Search word error correction method, system and equipment and readable storage medium
By using the combination method of error correction recall model and sorting model in search engines, the problem of low accuracy of search terms correction in the prior art is solved, and higher error correction accuracy and user experience are achieved.
Patent Information
- Application Number
- CN202411811411.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-10
- Publication Date
- 2025-03-28
AI Technical Summary
The existing search term correction methods have problems such as few recalls, overfitting and lack of semantics, resulting in low error correction accuracy.
By obtaining the search terms entered by the user, inputting the error correction recall model for correction, multiple error correction candidate words are obtained, and by determining the sorting characteristics, inputting the sorting model for sorting, obtaining the target error correction candidate words.
The accuracy of error correction of search terms is improved, and the accuracy and efficiency of error correction are enhanced by combining error correction recall model and sorting model.
Smart Images
Figure CN119719455A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of information retrieval technology, and more specifically, to a search term error correction method, system, device and readable storage medium. Background Art
[0002] With the rapid development of the Internet, users can easily obtain the information they want through search engines. Information retrieval is one of the most important ways to obtain and query information. However, while information retrieval brings convenience to users in obtaining information, it also has some problems. For example, due to negligence or other reasons, users may enter incorrect search terms (queries) when searching, causing the information returned by the search engine to deviate from the user's true intention, thus affecting the user experience. This requires the search engine to be able to correct the incorrect search terms entered by the user into the search terms that the user actually needs.
[0003] However, existing search term correction methods generally have problems such as low recall, overfitting and lack of semantics, that is, the accuracy of search term correction is not high. Summary of the invention
[0004] The embodiments of the present application provide a search term error correction method, system, device and readable storage medium, which can improve the error correction accuracy of the search term.
[0005] In view of this, an embodiment of the present application provides a search term error correction method, comprising:
[0006] Get the search term entered by the user;
[0007] Inputting the search term into the error correction recall model for error correction, and obtaining a plurality of error correction candidate terms for the search term;
[0008] Determining ranking features of the plurality of error correction candidate words;
[0009] The multiple error correction candidate words and the sorting features of the multiple error correction candidate words are input into a sorting model to sort the multiple error correction candidate words to obtain target error correction candidate words.
[0010] Optionally, the step of inputting the search term into an error correction recall model for prediction to obtain a plurality of error correction candidate terms for the search term includes:
[0011] Inputting the search word into an error detection module to obtain the correct probability of each character in the search word, wherein the error detection module includes a bidirectional gated recurrent unit layer and a text convolutional neural network layer;
[0012] Inputting the search word and the correct probability of each character into an error correction module to obtain an original error correction candidate word and the correct probability of the original error correction candidate word, wherein the error correction module includes an encoder layer and a linear layer;
[0013] The original error correction candidate words and the correct probabilities of the original error correction candidate words are input into a decoder to obtain a plurality of error correction candidate words of the search word.
[0014] Optionally, the step of inputting the original error correction candidate words and the correct probabilities of the original error correction candidate words into a decoder to obtain a plurality of error correction candidate words for the search word comprises:
[0015] Inputting the original error correction candidate word and the correct probability of the original error correction candidate word into a decoder for decoding;
[0016] Based on the decoder and the beam search algorithm, a plurality of error correction candidate words are determined, and the correct probabilities of the plurality of error correction candidate words are greater than a first threshold.
[0017] Optionally, the sorting features include low-order features and high-order features; the low-order features include at least one of the perplexity of the search term, the perplexity of the multiple error correction candidate words, pinyin differences and text similarities; the high-order features include at least one of the model score of the error correction recall model and the model score of the deep perplexity model; the deep perplexity model is trained based on a skip-word model and the user's interest point data.
[0018] Optionally, the method further comprises:
[0019] Based on user interest point data, we build character-level replacement prediction tasks, replacement word prediction tasks, error detection tasks, and error correction tasks;
[0020] Pre-train the error correction recall model to be trained based on the replacement prediction task and the replacement word prediction task to obtain an initial error correction recall model;
[0021] The initial error correction recall model is trained based on the error detection task and the error correction task to obtain the error correction recall model.
[0022] Optionally, the method further comprises:
[0023] Build a sorting model training dataset based on user interest point data;
[0024] The ranking model to be trained is trained using a gradient boosting decision tree paradigm based on the ranking model training data set to obtain a ranking model.
[0025] Optionally, the sorting model to be trained is a lightweight gradient boosting machine model.
[0026] The present application also provides a search term error correction system, including:
[0027] The acquisition module is used to obtain the search term input by the user;
[0028] An error correction recall module, used for inputting the search word into an error correction recall model to obtain multiple error correction candidate words of the search word, wherein the error correction recall model comprises an error detection module, an error correction module and a decoder, wherein the error detection module comprises a bidirectional gated recurrent unit layer and a text convolutional neural network layer;
[0029] A determination module, used to determine the ranking features of the plurality of error correction candidate words;
[0030] A sorting module is used to sort the multiple error correction candidate words based on the multiple error correction candidate words, the sorting features of the multiple error correction candidate words and the sorting model to obtain a target error correction candidate word.
[0031] The embodiment of the present application also provides a computer device, including: a memory, a processor, and a bus system;
[0032] Wherein, the memory is used to store programs;
[0033] The processor is used to execute the program in the memory to implement any one of the search word error correction methods described above;
[0034] The bus system is used to connect the memory and the processor so that the memory and the processor can communicate with each other.
[0035] An embodiment of the present application also provides a computer-readable storage medium storing instructions, which, when executed on a computer, enables the computer to execute any one of the search term error correction methods described above.
[0036] The embodiment of the present application provides a search term error correction method, including: obtaining a search term input by a user; inputting the search term into an error correction recall model for error correction, and obtaining multiple error correction candidate terms for the search term; determining the sorting features of the multiple error correction candidate terms; inputting the multiple error correction candidate terms and the sorting features of the multiple error correction candidate terms into a sorting model to sort the multiple error correction candidate terms, and obtaining target error correction candidate terms. It can be seen that the present application predicts multiple error correction candidate terms by inputting the search term input by the user into the error correction recall model, and by inputting the multiple error correction candidate terms and the sorting features of the multiple error correction candidate terms into the sorting model, a target error correction candidate term with higher accuracy can be obtained, thereby improving the error correction accuracy of the search term. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.
[0038] Figure 1 A flowchart of a search term error correction method provided in an embodiment of the present application;
[0039] Figure 2 A schematic diagram of a prediction process of an error correction recall model provided in an embodiment of the present application;
[0040] Figure 3 A structural diagram of a search term error correction system provided in an embodiment of the present application. DETAILED DESCRIPTION
[0041] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0042] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.
[0043] With the rapid development of the Internet, users can easily obtain the information they want through search engines. Information retrieval is one of the most important ways to obtain and query information. However, while information retrieval brings convenience to users in obtaining information, it also has some problems. For example, due to negligence or other reasons, users may enter incorrect search terms (queries) when searching, causing the information returned by the search engine to deviate from the user's true intention, thus affecting the user experience. This requires the search engine to be able to correct the incorrect search terms entered by the user into the search terms that the user actually needs.
[0044] However, existing search term correction methods generally have problems such as low recall, overfitting and lack of semantics, that is, the accuracy of search term correction is not high.
[0045] For example, the existing technology usually adopts an error detection and correction solution. In the error detection module, dictionary matching, language model scoring, machine learning classification and other methods are usually used to determine whether the search term needs to be corrected; in the error correction module, multiple strategies such as edit distance, perplexity set dictionary, pre-trained language model, etc. are usually used to recall candidates, and the perplexity is calculated through the N-Gram language model, and the machine learning model sorting and other strategies are used to sort them, and n error correction results are returned.
[0046] However, this solution has the following disadvantages:
[0047] In the freight address retrieval scenario, the names of user points of interest (POI) are usually short and concise, have a certain geographical background, and are different from the usual expression habits. Solutions based on traditional vocabulary or machine learning methods have weak generalization capabilities and cannot solve long-tail problems and semantic-related problems. They are prone to problems such as low recall, overfitting, and lack of semantics. The low recall problem means that the correct candidates are not recalled due to insufficient confusion sets / insufficient model capabilities. The overfitting problem means that the model learns a single mapping between historical erroneous characters and corresponding correct characters, which makes it difficult to select the correct candidates in the recall and sorting stages. The lack of semantics problem means that the semantic background, especially geographical knowledge, cannot be used, resulting in knowledge errors in the error correction results. In other words, the existing technology has a low error correction accuracy rate for search terms.
[0048] Therefore, in response to the above problems, the embodiments of the present application provide a search term error correction method, system, device and readable storage medium, which can improve the accuracy of search term error correction.
[0049] See also Figure 1 , a search word error correction method provided in an embodiment of the present application includes the following steps.
[0050] S101: Obtain a search term input by a user.
[0051] In this embodiment, the user can enter the search term or search sentence composed of the search term to be retrieved in the input box of the search engine, so as to obtain the required information through the search engine. After the user enters the search term or search sentence to be retrieved, the search term entered by the user can be obtained. Specifically, the search term obtained can be a search term that the user may enter incorrectly, that is, a search term that may contain errors.
[0052] S102: Input the search term into the error correction recall model for error correction, and obtain multiple error correction candidate terms for the search term.
[0053] Among them, the error correction recall model includes an error detection module, an error correction module and a decoder. The error detection module includes a bidirectional gated recurrent unit layer and a text convolutional neural network layer.
[0054] In the embodiment of the present application, an error correction recall model may be pre-trained. It is understandable that the error correction recall model may be trained based on a pre-trained language model, that is, the pre-trained language model may be used as the error correction recall model to be trained, such as a BERT (Bidirectional Encoder Representations from Transformers) model, wherein the BERT model is a pre-trained deep neural network model based on the Transformer architecture, which is intended to solve the language representation problem in natural language processing.
[0055] In a possible implementation of the embodiment of the present application, the method further includes: constructing a character-level replacement prediction task, a replacement word prediction task, an error detection task, and an error correction task based on user interest point data; pre-training the error correction recall model to be trained based on the replacement prediction task and the replacement word prediction task to obtain an initial error correction recall model; training the initial error correction recall model based on the error detection task and the error correction task to obtain an error correction recall model. In the embodiment of the present application, the error correction recall model can be obtained by training based on the pre-training and fine-tuning method. Specifically, the error correction recall model can be trained by the pre-training and fine-tuning method, wherein the pre-training strategy can be: pre-training based on the pre-stored POI data and the error correction recall model before pre-training, performing different types of replacements on random position characters in the correct data, for example, 60% are replaced with homophones, 20% are replaced with similar characters, 10% are replaced with similar characters, 5% are replaced with random characters, and 5% retain the original position characters to design character-level replacement prediction tasks and replacement word prediction tasks. Then, based on the replacement prediction task and the replacement word prediction task, the error correction recall model to be trained can be jointly trained using BCE loss and CE loss weighted to obtain the initial error correction recall model. Among them, BCE loss refers to binary cross-entropy loss (Binary Cross-Entropy Loss), also referred to as BCE Loss, which is a loss function commonly used in binary classification problems and is used to evaluate the difference between the probability distribution of the model output and the actual label. CE loss (Cross-Entropy Loss) refers to cross-entropy loss, which is a loss function commonly used in classification problems. Cross-entropy loss can measure the degree of difference between two different probability distributions in the same random variable. When the two probability distributions are closer, the smaller the cross-entropy loss, the more accurate the model prediction result. The fine-tuning strategy can be: construct error detection tasks and error correction tasks based on the user's search word-click data in the pre-stored user interest point data, and then use BCE loss and CE loss weighted to jointly train the initial error correction recall model to obtain the error correction recall model. In the training data of the error correction recall model, based on the user's historical search terms and corresponding click data, possible confusion sets such as similar sounds, homophones, and similar shapes can be mined through regular expressions, pinyin matching, etc. as part of the training data for the error correction recall model.
[0056] In an embodiment of the present application, the trained error correction recall model may include an error detection module, an error correction module and a decoder, wherein the error detection module may include a bidirectional gated recurrent unit layer (Bi-directional Gated Recurrent Unit, BiGRU) layer and a text convolutional neural network layer (Text Convolutional Neural Networks, TextCNN) layer. The BiGRU layer is a variant of the recurrent neural network (RNN), which is intended to solve the gradient vanishing or gradient exploding problems that are prone to occur in traditional RNNs when processing long sequences, and at the same time introduces the bidirectional processing capability of sequence data. The TextCNN layer refers to a specific layer that uses a convolutional neural network (CNN) to process text data, which is an application of the CNN model in the field of text classification. The advantages of the TextCNN layer are its simple structure, small amount of calculation, fast training speed, and the ability to achieve good results in text classification tasks by introducing trained word vectors. In addition, TextCNN also has a strong feature extraction capability, which can capture local features in the text and combine them into more advanced feature representations through convolution and pooling operations. Therefore, since the error correction recall model includes a bidirectional gated recurrent unit layer and a text convolutional neural network layer, the ability of the error correction recall model to utilize sequence information and local information is improved, that is, the prediction accuracy of the error correction recall model is improved. Therefore, by inputting the search term entered by the user into the error correction recall model to predict multiple error correction candidate words, and by inputting the multiple error correction candidate words and the sorting features of the multiple error correction candidate words into the sorting model, more accurate target error correction candidate words can be obtained, thereby improving the error correction accuracy of the search term.
[0057] In a possible implementation of the embodiment of the present application, the error detection module also includes a linear layer and an activation function, and the error correction module includes an encoder layer and a linear layer. In the embodiment of the present application, the error detection module can be specifically composed of two BiGRU layers, a TextCNN layer with a convolution kernel size of [3,5], a linear layer and a sigmoid activation function; the error correction module can be specifically composed of four encoder layers such as a BertEncoder layer and a linear layer, thereby improving the ability of the error correction recall model to utilize sequence information and local information, that is, improving the prediction accuracy of the error correction recall model.
[0058] In one possible implementation of the embodiment of the present application, a search word is input into an error correction recall model for prediction to obtain multiple error correction candidate words for the search word, including: inputting the search word into an error detection module to obtain the correct probability of each character in the search word; inputting the search word and the correct probability of each character into the error correction module to obtain the original error correction candidate word and the correct probability of the original error correction candidate word; inputting the original error correction candidate word and the correct probability of the original error correction candidate word into a decoder to obtain multiple error correction candidate words for the search word. In the embodiment of the present application, Figure 2 As shown, after obtaining the search term input by the user, the search term can be first input into the error detection module to obtain the correct probability of each character in the search term. Then the search term and the correct probability p of each character, as well as the placeholder [MASK] token and the corresponding probability weighted value 1-p can be input into the error correction module to obtain the original error correction candidate word and the correct probability of the original error correction candidate word at each character position of the search term. Finally, the original error correction candidate word and the correct probability of the original error correction candidate word can be input into the decoder, so that the decoder can determine multiple error correction candidate words of the search term from multiple original error correction candidate words, so that by inputting the sorting features of multiple error correction candidate words and multiple error correction candidate words into the sorting model, a target error correction candidate word with higher accuracy can be obtained, thereby improving the error correction accuracy of the search term. Among them, [MASK] token refers to a special placeholder or marker used to temporarily replace or hide certain parts of the input sequence during training or inference.
[0059] In a possible implementation method of the embodiment of the present application, the original error correction candidate words and the correct probabilities of the original error correction candidate words are input into the decoder to obtain multiple error correction candidate words of the search word, including: inputting the original error correction candidate words and the correct probabilities of the original error correction candidate words into the decoder for decoding; determining multiple error correction candidate words based on the decoder and the beam search algorithm, and the correct probabilities of the multiple error correction candidate words are greater than the first threshold. In the embodiment of the present application, the original error correction candidate words and the correct probabilities of the original error correction candidate words can be first input into the decoder, so that after the decoder is input with the original error correction candidate words and the correct probabilities of the original error correction candidate words, the multiple original error correction candidate words with the largest correct probabilities can be determined by the beam search algorithm as multiple error correction candidate words of the search word, so that multiple error correction candidate words can be predicted in the error correction recall model, so that the multiple error correction candidate words can be further obtained by inputting the sorting features of the multiple error correction candidate words and the multiple error correction candidate words into the sorting model, thereby improving the error correction accuracy of the search word. The first threshold can be set according to the actual situation, and the embodiment of the present application is not limited here.
[0060] The beam search algorithm is a heuristic graph search algorithm, which is usually used to solve problems with a large solution space to reduce the space and time occupied by the search. The beam search algorithm can specifically include the following steps:
[0061] Step 1: Insert the initial node into the list and set the cluster width.
[0062] Step 2: Take a node from the list and expand it to generate its child nodes.
[0063] Step 3: Sort the generated child nodes according to the heuristic function, and select the first m (m is the cluster width) best nodes as candidate nodes for the next layer of expansion.
[0064] Step 4: Repeat step 2 and step 3 until the target node is found or the list is empty.
[0065] Step 5: When the target node is found, the algorithm ends and outputs the optimal solution. If the list is empty and the target node is not found, the algorithm ends and outputs the information that no solution was found.
[0066] S103: Determine the ranking features of multiple error correction candidate words.
[0067] In an embodiment of the present application, after a search term is input into the error correction recall model and multiple error correction candidates of the search term are obtained, the ranking feature of each of the multiple error correction candidates can be further confirmed, so that the multiple error correction candidates can be ranked by the ranking model to obtain a target error correction candidate with higher accuracy, thereby improving the error correction accuracy of the search term. The ranking feature can be a feature required by the ranking model when ranking multiple error correction candidates.
[0068] In a possible implementation of an embodiment of the present application, the ranking features include low-order features and high-order features; the low-order features include at least one of the perplexity of the search term, the perplexity of the multiple error correction candidate words, the pinyin difference and the text similarity; the high-order features include at least one of the model score of the error correction recall model and the model score of the deep perplexity model; the deep perplexity model is trained based on the word skipping model and the user's interest point data. In an embodiment of the present application, the ranking features of the error correction candidate words can be composed of at least one of the perplexity of the search term, the perplexity of the multiple error correction candidate words, the pinyin difference and the text similarity, and at least one of the model score of the error correction recall model and the model score of the deep perplexity model. Since the ranking features of error correction candidate words include not only low-level features such as the perplexity of search terms, the perplexity of multiple error correction candidate words, pinyin differences and text similarity, but also high-level features such as the model score of the error correction recall model and the model score of the deep perplexity model, the ranking model can improve the effect of the ranking model under the premise of low latency when sorting multiple error correction candidate words based on the ranking features of multiple error correction candidate words, making the error correction results more accurate and helping to meet user retrieval needs. Among them, perplexity is an indicator to measure the performance of language models, especially widely used in the field of natural language processing. Its core idea is to evaluate the model's predictive ability for a set of sample data, that is, the fluency and rationality of the text generated by the model. Specifically, perplexity measures the uncertainty or confusion of the model for a set of data, reflecting the degree of confusion of the model when predicting the next word or character.
[0069] S104: Input multiple error correction candidate words and ranking features of the multiple error correction candidate words into a ranking model to sort the multiple error correction candidate words to obtain a target error correction candidate word.
[0070] In an embodiment of the present application, after determining multiple error correction candidate words and the ranking features of the multiple error correction candidate words, the multiple error correction candidate words and the ranking features of the multiple error correction candidate words can be input into a ranking model, so that the ranking model can sort the multiple error correction candidate words based on the ranking features of the multiple error correction candidate words, thereby being able to determine one or more target error correction candidate words with a higher probability of being correct from the multiple error correction candidate words, thereby improving the error correction accuracy of the search words.
[0071] In a possible implementation of an embodiment of the present application, the method further includes: constructing a sorting model training data set based on user interest point data; training the sorting model to be trained in a gradient boosting decision tree paradigm based on the sorting model training data set to obtain a sorting model. Among them, the sorting model can be obtained by training based on a lightweight gradient boosting machine model. In an embodiment of the present application, taking into account the limitation of online delay, in order to improve the error correction efficiency of search terms, a sorting model can be trained based on a lightweight gradient boosting machine model, that is, a lightGBM model. Specifically, the lightGBM model can be used as the sorting model to be trained, and then a sorting model training data set can be constructed based on user interest point data, and the sorting model to be trained is trained in a gradient boosting decision tree GBDT paradigm based on the sorting model training data set, and Bayesian optimization is used for hyper-parameter tuning based on 1000 trees. The sorting model training data set can be constructed based on the user's search term-click data in the user's interest point data. Specifically, an original data set with a ratio of correct search terms to incorrect search terms of 1:3 can be constructed, and the recall results of the error correction recall model can be negatively sampled to construct a sorting data set with a ratio of correct error correction candidate words to incorrect error correction candidate words of 1:4.
[0072] It can be seen from this that the embodiment of the present application provides a search term correction method, which predicts multiple error correction candidate words by inputting the search term input by the user into the error correction recall model, and inputs the multiple error correction candidate words and the sorting features of the multiple error correction candidate words into the sorting model. More accurate target error correction candidate words can be obtained, thereby improving the error correction accuracy of the search term.
[0073] See also Figure 3 , the embodiment of the present application also provides a search term error correction system, the system comprising:
[0074] The acquisition module 301 is used to acquire the search term input by the user;
[0075] The error correction recall module 302 is used to input the search word into the error correction recall model for error correction, and obtain multiple error correction candidate words of the search word;
[0076] A determination module 303 is used to determine the ranking features of the plurality of error correction candidate words;
[0077] The sorting module 304 is used to input the multiple error correction candidate words and the sorting features of the multiple error correction candidate words into the sorting model to sort the multiple error correction candidate words to obtain the target error correction candidate words.
[0078] Optionally, the error correction and recall module 302 is specifically configured to:
[0079] Inputting the search word into an error detection module to obtain the correct probability of each character in the search word, wherein the error detection module includes a bidirectional gated recurrent unit layer and a text convolutional neural network layer;
[0080] Inputting the search word and the correct probability of each character into an error correction module to obtain an original error correction candidate word and the correct probability of the original error correction candidate word, wherein the error correction module includes an encoder layer and a linear layer;
[0081] The original error correction candidate words and the correct probabilities of the original error correction candidate words are input into a decoder to obtain a plurality of error correction candidate words of the search word.
[0082] Optionally, the error correction and recall module 302 is specifically configured to:
[0083] Inputting the original error correction candidate word and the correct probability of the original error correction candidate word into a decoder for decoding;
[0084] Based on the decoder and the beam search algorithm, a plurality of error correction candidate words are determined, and the correct probabilities of the plurality of error correction candidate words are greater than a first threshold.
[0085] Optionally, the sorting features include low-order features and high-order features; the low-order features include at least one of the perplexity of the search term, the perplexity of the multiple error correction candidate words, pinyin differences and text similarities; the high-order features include at least one of the model score of the error correction recall model and the model score of the deep perplexity model; the deep perplexity model is trained based on a skip-word model and the user's interest point data.
[0086] Optionally, the system further comprises:
[0087] A building module is used to build character-level replacement prediction tasks, replacement word prediction tasks, error detection tasks, and error correction tasks based on user interest point data;
[0088] A pre-training module, used to pre-train the error correction recall model to be trained based on the replacement prediction task and the replacement word prediction task to obtain an initial error correction recall model;
[0089] The training module is used to train the initial error correction recall model based on the error detection task and the error correction task to obtain the error correction recall model.
[0090] Optionally, the system further comprises:
[0091] A construction module is used to construct a sorting model training data set based on user interest point data;
[0092] The training module is used to train the sorting model to be trained in a gradient boosting decision tree paradigm based on the sorting model training data set to obtain the sorting model.
[0093] Optionally, the sorting model to be trained is a lightweight gradient boosting machine model.
[0094] It can be seen from this that the embodiment of the present application provides a search term correction system, which predicts multiple error correction candidate words by inputting the search term input by the user into the error correction recall model, and inputs the multiple error correction candidate words and the sorting features of the multiple error correction candidate words into the sorting model. More accurate target error correction candidate words can be obtained, thereby improving the error correction accuracy of the search term.
[0095] The embodiment of the present application also provides a computer device, including: a memory, a processor, and a bus system;
[0096] Wherein, the memory is used to store programs;
[0097] The processor is used to execute the program in the memory to implement any one of the search word error correction methods described above;
[0098] The bus system is used to connect the memory and the processor so that the memory and the processor can communicate with each other.
[0099] An embodiment of the present application also provides a computer-readable storage medium storing instructions, which, when executed on a computer, enables the computer to execute any one of the search term error correction methods described above.
[0100] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.
[0101] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A search word error correction method, characterized in that: The method comprises: Get the search term entered by the user; Inputting the search term into the error correction recall model for prediction to obtain multiple error correction candidate terms for the search term; Determining ranking features of the plurality of error correction candidate words; The multiple error correction candidate words and the sorting features of the multiple error correction candidate words are input into a sorting model to sort the multiple error correction candidate words to obtain target error correction candidate words.
2. The method according to claim 1, characterized in that The inputting the search word into the error correction recall model for prediction to obtain multiple error correction candidate words of the search word includes: Inputting the search word into an error detection module to obtain the correct probability of each character in the search word, wherein the error detection module includes a bidirectional gated recurrent unit layer and a text convolutional neural network layer; Inputting the search word and the correct probability of each character into an error correction module to obtain an original error correction candidate word and the correct probability of the original error correction candidate word, wherein the error correction module includes an encoder layer and a linear layer; The original error correction candidate words and the correct probabilities of the original error correction candidate words are input into a decoder to obtain a plurality of error correction candidate words of the search word.
3. The method according to claim 2, characterized in that The step of inputting the original error correction candidate words and the correct probabilities of the original error correction candidate words into a decoder to obtain a plurality of error correction candidate words of the search word comprises: Inputting the original error correction candidate word and the correct probability of the original error correction candidate word into a decoder for decoding; Based on the decoder and the beam search algorithm, a plurality of error correction candidate words are determined, and the correct probabilities of the plurality of error correction candidate words are greater than a first threshold.
4. The method according to claim 1, characterized in that The sorting features include low-order features and high-order features; the low-order features include at least one of the perplexity of the search term, the perplexity of the multiple error correction candidate words, pinyin differences and text similarities; the high-order features include at least one of the model score of the error correction recall model and the model score of the deep perplexity model; the deep perplexity model is trained based on a skip-word model and the user's interest point data.
5. The method according to claim 1, characterized in that The method further comprises: Based on user interest point data, we build character-level replacement prediction tasks, replacement word prediction tasks, error detection tasks, and error correction tasks; Pre-train the error correction recall model to be trained based on the replacement prediction task and the replacement word prediction task to obtain an initial error correction recall model; The initial error correction recall model is trained based on the error detection task and the error correction task to obtain the error correction recall model.
6. The method according to claim 1, characterized in that The method further comprises: Build a sorting model training dataset based on user interest point data; The ranking model to be trained is trained using a gradient boosting decision tree paradigm based on the ranking model training data set to obtain a ranking model.
7. The method according to claim 6, characterized in that The sorting model to be trained is a lightweight gradient boosting machine model.
8. A search word error correction system, characterized in that: The system comprises: The acquisition module is used to obtain the search term input by the user; An error correction recall module, used for inputting the search word into an error correction recall model to obtain multiple error correction candidate words of the search word, wherein the error correction recall model comprises an error detection module, an error correction module and a decoder, wherein the error detection module comprises a bidirectional gated recurrent unit layer and a text convolutional neural network layer; A determination module, used to determine the ranking features of the plurality of error correction candidate words; A sorting module is used to sort the multiple error correction candidate words based on the multiple error correction candidate words, the sorting features of the multiple error correction candidate words and the sorting model to obtain a target error correction candidate word.
9. A computer device, characterized in that: include: Memory, processor, and bus system; Wherein, the memory is used to store programs; The processor is used to execute the program in the memory to implement the method according to any one of claims 1 to 7; The bus system is used to connect the memory and the processor so that the memory and the processor can communicate with each other.
10. A computer-readable storage medium, characterized in that: The device stores instructions which, when executed on a computer, cause the computer to execute the method according to any one of claims 1 to 7.