Intelligent voice system for purchasing based on ASR and NLP technologies

Abstract

The invention relates to the technical field of natural language processing, and provides an ASR and NLP technology-based intelligent voice system for purchasing, which comprises an integrated application layer, a voice service layer and a data storage layer, the integrated application layer comprises an external APP, an internal purchasing subsystem and a voice assistant; the voice service layer comprises an intelligent voice model integrating an ASR voice recognition model and a UIE information extraction model; the data storage layer comprises a database component and a knowledge graph and is used for data storage and information query; the voice service layer is connected through the integrated application layer, voice data are input, an intelligent voice model of the voice service layer carries out semantic understanding on voice, and result information is output by querying the data storage layer. The intelligent voice technology and natural language processing (intelligent questions and answers) can be applied to internal purchase management and external client communication, internal company and client service voices are effectively utilized, and the improvement of operation intelligence and the improvement of service level are facilitated.

Description

Technical Field

[0001] This application belongs to the field of natural language processing technology, and specifically relates to an intelligent voice system for procurement based on ASR and NLP technologies. Background Art

[0002] In today's wave of digital transformation, intelligent voice technology and natural language processing are gradually changing the way businesses operate and provide services. In the tobacco procurement sector in particular, how to effectively utilize these advanced technologies to improve internal management and customer service has become a critical issue.

[0003] In the field of tobacco procurement, existing intelligent voice technology, whether with customers or within the company, mostly relies on proprietary interfaces provided by communication software (WeChat, DingTalk, official accounts, etc.). Different software cannot communicate with each other, and improvements and optimizations rely on updates from software suppliers, making it impossible to perform targeted optimizations in the tobacco procurement field. Natural language processing also relies on model suppliers, making it impossible to perform targeted optimizations in the tobacco procurement field, and unable to associate with the voice received by communication software, making it impossible to coordinate and optimize, resulting in a huge waste of internal voice resources and hindering the advancement of intelligent operations and the improvement of service levels in the tobacco procurement field. Summary of the Invention

[0004] This application provides an intelligent voice system for procurement based on ASR and NLP technologies to solve or partially solve the problems raised in the above background technology.

[0005] This application provides an intelligent voice system for procurement based on ASR and NLP technologies, including: an integrated application layer, a voice service layer, and a data storage layer;

[0006] Integrated application layer, including external apps, internal procurement subsystems, and voice assistants;

[0007] The voice service layer includes an intelligent voice model that integrates the ASR speech recognition model and the UIE information extraction model;

[0008] The data storage layer includes database components and knowledge graphs for data storage and information query;

[0009] The voice service layer is connected through the integrated application layer, and voice data is input. The intelligent voice model of the voice service layer performs semantic understanding of the voice, queries the data storage layer, and outputs the result information.

[0010] Preferably, the external APP includes at least WeChat for Business, WeChat Official Account, and DingTalk;

[0011] The internal procurement subsystem includes at least user management and order management functions;

[0012] The voice assistant includes at least APP, mini program, official account, and H5.

[0013] Preferably, the database components include a Redis in-memory database, a Mysql relational database, and an OSS object storage service;

[0014] The content of the knowledge graph includes tobacco information, procurement information, and a combination of the two.

[0015] Preferably, the construction and training method of the intelligent sound model is as follows:

[0016] S1: Preprocessing of speech and text data;

[0017] S2: Build an ASR speech recognition model based on multiple neural network models and a UIE information extraction model based on the BERT model, and train and optimize them;

[0018] S3: Jointly optimize the ASR speech recognition model and UIE information extraction model.

[0019] Preferably, in step S1, the method for preprocessing the voice data is as follows:

[0020] Perform noise reduction processing on voice data to remove background noise;

[0021] Perform VAD voice activity detection on the voice data to remove silent segments and non-speech segments in the voice data;

[0022] Extract features from speech data and convert speech signals into feature vectors suitable for neural network processing;

[0023] The method for preprocessing text data is as follows:

[0024] Perform data cleaning on text data to remove meaningless characters, symbols, and special words;

[0025] Perform word segmentation on text data;

[0026] Label text data.

[0027] Preferably, in step S2, the training and optimization method of the ASR speech recognition model is as follows:

[0028] S211: Build one or more deep learning models based on DNN, RNN / LSTM, and CNN models;

[0029] S212: Divide the preprocessed speech data into a training set, a validation set, and a test set;

[0030] S213: Set the loss function number and optimization algorithm, and use the training set data to train the deep learning model;

[0031] S214: During the training process, the output and loss values ​​of the model are calculated using forward propagation, and then the gradient is calculated through backpropagation. The model parameters are updated using an optimization algorithm so that the loss of the corresponding model on the training set is gradually reduced and the performance on the validation set reaches a preset level.

[0032] S215: Perform fusion calculation on the outputs of each deep learning model.

[0033] Preferably, in step S2, the training and optimization method of the UIE information extraction model is as follows:

[0034] Construct a UIE model based on the BERT model, and add an output layer structure suitable for information extraction tasks on the basis of the BERT model; set the loss function and optimization algorithm, and use the preprocessed text data to train and optimize the UIE information extraction model.

[0035] Preferably, in step S3, the method for jointly optimizing the ASR speech recognition model and the UIE information extraction model includes adjusting the decoding strategy of the ASR speech recognition model so that its speech recognition text adapts to the UIE model for information extraction. The specific method is as follows:

[0036] S311: Obtain the keyword set V of the UIE information extraction model for different information extraction tasks key ;

[0037] S312: In the decoding process of the ASR speech recognition model, a keyword weighting mechanism is introduced to output a candidate sequence T for each text. candidate , calculate the probability weighted score of the keyword, the formula is:

[0038]

[0039] Among them, λ is the keyword weight coefficient, P(v|T candidate ) is the word v in the sequence candidate T candidate Probability of occurrence;

[0040] S313: Calculate the final probability score S of each text output sequence candidate final , select the one with the largest probability as the optimal text output sequence, as shown below:

[0041] S final (T candidate )=S asr (T eandidate) +S key (T candidate )

[0042] Among them, S asr (T candidate ) is the sequence candidate T candidate Original decoding score.

[0043] Preferably, in the beam search decoding process of the ASR speech recognition model, according to the final probability score S final The candidate sequences are screened and expanded, and sequences containing more keywords and with higher text probabilities are retained first.

[0044] Preferably, in step S3, the method for jointly optimizing the ASR speech recognition model and the UIE information extraction model further comprises setting a joint loss function for the ASR speech recognition model and the UIE information extraction model, and achieving collaborative optimization between the two by simultaneously updating the parameters of the ASR speech recognition model and the UIE information extraction model;

[0045] The setting method of the joint loss function is as follows:

[0046] S321: Setting the connection time classification loss function of the ASR speech recognition model

[0047]

[0048] Among them, T is the sequence length, C is the number of character categories, and y t,c is the indicator variable of the true label at position t and category c, is the predicted probability;

[0049] S322: Setting the cross entropy loss function of the UIE information extraction model

[0050]

[0051] Among them, N is the number of samples, M is the number of labels at each position in the sequence, K is the number of entity label types, and y i,j,k An indicator variable indicating that the true label of the i-th sample position j is k, is the probability that the model predicts that the position is label k;

[0052] S323: Setting the joint loss function

[0053] L joint =αL asr +βL uie

[0054] Among them, ɑ and β are balance weight coefficients;

[0055] S324: Setting up the stochastic gradient optimization algorithm

[0056]

[0057] Among them, θ represents the model parameters, η is the learning rate, is the gradient of the joint loss function with respect to the parameters.

[0058] Compared with the prior art, this application has the following beneficial effects:

[0059] (1) This application sets up an intelligent voice model that integrates an ASR speech recognition model and a UIE information extraction model, and associates the internal procurement subsystem with the external communication application, so that intelligent voice technology and natural language processing (intelligent question and answer) can be applied to internal procurement management and external customer communication, and the company's internal and customer service voices are effectively utilized, which is conducive to the improvement of business intelligence and service levels.

[0060] (2) This application adjusts the decoding strategy of the ASR speech recognition model and sets the joint loss function and optimization algorithm of the ASR speech recognition model and the UIE information extraction model, comprehensively considers the joint loss of speech recognition accuracy and information extraction accuracy, and achieves collaborative optimization between the two models by simultaneously updating the parameters of the two models, thereby improving the performance of the entire speech recognition and information extraction system. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] The present application is further described below with reference to the accompanying drawings and examples.

[0062] Figure 1 This is a schematic diagram of the system composition of this application. DETAILED DESCRIPTION

[0063] For example, certain words are used in the specification and claims to refer to specific components. Those skilled in the art should understand that hardware manufacturers may use different terms to refer to the same component. This specification and claims do not use differences in names as a way to distinguish components, but use differences in the functions of the components as the criteria for distinction. For example, "including" mentioned throughout the specification and claims is an open term, so it should be interpreted as "including but not limited to". "Approximately" means that within an acceptable error range, those skilled in the art can solve the technical problem within a certain error range and basically achieve the technical effect.

[0064] In the description of the present application, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "horizontal", etc., indicating directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings and are only used to facilitate the description of the present application and simplify the description. They do not indicate or imply that the devices or elements referred to must have a specific direction, be constructed and operate in a specific direction. Therefore, they should not be understood as limiting the present application.

[0065] In this application, unless otherwise specified or limited, the terms "mounted," "connected," "connect," "fixed," etc. should be understood broadly. For example, they can refer to fixed connection, detachable connection, or integral connection; mechanical connection or electrical connection; direct connection or indirect connection through an intermediate medium; or internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in this application based on specific circumstances.

[0066] Example 1

[0067] like Figure 1 As shown, this application provides an intelligent voice system for procurement based on ASR and NLP technologies, including: an integrated application layer, a voice service layer, and a data storage layer;

[0068] Integrated application layer, including external apps, internal procurement subsystems, and voice assistants;

[0069] The voice service layer includes an intelligent voice model that integrates the ASR speech recognition model and the UIE information extraction model;

[0070] The data storage layer includes database components and knowledge graphs for data storage and information query;

[0071] The voice service layer is connected through the integrated application layer, and voice data is input. The intelligent voice model of the voice service layer performs semantic understanding of the voice, queries the data storage layer, and outputs the result information.

[0072] Specifically, the external APP includes at least enterprise WeChat, WeChat official account, and DingTalk; the internal procurement subsystem includes at least user management, order management, and indicator retrieval functions; the voice assistant includes at least APP, mini-program, official account, and H5; the database components include Redis memory database, MySQL relational database, and OSS object storage service; the content of the knowledge graph includes tobacco information, procurement information, and a combination of the two.

[0073] Specifically, this application uses H5 (HTML5) to open up the entrance to conversational interactions through channels such as corporate WeChat, WeChat work accounts, and DingTalk, and adopts Redis, Mysql, and OSS as storage to ensure data persistence. Through the voice recognition technology of the intelligent voice model and the UIE information extraction model, the voice is first intelligently recognized, and then the recognized text is semantically understood. Finally, the database and knowledge graph are searched, and the result information based on semantic understanding is output to realize functions such as intelligent information retrieval, natural language question and answer, FAQ question and answer, and internal information query.

[0074] Specifically, the construction and training method of the intelligent sound model is as follows:

[0075] S1: Preprocess the speech and text data;

[0076] S2: Construct an ASR speech recognition model based on multiple neural network models and a UIE information extraction model based on the BERT model, and train and optimize them;

[0077] S3: Jointly optimize the ASR speech recognition model and the UIE information extraction model.

[0078] Furthermore, in the step S1, the method for preprocessing the speech data is as follows:

[0079] Perform noise reduction on the speech data. Use noise reduction algorithms such as spectral subtraction and wavelet transform noise reduction to remove background noise in the speech and improve the purity of the speech signal. For example, in the customer service call scenario, there may be background noises such as keyboard sounds and office environment noises, which need to be reduced through noise reduction to minimize their impact on speech recognition;

[0080] Perform voice activity detection (VAD) on the speech to distinguish between speech and non-speech parts, remove the silent segments and non-speech segments in the speech data, further improve the effectiveness of the speech data, and reduce the interference of invalid data on the model training and recognition process;

[0081] Extract the features of the speech, convert the speech signal into a feature vector suitable for deep learning model processing. Commonly used speech features include Mel Frequency Cepstral Coefficients (MFCC), spectrogram, etc. By extracting these features, the speech signal is converted from a time-domain or frequency-domain representation to a sequence of feature vectors that can reflect the essential features of the speech, providing input for subsequent deep learning model processing.

[0082] Specifically, the method for preprocessing the text data is as follows:

[0083] Clean the text data, remove meaningless characters, symbols, stop words, etc., such as "、", ",", "的", "了", etc., to reduce the complexity of the data and improve the running efficiency of the model;

[0084] Perform word segmentation on the text, cut the continuous text into individual words or phrases. For example, cut "我要查询公开招标采购方式" into words such as "我", "要", "查询", "公开招标", "采购方式", etc., to facilitate subsequent morphological analysis and syntactic analysis;

[0085] The text data is labeled. Preferably, in addition to traditional BIO labeling, the text is emotionally labeled according to the emotional tendency of the text content, and the text is labeled as positive, negative or neutral emotion categories. For example, "This procurement supplier's service is good, I am very satisfied" is labeled as positive emotion, "This process is very complicated, I am very disappointed" is labeled as negative emotion, and "Please ask for a competitive negotiation procurement document template" is labeled as neutral emotion.

[0086] Specifically, in step S2, the training and optimization method of the ASR speech recognition model is as follows:

[0087] S211: Build one or more deep learning models based on DNN, RNN / LSTM, and CNN models;

[0088] S212: Divide the preprocessed speech data into a training set, a validation set, and a test set;

[0089] S213: Set the loss function number and optimization algorithm, and use the training set data to train the deep learning model;

[0090] S214: During the training process, the output and loss values ​​of the model are calculated using forward propagation, and then the gradient is calculated through backpropagation. The model parameters are updated using an optimization algorithm so that the loss of the corresponding model on the training set is gradually reduced and the performance on the validation set reaches a preset level.

[0091] S215: Perform fusion calculation on the outputs of each deep learning model.

[0092] DNN model construction: Determine the structure of the DNN model, including the input layer, multiple hidden layers, and the output layer. The input layer receives the preprocessed speech feature vector. The hidden layers use activation functions (such as ReLU) to enhance the model's nonlinear fitting capabilities. The output layer is designed based on the needs of the speech recognition task. For example, in speech recognition for a specific vocabulary, the number of neurons in the output layer can correspond to the size of the vocabulary, and the softmax activation function can be used to output the probability distribution of each word.

[0093] RNN / LSTM model construction: Given the temporal nature of speech signals, RNN / LSTM models can better capture the temporal dependencies of speech. When constructing an RNN / LSTM model, the number of network layers and the number of neurons in each layer must be determined. The LSTM unit controls the flow and memory of information through a structure consisting of forget gates, input gates, and output gates, effectively addressing the vanishing and exploding gradient issues common in traditional RNNs. The RNN / LSTM model sequentially inputs a sequence of speech feature vectors, outputting corresponding state information at each time step for subsequent speech recognition decisions.

[0094] CNN model construction: For speech feature images (such as spectrograms), the CNN model can automatically learn local spatial features and hierarchical feature representations; design the CNN model's convolutional layer, pooling layer, and fully connected layer structures. The convolutional layer uses convolution kernels of varying sizes and numbers to perform convolution operations on the input feature image to extract local features; the pooling layer is used to reduce the spatial dimension of the feature image while retaining important feature information; the fully connected layer integrates the convolutional and pooling features to output the speech recognition results. For example, when processing a spectrogram, the convolutional layer can extract the frequency and temporal local features in the spectrogram, the pooling layer can downsample these features, and the fully connected layer maps the features to the output space of the speech text.

[0095] The preprocessed speech data is divided into a training set, a validation set, and a test set. Random division is usually used to ensure that the training set contains enough samples for model training. The validation set is used to evaluate the model and adjust hyperparameters during training. The test set is used to finally evaluate the performance of the model. For example, the data can be divided into 70% training set, 15% validation set, and 15% test set ratios.

[0096] Select appropriate loss functions and optimization algorithms. In speech recognition tasks, commonly used loss functions include cross-entropy loss functions, and optimization algorithms such as stochastic gradient descent (SGD) and Adam are used to minimize the loss function and update the model parameters. During training, the output and loss value of the model are calculated through forward propagation, and then the gradient is calculated through backpropagation, and the model parameters are updated using the optimization algorithm.

[0097] Perform multiple rounds of iterative training and gradually adjust the model parameters to gradually reduce the model loss on the training set and achieve a better performance on the validation set. During the training process, an early stopping strategy can be adopted. When the performance on the validation set no longer improves after several rounds of iterations, stop training to avoid overfitting of the model.

[0098] Evaluate and analyze the trained deep learning models (such as DNN, RNN / LSTM, and CNN) to identify their respective advantages, disadvantages, and applicable scenarios. For example, the DNN model may have a high recognition accuracy when processing simple speech feature patterns, but its ability to capture temporal features is relatively weak; the RNN / LSTM model performs better in processing long-term speech dependencies; and the CNN model has a strong ability to extract local and hierarchical features in speech feature images.

[0099] Model fusion technology is used to combine multiple deep learning models to improve speech recognition performance. Fusion strategies such as simple averaging and weighted averaging can be used to weight the outputs of each model to obtain the final speech recognition result. Alternatively, more complex fusion methods such as stacking fusion and voting fusion can be used to train a meta-learner to comprehensively judge the outputs of each model, further improving the accuracy and robustness of speech recognition.

[0100] Preferably, in step S2, the training and optimization method of the UIE information extraction model is as follows:

[0101] Construct a UIE model based on the BERT model, and add an output layer structure suitable for information extraction tasks on the basis of the BERT model; set the loss function and optimization algorithm, and use the preprocessed text data to train and optimize the UIE information extraction model.

[0102] The BERT (Bidirectional Encoder Representations from Transformers) model is a pre-trained model that is fine-tuned to adapt to specific information extraction tasks. On the basis of the pre-trained model, an output layer structure suitable for information extraction tasks is added, such as a sequence annotation layer for entity recognition (using annotation schemes such as BIO) or a classification layer for relationship extraction. The labeled information extraction training data is used to adjust the model parameters so that the model can accurately extract the required structured information from the text after speech recognition. During the training process, loss functions and optimization algorithms are also used to guide the model's learning process. For example, the conditional random field (CRF) loss function can be used for sequence labeling tasks, and the cross entropy loss function can be used for classification tasks.

[0103] Furthermore, the training and optimization method of the UIE information extraction model is further elaborated: lexical analysis model construction and training, syntactic analysis model construction and training, sentiment analysis model construction and training, model fusion and optimization.

[0104] Lexical analysis model construction and training: Build a lexical analysis model based on word vectors. For example, models such as Word2Vec and GloVe are used to map words into a vector space, capturing the semantic similarity and correlation between words. By training the model, semantically similar words have similar vector representations in the vector space. For example, "direct procurement" and "direct procurement applicable conditions" are relatively close in the vector space.

[0105] The model is trained and optimized using preprocessed text data. By adjusting the model parameters, the model can accurately perform lexical analysis on vocabulary, including tasks such as part-of-speech tagging and named entity recognition. For example, in the part-of-speech tagging task, the model can accurately distinguish the different parts of speech of "supply" in "procurement supplier selection" (noun) and "fixed-point supply" (verb).

[0106] Syntactic analysis model construction and training: Building a syntactic analysis model based on deep learning, such as a model using the Transformer architecture, can capture long-range dependencies and sentence structure information in the text. By training the model, it can generate a syntactic tree for the text, showing the syntactic relationships between each word in the sentence, such as subject-verb relationships and verb-object relationships.

[0107] The model is trained and validated using text data labeled with syntactic relations. By minimizing the loss function and adjusting the model parameters, the model's recognition accuracy of syntactic relations is improved. For example, when processing the sentence "In XX year, XX Tobacco Company encountered XXXX problems during XX procurement inspection", the model can recognize that "XX Tobacco Company" is the subject, "appeared" is the predicate, and "in XX procurement inspection" is the adverbial, and generate the corresponding syntactic tree structure.

[0108] Sentiment analysis model construction and training: Constructing a sentiment analysis model, such as one based on a recurrent neural network (RNN), can process text sequence information and capture the contextual dependencies of sentiment expressions in the text. By training the model, the model can accurately determine the sentiment tendency of the text based on the words and phrases in the text.

[0109] The sentiment analysis model is trained and optimized using text data labeled with sentiment tendencies. Model performance is evaluated using methods such as cross-validation. The model's hyperparameters (such as learning rate, hidden layer size, etc.) are adjusted to improve the model's classification accuracy for sentiment polarity. For example, when processing text such as "This procurement feedback has not been resolved," the model can accurately determine that its sentiment tendency is negative.

[0110] Model fusion and optimization: This integrates lexical analysis, syntactic analysis, and sentiment analysis models to build a comprehensive NLP model that comprehensively analyzes natural language text input by users. For example, lexical analysis first processes the text at the lexical level to identify key entities and parts of speech. A syntactic analysis model then analyzes sentence structure and determines the relationships between words. Finally, a sentiment analysis model is used to determine the emotional orientation of the text, thereby more accurately understanding the user's intent.

[0111] Perform joint optimization of the model, comprehensively consider the performance indicators of lexical, syntactic and sentiment analysis tasks (such as accuracy, recall rate, etc.), adjust the parameters and structure of the model, so that the comprehensive model can achieve a good performance balance in various tasks. For example, by adjusting the weight of the model's loss function, the model will not overly rely on syntactic structure information when processing sentiment analysis tasks, while not ignoring the results of lexical analysis.

[0112] Specifically, in step S3, the method for jointly optimizing the ASR speech recognition model and the UIE information extraction model includes adjusting the decoding strategy of the ASR speech recognition model so that its speech recognition text adapts to the UIE model for information extraction. The specific method is as follows:

[0113] S311: Obtain the keyword set V of the UIE information extraction model for different information extraction tasks key , such as entity words such as cigarette specification names, cigarette brand names, cigarette supplier names, and specific field terms;

[0114] S312: In the decoding process of the ASR speech recognition model, a keyword weighting mechanism is introduced to output a candidate sequence T for each text. candidate , calculate the probability weighted score of the keyword, the formula is:

[0115]

[0116] Among them, λ is the keyword weight coefficient, P(v|T candidate ) is the word v in the sequence candidate T candidate Probability of occurrence;

[0117] S313: Calculate the final probability score S of each text output sequence candidate final , select the one with the largest probability as the optimal text output sequence, as shown below:

[0118] S final (T candidate )=S asr (T candidate )+S key (T candidate )

[0119] Among them, S asr (T candidate ) is the sequence candidate T candidate Original decoding score.

[0120] Preferably, in the beam search decoding process of the ASR speech recognition model, according to the final probability score S finalThe candidate sequences are screened and expanded, and sequences containing more keywords and with higher text probabilities are retained first, so that the text output by ASR is more conducive to information extraction by the UIE information extraction model.

[0121] Specifically, in step S3, the method for jointly optimizing the ASR speech recognition model and the UIE information extraction model further includes setting a joint loss function for the ASR speech recognition model and the UIE information extraction model, and achieving collaborative optimization between the two by simultaneously updating the parameters of the ASR speech recognition model and the UIE information extraction model;

[0122] The setting method of the joint loss function is as follows:

[0123] S321: Setting the connection time classification loss function of the ASR speech recognition model

[0124]

[0125] Among them, T is the sequence length, C is the number of character categories, and y t,c is the indicator variable of the true label at position t and category c, is the predicted probability;

[0126] S322: Setting the cross entropy loss function of the UIE information extraction model

[0127]

[0128] Among them, N is the number of samples, M is the number of labels at each position in the sequence, K is the number of entity label types, and y i,j,k An indicator variable indicating that the true label of the i-th sample position j is k, is the probability that the model predicts that the position is label k;

[0129] S323: Setting the joint loss function

[0130] L joint =αL asr +βL uie

[0131] Among them, ɑ and β are balance weight coefficients;

[0132] S324: Setting up the stochastic gradient optimization algorithm

[0133]

[0134] Among them, θ represents the model parameters, η is the learning rate, is the gradient of the joint loss function with respect to the parameters.

[0135] In the model evaluation and iterative optimization stage, the speech recognition accuracy of the ASR model (such as character error rate CER, etc.) and the information extraction accuracy of the UIE model (such as entity F1 value, relationship accuracy, etc.) are evaluated on the validation set. Based on the evaluation results, hyperparameters such as the balance weight parameters α and β, and the learning rate η are adjusted, or the model architecture is fine-tuned. Then, the joint training process is repeated until satisfactory performance indicators are achieved.

[0136] The above describes the implementation methods of the present application in detail in conjunction with the accompanying drawings, but the present application is not limited to the above implementation methods. Various changes can be made within the scope of knowledge possessed by ordinary technicians in the relevant technical field without departing from the purpose of the present application.

Claims

1. Intelligent voice system for procurement based on ASR and NLP technology, characterized by: include: Integrate application layer, voice service layer, and data storage layer; Integrated application layer, including external apps, internal procurement subsystems, and voice assistants; The voice service layer includes an intelligent voice model that integrates the ASR speech recognition model and the UIE information extraction model; The data storage layer includes database components and knowledge graphs for data storage and information query; The voice service layer is connected through the integrated application layer, and voice data is input. The intelligent voice model of the voice service layer performs semantic understanding of the voice, queries the data storage layer, and outputs the result information.

2. The intelligent voice system for procurement based on ASR and NLP technology according to claim 1, characterized in that: The external APPs include at least WeChat for Business, WeChat Official Account, and DingTalk; The internal procurement subsystem includes at least user management and order management functions; The voice assistant includes at least APP, mini program, official account, and H5.

3. The intelligent voice system for procurement based on ASR and NLP technology according to claim 1, characterized in that: The database components include Redis in-memory database, MySQL relational database and OSS object storage service; The content of the knowledge graph includes tobacco information, procurement information, and a combination of the two.

4. The intelligent voice system for procurement based on ASR and NLP technology according to claim 1, characterized in that: The construction and training method of the intelligent sound model is as follows: S1: Preprocessing of speech and text data; S2: Build an ASR speech recognition model based on multiple neural network models and a UIE information extraction model based on the BERT model, and train and optimize them; S3: Jointly optimize the ASR speech recognition model and UIE information extraction model.

5. The intelligent voice system for procurement based on ASR and NLP technology according to claim 4, characterized in that: In step S1, the method for preprocessing the voice data is as follows: Perform noise reduction processing on voice data to remove background noise; Perform VAD voice activity detection on the voice data to remove silent segments and non-speech segments in the voice data; Extract features from speech data and convert speech signals into feature vectors suitable for neural network processing; The method for preprocessing text data is as follows: Perform data cleaning on text data to remove meaningless characters, symbols, and special words; Perform word segmentation on text data; Label the text data.

6. The intelligent voice system for procurement based on ASR and NLP technology according to claim 4, characterized in that: In step S2, the training and optimization method of the ASR speech recognition model is as follows: S211: Build one or more deep learning models based on DNN, RNN / LSTM, and CNN models; S212: Divide the preprocessed speech data into a training set, a validation set, and a test set; S213: Set the loss function number and optimization algorithm, and use the training set data to train the deep learning model; S214: During the training process, the output and loss values ​​of the model are calculated using forward propagation, and then the gradient is calculated through backpropagation. The model parameters are updated using an optimization algorithm so that the loss of the corresponding model on the training set is gradually reduced and the performance on the validation set reaches a preset level. S215: Perform fusion calculation on the outputs of each deep learning model.

7. The intelligent voice system for procurement based on ASR and NLP technology according to claim 6, characterized in that: In step S2, the training and optimization method of the UIE information extraction model is as follows: Construct a UIE model based on the BERT model, and add an output layer structure suitable for information extraction tasks on the basis of the BERT model; set the loss function and optimization algorithm, and use the preprocessed text data to train and optimize the UIE information extraction model.

8. The intelligent voice system for procurement based on ASR and NLP technology according to claim 4, characterized in that: In step S3, the method for jointly optimizing the ASR speech recognition model and the UIE information extraction model includes adjusting the decoding strategy of the ASR speech recognition model so that its speech recognition text adapts to the UIE model for information extraction. The specific method is as follows: S311: Obtain the keyword set V of the UIE information extraction model for different information extraction tasks key ; S312: In the decoding process of the ASR speech recognition model, a keyword weighting mechanism is introduced to output a candidate sequence T for each text. candidate , calculate the probability weighted score of the keyword, the formula is: Among them, λ is the keyword weight coefficient, P(v|T candidate ) is the word v in the sequence candidate T candidate Probability of occurrence; S313: Calculate the final probability score S of each text output sequence candidate final , select the one with the largest probability as the optimal text output sequence, as shown below: S final (T candidate )=S asr (T candidate )+S key (T candidate ) Among them, S asr (T candidate ) is the sequence candidate T candidate Original decoding score.

9. The intelligent voice system for procurement based on ASR and NLP technology according to claim 8, characterized in that: In the beam search decoding process of the ASR speech recognition model, according to the final probability score S final The candidate sequences are screened and expanded, and sequences containing more keywords and with higher text probabilities are retained first.

10. The intelligent voice system for procurement based on ASR and NLP technology according to any one of claims 4 or 8, characterized in that: In step S3, the method for jointly optimizing the ASR speech recognition model and the UIE information extraction model further includes setting a joint loss function for the ASR speech recognition model and the UIE information extraction model, and achieving collaborative optimization between the two by simultaneously updating the parameters of the ASR speech recognition model and the UIE information extraction model; The setting method of the joint loss function is as follows: S321: Setting the connection time classification loss function of the ASR speech recognition model Among them, T is the sequence length, C is the number of character categories, and y t,c is the indicator variable of the true label at position t and category c, is the predicted probability; S322: Setting the cross entropy loss function of the UIE information extraction model Among them, N is the number of samples, M is the number of labels at each position in the sequence, K is the number of entity label types, and y i,j,k An indicator variable indicating that the true label of the i-th sample position j is k, is the probability that the model predicts that the position is label k; S323: Setting the joint loss function L joint =αL ast +βL uie Among them, ɑ and β are balance weight coefficients; S324: Setting up the stochastic gradient optimization algorithm Among them, θ represents the model parameters, η is the learning rate, is the gradient of the joint loss function with respect to the parameters.

Images