Information query method and semiconductor process equipment

Abstract

The invention provides an information query method and semiconductor process equipment, and the information query method comprises the steps: determining a first context vector of a target query word and second context vectors of a plurality of candidate words based on a target context vector matrix, calculating a first similarity between the first context vector and each second context vector, and meanwhile, calculating a second similarity between the first context vector and each second context vector; obtaining a first semantic vector of the target query word and second semantic vectors of the plurality of candidate words, and calculating a second similarity between the first semantic vector and each second semantic vector, so as to determine a target similarity between the target query word and the plurality of candidate words based on the plurality of first similarities and the plurality of second similarities, and determining the target word from the plurality of candidate words based on the target similarity between the target query word and the plurality of candidate words to determine the target information in the target information base based on the target word, so that the accuracy of the information query result and the information query efficiency can be effectively improved.

Description

Technical Field

[0001] This application relates to the field of semiconductor process equipment, specifically to data processing technology in the field of semiconductor process equipment, and more specifically to an information query method and semiconductor process equipment. Background Technology

[0002] With the continuous advancement of semiconductor technology, the complexity of semiconductor process equipment is increasing, generating a greater amount of information, such as operational and alarm data. Querying this information allows users to analyze and statistically assess the operational status of semiconductor process equipment and identify faults. This enables timely and effective repair or adjustment of the equipment, reducing downtime in the semiconductor production process and maintaining continuous production line operation.

[0003] Currently, the method of semantic matching based on the user's input query terms is usually used to retrieve relevant information, which cannot guarantee query efficiency and accuracy of query results. Summary of the Invention

[0004] To address the aforementioned technical problems, this application provides an information retrieval method and semiconductor process equipment, thereby improving both the accuracy and efficiency of information retrieval results.

[0005] To achieve the above technical objectives, the embodiments of this application provide the following technical solutions:

[0006] Firstly, the embodiments of this specification provide an information query method, including:

[0007] Based on the target context vector matrix, a first context vector of the target query term and second context vectors of multiple candidate terms are determined. Based on the first context vector and the multiple second context vectors, multiple first similarities are calculated. The first similarity includes the similarity between the first context vector and the second context vector. The target context vector matrix includes the context vectors of multiple text words, which are obtained by segmenting each piece of information in the target information database of semiconductor process equipment.

[0008] Obtain the first semantic vector of the target query term and the second semantic vectors of multiple candidate terms, and calculate multiple second similarities based on the first semantic vector and the multiple second semantic vectors; the second similarity includes the similarity between the first semantic vector and the second semantic vector.

[0009] Based on multiple first similarities and multiple second similarities, the target similarity between the target query term and multiple candidate terms is determined, and based on the target similarity between the target query term and multiple candidate terms, the target term is determined from the multiple candidate terms, and the target information in the target information database is determined based on the target term.

[0010] Optionally, based on the target context vector matrix, the first context vector of the target query term is determined, including:

[0011] Based on the target query term, determine the words to be matched in each of the text terms;

[0012] The first context vector of the target query term is determined based on the mean of the context vectors of each of the terms to be matched.

[0013] Optionally, determining the target similarity between the target query term and the multiple candidate terms based on multiple first similarities and multiple second similarities includes:

[0014] Based on the word frequency of the target query term, a first weight coefficient of the first similarity and a second weight coefficient of the second similarity are determined. The first weight coefficient is negatively correlated with the word frequency of the target query term, and the second weight coefficient is negatively correlated with the first weight coefficient.

[0015] The first similarity is weighted based on the first weight coefficient, and the second similarity is weighted based on the second weight coefficient;

[0016] The target similarity between the target query term and the candidate term is determined based on the weighted first similarity and the weighted second similarity.

[0017] Optionally, determining the target similarity between the target query term and the candidate term based on the weighted first similarity and the weighted second similarity includes:

[0018] The weighted first similarity and the weighted second similarity are input into the similarity recognition model to obtain the target similarity between the target query term and the candidate term; wherein, the similarity recognition model is constructed by the following method:

[0019] The information of each sample in the sample dataset is segmented into words to obtain multiple sample words;

[0020] The model set is iteratively updated based on each of the sample words, and the model set includes multiple candidate similarity recognition models;

[0021] When the iterative update termination condition is met, the similarity recognition model is determined based on the loss value of each candidate similarity recognition model in the model set.

[0022] Optionally, the iterative update of the model set based on each of the sample words includes:

[0023] In the current iteration update, the loss value of each candidate similarity recognition model is determined based on each of the sample words;

[0024] Based on the loss values ​​of each candidate similarity recognition model, the target similarity recognition model among the candidate similarity recognition models is determined, and each candidate similarity recognition model other than the target similarity recognition model in the model set is updated to obtain the update result of the model set in the current iteration update.

[0025] Optionally, updating each candidate similarity recognition model other than the target similarity recognition model in the model set includes:

[0026] Delete all candidate similarity recognition models in the model set other than the target similarity recognition model;

[0027] The model update is repeated a specified number of times, and the result of each model update is added to the model set as a candidate similarity recognition model; the model update includes updating the model parameters of the target similarity model.

[0028] Optionally, determining the loss value of each candidate similarity recognition model based on each of the sample words includes:

[0029] Based on the candidate similarity recognition model, a first sample similarity and a second sample similarity are determined for each word set corresponding to multiple sample words; the word set corresponding to the sample word includes the sample word, the synonyms of the sample word, and the non-synonyms of the sample word, the first sample similarity includes the similarity between the sample word and the synonyms of the sample word, and the second sample similarity includes the similarity between the sample word and the non-synonyms of the sample word;

[0030] The loss value of the candidate similarity recognition model is determined based on the first sample similarity and the second sample similarity of each of the aforementioned word sets.

[0031] Optionally, determining the target information in the target information database based on the target word includes:

[0032] Based on the correspondence between the text words and the documents in the target information database, and the target words, the documents containing the target words are determined, and the documents containing the target words are taken as the target documents;

[0033] Based on the location information of the target word in the target document, the target information in the target document is determined.

[0034] Secondly, embodiments of this specification provide a semiconductor process apparatus, including a process chamber and a controller. The controller includes at least one processor and at least one memory, in which a computer program is stored. When the computer program is executed by the processor, it implements the information query method as described in any of the preceding claims.

[0035] Thirdly, embodiments of this specification provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the information query method as described in any of the preceding claims.

[0036] Fourthly, embodiments of this specification provide a computer program product or a computer program, the computer program product including a computer program stored in a computer-readable storage medium; the processor of the computer device reads the computer program from the computer-readable storage medium, and when the processor executes the computer program, it implements the information query method as described in any of the preceding claims.

[0037] As can be seen from the above technical solutions, the embodiments of this application provide an information query method and a semiconductor process equipment. The information query method, based on a target context vector matrix, determines a first context vector of the target query term and second context vectors of multiple candidate words. It then calculates multiple first similarities based on the first context vector and the multiple second context vectors. The first similarity includes the similarity between the first context vector and the second context vector. The target context vector matrix includes context vectors of multiple text words, which are obtained by segmenting information in the target information database of the semiconductor process equipment. Simultaneously, it obtains a first semantic vector of the target query term and second semantic vectors of multiple candidate words, and calculates multiple second similarities based on the first semantic vector and the multiple second semantic vectors. The second similarity includes the first semantic vector... The similarity with the second semantic vector is used to determine the target similarity between the target query word and multiple candidate words based on multiple first similarities and multiple second similarities. The target word is then determined from the multiple candidate words based on the target similarity between the target query word and multiple candidate words. The target information in the target information database is then determined based on the target word. Thus, by combining the context vector similarity between the target query word and each candidate word as well as the semantic vector similarity between the target query word and each candidate word, the accuracy of the target word determination result can be effectively improved when the target query word contains rare words and / or multiple semantic words. This improves the accuracy of the information query result. Moreover, the context vector does not participate in the training of the semantic model used to obtain the first semantic vector and the second semantic vector. Therefore, when the target query word contains rare words, it is not necessary to retrain the semantic model, thus improving the efficiency of information query. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0039] Figure 1 A flowchart illustrating an information query method provided for one embodiment of this specification;

[0040] Figure 2 A flowchart illustrating another information query method provided as one embodiment of this specification;

[0041] Figure 3 This is a schematic diagram of a semiconductor process apparatus provided for one embodiment of this specification. Detailed Implementation

[0042] Unless otherwise defined, the technical or scientific terms used in the embodiments of this specification shall have the ordinary meaning understood by one of ordinary skill in the art to which this specification pertains. The terms "first," "second," and similar terms used in the embodiments of this specification do not indicate any order, quantity, or importance, but are merely used to avoid confusion of constituent elements.

[0043] Unless the context otherwise requires, throughout this specification, "a plurality of" means "at least two," and "including" is interpreted as open-ended or encompassing, that is, "including, but not limited to." In the description of this specification, terms such as "one embodiment," "some embodiments," "exemplary embodiment," "example," "specific example," or "some examples" are intended to indicate that a particular feature, structure, material, or characteristic associated with that embodiment or example is included in at least one embodiment or example of this specification. The illustrative representations of the above terms do not necessarily refer to the same embodiment or example.

[0044] The technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this specification.

[0045] Overview

[0046] As described in the background section, with the continuous advancement of semiconductor technology, the complexity of semiconductor process equipment is increasing, and the amount of information generated is also growing, such as operating information and alarm information. By querying the information generated by semiconductor process equipment, users can easily analyze and statistically analyze the operating status of the equipment and identify faults. This allows for timely and effective repair or adjustment of the equipment, reducing downtime in the semiconductor production process and maintaining continuous production line operation.

[0047] Currently, semantic matching is typically used to retrieve relevant information based on the query terms entered by the user. However, when the query terms contain uncommon words (i.e., words that appear less frequently or not at all in the training samples during the training of the semantic model used for semantic matching) or polysemous words, the accuracy of semantic matching is low, thus failing to guarantee the accuracy of the query results.

[0048] In addition, some existing solutions consider contextual information during semantic matching. However, these solutions train the semantic model by concatenating the context into the training samples. This allows the trained semantic model to consider contextual information during semantic recognition, thereby improving the accuracy of semantic matching. However, when rare characters or polysemous words appear, the semantic model needs to be retrained, which not only incurs high training costs but also has poor scalability and cannot guarantee query efficiency.

[0049] Therefore, it is necessary to provide an information retrieval method that can simultaneously improve the efficiency and accuracy of information retrieval results.

[0050] To address the shortcomings of traditional methods in ensuring efficient and accurate information retrieval, this application's technical solution determines a first context vector for the target query term and second context vectors for multiple candidate words based on a target context vector matrix. Multiple first similarities are then calculated based on these first and second context vectors. The first similarity includes the similarity between the first and second context vectors. The target context vector matrix includes context vectors for multiple text words, obtained by segmenting information from a target information database of semiconductor process equipment. Simultaneously, a first semantic vector for the target query term and second semantic vectors for multiple candidate words are obtained. Multiple second similarities are then calculated based on these first and second semantic vectors. The second similarity includes the similarity between the first and second context vectors. The method uses semantic vector similarity to determine the target similarity between the target query term and multiple candidate terms based on multiple first similarity and multiple second similarity. Then, it identifies the target term from the candidate terms based on the target similarity between the target query term and the multiple candidate terms. Finally, it identifies the target information in the target information database based on the target term. By combining the context vector similarity between the target query term and each candidate term, as well as the semantic vector similarity between the target query term and each candidate term, the accuracy of target term identification can be effectively improved when the target query term contains rare words and / or multiple semantic words. This improves the accuracy of information retrieval results. Furthermore, the context vector is not involved in the training of the semantic model used to obtain the first and second semantic vectors, so there is no need to retrain the semantic model when the target query term contains rare words, thus improving information retrieval efficiency.

[0051] Based on the above inventive concept, the information query method provided in the embodiments of this specification will be described exemplarily below.

[0052] Exemplary methods

[0053] refer to Figure 1 This specification provides an information query method, which includes:

[0054] S101. Based on the target context vector matrix, determine the first context vector of the target query term and the second context vectors of multiple candidate terms, and calculate multiple first similarities based on the first context vector and the multiple second context vectors; the first similarity includes the similarity between the first context vector and the second context vector; the target context vector matrix includes the context vectors of multiple text words, and the text words are obtained by segmenting each piece of information in the target information database of semiconductor process equipment.

[0055] Specifically, semiconductor process equipment includes etching equipment, deposition equipment, and resist stripping equipment. The target information database for semiconductor process equipment can include databases for storing information related to target query terms, such as databases for storing alarm information or databases for storing equipment operation information. Target query terms can include terms entered by the user for querying.

[0056] During implementation, new information will be continuously generated during the operation of semiconductor process equipment; that is, the information in the target information database is dynamically added. A single piece of information can be a sentence. When new information is generated, it can be stored in the corresponding information database. Simultaneously, the information can be segmented into multiple text words, and the target context vector matrix can be updated based on the segmentation results. The target context vector matrix can include the context vectors corresponding to each text word in the target information database.

[0057] For any given text word, its context vector includes the frequencies of all other text words in the target information database that appear in the context of that text word. The frequency of text word j in the context of text word i can be the ratio of the total number of times text word j appears in the context of text word i to the total number of text words in the context of text word i in the target information database. For example, the value M[id(i), id(j)] in the id(i)-th row and id(j)-th column of the target context vector matrix M can represent the frequency of text word j in the context of text word i. The calculation method for M[id(i), id(j)] is shown in equation (1):

[0058]

[0059] In the formula, D represents all information in the target information database; Context(s,i,k) represents all text words in information s that appear within the range of the k words before and the k words after text word i. These text words can be represented as the text words in the context of text word i. The value of k is a preset value, which can be determined based on the number of segments or the number of information entries, or it can be an empirical value. The setting of k ensures that when determining the context of any text word in the information, it covers all text words in the entire information. For example, k can be set to 5; 1 j＝＝c The value is 1 when text word c is the same as text word j, and 0 when text word c is different from text word j. That is, when text word c is the same as text word j, it means that text word j appears once in the context of text word i, and the value is incremented by 1; otherwise, it is not incremented. |Context(s,i,k)| represents the number of text words in information s that appear in the range of the k words before and the k words after text word i.

[0060] In the process of determining the first context vector of the target query term based on the target context vector matrix, the words to be matched in each text word of the target information database can be obtained based on the target query term, and the first context vector of the target query term can be determined based on the mean of the context vectors of each word to be matched. The words to be matched may include text words that contain some or all of the words in the target query term, and may also include synonyms of some or all of the words in the target query term, thereby effectively ensuring the validity of the determination result of the first context vector.

[0061] Candidate words can include text words in the target information database that need to be matched with the target query word in terms of similarity. They can also include all text words in the target information database, or can be filtered from all text words in the target information database based on preset conditions. For example, the semantic similarity between the target query word and each text word in the target information database can be obtained, and text words with a semantic similarity greater than a preset similarity threshold can be used as candidate words. Alternatively, candidate words can be filtered from all text words in the target information database based on the information category corresponding to the target query word; for example, text words in documents with the same information category as the target query word can be used as candidate words.

[0062] For any candidate word, its context vector in the target context vector matrix can be used as its second context vector. It is understood that candidate words can also include combinations of multiple text words; when a candidate word is a combination of multiple text words, its second context vector can be the average of the context vectors of the multiple text words corresponding to that candidate word.

[0063] After determining the first context vector of the target query term and the second context vectors of each candidate term, for any candidate term, the similarity between the first context vector of the target query term and the second context vector of the candidate term can be calculated to obtain the first similarity. Thus, the first similarity between the target query term and each candidate term can be obtained. The first similarity can be cosine similarity, representing the contextual similarity between the target query term and the candidate term. It is understandable that it is also possible to obtain only the first similarity between the target query term and a subset of candidate terms; the specific settings can be configured according to actual needs.

[0064] S102. Obtain the first semantic vector of the target query term and the second semantic vectors of multiple candidate terms, and calculate multiple second similarities based on the first semantic vector and the multiple second semantic vectors; the second similarity includes the similarity between the first semantic vector and the second semantic vector.

[0065] Specifically, the target query term can be input into a pre-trained semantic model to obtain the first semantic vector of the target query term. At the same time, for any candidate term, the candidate term can be input into the pre-trained semantic model to obtain the second semantic vector of the candidate term. The semantic model can be a deep learning semantic model.

[0066] After determining the first semantic vector of the target query term and the second semantic vectors of each candidate term, for any candidate term, the similarity between the first semantic vector of the target query term and the second semantic vector of the candidate term can be calculated to obtain the second similarity. Thus, the second similarity between the target query term and each candidate term can be obtained. The second similarity can be cosine similarity, and it represents the semantic similarity between the target query term and the candidate term. It is understandable that it is also possible to obtain only the second similarity between the target query term and a subset of candidate terms; the specific settings can be configured according to actual needs.

[0067] Furthermore, there is no restriction on the execution order of steps S101 and S102. Step S101 can be executed first and then step S102, or step S102 can be executed first and then step S101, or steps S101 and S102 can be executed simultaneously.

[0068] S103. Based on multiple first similarities and multiple second similarities, determine the target similarity between the target query term and multiple candidate terms, and based on the target similarity between the target query term and multiple candidate terms, determine the target term from the multiple candidate terms, and determine the target information in the target information database based on the target term.

[0069] Specifically, for any candidate word, the target similarity between the target query word and the candidate word can be determined based on the first similarity and the second similarity. For example, the first and second similarities can be weighted and summed to obtain the target similarity between the target query word and the candidate word. Alternatively, the first and second similarities can be input into a pre-built similarity recognition model to obtain the target similarity between the target query word and the candidate word. The similarity recognition model can be a function model or a machine learning model. Alternatively, the target similarity between the target query word and the candidate word can be determined based on the first similarity, the second similarity, and a preset correspondence. The preset correspondence can include the correspondence between the first similarity, the second similarity, and the target similarity.

[0070] Therefore, the target similarity between the target query term and each candidate term can be obtained. In implementation, the target term can be determined from multiple candidate terms based on the target similarity between the target query term and each candidate term. The target term can include the candidate term with the highest target similarity to the target query term. It can be understood that the target term can also include candidate terms with the second highest or second-highest target similarity to the target query term. The specific settings can be configured according to actual needs.

[0071] After identifying the target words, target information in the target information database can be determined based on the target words. The target information can be information in the target information database that includes the target words.

[0072] Therefore, by combining the context vector similarity between the target query term and each candidate term, as well as the semantic vector similarity between the target query term and each candidate term, the accuracy of the target term determination can be effectively improved when the target query term contains uncommon words and / or multiple semantic words, thereby improving the accuracy of the information query results. Furthermore, since the context vector does not participate in the training of the semantic model used to obtain the first and second semantic vectors, there is no need to retrain the semantic model when the target query term contains uncommon words, thus improving the efficiency of information query.

[0073] To quickly and accurately determine the first context vector of the target query term, in one embodiment of this specification, the first context vector of the target query term is determined based on the target context vector matrix, including:

[0074] Based on the target query term, determine the words to be matched in each of the text terms;

[0075] The first context vector of the target query term is determined based on the mean of the context vectors of each of the terms to be matched.

[0076] Specifically, in the process of determining the first context vector of the target query term, it can be determined whether the target query term meets the segmentation conditions. For example, if the target query term is not the smallest segmentation unit, then the target query term meets the segmentation conditions; if the target query term is the smallest segmentation unit, then the target query term does not meet the segmentation conditions.

[0077] If the target query term meets the word segmentation criteria, it can be segmented to obtain multiple sub-query terms. The matching terms for each sub-query term are then determined from the text words in the target information database. For any sub-query term, the matching terms can include text words containing that sub-query term, as well as synonyms of that sub-query term; the specific settings can be configured according to actual needs.

[0078] If the target query term does not meet the word segmentation criteria, the corresponding matchable terms can be determined from the text terms in the target information database. The matchable terms can include text terms containing the target query term, as well as synonyms of the target query term, which can be set according to actual needs.

[0079] In practice, the context vectors of each word to be matched can be obtained from the target context vector matrix, and the average of the context vectors of each word to be matched can be used as the first context vector of the target query word, thereby enabling the first context vector of the target query word to be determined quickly and accurately.

[0080] To further improve the accuracy of information query results, in one embodiment of this specification, determining the target similarity between the target query term and the multiple candidate terms based on multiple first similarities and multiple second similarities includes:

[0081] Based on the word frequency of the target query term, a first weight coefficient of the first similarity and a second weight coefficient of the second similarity are determined. The first weight coefficient is negatively correlated with the word frequency of the target query term, and the second weight coefficient is negatively correlated with the first weight coefficient.

[0082] The first similarity is weighted based on the first weight coefficient, and the second similarity is weighted based on the second weight coefficient;

[0083] The target similarity between the target query term and the candidate term is determined based on the weighted first similarity and the weighted second similarity.

[0084] Specifically, the term frequency of the target query term can include the average term frequency of each word to be matched. For any word to be matched, the term frequency of that word can include the ratio of the number of times that word appears in each piece of information in the target information database to the total number of text words in each piece of information in the target information database. For example, the term frequency f of the target query term m It can be shown in equation (2):

[0085] f m =∑ g∈m f g / |m| (2)

[0086] In the formula, f g Let g be the frequency of the word to be matched, m be the set of all words to be matched, including the target query word, and |m| be the number of words to be matched corresponding to the target query word.

[0087] The first weight coefficient of the first similarity is used to characterize the magnitude of the weight of the first similarity, and the second weight coefficient is used to characterize the magnitude of the weight of the second similarity. The first weight coefficient can be negatively correlated with the word frequency of the target query term; that is, the lower the word frequency of the target query term, the larger the first weight coefficient, and vice versa. Simultaneously, the second weight coefficient can be negatively correlated with the first weight coefficient, and the sum of the first and second weight coefficients can be 1. The values ​​of both the first and second weight coefficients are in the range (0,1). For example, the first weight coefficient α... m It can be shown in equation (3):

[0088]

[0089] Second weighting coefficient β m It can be shown in equation (4):

[0090] β m =1-α m (4).

[0091] The lower the frequency of the target query term, the lower the frequency of each matching word corresponding to the target query term in the target information database, and the higher the probability that the target query term contains rare words. In this case, by increasing the first weight coefficient and decreasing the second weight coefficient, the matching degree of contextual information can be effectively improved and the matching degree of semantic information can be weakened during the information query process. Thus, when the target query term contains rare words, the impact of inaccurate semantic vector recognition of the target query term on the information query results can be effectively reduced. At the same time, by shifting the focus of target word matching to context matching, the probability of the target word appearing in the context of the target query term can be increased, thereby effectively locking the range of information query results and improving the accuracy of information query results. Moreover, there is no need to retrain the semantic model, ensuring the efficiency of information query.

[0092] After determining the first and second weight coefficients, the first weight coefficient can be multiplied by the first similarity to obtain a weighted first similarity, and the second weight coefficient can be multiplied by the second similarity to obtain a weighted second similarity. The target similarity between the query term and candidate terms is then determined based on these weighted first and second similarities. For example, the sum of the weighted first and second similarities can be used as the target similarity. Alternatively, the weighted first and second similarities can be input into a pre-built similarity recognition model to obtain the target similarity. Furthermore, the target similarity can be determined based on the weighted first and second similarities and a preset correspondence, which can include the correspondence between the first, second, and target similarities. This allows for the rapid and accurate determination of the target similarity between the query term and candidate terms.

[0093] In one feasible implementation, determining the target similarity between the target query term and the candidate terms based on the weighted first similarity and the weighted second similarity includes:

[0094] The weighted first similarity and the weighted second similarity are input into the similarity recognition model to obtain the target similarity between the target query term and the candidate term; wherein, the similarity recognition model is constructed by the following method:

[0095] The information of each sample in the sample dataset is segmented into words to obtain multiple sample words;

[0096] The model set is iteratively updated based on each of the sample words, and the model set includes multiple candidate similarity recognition models;

[0097] When the iterative update termination condition is met, the similarity recognition model is determined based on the loss value of each candidate similarity recognition model in the model set.

[0098] Specifically, the similarity recognition model can be a function model or a machine learning model, depending on the actual needs. The weighted first similarity and weighted second similarity can be input into the similarity recognition model to obtain the target similarity between the target query term and the candidate terms, thus enabling quick and accurate determination of the target similarity between the target query term and the candidate terms.

[0099] The methods for constructing a similarity recognition model may include:

[0100] First, the sample information in the sample dataset is segmented into multiple sample words. The sample dataset can include information from the target information database, and the sample words can include multiple text words obtained by segmenting the information from the target information database.

[0101] Secondly, the model set can be iteratively updated based on each sample word. The model set can include multiple candidate similarity recognition models. For example, before the iterative update, multiple candidate similarity recognition models can be randomly obtained. These multiple candidate similarity recognition models can be different types of similarity recognition models, or they can be similarity recognition models of the same type but with different parameter values.

[0102] In each iteration, for any candidate similarity recognition model in the model set, the loss value of the candidate similarity recognition model can be calculated based on multiple sample words, and each candidate similarity recognition model can be updated according to the loss value. The update method can include replacing the candidate similarity recognition model in the model set with a new candidate similarity recognition model, or optimizing the parameter values ​​of the candidate similarity recognition model in the model set. The specific settings can be set according to actual needs.

[0103] Furthermore, when the iteration update termination condition is met, the final similarity recognition model can be determined based on the loss values ​​of each candidate similarity recognition model in the model set. For example, the candidate similarity recognition model with the smallest loss value can be used as the final similarity recognition model. The iteration update termination condition can be set according to actual needs. For example, it can be determined that the iteration update termination condition is met when the number of iteration updates reaches a preset number, or when the loss value of a candidate similarity recognition model is less than a preset loss value. This effectively ensures the reliability of the similarity recognition model, thereby ensuring the accuracy of the target similarity determination result between the target query term and the candidate terms, and further improving the accuracy of information query results.

[0104] In one feasible implementation, the iterative update of the model set based on each of the sample words includes:

[0105] In the current iteration update, the loss value of each candidate similarity recognition model is determined based on each of the sample words;

[0106] Based on the loss values ​​of each candidate similarity recognition model, the target similarity recognition model among the candidate similarity recognition models is determined, and each candidate similarity recognition model other than the target similarity recognition model in the model set is updated to obtain the update result of the model set in the current iteration update.

[0107] Specifically, in the current iteration, for any candidate similarity recognition model, the loss value of the candidate similarity recognition model can be determined based on each sample word. This can be achieved by randomly combining each sample word pairwise to obtain multiple training samples, and then labeling the similarity of each training sample. The similarity of the training sample is the similarity between the two sample words in that training sample. In determining the loss value of the candidate similarity recognition model, the predicted similarity value for each training sample can be calculated using the candidate similarity recognition model. Based on the predicted similarity values ​​and labeled similarity values ​​of each training sample, the loss value of the candidate similarity recognition model is determined using a loss function. Furthermore, for any sample word, one or more word sets corresponding to that sample word can be constructed. Each word set can include the sample word, a synonym of the sample word, and a non-synonym of the sample word. For any word set, the candidate similarity recognition model can calculate the first sample similarity and the second sample similarity of the word set. Based on the first and second sample similarities, the sub-loss value of the word set is determined, and the loss value of the candidate similarity recognition model is determined based on the sub-loss values ​​of each word set. The first sample similarity can include the similarity between the sample word and its synonym, and the second sample similarity can include the similarity between the sample word and its non-synonym.

[0108] After obtaining the loss values ​​of each candidate similarity recognition model in the model set, the target similarity recognition model can be determined based on the loss values ​​of each candidate similarity recognition model. Then, the candidate similarity recognition models other than the target similarity recognition model in the model set are updated to obtain the update result of the model set in the current iteration. The loss value of the target similarity recognition model can be less than the loss values ​​of the candidate similarity recognition models other than the target similarity recognition model in the model set. Thus, the similarity prediction accuracy of each candidate similarity recognition model in the model set can be effectively improved through iterative updates.

[0109] In practice, a new candidate similarity recognition model can be generated based on the target similarity recognition model, and the candidate similarity recognition model other than the target similarity recognition model in the model set can be replaced with the new candidate similarity recognition model. In addition, the model parameters of the candidate similarity recognition model other than the target similarity recognition model in the model set can be updated directly.

[0110] In one feasible implementation, updating each candidate similarity recognition model other than the target similarity recognition model in the model set includes:

[0111] Delete all candidate similarity recognition models in the model set other than the target similarity recognition model;

[0112] The model update is repeated a specified number of times, and the result of each model update is added to the model set as a candidate similarity recognition model; the model update includes updating the model parameters of the target similarity model.

[0113] Specifically, during the process of updating candidate similarity recognition models other than the target similarity recognition model in the model set, all candidate similarity recognition models other than the target similarity recognition model can be deleted from the model set, that is, only the target similarity recognition model is retained, and the model update is repeated a specified number of times. In any model update, the model parameters of the target similarity model can be updated to obtain a new candidate similarity recognition model, and this new candidate similarity recognition model is added to the model set. During the process of updating the model parameters of the target similarity model, the model parameters can be randomly updated within the value range of the model parameters, and the model parameters can include symbols and / or variables.

[0114] Understandably, the number of target similarity recognition models can be one or more. When there are multiple target similarity recognition models, in any model update, one target similarity recognition model can be randomly selected to perform the model update. Alternatively, the number of times to perform the model update for each target similarity recognition model can be determined based on the specified number of times and the number of target similarity recognition models.

[0115] Therefore, by repeatedly performing model updates a specified number of times, multiple new candidate similarity recognition models can be obtained. The specified number of times can be set according to actual needs; for example, it can be the target number of candidate similarity recognition models minus the number of target similarity recognition models in the model set.

[0116] In each iteration, by removing candidate similarity recognition models with poor similarity prediction accuracy from the model set and optimizing candidate similarity recognition models with high similarity prediction accuracy, the similarity prediction accuracy of the final similarity recognition model can be effectively guaranteed, thereby improving the accuracy of information query results. At the same time, by optimizing the model parameters of candidate similarity recognition models with high similarity prediction accuracy in each iteration, the final similarity recognition model has strong extrapolation generalization ability, and can still accurately predict similarity when faced with new text words or data, without the need to retrain the similarity recognition model.

[0117] In one feasible implementation, determining the loss value of each candidate similarity recognition model based on each of the sample words includes:

[0118] Based on the candidate similarity recognition model, a first sample similarity and a second sample similarity are determined for each word set corresponding to multiple sample words; the word set corresponding to the sample word includes the sample word, the synonyms of the sample word, and the non-synonyms of the sample word, the first sample similarity includes the similarity between the sample word and the synonyms of the sample word, and the second sample similarity includes the similarity between the sample word and the non-synonyms of the sample word;

[0119] The loss value of the candidate similarity recognition model is determined based on the first sample similarity and the second sample similarity of each of the aforementioned word sets.

[0120] Specifically, after segmenting each sample information in the sample dataset to obtain multiple sample words, one or more word sets can be constructed for each sample word. For any word set, the word set can include the sample word, a synonym of the sample word, and a non-synonym of the sample word. For example, for any sample word e, H synonyms of the sample word can be randomly selected from the sample dataset to obtain the synonym set P = {p1, p2, ..., p...} H}, and randomly select R non-synonyms of the sample word from the sample dataset to obtain the non-synonym set Q = {q1, q2, ..., q} of the sample word. R R can be the same as or different from H. Based on the set of synonyms and non-synonyms of the sample word, we can obtain the H*R word set corresponding to the sample word: (e,p1,q1), (e,p1,q2), ..., (e,p1,q3). R ), (e,p2,q1), (e,p2,q2),…, (e,p2,q R ), ..., (e,p H ,q1), (e,p H ,q2), …,(e,pH ,q R ).

[0121] For any candidate similarity recognition model F t In determining the loss value of the candidate similarity recognition model, the first sample similarity and the second sample similarity corresponding to each word set can be calculated using the candidate similarity recognition model. The first sample similarity is the similarity between the sample word and its synonyms in the word set, and the second sample similarity is the similarity between the sample word and its non-synonyms in the word set.

[0122] For the word set (e, p) h ,q r This allows us to calculate the synonyms p of the sample word e and the sample word e. h First context vector similarity The sample word 'e' and its synonym 'p' h First semantic vector similarity The sample word 'e' and its non-synonym 'q' r Second context vector similarity And the sample word e and its non-synonym q r Second semantic vector similarity Based on the word frequency of the sample word e in the sample dataset, the first weight coefficient α is calculated. e Second weighting coefficient β e ,Will and Substitute into the candidate similarity recognition model F t This will give us the word set (e, p) h ,q r The first sample similarity corresponding to ) Will and Substitute into the candidate similarity recognition model F t This will give us the word set (e, p) h ,q r The second sample similarity corresponding to )

[0123] Candidate similarity recognition model F t loss value L t It can be shown in equation (5):

[0124]

[0125] In the formula, V is the set of all sample words, |V| is the total number of sample words, and max() is the function that takes the maximum value.

[0126] When the similarity of the first sample is greater than that of the second sample, it indicates that the candidate similarity recognition model F... t The calculated sample word e and its synonym p h The similarity is greater than that between sample word e and non-synonyms of sample word e, q. r Similarity, candidate similarity recognition model F t loss value L t It will approach 0, representing the candidate similarity recognition model F. t The similarity prediction accuracy is high, which can effectively ensure the reliability of the determination result of the similarity recognition model. At the same time, in the process of building the similarity recognition model, there is no need to label the similarity between two sample words in the sample dataset, thus effectively improving the construction efficiency of the similarity recognition model.

[0127] To further improve information retrieval efficiency, in one embodiment of this specification, determining the target information in the target information database based on the target word includes:

[0128] Based on the correspondence between the text words and the documents in the target information database, and the target words, the documents containing the target words are determined, and the documents containing the target words are taken as the target documents;

[0129] Based on the location information of the target word in the target document, the target information in the target document is determined.

[0130] Specifically, when new information is generated, it is typically inserted into a data table in the target information database. As information is continuously inserted, the data volume in the table becomes extremely large. Simultaneously, regular backups of the information in the target database are necessary. Therefore, information in the target database can be exported into multiple documents according to information category, time of generation, etc., allowing for document-based information retrieval during information searches. Currently, information searches within documents are usually performed using document search functions. With large amounts of information, the search efficiency is low, and it cannot guarantee that the search results match the user's intent, thus compromising the accuracy of the search results. Furthermore, when the target query contains uncommon or polysemous words, users need to consult related information one by one and identify semantically similar information, further compromising search efficiency.

[0131] To address this, the correspondence between each text word and a document in the target information database can be stored, along with the positional information of each text word within different documents. After a target word is matched, the documents containing the target word can be identified based on the target word and the correspondence between each text word and a document. These documents are then designated as the target documents. Furthermore, based on the positional information of the target word within the target documents, the target information within those documents can be determined. This target information can be information containing the target word. For example, the correspondence between each text word and a document in the target information database can be stored using an inverted index table. One side of the inverted index table contains the text word, and the other side contains the document containing that text word. Each document's index entry contains the position of the text word within that document. Thus, information related to the target query word can be quickly and effectively retrieved based on the target word. Furthermore, considering that the target word is determined based on the context vector similarity and semantic vector similarity between the target query word and each candidate word, even when the target query word contains uncommon or polysemous words, the target word can still be quickly and effectively matched, and related information can be automatically retrieved. This eliminates the need for users to manually search for related information or distinguish semantically similar information, greatly improving information retrieval efficiency.

[0132] The following uses alarm information query as an example to explain in detail the implementation process of the information query method in this application. The information query process can be as follows: Figure 2 As shown, it includes:

[0133] S201. Perform preprocessing and word segmentation on the target query term and candidate terms, and then execute steps S202 and S204; wherein, preprocessing may include text cleaning, tag removal, link removal, case conversion, etc.

[0134] S202. Determine the first context vector of the target query term and the second context vector of each candidate term based on the target context vector matrix;

[0135] S203. Based on the first context vector of the target query term and the second context vector of each candidate term, calculate the first similarity between the target query term and each candidate term using a cosine scorer, and then proceed to step S206.

[0136] S204. Obtain the first semantic vector of the target query word and the second semantic vector of each candidate word through a pre-trained semantic model;

[0137] S205. Based on the first semantic vector of the target query term and the second semantic vector of each candidate term, calculate the second similarity between the target query term and each candidate term using a cosine scorer, and then proceed to step S206.

[0138] S206. Determine the first weight coefficient of the first similarity and the second weight coefficient of the second similarity based on the word frequency of the target query term, and perform weighted processing on the first similarity based on the first weight coefficient, and perform weighted processing on the second similarity based on the second weight coefficient.

[0139] S207. Input the weighted first similarity and the weighted second similarity into the similarity recognition model to obtain the target similarity between the target query term and each candidate term;

[0140] S208. Based on the target similarity between the target query term and each candidate term, determine the target term from multiple candidate terms, and query the target information from each document in the target information database based on the target term.

[0141] Assume the target information database contains the following 5 alarm messages:

[0142] Alarm message 1: "HeaterSetTemp / BsNotWork".

[0143] Alarm message 2: " / Heater / SetTempAlarm".

[0144] Alarm message 3: "PrepTransAlarm".

[0145] Alarm message 4: " / Heater / NotSetAlarm".

[0146] Alarm message 5: "Bs / NotWorkAlarm".

[0147] Word segmentation of alarm message 1 yields Heat, Set, Temp, Bs, Not, and Work; word segmentation of alarm message 2 yields Heat, Set, Temp, and Alarm; word segmentation of alarm message 3 yields Prep, Trans, and Alarm; word segmentation of alarm message 4 yields Heat, Not, Set, and Alarm; and word segmentation of alarm message 5 yields Bs, Not, Work, and Alarm. Therefore, the text words can include Heat, Set, Temp, Bs, Not, Work, Alarm, Prep, and Trans. The context of each text word in each alarm message is determined according to k=5, and the resulting target context vector matrix M is shown in Table 1.

[0148] Table 1

[0149] Heater Set Temp Bs Not Work Alarm Prep Trans Heater 0 3 / 11 2 / 11 1 / 11 2 / 11 1 / 11 2 / 11 0 / 11 0 / 11 Set 3 / 11 0 2 / 11 1 / 11 2 / 11 1 / 11 2 / 11 0 / 11 0 / 11 Temp 2 / 8 2 / 8 0 1 / 8 1 / 8 1 / 8 1 / 8 0 / 8 0 / 8 Bs 1 / 8 1 / 8 1 / 8 0 2 / 8 2 / 8 1 / 8 0 / 8 0 / 8 Not 2 / 11 2 / 11 1 / 11 2 / 11 0 2 / 11 2 / 11 0 / 11 0 / 11 Work 1 / 8 1 / 8 1 / 8 2 / 8 2 / 8 0 1 / 8 0 / 8 0 / 8 Alarm 2 / 11 2 / 11 1 / 11 1 / 11 2 / 11 1 / 11 0 1 / 11 1 / 11 Prep 0 / 2 0 / 2 0 / 2 0 / 2 0 / 2 0 / 2 1 / 2 0 1 / 2 Trans 0 / 2 0 / 2 0 / 2 0 / 2 0 / 2 0 / 2 1 / 2 1 / 2 0

[0150] It is understandable that in Table 1, the i-th row M[id(i)] represents the context vector of text word i.

[0151] Exemplary device

[0152] This specification also provides a semiconductor process apparatus, including a process chamber 20 and a controller. The controller includes at least one processor and at least one memory, in which a computer program is stored. When the computer program is executed by the processor, it implements the information query method as described in any of the above embodiments.

[0153] For example, the controller can be a host computer, which may include a human-computer interaction device. Users can input target query terms through the human-computer interaction device, the host computer performs information query based on the target query terms, and displays the target information obtained through the human-computer interaction device.

[0154] Among them, semiconductor process equipment 200 can be etching equipment, for reference Figure 3 When the semiconductor process equipment 200 is an etching equipment, it may also include an air intake assembly 20A and an air extraction assembly ( Figure 3 (Not shown in the diagram) Upper electrode assembly 20B and lower electrode assembly 20C. The controller can open the valve of the air inlet assembly 20A to introduce the corresponding process gas into the process chamber 20; the controller can also control the flow rate of the process gas by controlling the opening and closing degree of the valve of the air inlet assembly 20A. The controller can also control the pressure inside the process chamber 20 by controlling the air extraction assembly to extract air from the inside of the process chamber 20.

[0155] The upper electrode assembly 20B includes an RF coil 21, an upper RF power supply 23, and an upper matching unit 25. The controller is also used to control the upper RF power supply 23 to provide upper electrode power to the RF coil 21 through the upper matching unit 25, so that the RF coil 21 excites the process gas inside the process chamber 20 to generate plasma.

[0156] The lower electrode assembly 20C includes a wafer carrier 22, a lower RF power supply 24, and a lower matching unit 26. The controller is also used to control the lower RF power supply 24 to provide lower electrode power to the lower electrode of the wafer carrier 22 through the lower matching unit 26, so that the lower electrode of the wafer carrier 22 is provided with RF bias, so that the plasma has bombardment capability to perform anisotropic etching of the gate dielectric layer during the fabrication of the field-effect transistor 100.

[0157] The semiconductor process equipment 200 in this application embodiment can be any one of inductively coupled plasma (ICP) etching equipment, capacitively coupled plasma etching equipment, inductively coupled plasma reactive ion etching equipment, and reactive ion etching equipment. This application embodiment does not limit the type of semiconductor process equipment 200.

[0158] Exemplary computer program products and storage media

[0159] In addition to the methods and devices described above, the information query methods provided in the embodiments of this specification can also be computer program products, which include computer program instructions that, when executed by a processor, cause the processor to perform the steps in the information query methods according to various embodiments of this specification as described in the "Exemplary Methods" section above.

[0160] The computer program product described herein can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments described herein. These programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0161] Furthermore, embodiments of this specification also provide a computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor of the steps in the information query methods according to various embodiments of this specification as described in the "Exemplary Methods" section above.

[0162] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this specification can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0163] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0164] The embodiments described above are merely illustrative of several implementation methods outlined in this specification. While the descriptions are specific and detailed, they should not be construed as limiting the scope of the solutions provided in this specification. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this specification, and these all fall within the scope of protection of this specification. Therefore, the scope of protection for this patent should be determined by the appended claims.

Claims

1. An information retrieval method, characterized in that, include: Based on the target context vector matrix, a first context vector of the target query term and second context vectors of multiple candidate terms are determined. Based on the first context vector and the multiple second context vectors, multiple first similarities are calculated. The first similarity includes the similarity between the first context vector and the second context vector. The target context vector matrix includes the context vectors of multiple text words, which are obtained by segmenting each piece of information in the target information database of semiconductor process equipment. Obtain the first semantic vector of the target query term and the second semantic vectors of multiple candidate terms, and calculate multiple second similarities based on the first semantic vector and the multiple second semantic vectors; the second similarity includes the similarity between the first semantic vector and the second semantic vector. Based on multiple first similarities and multiple second similarities, the target similarity between the target query term and multiple candidate terms is determined, and based on the target similarity between the target query term and multiple candidate terms, the target term is determined from the multiple candidate terms, and the target information in the target information database is determined based on the target term.

2. The method according to claim 1, characterized in that, Based on the target context vector matrix, the first context vector of the target query term is determined, including: Based on the target query term, determine the words to be matched in each of the text terms; The first context vector of the target query term is determined based on the mean of the context vectors of each of the terms to be matched.

3. The method according to claim 1, characterized in that, The step of determining the target similarity between the target query term and the multiple candidate terms based on multiple first similarities and multiple second similarities includes: Based on the word frequency of the target query term, a first weight coefficient of the first similarity and a second weight coefficient of the second similarity are determined. The first weight coefficient is negatively correlated with the word frequency of the target query term, and the second weight coefficient is negatively correlated with the first weight coefficient. The first similarity is weighted based on the first weight coefficient, and the second similarity is weighted based on the second weight coefficient; The target similarity between the target query term and the candidate term is determined based on the weighted first similarity and the weighted second similarity.

4. The method according to claim 3, characterized in that, The determination of the target similarity between the target query term and the candidate terms based on the weighted first similarity and the weighted second similarity includes: The weighted first similarity and the weighted second similarity are input into the similarity recognition model to obtain the target similarity between the target query term and the candidate term; wherein, the similarity recognition model is constructed by the following method: The information of each sample in the sample dataset is segmented into words to obtain multiple sample words; The model set is iteratively updated based on each of the sample words, and the model set includes multiple candidate similarity recognition models; When the iterative update termination condition is met, the similarity recognition model is determined based on the loss value of each candidate similarity recognition model in the model set.

5. The method according to claim 4, characterized in that, The iterative update of the model set based on each of the sample words includes: In the current iteration update, the loss value of each candidate similarity recognition model is determined based on each of the sample words; Based on the loss values ​​of each candidate similarity recognition model, the target similarity recognition model among the candidate similarity recognition models is determined, and each candidate similarity recognition model other than the target similarity recognition model in the model set is updated to obtain the update result of the model set in the current iteration update.

6. The method according to claim 5, characterized in that, Updating each candidate similarity recognition model in the model set other than the target similarity recognition model includes: Delete all candidate similarity recognition models in the model set other than the target similarity recognition model; The model update is repeated a specified number of times, and the result of each model update is added to the model set as a candidate similarity recognition model; the model update includes updating the model parameters of the target similarity model.

7. The method according to claim 5, characterized in that, The step of determining the loss value of each candidate similarity recognition model based on each of the sample words includes: Based on the candidate similarity recognition model, a first sample similarity and a second sample similarity are determined for each word set corresponding to multiple sample words; the word set corresponding to the sample word includes the sample word, the synonyms of the sample word, and the non-synonyms of the sample word, the first sample similarity includes the similarity between the sample word and the synonyms of the sample word, and the second sample similarity includes the similarity between the sample word and the non-synonyms of the sample word; The loss value of the candidate similarity recognition model is determined based on the first sample similarity and the second sample similarity of each of the aforementioned word sets.

8. The method according to any one of claims 1 to 7, characterized in that, The step of determining the target information in the target information database based on the target words includes: Based on the correspondence between the text words and the documents in the target information database, and the target words, the documents containing the target words are determined, and the documents containing the target words are taken as the target documents; Based on the location information of the target word in the target document, the target information in the target document is determined.

9. A semiconductor process apparatus, comprising a process chamber and a controller, characterized in that, The controller includes at least one processor and at least one memory, the memory storing a computer program that, when executed by the processor, implements the information query method as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the information query method as described in any one of claims 1 to 8.

Images