Hot word recognition method and device, electronic equipment and storage medium
By combining the first and second training sets to train the hot word recognition and hot event classification models, and using the hot event classification model to provide mutual supervision signals, the problem of limited training sample size in hot word recognition is solved, and the accuracy and efficiency of hot word recognition are improved.
Patent Information
- Application Number
- CN202510748645.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-05
- Publication Date
- 2025-10-17
AI Technical Summary
Existing hot word recognition methods have high annotation costs in massive training corpora, and due to the uncertainty and ambiguity of hot words, the training sample size is limited, affecting the accuracy of hot word recognition.
By obtaining the first training set and the second training set, the hot word recognition model and the hot event classification model are trained respectively. The hot event classification model is used to provide mutual supervision signals, and the hot word recognition model is updated to determine the new hot word annotations, thereby reducing the workload of hot word annotation and improving accuracy.
Iterative optimization of the hot word recognition model is achieved, which reduces the workload of hot word annotation and improves the accuracy of hot event mining.
Smart Images
Figure CN120804802A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of data processing, in particular to a hot word recognition method and device, electronic equipment and storage medium. BACKGROUND
[0002] Hot event mining and hot word mining are important tasks in the field of text data analysis, and their practical significance is very extensive. Hot event mining aims to dig out the topics or events that attract public attention in the current society from a large amount of text data, so as to provide decision basis and public opinion analysis for news media, government agencies, enterprise decision makers, etc. Hot word mining refers to identifying frequently occurring and representative words or phrases from a large amount of text data, which can help information consumers better understand and grasp the main content and trend of the current hot topic, and also help enterprises monitor and analyze the attention of their own brand, product, service, etc.
[0003] Hot word mining is one of the important tasks in the field of natural language processing and text mining, and its goal is to automatically discover the most commonly used and most representative key and hot words from large-scale text data. At present, the mainstream methods of hot word mining mainly include the following: (1) statistical-based method: this method is the most basic hot word mining method, which determines which words are the most common by counting the frequency of words in the text (such as TF-IDF score), and then determines which are hot words. (2) Topic model-based method: this method determines which words are related to the topic by mining the underlying topics in the text (such as using LDA model). (3) Deep learning-based method: this method determines the position of hot words through sequence labeling, and uses LSTM, LSTM+CRF, etc. Model to complete the sequence labeling task, so as to extract hot words.
[0004] Although the deep learning-based method achieves the best results, the data threshold of this method is high, which requires a large amount of training corpus, and due to the uncertainty and ambiguity of hot words, it is too high to annotate all hot words in a large training corpus. Therefore, the annotated samples that can actually be used for hot word recognition model training are usually very limited in size. SUMMARY
[0005] In view of the above problems, a hot word recognition method and device, electronic equipment and storage medium are provided to overcome the above problems or at least partially solve the above problems, which include:
[0006] A hot word recognition method, the method comprising:
[0007] obtain a first training set and a second training set, the first training set being first training samples of hotword annotation and hotspot event type annotation, and the second training set being second training samples of hotspot event type annotation;
[0008] train a model according to the first training set to obtain a hotword recognition model and a first hotspot event classification model, and train a model according to the second training set to obtain a second hotspot event classification model;
[0009] determine new hotword annotation of the second training set based on the hotword recognition model, the first hotspot event classification model and the second hotspot event classification model;
[0010] update the hotword recognition model and the first hotspot event classification model based on the new hotword annotation;
[0011] use the updated hotword recognition model and the first hotspot event classification model to extract hotwords from the obtained events to be classified.
[0012] Optionally, the determining of the new hotword annotation of the second training set based on the hotword recognition model, the first hotspot event classification model and the second hotspot event classification model comprises:
[0013] input the second training set into the hotword recognition model to obtain a candidate hotword list;
[0014] input the candidate hotword list into the first hotspot event classification model to output a first classification result and a classification confidence;
[0015] input the second training set into the second hotspot event classification model to output a second classification result;
[0016] determine the new hotword annotation of the second training set from the candidate hotword list based on the first classification result, the second classification result and the classification confidence.
[0017] Optionally, the determining of the new hotword annotation of the second training set from the candidate hotword list based on the first classification result, the second classification result and the classification confidence comprises:
[0018] determine whether the first classification result matches the second classification result, and determine whether the classification confidence is greater than a preset confidence threshold;
[0019] determine the new hotword annotation of the second training set from the candidate hotword list according to the determination result.
[0020] Optionally, the determining of the new hotword annotation of the second training set from the candidate hotword list according to the determination result comprises:
[0021] When it is determined that the first classification result matches the second classification result and the classification confidence is greater than a preset confidence threshold, a candidate hot word in the candidate hot word list is determined as a newly added hot word label of the second training set.
[0022] Optionally, the determining the newly added hot word label of the second training set from the candidate hot word list according to the determination result comprises:
[0023] When it is determined that the first classification result does not match the second classification result and / or the classification confidence is not greater than the preset confidence threshold, a candidate hot word with the lowest confidence is deleted from the candidate hot word list.
[0024] Based on the updated candidate hot word list, the candidate hot word list is input into the trained first hot event classification model to output a first classification result and a classification confidence.
[0025] Optionally, the model training according to the first training set to obtain a hot word recognition model and a first hot event classification model comprises:
[0026] Feature extraction is performed on each sample word in the first training set according to a plurality of preset feature types to obtain a plurality of candidate feature data.
[0027] The plurality of candidate feature data are spliced into a target feature vector.
[0028] The target feature vector is input into an initial hot word recognition model to obtain hot word prediction data of each sample word.
[0029] Based on the hot word prediction data and actual hot word data corresponding to the first sample word, model parameter updating is performed on the initial hot word recognition model.
[0030] Optionally, the updating the hot word recognition model and the first hot event classification model based on the newly added hot word label comprises:
[0031] The first training sample is updated based on the newly added hot word label.
[0032] The hot word recognition model and the first hot event are trained according to the updated first training sample.
[0033] A hot word recognition device, the device comprising:
[0034] The training set acquisition module is configured to acquire a first training set and a second training set, the first training set being first training samples of hotword annotation and hotspot event type annotation, and the second training set being second training samples of hotspot event type annotation;
[0035] The model training module is configured to train a model according to the first training set to obtain a hotword recognition model and a first hotspot event classification model, and train a model according to the second training set to obtain a second hotspot event classification model;
[0036] The new hotword annotation module is configured to determine new hotword annotation of the second training set based on the hotword recognition model, the first hotspot event classification model and the second hotspot event classification model.
[0037] The model updating module is configured to update the hotword recognition model and the first hotspot event classification model based on the new hotword annotation.
[0038] The hotword extraction module is configured to extract hotwords from the acquired events to be classified by using the updated hotword recognition model and the first hotspot event classification model.
[0039] An electronic device includes a processor, a memory, and a computer program stored on the memory and capable of running on the processor, and the computer program is executed by the processor to implement the hotword recognition method.
[0040] A computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the hotword recognition method.
[0041] The embodiments of the present application have the following advantages:
[0042] After the hotword recognition model, the first hotspot event classification model and the second hotspot event classification model are trained based on the first training set and the second training set, the new hotword annotation of the second training set can be determined based on the hotword recognition model, the first hotspot event classification model and the second hotspot event classification model, the hotword recognition model and the first hotspot event classification model can be updated, and the updated hotword recognition model and the first hotspot event classification model can be used to extract hotwords from the acquired events to be classified, so that the training strategy of using the hotspot event classification model to provide mutual supervision signals for the hotword recognition model to help the hotword recognition model to iteratively optimize is realized, the workload of hotword annotation can be reduced, and the hotspot event mining accuracy can be ensured to be higher. BRIEF DESCRIPTION OF DRAWINGS
[0043] In order to more clearly illustrate the technical solutions of the present application, the drawings needed to be used in the description of the present application will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort on the basis of these drawings.
[0044] Figure 1 is a step flow chart of a hot word recognition method provided by an embodiment of the present application;
[0045] Figure 2 is a step flow chart of another hot word recognition method provided by an embodiment of the present application;
[0046] Figure 3a is a schematic diagram of a hot word extraction method for hot event mining provided by an embodiment of the present application;
[0047] Figure 3b is a hot word recognition model provided by an embodiment of the present application to provide a supervision signal;
[0048] Figure 4 is a structural schematic diagram of a hot word recognition device provided by an embodiment of the present application. DETAILED DESCRIPTION
[0049] In order to make the above-mentioned purposes, features and advantages of the present application more apparent and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments. Obviously, the described embodiments are some embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without any creative effort fall within the scope of protection of the present application.
[0050] Referring to Figure 1 , a step flow chart of a hot word recognition method provided by an embodiment of the present application is shown, which can specifically include the following steps:
[0051] Step S101, a first training set and a second training set are obtained, the first training set is a first training sample with hot word annotation and hot event type annotation, and the second training set is a second training sample with hot event type annotation;
[0052] In actual application, the first training set can be a text containing hot words, and the text is annotated with hot word and hot event type, and the second training set can be a text annotated with hot event type.
[0053] The first training set and the second training set can be selected from various website comments and online and offline user evaluations and annotated correspondingly to serve as the training samples used in the embodiments of the present application.
[0054] Step S102, model training according to the first training set to obtain a hot word recognition model and a first hot event classification model, and model training according to the second training set to obtain a second hot event classification model;
[0055] In the embodiment of the present application, the initial training model of the hot word recognition model and the first hot event classification model and the second hot event classification model to be trained can be selected in advance. For example, BiLSTM (Bidirectional Long Short-Term Memory, bidirectional long short-term memory network) + CRF (Conditional Random Field, conditional random field) can be selected as the initial training model of the hot word recognition model in the embodiment of the present application. By training in combination with the training set selected in the embodiment of the present application, the hot word recognition model actually needed in the hot word extraction scene in the embodiment of the present application can be obtained. Similarly, the Text CNN (Text Convolutional Neural Network, text convolutional neural network) model can be selected as the initial model of the first hot event classification model and the second hot event classification model for training, and then the first training set is used to train the first hot event classification model, and the second training set is used to train the second hot event classification model.
[0056] In an embodiment of the present application, model training according to the first training set to obtain a hot word recognition model includes the following process:
[0057] Sub-step S11, performing feature extraction on the sample words of each first training sample in the first training set according to a plurality of preset feature types to obtain a plurality of candidate feature data;
[0058] In the embodiment of the present application, the first training set can be used to train the hot word recognition model. Specifically, the sample words in the first training set can be first subjected to feature extraction, and the feature extraction can be performed according to the preset feature type in the extraction process. Through feature extraction, important information in the sample words can be extracted to facilitate the determination of hot words in the word group. The preset feature type includes any one of the following: semantic type, position type, dependency syntax analysis type, and hot event type.
[0059] The specific extraction method for the above-mentioned preset feature type is as follows:
[0060] (1) Semantic feature extraction: ω ki The corresponding d e dimensional word vector is used as a semantic feature, denoted as e ki .
[0061] (2) Position feature extraction: ωki In D k The position in D ki is encoded into a feature vector, denoted as p ki , where the calculation method of p ki [j] is formula (1), t is the word ω k The position in D p , d ki is the dimension of the position feature vector, j represents the dimension of p p , l=0,1,2,…,d k / 2-1.
[0062]
[0063] Where,
[0064] (3) Dependency syntax feature extraction: the dependency syntax tree is used to analyze D ki , and the one-hot encoding of the depth of ω k in the corresponding dependency syntax tree is used as a feature vector, denoted as r ki .
[0065] (4) Hot event type feature extraction: the one-hot encoding of the corresponding hot event type of D k is used as a feature vector, denoted as t ki .
[0066] After feature extraction, multiple candidate feature data can be obtained, and each candidate feature data can be feature vector data.
[0067] Sub-step S12, the multiple candidate feature data are vector spliced to obtain a target feature vector;
[0068] In order to accurately determine the hot word judgment of the word group, different candidate feature data can be spliced to generate a target feature vector, which can represent the features of the word group from different angles, for example, the final feature vector of ω ki is V ki =[e ki , p ki , r ki , t ki ], and the total dimension is d.
[0069] Sub-step S13, input the target feature vector into the constructed initial hot word recognition model to obtain hot word prediction data of each sample word;
[0070] After obtaining the target feature vector, the target feature vector can be input into the initial hot word recognition model, and then the corresponding hot word prediction data is output. That is, the result of predicting whether the word group is a hot word.
[0071] In sub-step S14, the initial hotword recognition model is updated based on the hotword prediction data and the actual hotword data corresponding to the first sample word.
[0072] After obtaining the hotword prediction data, the model parameter value to be modified is determined by comparing the hotword prediction data with the actual hotword data of the first sample word. The modified model parameter value can make the model prediction result of the hotword recognition model close to the actual labeled hotword data.
[0073] In an embodiment of the present application, the specific process of training the first hot event classification model is as follows: for each sample D k in the first training set, only the hotwords (denoted as DW k ) in D k are retained, DW k is input into the TextCNN model for training, and the first hot event classification model after training is obtained. The first hot event classification model takes a hotword list as input and outputs the corresponding hot event type. In the training process, by inputting the hotword list into the first hot event classification model, the predicted hot event type can be obtained through the processing of the parameters in the model. The predicted hot event type is compared with the labeled hot event type in the first training set, and the model parameters are modified so that the data output by the subsequent first hot event classification model is closer to the labeled hot event type of the first training set.
[0074] According to the above training method, the corresponding hotword recognition model and the first hot event classification model can be obtained after multiple iterations.
[0075] Similarly, the second hot event classification model can also be trained using the second training set. The specific process of training the second hot event classification model is as follows: for each sample O k in the second training set ES, O k is directly input into the TextCNN model for training, and the second hot event classification model after training is obtained. The second hot event classification model takes a sentence as input and outputs the corresponding hot event type.
[0076] In the embodiment of the present application, the hot word recognition model WC is constructed, and the constructed hot word recognition model WC is trained using the first training set WS to obtain the trained hot word recognition model WC; wherein the first training set WS has a small amount of training samples with hot word annotation and hot event type annotation; meanwhile, the first hot event classification model WEC and the second hot event classification model SEC can be constructed, the first hot event classification model WEC is trained using the first training set WS, and the second hot event classification model SEC is trained using the second training set ES to obtain the trained first hot event classification model WEC and the trained second hot event classification model SEC; wherein the second training set ES can include a large number of training samples with only hot event type annotation.
[0077] In step S103, the new hot word annotation of the second training set is determined based on the hot word recognition model, the first hot event classification model and the second hot event classification model.
[0078] After the hot word recognition model, the first hot event classification model and the second hot event classification model are trained, the second training set can be used as the recognition subject, and the hot word recognition model, the first hot event classification model and the second hot event classification model are applied to perform hot word annotation of the second training set, i.e. to determine the hot word corresponding to the word group in the second training set, thereby expanding the hot word annotation and reducing the amount of original hot word annotation.
[0079] In step S104, the hot word recognition model and the first hot event classification model are updated based on the new hot word annotation.
[0080] After the new hot word annotation is determined, the hot word recognition model and the first hot event classification model can be retrained based on the new hot word to improve the accuracy of model recognition.
[0081] In step S105, the updated hot word recognition model and the first hot event classification model are used to extract hot words from the obtained events to be classified.
[0082] Then, the updated hot word recognition model and the first hot event classification model are used to perform actual hot word extraction of the events to be classified, and accurate hot words are obtained.
[0083] In an embodiment of the present application, updating the hot word recognition model and the first hot event classification model based on the new hot word annotation includes: updating the first training sample based on the new hot word annotation, i.e. supplementing the new hot word annotation into the first training sample; and then the hot word recognition model and the first hot event classification model can be trained according to the updated first training sample.
[0084] In the embodiment of the present application, after the hot word recognition model, the first hot event classification model, and the second hot event classification model are trained based on the first training set and the second training set, the newly added hot word labels of the second training set can be determined based on the hot word recognition model, the first hot event classification model, and the second hot event classification model, so as to update the hot word recognition model and the first hot event classification model, and the updated hot word recognition model and the first hot event classification model can be used to extract hot words from the obtained events to be classified, thereby realizing the training strategy of using the hot event classification model to provide mutual supervision signals for the hot word recognition model to help the hot word recognition model to iteratively optimize, and the workload of hot word labeling can be reduced and the hot event mining accuracy can be ensured to be higher.
[0085] Referring to Figure 2 , a step flowchart of another hot word recognition method provided by an embodiment of the present application is shown, which can specifically include the following steps:
[0086] In step S201, a first training set and a second training set are obtained, the first training set is a first training sample with hot word labels and hot event type labels, and the second training set is a second training sample with hot event type labels;
[0087] In step S202, a hot word recognition model and a first hot event classification model are obtained by model training according to the first training set, and a second hot event classification model is obtained by model training according to the second training set;
[0088] In step S203, the second training set is input into the hot word recognition model to obtain a candidate hot word list;
[0089] In step S204, the candidate hot word list is input into the first hot event classification model to output a first classification result and a classification confidence;
[0090] The classification confidence can be calculated according to the first hot event classification model after training, that is, the classification confidence of the first hot event classification model when performing classification can be calculated by extracting data in the second training set. The classification confidence can be calculated based on the prediction result of the hot event classification model and the actual classification result of the second training set.
[0091] In step S205, the second training set is input into the second hot event classification model to output a second classification result;
[0092] In step S206, the newly added hot word labels of the second training set are determined from the candidate hot word list based on the first classification result, the second classification result, and the classification confidence.
[0093] In the embodiment of the present application, the first hot event classification model WEC after training and the second hot event classification model SEC after training can also be used to classify hot event results, and the hot word recognition confidence of the hot word recognition model WC after training, and then based on the hot word recognition confidence, a supervision signal is constantly provided for the hot word recognition model WC to update the hot word recognition model.
[0094] In an embodiment of the present application, based on the first classification result, the second classification result and the classification confidence, the newly added hot word label of the second training set is determined from the candidate hot word list, including: judging whether the first classification result and the second classification result match, and judging whether the classification confidence is greater than the pre-set confidence threshold; according to the judgment result, the newly added hot word label of the second training set is determined from the candidate hot word list. The pre-set confidence threshold can be set according to actual needs, for example, the confidence threshold can be 0.8.
[0095] In an embodiment of the present application, when it is determined that the first classification result and the second classification result match, and the classification confidence is greater than the pre-set confidence threshold, the candidate hot word in the candidate hot word list is determined as the newly added hot word label of the second training set.
[0096] In another embodiment of the present application, when it is determined that the first classification result and the second classification result do not match and / or the classification confidence is not greater than the pre-set confidence threshold, the candidate hot word with the lowest confidence is deleted from the candidate hot word list; based on the updated candidate hot word list, the candidate hot word list is input into the first hot event classification model trained to output the first classification result and the classification confidence.
[0097] In the embodiment of the present application, the hot word list in a sentence can better reflect the hot event type of the sentence, so by comparing the output of the normal hot event classification model based on the sentence with the output of the hot event classification model based only on the hot word, if the difference between the two is too large, it means that the hot word list is not accurate, and on this basis, the hot word list is constantly adjusted to reduce the output difference, so as to obtain a more accurate hot word list.
[0098] Step S207, updating the hot word recognition model and the first hot event classification model based on the newly added hot word label;
[0099] Step S208, using the updated hot word recognition model and the first hot event classification model to extract hot words from the obtained events to be classified.
[0100] In the embodiment of the present application, the hot event classification model can be used to provide mutual supervision signals for the hot word recognition model to help the hot word recognition model to iteratively optimize the training strategy, and can reduce the workload of hot word labeling and ensure higher accuracy of hot event mining.
[0101] With reference to Figure 3a , a schematic diagram of a hot word extraction method for hotspot event mining in an embodiment of the present application is shown, which can specifically include the following processes:
[0102] (1) A hot word recognition model WC is constructed, and the constructed hot word recognition model WC is trained using a first training set WS to obtain a trained hot word recognition model WC; wherein the first training set WS has a small amount of training samples with hot word annotation and hotspot event type annotation;
[0103] The specific training process of the hot word recognition model WC is as follows:
[0104] For each word ω k of each sample D ki in the first training set WS, features are extracted from four aspects of semantics, position, dependency syntax analysis, and hotspot event type, and the extracted features are spliced to obtain the final feature vector of each word;
[0105] Specifically, the feature extraction process is as follows:
[0106] Semantic feature extraction: the d ki -dimensional word vector corresponding to ω e in the word2vec model is used as the semantic feature, denoted as e ki ;
[0107] Position feature extraction: the position of ω ki in D k is encoded into a feature vector using the sine position encoding, denoted as p ki , wherein the calculation method of p ki [j] is as shown in formula (1), t is the position of word ω ki in D k , d p is the dimension of the position feature vector, j represents the dimension of p ki , and l=0,1,2,…,d p / 2–1;
[0108]
[0109] wherein,
[0110] Dependency syntax feature extraction: D k is analyzed using the dependency syntax tree, and the one-hot encoding of the depth of ω ki in the corresponding dependency syntax tree of D k is used as the feature vector, denoted as r ki ;
[0111] Hot event type feature extraction: adopt D k The one-hot encoding corresponding to the hot event type is taken as a feature vector, denoted as t ki ;
[0112] Feature splicing: ω ki The final feature vector of V ki = [e ki , p ki , r ki , t ki ], and the total dimension is d.
[0113] BiLSTM+CRF is adopted as the hot word recognition model, the input of which is a feature matrix V k composed of feature vectors V ki of all words in D k , wherein V k [i] = V ki , and the output is the recognition result (whether it is a hot word) of each word.
[0114] All labeled hot word samples in the first training set WS are used to train the hot word recognition model, and an initial hot word recognition model WC is obtained, which is used to recognize all hot words in a sentence.
[0115] (2) Construct a first hot event classification model WEC and a second hot event classification model SEC, respectively train the first hot event classification model WEC using the first training set WS and train the second hot event classification model SEC using the second training set ES, to obtain the trained first hot event classification model WEC and the trained second hot event classification model SEC; wherein the second training set ES includes a large number of training samples only labeled with hot event types;
[0116] The first hot event classification model WEC and the second hot event classification model SEC in this embodiment are both TextCNN models.
[0117] The specific process of training the first hot event classification model WEC in this embodiment is as follows:
[0118] For each sample D k in the first training set WS, only all hot words (denoted as DW k ) in D k are retained, and DW k is input into the TextCNN model for training, to obtain the trained first hot event classification model WEC, which takes a hot word list as input and outputs the corresponding hot event type.
[0119] The specific process of training the second hot event classification model SEC in the embodiment is as follows:
[0120] For each sample O k in the second training set ES, the following operations are performed: k The sample O k is directly input into the TextCNN model for training, and the second hot event classification model SEC after training is obtained. The SEC is a model that takes a sentence as input and outputs the corresponding hot event type.
[0121] (3) The hot event classification results of the first hot event classification model WEC after training and the second hot event classification model SEC after training, and the hot word recognition confidence of the hot word recognition model WC after training are obtained. The hot word recognition model WC is constantly provided with a supervision signal to update the hot word recognition model;
[0122] As shown in FIG. 4, the specific process of providing a supervision signal for the hot word recognition model in the embodiment of the application is as follows: Figure 3b
[0123] (1) Sample labeling and screening: For each sample O k in the second training set ES, the following operations are performed:
[0124] (2) Hot word recognition: The hot word recognition model WC is used to perform hot word recognition on O k , and a recognized hot word list DWL k and the recognition confidence of each hot word in DWL k are obtained.
[0125] (3) Hot event classification: The first hot event classification model WEC after training is used to classify DWL k , and a classification result W rk and a classification confidence W pk are obtained. The second hot event classification model after training is used to classify O k , and a classification result S rk is obtained.
[0126] (4) Hot word labeling: If W rk =S rk , and W pk >0.8, all words in DWL k are labeled as hot words, and the labeled sample O k is put into the first training set WS. Otherwise, one hot word with the lowest recognition confidence in DWL k is deleted, and step (2) is returned until the sample O k is put into WS or DWL k is empty.
[0127] Wherein, the hot word recognition model WC and the first hot event classification model WEC are iteratively trained.
[0128] Specifically,
[0129] The hot word recognition model WC is updated: the hot word recognition model WC is retrained by using the updated training sample set WS;
[0130] The first hot event classification model WEC is updated: the hot event classification model WEC based on the hot word is retrained by using the updated training sample set WS;
[0131] The hot word recognition model WC and the first hot event classification model WEC are iteratively trained until all samples in ES are processed, or the number of iterations exceeds a specified threshold.
[0132] An event to be classified is obtained, the event to be classified is segmented, and all words segmented from the event to be classified are input into the updated hot word recognition model to obtain hot words in the event to be classified; and the hot words in the event to be classified are input into the trained first hot event classification model WEC to obtain a hot event type corresponding to the event to be classified.
[0133] It should be noted that, for the method embodiment, in order to simply describe, it is expressed as a series of action combinations, but those skilled in the art should know that the embodiment of the present application is not limited by the order of the described actions, because according to the embodiment of the present application, certain steps can be performed in other order or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification all belong to preferred embodiments, and the actions involved are not necessarily necessary for the embodiment of the present application.
[0134] Referring to Figure 4 , a structural schematic diagram of a hot word recognition device provided by an embodiment of the present application is shown, which can specifically include the following modules:
[0135] The training set acquisition module 401 is configured to acquire a first training set and a second training set, the first training set being a first training sample with hot word annotation and hot event type annotation, and the second training set being a second training sample with hot event type annotation;
[0136] The model training module 402 is configured to train a hot word recognition model and a first hot event classification model according to the first training set, and train a second hot event classification model according to the second training set;
[0137] The new hot word annotation module 403 is configured to determine new hot word annotation of the second training set based on the hot word recognition model, the first hot event classification model, and the second hot event classification model.
[0138] The model updating module 404 is configured to update the hot word recognition model and the first hot event classification model based on the newly-added hot word label.
[0139] The hot word extraction module 405 is configured to extract hot words from the acquired event to be classified by using the updated hot word recognition model and the first hot event classification model.
[0140] In an embodiment of the present application, the newly-added hot word labeling module 403 comprises:
[0141] The candidate hot word list submodule is configured to input the second training set into the hot word recognition model to obtain a candidate hot word list.
[0142] The first classification submodule is configured to input the candidate hot word list into the first hot event classification model to output a first classification result and a classification confidence.
[0143] The second classification submodule is configured to input the second training set into the second hot event classification model to output a second classification result.
[0144] The newly-added hot word labeling submodule is configured to determine the newly-added hot word label of the second training set from the candidate hot word list based on the first classification result, the second classification result and the classification confidence.
[0145] In an embodiment of the present application, the newly-added hot word labeling submodule comprises:
[0146] The judging unit is configured to judge whether the first classification result matches the second classification result and whether the classification confidence is greater than a preset confidence threshold.
[0147] The newly-added hot word labeling unit is configured to determine the newly-added hot word label of the second training set from the candidate hot word list according to the judgment result.
[0148] In an embodiment of the present application, the newly-added hot word labeling unit can comprise:
[0149] The newly-added hot word labeling subunit is configured to determine the candidate hot word in the candidate hot word list as the newly-added hot word label of the second training set when it is determined that the first classification result matches the second classification result and the classification confidence is greater than the preset confidence threshold.
[0150] In an embodiment of the present application, the newly-added hot word labeling unit can comprise:
[0151] The candidate hot word deletion unit is configured to delete a candidate hot word with the lowest confidence from the candidate hot word list when it is determined that the first classification result does not match the second classification result and / or the classification confidence is not greater than the preset confidence threshold.
[0152] The execution sub-unit is configured to input the updated candidate hot word list into the trained first hot event classification model to output a first classification result and a classification confidence.
[0153] In an embodiment of the present application, the model training module 402 can include:
[0154] The candidate feature data acquisition sub-module is configured to perform feature extraction on sample words of each first training sample in the first training set according to a plurality of preset feature types to obtain a plurality of candidate feature data.
[0155] The vector splicing sub-module is configured to splice the plurality of candidate feature data to obtain a target feature vector.
[0156] The hot word prediction sub-module is configured to input the target feature vector into the constructed initial hot word recognition model to obtain hot word prediction data of each sample word.
[0157] The model iterative training sub-module is configured to update model parameters of the initial hot word recognition model based on the hot word prediction data and actual hot word data corresponding to the first sample words.
[0158] In an embodiment of the present application, the model updating module 404 can include:
[0159] The first training sample updating sub-module is configured to update the first training sample based on the newly added hot word label.
[0160] The updated training sub-module is configured to train the hot word recognition model and the first hot event according to the updated first training sample.
[0161] In an embodiment of the present application, after the hot word recognition model, the first hot event classification model, and the second hot event classification model are trained based on the first training set and the second training set, the newly added hot word label of the second training set can be determined based on the hot word recognition model, the first hot event classification model, and the second hot event classification model, thereby updating the hot word recognition model and the first hot event classification model. The updated hot word recognition model and the first hot event classification model can be used to extract hot words from the obtained events to be classified, thereby implementing a training strategy of using the hot event classification model to provide mutual supervision signals for the hot word recognition model to help the hot word recognition model to iteratively optimize, and reducing the workload of hot word labeling and ensuring higher hot event mining accuracy.
[0162] An embodiment of the present application further provides an electronic device, which can comprise a processor, a memory, and a computer program stored in the memory and capable of running on the processor, and the computer program is executed by the processor to implement the hotword recognition method.
[0163] An embodiment of the present application further provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the hotword recognition method.
[0164] For the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts are described in the part of the method embodiment.
[0165] Each embodiment in the specification is described in a progressive manner, and each embodiment mainly describes the difference from other embodiments, and the same and similar parts between each embodiment can be referred to each other.
[0166] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, device, or computer program product. Therefore, the embodiments of the present application can adopt a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the embodiments of the present application can adopt the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program codes.
[0167] The embodiments of the present application are described with reference to the flowcharts and / or block diagrams according to the method, terminal device (system), and computer program product of the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, so that the instructions executed by the computer or other programmable data processing terminal device produce a device implemented in the flowcharts and / or block diagrams. Figure 1 The function of one flow or multiple flows and / or blocks Figure 1 The device that implements the function of one block or multiple blocks.
[0168] These computer program instructions can also be stored in a computer readable storage medium, which can guide the computer or other programmable data processing terminal device to work in a specific way, so that the instructions stored in the computer readable storage medium produce a product including instruction devices, which implement the flowcharts and / or block diagrams. Figure 1 The function of one flow or multiple flows and / or blocksFigure 1 the function specified in one or more blocks.
[0169] These computer program instructions can also be loaded into computer or other programmable data processing terminal devices, so that a series of operation steps are performed on the computer or other programmable terminal devices to generate a computer implemented process, so that the instructions executed on the computer or other programmable terminal devices provide a process for implementing the flow Figure 1 one or more flows and / or blocks Figure 1 the function specified in one or more blocks.
[0170] Although the preferred embodiments of the present application have been described, those skilled in the art who understand the basic inventive concept can make further changes and modifications to the embodiments. Therefore, the appended claims are intended to cover all changes and modifications falling within the scope of the embodiments of the present application.
[0171] Finally, it should be noted that the relational terms herein, such as first and second, and the like, are used solely to distinguish one from another entity or action, without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, so that a process, method, article, or terminal device including a series of elements includes not only those elements but also other elements not expressly listed or inherent to such process, method, article, or terminal device. Without more limitations, an element defined by the statement "comprising a" does not exclude the existence of additional identical elements in the process, method, article, or terminal device including the element.
[0172] The above provides a hot word recognition method and device, electronic equipment, and storage medium, and the principle and implementation manner of the present application are described by applying specific examples. The above embodiment description is only used to help understand the method and core idea of the present application; meanwhile, for those skilled in the art, according to the idea of the present application, the specific implementation manner and application range can be changed, and the above description of the present application should not be understood as a limitation.
Claims
1. A hot word identification method, characterized in that: The method comprises: Obtain a first training set and a second training set, wherein the first training set is a first training sample labeled with hot words and hot event types, and the second training set is a second training sample labeled with hot event types; Performing model training according to the first training set to obtain a hot word recognition model and a first hot event classification model, and performing model training according to the second training set to obtain a second hot event classification model; Determining new hot word annotations for the second training set based on the hot word recognition model, the first hot event classification model, and the second hot event classification model; Update the hot word recognition model and the first hot event classification model based on the newly added hot word annotation; The updated hot word recognition model and the first hot event classification model are used to extract hot words from the acquired events to be classified.
2. The method according to claim 1, characterized in that The step of determining the newly added hot word annotations of the second training set based on the hot word recognition model, the first hot event classification model, and the second hot event classification model includes: Inputting the second training set into the hot word recognition model to obtain a candidate hot word list; Inputting the candidate hot word list into the first hot event classification model, and outputting a first classification result and classification confidence; Inputting the second training set into the second hot event classification model, and outputting a second classification result; Determine a new hot word label for the second training set from the candidate hot word list based on the first classification result, the second classification result, and the classification confidence.
3. The method according to claim 2, characterized in that The determining, based on the first classification result, the second classification result, and the classification confidence, of the newly added hot word labels for the second training set from the candidate hot word list includes: Determining whether the first classification result matches the second classification result, and determining whether the classification confidence is greater than a preset confidence threshold; Determine the new hot word labels for the second training set from the candidate hot word list according to the judgment result.
4. The method according to claim 3, characterized in that The step of determining the new hot word labeling of the second training set from the candidate hot word list according to the judgment result includes: When it is determined that the first classification result matches the second classification result and whether the classification confidence is greater than a preset confidence threshold, the candidate hot words in the candidate hot word list are determined as new hot word labels for the second training set.
5. The method according to claim 3, characterized in that The step of determining the new hot word labeling of the second training set from the candidate hot word list according to the judgment result includes: When it is determined that the first classification result does not match the second classification result and / or the classification confidence is not greater than a preset confidence threshold, deleting the candidate hot word with the lowest confidence from the candidate hot word list; Based on the updated candidate hot word list, the step of inputting the candidate hot word list into the trained first hot event classification model is executed, and a first classification result and classification confidence are output.
6. The method according to claim 1, characterized in that The hot word recognition model is obtained by performing model training according to the first training set, including: Extracting features from the sample words of each first training sample in the first training set according to a plurality of preset feature types to obtain a plurality of candidate feature data; Perform vector concatenation on the plurality of candidate feature data to obtain a target feature vector; Input the target feature vector into the constructed initial hot word recognition model to obtain hot word prediction data for each sample word; The model parameters of the initial hot word recognition model are updated based on the hot word prediction data and the actual hot word data corresponding to the first sample words.
7. The method according to any one of claims 1 to 6, characterized in that The updating of the hot word recognition model and the first hot event classification model based on the newly added hot word annotation includes: Update the first training sample based on the newly added hot word annotation; The hot word recognition model and the first hot event are trained according to the updated first training sample.
8. A hot word recognition device, characterized in that: The device comprises: A training set acquisition module is used to acquire a first training set and a second training set, wherein the first training set is a first training sample labeled with hot words and hot event types, and the second training set is a second training sample labeled with hot event types; A model training module, configured to perform model training according to the first training set to obtain a hot word recognition model and a first hot event classification model, and to perform model training according to the second training set to obtain a second hot event classification model; A new hot word annotation module is added, which is used to determine the new hot word annotation of the second training set based on the hot word recognition model, the first hot event classification model and the second hot event classification model; A model updating module, configured to update the hot word recognition model and the first hot event classification model based on the newly added hot word annotation; The hot word extraction module is used to extract hot words from the acquired events to be classified using the updated hot word recognition model and the first hot event classification model.
9. An electronic device, characterized in that: The method comprises a processor, a memory, and a computer program stored in the memory and capable of running on the processor, wherein when the computer program is executed by the processor, the hot word recognition method according to any one of claims 1 to 7 is implemented.
10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the hot word recognition method according to any one of claims 1 to 7 is implemented.