure 1The diagram shown is a flowchart of a preferred embodiment of the AI-based automatic response dialogue method of this application. The order of the steps in this flowchart can be changed, and some steps can be omitted, depending on different needs.

[0066] S10, preprocess the collected dialogue text to obtain a dialogue dataset, the dialogue dataset including multi-turn dialogue text, each turn of dialogue text including an input text and a response text.

[0067] In an optional embodiment, the preprocessing of the acquired dialogue text to obtain the dialogue dataset includes:

[0068] The dialogue text dataset is obtained by crawling dialogue text according to a preset method.

[0069] The dialogue text dataset is segmented into words to obtain a dialogue word dataset;

[0070] Statistical analysis of word frequencies in the aforementioned dialogue word dataset;

[0071] The dialogue dataset is obtained by filtering out low-frequency words from the dialogue word dataset.

[0072] In this optional embodiment, a large amount of Chinese dialogue text can be crawled from the Internet in advance using a web crawler tool. At the same time, it is combined with the existing English dialogue dataset DailyDialog to form a dialogue text dataset. DailyDialog is a multi-turn dialogue dataset for daily chat scenarios constructed by the dataset publisher by crawling English spoken dialogue websites. It contains dialogue texts from daily life, covers a lot of emotional information, and has a relatively natural dialogue pattern.

[0073] In this optional embodiment, in order to fully train the neural network model in subsequent processes, the more data in the dialogue text dataset obtained in this solution, the better. For example, the dialogue text dataset should contain at least 600,000 data entries.

[0074] In this optional embodiment, for the dialogue text in the dialogue text dataset, since there are no formal boundaries between words, in order to accurately obtain the semantic information expressed by the dialogue text, the dialogue text in the dialogue text dataset can first be segmented into words, thereby representing the dialogue text as a sequence of words. Specifically, since English text words are separated by spaces, word segmentation can be performed directly based on spaces. The nltk segmentation tool can be used for word segmentation, while the jieba segmentation tool can be used for Chinese text. In this scheme, the dialogue text dataset after word segmentation is used as the dialogue word dataset.

[0075] In this optional embodiment, to facilitate word frequency statistics on the dialogue word dataset, the segmented words in the dialogue word dataset can first be hashed to obtain the Unicode value of the words, and then the words can be re-encoded using a hash function, so that instant access can be achieved each time the word is accessed, so as to facilitate rapid statistics on the word frequency of each word.

[0076] In this optional embodiment, the word frequency is short for word frequency, which is the number of times a word appears in the dialogue text dataset. Words can be sorted in descending order of word frequency, and words with lower frequency rankings can be filtered out. In this scheme, the dialogue word dataset filtered of low-frequency words is used as the dialogue dataset. For example, the words in the dialogue word dataset have 10,000 rankings after word frequency sorting. The 100 words with the lowest frequency rankings can be filtered out, and the remaining 9,900 words are used as the dialogue dataset.

[0077] Thus, by performing a series of preprocessing operations on the collected dialogue text to obtain a dialogue dataset, more practical and effective data support can be provided for subsequent processes.

[0078] S11, Based on the dialogue generation model, obtain multiple response texts for each input text in the dialogue dataset to obtain a response candidate pool for each input text.

[0079] In this optional embodiment, the dialogue dataset includes multiple rounds of dialogue. In this scheme, each input text and corresponding response text during the interaction is considered a round of dialogue; that is, each round of dialogue includes one input text and one response text. All input texts and response texts constitute the dialogue dataset. The response text corresponding to the input text is selected from multiple alternative response texts generated based on the dialogue generation model. These alternative response texts constitute a response candidate pool; that is, each input text in a round of dialogue corresponds to one response candidate pool.

[0080] In this optional embodiment, a typical deployed open-domain dialogue system generally includes multiple response generators, including template-based response generators and different types of neural response generators. Each response generator also generates different responses to the same input text using different strategies, thus forming a response candidate pool.

[0081] In this optional embodiment, the response generator must generate the corresponding response through a dialogue generation model. The dialogue generation model can be a Seq2Seq (AttnSeq2Seq) with an attention mechanism, a hierarchical encoder-decoder (HRED), or a hierarchical latent variable encoder-decoder (VHRED). This solution does not impose any specific restrictions on this.

[0082] In this optional embodiment, different pre-trained dialogue generation models can be used to obtain multiple corresponding response texts for each input text, and all response texts corresponding to each input text can be used as a response candidate pool for that input text. The dialogue dataset can include multi-turn dialogues, and the input text is the user's input text in any turn of the dialogue dataset.

[0083] In this way, multiple response texts corresponding to the input text can be quickly generated through a pre-trained dialogue generation model, thereby forming a response candidate pool for the input text.

[0084] S12, label the response texts in all response candidate pools to obtain the dialogue training dataset.

[0085] In an optional embodiment, the step of annotating the response texts in all response candidate pools to obtain the dialogue training dataset includes:

[0086] Based on the dialogue generation model, obtain the response candidate pool corresponding to the user's input text in each round of dialogue in the dialogue dataset;

[0087] The reply text in the reply candidate pool is labeled according to the preset dialogue discrimination labels to obtain a reply text label set. The dialogue discrimination labels include global discrimination labels, local discrimination labels, security discrimination labels and scene discrimination labels.

[0088] The response candidate pool is traversed to obtain the response text label set generated in each round of dialogue, and all the input texts in the dialogue dataset and the response text label set corresponding to each input text are used as the dialogue training dataset.

[0089] In this optional embodiment, each reply text in the reply candidate pool can be labeled from different angles based on a plurality of preset dialogue discrimination labels. The dialogue discrimination labels include global discrimination labels, local discrimination labels, security discrimination labels, and scene discrimination labels.

[0090] In this optional embodiment, the global discrimination label is used to indicate whether the global context relationship between the labeled reply text and the entire dialogue text is consistent; the local discrimination label is used to indicate whether the local context relationship between the labeled reply text and the local dialogue text is consistent; the security discrimination label is used to indicate whether the labeled reply text contains offensive or intrusive statements; and the scene discrimination label is used to indicate whether the relationship between the labeled reply text and the current dialogue scene is consistent.

[0091] For example, let the dialogue between the user and the system in the current round be the t-th round of dialogue, where the input text of the user in the current round is u. t The historical dialogue text between the user and the system before the t-th round of dialogue is h. t The current user's input text u t The corresponding response candidate pool is i = 1, ..., M, where M is the total number of responses in the current round's response candidate pool. The M response texts in the current round's response candidate pool can be labeled based on the global label, local label, security label, and scenario label. The specific labeling criteria are as follows:

[0092] a) with h t Whether they are consistent; for example, if the reply text in the previous round of the conversation was "I like to eat apples", and the reply text in the current round is "I hate to eat apples", then the reply text in the current round is inconsistent with the global context.

[0093] If they match, the global discrimination label can be marked as A0; if they do not match, it can be marked as A1.

[0094] b) with u t Whether they are consistent; for example, if the user's input text in the current round of dialogue is "What food do you like to eat?", and the reply text in the current round is "I like to travel", then the reply text in the current round is inconsistent with the local context. If they are consistent, the local discrimination label can be marked as B0; if they are inconsistent, it can be marked as B1.

[0095] c) Is the response offensive? Examples include responses containing gender bias, offensive language, or insulting words.

[0096] If yes, the safety judgment label is marked as C1; otherwise, it is marked as C0.

[0097] d) Does it contain content related to the current application scenario? For example, in an intelligent customer service scenario, the generated response text should contain positive emotions; in a medical dialogue scenario, the generated response text should not contain medical errors, etc. If it does, the scenario discrimination label is marked as D0; otherwise, it is marked as D1.

[0098] In this optional embodiment, each reply text in the reply candidate pool, after being labeled, includes four tags: global discrimination tag, local discrimination tag, security discrimination tag, and scene discrimination tag.

[0099] In this optional embodiment, the response candidate pool corresponding to the user's input text in each round of dialogue can be obtained by traversing the dialogue dataset. Then, the response text in all response candidate pools is labeled to generate a response text label set corresponding to each response candidate pool. In this scheme, all response text label sets obtained after labeling and the input text corresponding to each response text label set are used as the dialogue training dataset.

[0100] In this way, the response text in the response candidate pool can be labeled from multiple perspectives using preset dialogue discrimination tags, and the set of response text tags corresponding to the user's input text in each round of dialogue can be obtained.

[0101] S13, Based on the dialogue training dataset and the preset neural network model, perform multi-task learning to obtain a scene response evaluation model.

[0102] In an optional embodiment, the step of obtaining a scene response evaluation model through multi-task learning based on the dialogue training dataset and a preset neural network model includes:

[0103] The preset neural network model is trained based on the target discrimination labels in the dialogue training dataset so that the neural network model can obtain the target discrimination score value of the response text, wherein the target discrimination label is any one of the discrimination labels;

[0104] The global discrimination label, local discrimination label, security discrimination label and scene discrimination label are traversed to enable the neural network model to perform multi-task learning, and after multi-task learning, the global discrimination score, local discrimination score, security discrimination score and scene discrimination score of the response text can be obtained simultaneously.

[0105] A neural network model capable of simultaneously acquiring the global discrimination score, local discrimination score, security discrimination score, and scene discrimination score of the response text is used as the scene response evaluation model.

[0106] In this optional embodiment, the preset neural network model can be a HAN model (Hierarchical Attention Networks), whose network structure mainly consists of four parts: word encoder, word attention, sentence encoder, and sentence attention.

[0107] In this optional embodiment, since each response text in the dialogue training dataset contains four different discriminant labels, this scheme uses any one of the discriminant labels as the target discriminant label to describe how to use the target discriminant label to train the neural network model. This enables the neural network model trained with the target discriminant label to obtain the target discriminant score value for any response text. The main training process is as follows:

[0108] First, the dialogue data labeled with target discrimination tags is mapped into word vectors through the embedding matrix of the word encoder to obtain dialogue word vectors. Then, the positive and negative context information of the dialogue word vectors is obtained through the bidirectional GRU network in the word encoder.

[0109] Next, the importance of each word vector is calculated through the word attention part, and then the importance of each word vector is normalized by a classification function such as softmax to obtain the dialogue word weight matrix composed of each word vector;

[0110] Since each sentence in the dialogue data labeled with target discrimination is composed of words, the corresponding sentence vector can be obtained by weighting and summing the word vectors using the dialogue word weight matrix.

[0111] Then, the bidirectional GRU network in the sentence encoder is used to obtain the forward and reverse context information of each dialogue sentence vector;

[0112] Similarly, the normalized weights of each dialogue sentence vector are calculated using the sentence attention, and the response text vector is obtained by weighted summation of each dialogue sentence vector.

[0113] Finally, the response text vector is classified based on the classifier in the neural network model to obtain the output target discrimination score value of the response text vector;

[0114] After obtaining the output of the neural network model, the classification loss between the output target discrimination score and the corresponding target discrimination label is calculated using the cross-entropy loss function. Then, the classification loss is further reduced by adjusting the parameters in the neural network model. The neural network model is iteratively trained by repeating the above training process. When the classification loss is 0, the neural network model can accurately obtain the target discrimination score of the response text.

[0115] In this optional embodiment, the formula for the cross-entropy loss function is:

[0116]

[0117] Where L represents the cross-entropy loss, and y represents the sample's label value. This represents the predicted value of the sample. For example, the target discriminant label is the global discriminant label. If the response text in the current round is consistent with the global context, the global discriminant label can be labeled as 1; otherwise, it is labeled as 0. Suppose the target discriminant score of the response text vector U is 0.8. In fact, the response text in the current round is consistent with the global context, so the global discriminant label value corresponding to the response text vector U is 1. Therefore, the cross-entropy loss corresponding to the response text vector U at this time can be calculated as L = -[1*ln0.8+(1-1)ln(1-0.8)] = 0.22.

[0118] In this optional embodiment, the above training process only describes how to train the neural network model using any one of the discriminant labels as the target discriminant label. Therefore, the neural network model can be trained simultaneously according to the four categories of discriminant labels through multi-task learning, so that the neural network model can ultimately obtain the global discriminant score, local discriminant score, security discriminant score, and scene discriminant score of the response text at the same time. In this scheme, the neural network model at this time is used as the scene response evaluation model. Here, multi-task learning is mainly used to distinguish it from single-task learning. Most machine learning tasks can only learn one task at a time, so they are called single-task learning, while multi-task learning can put multiple related tasks together for learning, thereby learning multiple tasks simultaneously.

[0119] In this way, by training the pre-set neural network model through multi-task learning, the training efficiency can be improved more quickly, and a scene response evaluation model generated by integrating four types of discrimination labels can be obtained.

[0120] S14, collect dialogue text in real time, and obtain the target response candidate pool of the target input text in the dialogue text through the dialogue generation model.

[0121] In this optional embodiment, for dialogue text acquired in real time, multiple pre-trained dialogue generation models can be used to generate multiple response texts for the target input text in the dialogue text, and these response texts can be used as the target response candidate pool for the target input text.

[0122] S15, Based on the scenario response evaluation model, score the response texts in the target response candidate pool to obtain the target response score value.

[0123] In an optional embodiment, the step of scoring the response texts in the target response candidate pool based on the scenario response evaluation model to obtain the target response score value includes:

[0124] Based on the scenario response evaluation model, the global discrimination score, local discrimination score, security discrimination score, and scenario discrimination score of each response text in the target response candidate pool are calculated;

[0125] The global discrimination score, local discrimination score, security discrimination score, and scene discrimination score of the response text are weighted according to the scene in which the target response candidate pool is located.

[0126] The target response score is obtained by weighted summing of the global discrimination score, local discrimination score, security discrimination score, scenario discrimination score, and the weights corresponding to each discrimination score of the response text. The target response score corresponds one-to-one with the response text.

[0127] In this optional embodiment, the global discrimination score, local discrimination score, security discrimination score, and scene discrimination score of each reply text in the target reply candidate pool can be calculated by the scene reply evaluation model.

[0128] In this optional embodiment, different weights can be pre-assigned to the global discrimination score, local discrimination score, security discrimination score, and scenario discrimination score of the response text according to the application scenario corresponding to the target response candidate pool. This allows the global discrimination score, local discrimination score, security discrimination score, and scenario discrimination score of the response text to reflect different importance according to the corresponding scenario. This facilitates the weighted summation of the global discrimination score, local discrimination score, security discrimination score, scenario discrimination score, and the weights corresponding to each discrimination score, thereby obtaining a response text that better meets the requirements of the current scenario.

[0129] For example, in both casual conversation and customer service scenarios, two response texts, P and Q, exist simultaneously. After calculation by the scenario response evaluation model, the global discrimination score, local discrimination score, security discrimination score, and scenario discrimination score of response texts P and Q are P = [0.6, 0.75, 0.1, 0.4] and Q = [0.8, 0.25, 0.7, 0.5], respectively. Due to the different scenarios, the weights of the global discrimination score, local discrimination score, security discrimination score, and scenario discrimination score of each response text in the casual conversation and customer service scenarios are [0.1, 0.5, 0.2, 0.2] and [0.2, 0.3, 0.3, 0.2], respectively. Therefore, the final target response scores for response texts P and Q in the casual conversation scenario are as follows:

[0130] P: 0.6*0.1+0.75*0.5+0.1*0.2+0.4*0.2=0.535;

[0131] Q: 0.8*0.1+0.25*0.5+0.7*0.2+0.5*0.2=0.445;

[0132] In a customer service scenario, the target response scores for text messages P and Q are as follows:

[0133] P: 0.6*0.2+0.75*0.3+0.1*0.3+0.4*0.2=0.455;

[0134] Q: 0.8*0.2+0.25*0.3+0.7*0.3+0.5*0.2=0.545;

[0135] Therefore, because the global discrimination score, local discrimination score, security discrimination score and scene discrimination score of the reply text are assigned different weights, even the same reply text will correspond to different target reply scores in different scenarios.

[0136] In this way, by assigning different weights to the global discrimination score, local discrimination score, security discrimination score, and scenario discrimination score of the response text in different scenarios, a target response score that is more in line with each scenario can be calculated, which makes it easier to more accurately select response text that meets the scenario requirements.

[0137] S16, Based on the target response score, select response texts that match the target input text to respond to the target input text.

[0138] In an optional embodiment, the step of filtering response texts that match the target input text based on the target response score to respond to the target input text includes:

[0139] The target response score values ​​of each response text in the target response candidate pool are sorted in descending order;

[0140] Select the response text corresponding to the target response score that ranks first as the response text that matches the target input text;

[0141] Reply to the target input text based on the response text that matches the target input text.

[0142] In this optional embodiment, the target reply score values ​​of each reply text in the obtained target reply candidate pool can be sorted in descending order, and the reply text ranked first, that is, the reply text with the highest target reply score value, can be used as the reply text that best matches the target input text. Therefore, the reply text can be directly used to reply to the corresponding user input text.

[0143] For example, in a casual conversation scenario, the target response scores for texts P and Q are 0.535 and 0.445, respectively. Therefore, text P is selected to respond to the user's input text in the current scenario. In a customer service scenario, the target response scores for texts P and Q are 0.455 and 0.545, respectively. Therefore, text Q is selected to respond to the user's input text in the current scenario.

[0144] In this way, the response text corresponding to the user's input text in the current scenario can be quickly matched according to the target response score of each response text in the target response candidate pool, thereby improving the efficiency of dialogue response.

[0145] Please see Figure 2 , Figure 2 This is a functional block diagram of a preferred embodiment of the AI-based automatic response dialogue device of this application. The AI-based automatic response dialogue device 11 includes a preprocessing module 110, an acquisition module 111, a labeling module 112, a learning module 113, a collection module 114, a scoring module 115, and a filtering module 116. The unit / module referred to in this application refers to a series of computer-readable instruction segments that can be executed by the processor 13 and perform a fixed function, and are stored in the memory 12. In this embodiment, the functions of each unit / module will be described in detail in subsequent embodiments.

[0146] In an optional embodiment, the preprocessing module 110 is used to preprocess the acquired dialogue text to obtain a dialogue dataset, the dialogue dataset including multi-turn dialogue text, each turn of dialogue text including an input text and a response text.

[0147] In an optional embodiment, the preprocessing of the acquired dialogue text to obtain the dialogue dataset includes:

[0148] The dialogue text dataset is obtained by crawling dialogue text according to a preset method.

[0149] The dialogue text dataset is segmented into words to obtain a dialogue word dataset;

[0150] Statistical analysis of word frequencies in the aforementioned dialogue word dataset;

[0151] The dialogue dataset is obtained by filtering out low-frequency words from the dialogue word dataset.

[0152] In an optional embodiment, the obtaining module 111 is used to obtain multiple response texts for each input text in the dialogue dataset based on the dialogue generation model to obtain a response candidate pool for each input text.

[0153] In an optional embodiment, the annotation module 112 is used to annotate the response texts in all response candidate pools to obtain a dialogue training dataset.

[0154] In an optional embodiment, the step of annotating the response texts in all response candidate pools to obtain the dialogue training dataset includes:

[0155] Based on the dialogue generation model, obtain the response candidate pool corresponding to the user's input text in each round of dialogue in the dialogue dataset;

[0156] The reply text in the reply candidate pool is labeled according to the preset dialogue discrimination labels to obtain a reply text label set. The dialogue discrimination labels include global discrimination labels, local discrimination labels, security discrimination labels and scene discrimination labels.

[0157] The response candidate pool is traversed to obtain the response text label set generated in each round of dialogue, and all the input texts in the dialogue dataset and the response text label set corresponding to each input text are used as the dialogue training dataset.

[0158] In an optional embodiment, the learning module 113 is used to perform multi-task learning based on the dialogue training dataset and a preset neural network model to obtain a scenario response evaluation model.

[0159] In an optional embodiment, the step of obtaining a scene response evaluation model through multi-task learning based on the dialogue training dataset and a preset neural network model includes:

[0160] The preset neural network model is trained based on the target discrimination labels in the dialogue training dataset so that the neural network model can obtain the target discrimination score value of the response text, wherein the target discrimination label is any one of the discrimination labels;

[0161] The global discrimination label, local discrimination label, security discrimination label and scene discrimination label are traversed to enable the neural network model to perform multi-task learning, and after multi-task learning, the global discrimination score, local discrimination score, security discrimination score and scene discrimination score of the response text can be obtained simultaneously.

[0162] A neural network model capable of simultaneously acquiring the global discrimination score, local discrimination score, security discrimination score, and scene discrimination score of the response text is used as the scene response evaluation model.

[0163] In an optional embodiment, the acquisition module 114 is used to acquire dialogue text in real time and obtain a target response candidate pool of the target input text in the dialogue text through a dialogue generation model.

[0164] In an optional embodiment, the scoring module 115 is used to score the response texts in the target response candidate pool based on the scenario response evaluation model to obtain the target response score value.

[0165] In an optional embodiment, training a preset neural network model based on target discrimination labels in the dialogue training dataset to enable the neural network model to obtain the target discrimination score value of the response text includes:

[0166] Based on a pre-defined neural network model, the dialogue data labeled with target discrimination tags is converted into dialogue word vectors;

[0167] Construct a dialogue word weight matrix based on the dialogue word vectors;

[0168] Based on the dialogue word weight matrix, the dialogue data labeled with target discrimination tags is transformed into dialogue sentence vectors;

[0169] The response text vector is constructed based on the dialogue statement vector;

[0170] The response text vector is classified based on the classifier in the neural network model to obtain the output target discrimination score value of the response text vector;

[0171] The classification loss between the output target discrimination score and the corresponding target discrimination label is calculated based on the cross-entropy loss function.

[0172] The neural network model is iteratively trained based on the classification loss so that the neural network model can obtain the target discrimination score value of the response text.

[0173] In an optional embodiment, the step of scoring the response texts in the target response candidate pool based on the scenario response evaluation model to obtain the target response score value includes:

[0174] Based on the scenario response evaluation model, the global discrimination score, local discrimination score, security discrimination score, and scenario discrimination score of each response text in the target response candidate pool are calculated;

[0175] The global discrimination score, local discrimination score, security discrimination score, and scene discrimination score of the response text are weighted according to the scene in which the target response candidate pool is located.

[0176] The target response score is obtained by weighted summing of the global discrimination score, local discrimination score, security discrimination score, scenario discrimination score, and the weights corresponding to each discrimination score of the response text. The target response score corresponds one-to-one with the response text.

[0177] In an optional embodiment, the filtering module 116 is used to filter out response texts that match the target input text based on the target response score to respond to the target input text.

[0178] In an optional embodiment, the step of filtering response texts that match the target input text based on the target response score to respond to the target input text includes:

[0179] The target response score values ​​of each response text in the target response candidate pool are sorted in descending order;

[0180] Select the response text corresponding to the target response score that ranks first as the response text that matches the target input text;

[0181] Reply to the target input text based on the response text that matches the target input text.

[0182] As can be seen from the above technical solutions, this application can use the collected dialogue dataset to train the neural network model to obtain a scene response evaluation model through multi-task learning. This scene response evaluation model can comprehensively consider whether the response text is reasonable from multiple perspectives, thereby filtering out response text that matches the user's input text and effectively improving the accuracy of automatic dialogue response.

[0183] Please see Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device 1 includes a memory 12 and a processor 13. The memory 12 is used to store computer-readable instructions, and the processor 13 is used to execute the computer-readable instructions stored in the memory to implement the artificial intelligence-based automatic response dialogue method described in any of the above embodiments.

[0184] In an alternative embodiment, the electronic device 1 further includes a bus and a computer program stored in the memory 12 and executable on the processor 13, such as an AI-based automated response dialog program.

[0185] Figure 3 Only electronic device 1 with memory 12 and processor 13 is shown. It will be understood by those skilled in the art that... Figure 3 The structure shown does not constitute a limitation on the electronic device 1, and may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0186] Combination Figure 1 The memory 12 in the electronic device 1 stores a plurality of computer-readable instructions to implement an artificial intelligence-based automatic response dialogue method, and the processor 13 can execute the plurality of instructions to achieve:

[0187] The collected dialogue text is preprocessed to obtain a dialogue dataset, which includes multi-turn dialogue text, each turn of dialogue text including an input text and a response text;

[0188] Based on the dialogue generation model, multiple response texts for each input text in the dialogue dataset are obtained to obtain a response candidate pool for each input text;

[0189] The dialogue training dataset is obtained by annotating the response texts in all response candidate pools.

[0190] Based on the dialogue training dataset and the preset neural network model, multi-task learning is performed to obtain a scenario response evaluation model;

[0191] Real-time acquisition of dialogue text, and acquisition of a target response candidate pool of target input text in the dialogue text through a dialogue generation model;

[0192] The target response score value is obtained by scoring the response texts in the target response candidate pool based on the scenario response evaluation model.

[0193] Based on the target response score, response texts that match the target input text are selected to respond to the target input text.

[0194] Specifically, the processor 13's implementation method for the above instructions can be found in [reference needed]. Figure 1 The descriptions of the relevant steps in the corresponding embodiments are not repeated here.

[0195] Those skilled in the art will understand that the schematic diagram is merely an example of electronic device 1 and does not constitute a limitation on electronic device 1. Electronic device 1 can be a bus-type structure or a star-type structure. Electronic device 1 may also include more or fewer other hardware or software than shown in the diagram, or different component arrangements. For example, electronic device 1 may also include input / output devices, network access devices, etc.

[0196] It should be noted that electronic device 1 is only an example. Other existing or future electronic products that are suitable for this application should also be included within the scope of protection of this application and are incorporated herein by reference.

[0197] The memory 12 includes at least one type of readable storage medium, which can be non-volatile or volatile. The readable storage medium includes flash memory, portable hard drives, multimedia cards, card-type memory (e.g., SD or DX memory), magnetic storage, magnetic disks, optical disks, etc. In some embodiments, the memory 12 can be an internal storage unit of the electronic device 1, such as a portable hard drive of the electronic device 1. In other embodiments, the memory 12 can also be an external storage device of the electronic device 1, such as a plug-in portable hard drive, Smart Media Card (SMC), Secure Digital (SD) card, Flash Card, etc., equipped on the electronic device 1. The memory 12 can be used not only to store application software and various types of data installed on the electronic device 1, such as the code of an AI-based automated response dialog program, but also to temporarily store data that has been output or will be output.

[0198] In some embodiments, the processor 13 may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits packaged with the same or different functions, including combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips. The processor 13 is the control unit of the electronic device 1, connecting various components of the electronic device 1 via various interfaces and lines. It executes programs or modules stored in the memory 12 (e.g., executing an AI-based automated response dialogue program) and calls data stored in the memory 12 to perform various functions and process data of the electronic device 1.

[0199] The processor 13 executes the operating system of the electronic device 1 and various installed applications. The processor 13 executes the applications to implement the steps in the various embodiments of the AI-based automatic response dialogue method described above, for example... Figure 1 The steps are shown.

[0200] For example, the computer program may be divided into one or more units / modules, which are stored in the memory 12 and executed by the processor 13 to complete this application. The one or more units / modules may be a series of computer-readable instruction segments capable of performing specific functions, which describe the execution process of the computer program in the electronic device 1. For example, the computer program may be divided into a preprocessing module 110, an acquisition module 111, an annotation module 112, a learning module 113, a collection module 114, a scoring module 115, and a filtering module 116.

[0201] The integrated unit implemented as a software functional module described above can be stored in a computer-readable storage medium. This software functional module, stored in a storage medium, includes several instructions to cause a computer device (which may be a personal computer, computer equipment, or network device, etc.) or processor to execute portions of the AI-based automatic response dialogue method described in the various embodiments of this application.

[0202] If the unit / module integrated in electronic device 1 is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware devices. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above.

[0203] The computer program includes computer program code, which may be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory, and other memory.

[0204] Furthermore, the computer-readable storage medium may primarily include a stored program area and a stored data area, wherein the stored program area may store the operating system, an application program required for at least one function, etc.; and the stored data area may store data created based on the use of blockchain nodes, etc.

[0205] The blockchain referred to in this application is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms. Essentially, a blockchain is a decentralized database, a chain of data blocks linked together using cryptographic methods. Each data block contains information about a batch of network transactions, used to verify the validity of the information (anti-counterfeiting) and generate the next block. A blockchain can include an underlying blockchain platform, a platform product service layer, and an application service layer.

[0206] The bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This bus can be divided into address bus, data bus, control bus, etc. For ease of representation, in... Figure 3 The symbol is represented by only one arrow, but this does not indicate that there is only one bus or one type of bus. The bus is configured to enable communication between the memory 12 and at least one processor 13, etc.

[0207] This application also provides a computer-readable storage medium (not shown) storing computer-readable instructions, which are executed by a processor in an electronic device to implement the AI-based automatic response dialogue method described in any of the above embodiments.

[0208] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0209] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0210] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0211] Furthermore, the word "comprising" clearly does not exclude other modules or steps, and the singular does not exclude the plural. Multiple modules or devices described in the specification can also be implemented by a single module or device through software or hardware. Terms such as "first," "second," etc., are used to indicate names and do not indicate any specific order.

[0212] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application and are not intended to limit it. Although this application has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of this application without departing from the spirit and scope of the technical solutions of this application.

Claims

1. An automatic response dialogue method based on artificial intelligence, characterized in that, The method includes: The collected dialogue text is preprocessed to obtain a dialogue dataset, which includes multi-turn dialogue text, each turn of dialogue text including an input text and a response text; Based on the dialogue generation model, multiple response texts for each input text in the dialogue dataset are obtained to obtain a response candidate pool for each input text; The process of labeling response texts in all response candidate pools to obtain a dialogue training dataset includes: obtaining a response candidate pool corresponding to the user's input text in each round of dialogue in the dialogue dataset based on a dialogue generation model; labeling the response texts in the response candidate pools according to preset dialogue discrimination labels to obtain a response text label set, wherein the dialogue discrimination labels include global discrimination labels, local discrimination labels, security discrimination labels, and scene discrimination labels; traversing the response candidate pools to obtain the response text label set generated in each round of dialogue, and using all input texts in the dialogue dataset and the response text label set corresponding to each input text as the dialogue training dataset; The process of obtaining a scene response evaluation model through multi-task learning based on the dialogue training dataset and a preset neural network model includes: training the preset neural network model with target discriminant labels in the dialogue training dataset so that the neural network model can obtain the target discriminant score value of the response text, wherein the target discriminant label is any one of the discriminant labels; traversing the global discriminant label, local discriminant label, safety discriminant label, and scene discriminant label to enable the neural network model to perform multi-task learning, and after multi-task learning, being able to simultaneously obtain the global discriminant score value, local discriminant score value, safety discriminant score value, and scene discriminant score value of the response text; and using the neural network model that can simultaneously obtain the global discriminant score value, local discriminant score value, safety discriminant score value, and scene discriminant score value of the response text as the scene response evaluation model. Real-time acquisition of dialogue text, and acquisition of a target response candidate pool of target input text in the dialogue text through a dialogue generation model; The target response score value is obtained by scoring the response texts in the target response candidate pool based on the scenario response evaluation model. Based on the target response score, response texts that match the target input text are selected to respond to the target input text.

2. The AI-based automatic response dialogue method as described in claim 1, characterized in that, The process of preprocessing the collected dialogue text to obtain the dialogue dataset includes: The dialogue text dataset is obtained by crawling dialogue text according to a preset method. The dialogue text dataset is segmented into words to obtain a dialogue word dataset; Statistical analysis of word frequencies in the aforementioned dialogue word dataset; The dialogue dataset is obtained by filtering out low-frequency words from the dialogue word dataset.

3. The AI-based automatic response dialogue method as described in claim 1, characterized in that, The step of training a preset neural network model based on the target discrimination labels in the dialogue training dataset to enable the neural network model to obtain the target discrimination score value of the response text includes: Based on a pre-defined neural network model, the dialogue data labeled with target discrimination tags is converted into dialogue word vectors; Construct a dialogue word weight matrix based on the dialogue word vectors; Based on the dialogue word weight matrix, the dialogue data labeled with target discrimination tags is transformed into dialogue sentence vectors; The response text vector is constructed based on the dialogue statement vector; The response text vector is classified based on the classifier in the neural network model to obtain the output target discrimination score value of the response text vector; The classification loss between the output target discrimination score and the corresponding target discrimination label is calculated based on the cross-entropy loss function. The neural network model is iteratively trained based on the classification loss so that the neural network model can obtain the target discrimination score value of the response text.

4. The AI-based automatic response dialogue method as described in claim 1, characterized in that, The step of scoring the response texts in the target response candidate pool based on the scenario response evaluation model to obtain the target response score value includes: Based on the scenario response evaluation model, the global discrimination score, local discrimination score, security discrimination score, and scenario discrimination score of each response text in the target response candidate pool are calculated; The global discrimination score, local discrimination score, security discrimination score, and scene discrimination score of the response text are weighted according to the scene in which the target response candidate pool is located. The target response score is obtained by weighted summing of the global discrimination score, local discrimination score, security discrimination score, scenario discrimination score, and the weights corresponding to each discrimination score of the response text. The target response score corresponds one-to-one with the response text.

5. The AI-based automatic response dialogue method as described in claim 1, characterized in that, The step of filtering response texts that match the target input text based on the target response score to respond to the target input text includes: The target response score values ​​of each response text in the target response candidate pool are sorted in descending order; Select the response text corresponding to the target response score that ranks first as the response text that matches the target input text; Reply to the target input text based on the response text that matches the target input text.

6. An automatic response dialogue device based on artificial intelligence, characterized in that, The device includes a preprocessing module, an acquisition module, an annotation module, a learning module, a data acquisition module, a scoring module, and a filtering module. The preprocessing module is used to preprocess the collected dialogue text to obtain a dialogue dataset, which includes multi-turn dialogue text, each turn of dialogue text including an input text and a response text. The obtaining module is used to obtain multiple response texts for each input text in the dialogue dataset based on the dialogue generation model to obtain a response candidate pool for each input text; The annotation module is used to annotate the response texts in all response candidate pools to obtain a dialogue training dataset. This includes: obtaining the response candidate pool corresponding to the user's input text in each round of dialogue based on the dialogue generation model; annotating the response texts in the response candidate pools with preset dialogue discrimination labels to obtain a response text label set, where the dialogue discrimination labels include global discrimination labels, local discrimination labels, security discrimination labels, and scene discrimination labels; traversing the response candidate pools to obtain the response text label set generated in each round of dialogue; and using all input texts in the dialogue dataset and the response text label sets corresponding to each input text as the dialogue training dataset. The learning module is used to perform multi-task learning based on the dialogue training dataset and a preset neural network model to obtain a scene response evaluation model. This includes: training the preset neural network model according to the target discriminant labels in the dialogue training dataset so that the neural network model can obtain the target discriminant score value of the response text, where the target discriminant label is any one of the discriminant labels; traversing the global discriminant label, local discriminant label, safety discriminant label, and scene discriminant label to enable the neural network model to perform multi-task learning, and after multi-task learning, being able to simultaneously obtain the global discriminant score value, local discriminant score value, safety discriminant score value, and scene discriminant score value of the response text; and using the neural network model that can simultaneously obtain the global discriminant score value, local discriminant score value, safety discriminant score value, and scene discriminant score value of the response text as the scene response evaluation model. The acquisition module is used to acquire dialogue text in real time and obtain the target response candidate pool of the target input text in the dialogue text through the dialogue generation model; The scoring module is used to score the response texts in the target response candidate pool based on the scenario response evaluation model to obtain the target response score value; The filtering module is used to filter out response texts that match the target input text based on the target response score, so as to respond to the target input text.

7. An electronic device, characterized in that, The electronic device includes: Memory, which stores computer-readable instructions; and The processor executes computer-readable instructions stored in the memory to implement the AI-based automated response dialogue method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the AI-based automatic response dialogue method as described in any one of claims 1 to 5.

Images