Label information determination method and device, storage medium, electronic equipment and program product

Abstract

The invention discloses a tag information determination method and device, a storage medium, electronic equipment and a program product. Relates to the field of artificial intelligence. Splitting the to-be-processed data into character sequences; the correlation degree between the correlation data and the character sequence is larger than a correlation degree threshold value; determining context information of the character sequence in the to-be-processed data based on the associated data; and determining label information of the character sequence based on the associated data and the contextual information. Through the text analysis method and device, the technical problem that the accuracy of text analysis is low in the prior art is solved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence, and in particular, to a method, device, storage medium, electronic device and program product for determining tag information. Background Art

[0002] Currently, entity recognition is the basis for building key applications such as knowledge graphs, intelligent search, and compliance analysis.

[0003] In related technologies, entities are usually extracted based on preset rules, such as regular expression matching. However, this method has limited flexibility and is difficult to adapt to the challenges of polysemy and semantic parsing of long sentences. In addition, building and maintaining rule sets is time-consuming and labor-intensive, difficult to scale and automate, and there is a technical problem of low accuracy in parsing text.

[0004] With respect to the technical problem of low accuracy in parsing text in related technologies, no effective solution has been proposed so far. Summary of the invention

[0005] The main purpose of the present application is to provide a method, device, storage medium, electronic device and program product for determining tag information to solve the technical problem of low accuracy in parsing text in related technologies.

[0006] In order to achieve the above-mentioned purpose, according to one aspect of the present application, a method for determining label information is provided. The method may include: obtaining data to be processed; splitting the data to be processed into character sequences; obtaining associated data whose degree of association with the character sequence is greater than an association threshold; determining context information of the character sequence in the data to be processed based on the associated data; and determining label information of the character sequence based on the associated data and the context information.

[0007] Optionally, splitting the data to be processed into character sequences includes: extracting text data from the data to be processed; and splitting the text data to obtain character sequences.

[0008] Optionally, obtaining associated data having an association degree with a character sequence greater than an association degree threshold includes: annotating the character sequence to obtain an annotated character sequence; converting the annotated character sequence into a word embedding vector; and using dictionary information to determine associated data having an association degree with the word embedding vector greater than an association degree threshold, wherein the dictionary information is used to provide additional semantics and / or characters to the word embedding vector.

[0009] Optionally, based on the associated data, context information of the character sequence in the data to be processed is determined, including: integrating the associated data and the word embedding vector to obtain integrated data; and obtaining context data of the integrated data.

[0010] Optionally, determining the label information of the character sequence based on the associated data and the context information includes: identifying the character sequence based on the associated data and the context information to obtain multiple initial label information of the character sequence; and determining the label information from the multiple initial label information.

[0011] Optionally, determining the label information from multiple initial label information includes: retrieving a conditional random field model to determine evaluation information of the initial label information; and determining the label information from multiple initial label information based on the evaluation information.

[0012] In order to achieve the above-mentioned purpose, according to one aspect of the present application, a device for determining label information is provided. The device may include: a first acquisition unit, used to acquire data to be processed; a splitting unit, used to split the data to be processed into character sequences; a second acquisition unit, used to acquire associated data whose degree of association with the character sequence is greater than an association threshold; a determination unit, used to determine context information of the character sequence in the data to be processed based on the associated data; and a recognition unit, used to determine label information of the character sequence based on the associated data and the context information.

[0013] Optionally, the splitting unit may further include: an extraction module for extracting text data from the data to be processed; and a splitting module for splitting the text data to obtain a character sequence.

[0014] Optionally, the second acquisition unit may also include: a labeling module, used to label the character sequence to obtain a labeled character sequence; a conversion module, used to convert the labeled character sequence into a word embedding vector; a first determination module, used to use dictionary information to determine associated data whose association with the word embedding vector is greater than an association threshold, wherein the dictionary information is used to provide additional semantics and / or characters to the word embedding vector.

[0015] Optionally, the determination unit may further include: an integration module, used to integrate the associated data and the word embedding vector to obtain integrated data; and an acquisition module, used to acquire context data of the integrated data.

[0016] Optionally, the recognition unit may further include: a recognition module, used to recognize the character sequence based on the associated data and context information to obtain multiple initial label information of the character sequence; and a second determination module, used to determine the label information from the multiple initial label information.

[0017] Optionally, the second determination module may include: a first sub-determination module, used to call the conditional random field model to determine evaluation information of the initial label information; and a second sub-determination module, used to determine the label information from multiple initial label information based on the evaluation information.

[0018] In order to achieve the above-mentioned purpose, according to another aspect of the present application, a computer-readable storage medium is further provided, which may include a stored executable program, wherein when the executable program is running, the device where the computer-readable storage medium is located can be controlled to execute the above-mentioned method.

[0019] In order to achieve the above-mentioned purpose, according to another aspect of the present application, an electronic device is further provided, which may include: a memory storing an executable program; and a processor for running the program, wherein the above-mentioned method is executed when the program is running.

[0020] In order to achieve the above objective, according to another aspect of the present application, a computer program product is provided. The computer program product may include computer instructions. When the computer instructions are executed by a processor, the steps of the above method are implemented.

[0021] In an embodiment of the present application, data to be processed is obtained; the data to be processed is split into a character sequence; associated data whose association with the character sequence is greater than an association threshold is obtained; based on the associated data, context information of the character sequence in the data to be processed is determined; based on the associated data and the context information, label information of the character sequence is determined. That is, in an embodiment of the present application, a configured interface protocol and an interface protocol to be configured are obtained, and the difference information between the two is determined. The difference information is converted using the configuration data of the target interface to obtain the configuration data. The configured interface protocol can be updated according to the configuration data to obtain the interface protocol to be configured, thereby achieving the technical effect of improving the accuracy of text parsing and solving the technical problem of low accuracy of text parsing. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0023] Figure 1 A hardware structure block diagram of a computer terminal (or mobile device) for implementing a method for determining tag information is shown;

[0024] Figure 2 is a flow chart of a method for determining tag information according to an embodiment of the present application;

[0025] Figure 3 is a flow chart of an entity extraction method according to an embodiment of the present application;

[0026] Figure 4 is a schematic diagram of an entity extraction model according to an embodiment of the present application;

[0027] Figure 5 is a schematic diagram of a device for determining tag information according to an embodiment of the present application;

[0028] Figure 6 It is a structural block diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0029] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present application.

[0030] It should be noted that the terms "first", "second", etc. 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 of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising 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.

[0031] First, some nouns or terms that appear in the description of the embodiments of the present application are subject to the following explanations:

[0032] Knowledge extraction refers to the process of automatically extracting useful information and knowledge from text or data. This process may include identifying entities, relationships, events, concepts, etc., and representing this information in a structured manner for further analysis and application.

[0033] Lattice: In natural language processing, Lattice refers to a data structure that can be used to represent multiple possible word segmentation and part-of-speech tagging results of a sentence. In Chinese word segmentation, Lattice can represent all possible word segmentation results of a sentence, as well as the part-of-speech tagging of each word.

[0034] Bidirectional Long Short-Term Memory (BiLSTM) is a deep learning model that can be used to process sequence data, such as text or time series. It can be composed of two LSTMs and can be used to understand the context of sequence data (such as character sequences), thereby achieving better performance in natural language processing and other sequence data tasks.

[0035] Conditional Random Field (CRF) is a probabilistic graphical model that can be used to annotate sequence data. In natural language processing, CRF can be used for tasks such as part-of-speech tagging and named entity recognition. CRF can consider the contextual information in the input sequence and annotate the sequence according to conditional probability, so as to better handle the dependencies and constraints in the annotation task and improve the accuracy of annotation.

[0036] It should be noted that the collected information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for display, data for analysis, etc.) involved in this application are information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of relevant data are in compliance with relevant laws, regulations and standards, necessary confidentiality measures are taken, and public order and good customs are not violated, and corresponding operation entrances are provided for users to choose to authorize or refuse. For example, an interface is set up between this system and relevant users or institutions to provide users with corresponding operation entrances for users to choose to agree or refuse the results of automated decision-making; if the user chooses to refuse, the expert decision-making process will be entered.

[0037] Example 1

[0038] According to an embodiment of the present application, a method embodiment for determining label information is also provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0039] The method embodiment provided in the first embodiment of the present application can be executed in a mobile terminal, a computer terminal or a similar computing device. Figure 1 FIG. 1 shows a hardware structure block diagram of a computer terminal (or mobile device) for implementing a method for determining tag information. Figure 1As shown, the computer terminal 10 (or mobile device) may include one or more (102a, 102b, ..., 102n are used to illustrate) processors 102 (the processor 102 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 104 for storing data, and a transmission device 106 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the BUS bus), a network interface, a power supply and / or a camera. It can be understood by those skilled in the art that Figure 1 The structure shown is only for illustration and does not limit the structure of the above electronic device. Figure 1 More or fewer components as shown, or with Figure 1 Different configurations are shown.

[0040] It should be noted that the one or more processors 102 and / or other data processing circuits described above may generally be referred to herein as "data processing circuits". The data processing circuits may be embodied in whole or in part as software, hardware, firmware, or any other combination thereof. In addition, the data processing circuit may be a single independent processing module, or may be incorporated in whole or in part into any of the other components in the computer terminal 10 (or mobile device). As described in the embodiments of the present application, the data processing circuit acts as a processor control (e.g., selection of a variable resistor terminal path connected to an interface).

[0041] The memory 104 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the method for determining tag information in the embodiment of the present application. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, that is, the above-mentioned method for determining tag information is realized. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include a memory remotely arranged relative to the processor 102, and these remote memories may be connected to the computer terminal 10 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0042] The transmission device 106 is used to receive or send data via a network. The specific example of the above network may include a wireless network provided by a communication provider of the computer terminal 10. In one example, the transmission device 106 includes a network adapter (Network Interface Controller, referred to as NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 106 can be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0043] The display may be, for example, a touch screen liquid crystal display (LCD) that enables a user to interact with a user interface of the computer terminal 10 (or mobile device).

[0044] Under the above operating environment, this application provides Figure 2 The method for determining the label information shown. Figure 2 is a flowchart of a method for determining tag information according to an embodiment of the present application. Figure 2 As shown, the method may include the following steps:

[0045] Step S201, obtaining data to be processed.

[0046] In this embodiment, the data to be processed may be entities that need knowledge extraction, data containing characters and words, for example, raw text data that need entity extraction analysis, for example, "Wang applied for a 1 million yuan mortgage at the Guangzhou branch of the bank". In the entity recognition task, the data to be processed may be any text containing potential entities, such as bank transaction records, customer service dialogues, financial reports, etc. It should be noted that this is only an example, and there is no specific limitation on the content of the data to be processed.

[0047] Optionally, the data to be processed may be obtained from a bank system, and the data to be processed may include part of the bank's customer data information.

[0048] Optionally, the first step of the entity extraction process may be to obtain text data that requires entity recognition. Thus, the data to be processed may be obtained from files such as transaction data, conversation data, and reports.

[0049] For example, assuming that the banking system needs to identify entities in customer transaction records, then a specific transaction record text, such as "Mr. Wang made a time deposit of 1 million yuan at the bank's Guangzhou branch", is the data to be processed.

[0050] Step S202, splitting the data to be processed into character sequences.

[0051] In this embodiment, the above character sequence may be a split text.

[0052] Optionally, after obtaining the data to be processed, the data to be processed can be decomposed into a series of characters or morphemes to facilitate the model to analyze them one by one. For example, in the data to be processed "Wang Mou applied for a housing loan of 1 million yuan at the Guangzhou branch of the bank", the text can be divided into "Wang Mou", "at", "in", "industry", "business", "bank", "bank", "Guang", "zhou", "branch", "bank", "apply", "a", "a", "billion yuan", "of", "house", "loan". The above character sequence can be the basic unit of model analysis, which is helpful to identify the boundaries of entities, especially the boundaries of Chinese text, where the above entity may be composed of a single character or multiple characters.

[0053] Optionally, character-level processing helps the model capture entity boundaries and has a better recognition effect on phrases and compound words. Therefore, after obtaining the data to be processed, the data to be processed can be converted into a character sequence so that the model can process and analyze it character by character.

[0054] Step S203: acquiring associated data whose degree of association with the character sequence is greater than a threshold value of the degree of association.

[0055] In this embodiment, the above-mentioned associated data may be additional information related to the character sequence, and may be obtained from a dictionary or a knowledge base. By using the associated data to identify the character sequence, the accuracy of the data to be processed may be improved. The above-mentioned association threshold may be a preset value. It should be noted that this is only an example, and no specific limitation is imposed on the meaning of the associated data and the method for determining the association threshold.

[0056] Optionally, this embodiment may filter out additional information that is highly associated with the current character sequence, such as professional terms, entity names, etc., based on the dictionary of the banking system.

[0057] In this embodiment, the model is provided with prior knowledge of similar entity types through dictionary information, which helps to improve the accuracy of entity recognition.

[0058] For example, in entity recognition in a banking system, associated data may include entity type information corresponding to "bank", "Guangzhou branch", "Wang" and "1 million yuan" in the dictionary. The associated data can be used to provide context and domain-specific knowledge, which can help the model determine what entity type a character sequence belongs to. In particular, in the case of polysemous words, associated data can provide key disambiguation information.

[0059] Optionally, the embodiment can search for entity information that is highly relevant to the current character sequence from an external dictionary or knowledge base, thereby providing additional context and semantic clues to the model. In entity recognition tasks, especially when processing domain-specific text, obtaining highly relevant dictionary data can significantly improve the performance and accuracy of the model.

[0060] Optionally, the above-mentioned associated data may refer to entity information related to the current character sequence, and the information may be obtained from a pre-built dictionary or knowledge base. In a banking scenario, the dictionary may contain the names of various financial entities, such as the names of different banks, financial product names, customer names, etc., as well as their entity types.

[0061] Optionally, based on the degree of match between the character sequence and the entries in the dictionary, the degree of association between the character sequence and the multiple entries can be determined to obtain at least one piece of associated data with a degree of association greater than a threshold value of the degree of association. Among them, the above calculation method may include exact matching based on the dictionary, approximate matching based on the edit distance, matching based on semantic similarity, etc. For example, when processing "Guangzhou Branch of the Bank", the institutional entity information that matches it in the dictionary will be selected, and may also include some common names and asset names, but only those with a high enough degree of association will be retained. It should be noted that the method for determining associated data here is only for illustration and no specific limitation is imposed.

[0062] Optionally, the above-mentioned associated data may include professional terms in the field, entity types, and patterns or location information of these entities that may appear in the text. In the bank entity recognition scenario, the associated data may come from the bank's internal dictionary and may include common entity types such as names of people, organizations, products, transaction types, and their examples. These data are used to guide the model on how to understand the meaning of specific words in the banking context.

[0063] Step S204: determining context information of the character sequence in the data to be processed based on the associated data.

[0064] In this embodiment, the above-mentioned context information may refer to the position of the character sequence in the text and the relationship with the surrounding words, which may be used to understand the semantics and function of the entity. The above-mentioned context information may enable the model to understand the meaning of the entity in a specific context and overcome problems such as polysemy.

[0065] Optionally, the filtered associated data may be used to analyze the role and meaning of the character sequence in the data to be processed, such as the type and position of the entity, to determine context information of the character sequence in the data to be processed.

[0066] Optionally, in an entity recognition task, the model needs to identify entities in the text and determine their types, such as person names, organization names, asset names, etc. However, the meaning of an entity in the text usually depends on the context of the entity. For example, "bank" may refer to a bank or a bank product in different contexts. Therefore, it is very necessary to use associated data to judge the context information.

[0067] Optionally, this embodiment can look up entity information highly relevant to the character sequence in the data to be processed from a dictionary. For example, for "Guangzhou Branch of [Bank Name]", the model can find the entries of "[Bank Name]" and "Guangzhou Branch" in the dictionary and obtain entity information with an association degree higher than the association degree threshold to obtain associated data. Further, the relative position of the character sequence in the data to be processed can be analyzed, and based on the associated data, the association between words and potential semantic relationships can be determined. For example, if "Guangzhou Branch of [Bank Name]" appears immediately after the name "Wang Mou", it can be inferred that "Guangzhou Branch of [Bank Name]" is more likely to be the bank institution involved by "Wang Mou", thereby obtaining the context data of the character sequence in the data to be processed. Finally, the context information and the associated data can be used to guide the judgment of the entity type. For example, in the sentence "Wang Mou handled a 1 million yuan time deposit at the Guangzhou Branch of [Bank Name]", by analyzing the relationship between "Guangzhou Branch of [Bank Name]" and "Wang Mou" and "1 million yuan", the model can more accurately identify "Guangzhou Branch of [Bank Name]" as an organization entity, "Wang Mou" as a person name entity, and "1 million yuan" as an asset entity.

[0068] For example, obtain the data to be processed: "Wang Mou handled a 1 million yuan time deposit at the Guangzhou Branch of [Bank Name]." Obtain the entries of "Wang Mou", "[Bank Name]", "Guangzhou Branch", "1 million yuan", etc. from the dictionary, which are respectively associated with person names, organization names, and asset names. The position and relationship of each entry in the sentence can be analyzed through a Lattice structure. For example, the verb "at" between "Wang Mou" and "Guangzhou Branch of [Bank Name]" indicates that there is a certain association between Wang Mou and the Guangzhou Branch of [Bank Name], and "1 million yuan" appears immediately after "handled", suggesting that it may be the amount of a certain transaction. Therefore, based on the above analysis, the role of "Wang Mou" as a person name entity in the context, "Guangzhou Branch of [Bank Name]" as an organization entity related to Wang Mou's transaction, and "1 million yuan" as the asset entity of the transaction can be determined. These context information will be used in the training of the model and the decision-making process of entity recognition.

[0069] In this embodiment, the understanding of the context of the character sequence by the model is enhanced by associated data, which is crucial for entity recognition and knowledge extraction. This context information can refer to the position of the character sequence in the text and its relationship with the surrounding words, and these relationships can help determine the semantics and entity type of the character sequence. By using associated data, that is, information from a dictionary or other knowledge bases, the model can more accurately grasp the boundaries and types of entities.

[0070] Step S205: Determine the label information of the character sequence based on the associated data and context information.

[0071] In this embodiment, the above label information is the prediction of the entity type of the character sequence by the model, which can be used to identify the meaning or entity type of the character sequence and is the basis for entity extraction and semantic analysis. It can include BIO (BeginInside, simply referred to as Outside) annotation information, can include labels, and the probability that the character sequence is such a label. It should be noted that this is only an example, and the type of label information is not specifically limited here.

[0072] For example, obtain the data to be processed: Wang Mou handled a 1 million yuan fixed deposit in Guangzhou Branch of Industrial and Commercial Bank. Split the data to be processed to obtain the character sequences: "Wang Mou", "in", "in", "dustrial", "and", "commercial", "bank", "Guang", "zhou", "branch", "applied", "for", "a", "1 million yuan", "housing", "loan". Obtain the associated data of the character sequences. Based on the associated data, the context data of the character sequences in the data to be processed can be determined. Further, based on the associated data and context data, the corresponding labels of the character sequences can be determined: "Wang - B - Dname", "Mou - I - Dname", "Long - E - Dname", "in - B - Org", "dustrial - I - Org", "and - I - Org", "commercial - I - Org", "bank - E - Org", "Guang - B - Org", "zhou - I - Org", "branch - I - Org", "bank - E - Org", "1 - B - Asset", "0 - I - Asset", "0 - I - Asset", "million - E - Asset", "yuan - O", "housing - B - Asset", "loan - E - Asset", and / or the probability that the character sequence is the above label.

[0073] Optionally, obtain the data to be processed, enhance the model's understanding of entities through character sequence decomposition, utilize the associated data and context information, and finally output the label information corresponding to the character sequence, thereby completing the recognition of entities in the character sequence. Through the above method, key entities such as customer names, account information, amounts, etc. can be accurately recognized, thus supporting more accurate data analysis and decision-making support.

[0074] In the above steps S201 to S205, the data to be processed is obtained; the data to be processed is split into a character sequence; the associated data whose association with the character sequence is greater than the association threshold is obtained; based on the associated data, the context information of the character sequence in the data to be processed is determined; based on the associated data and the context information, the label information of the character sequence is determined. That is, in the embodiment of the present application, the configured interface protocol and the interface protocol to be configured are obtained, the difference information between the two is determined, and the difference information is converted using the configuration data of the target interface to obtain the configuration data. The configured interface protocol can be updated according to the configuration data to obtain the interface protocol to be configured, thereby achieving the technical effect of improving the accuracy of text parsing and solving the technical problem of low accuracy of text parsing.

[0075] The embodiments of the present application are described in detail below in combination with the above steps.

[0076] As an optional implementation, step S202, splitting the data to be processed into character sequences, includes: extracting text data from the data to be processed; and splitting the text data to obtain character sequences.

[0077] In this embodiment, after obtaining the data to be processed, preprocessing can be performed on the data to be processed, such as removing irrelevant characters, formatting, etc., to ensure that the quality and structure of the data are suitable for subsequent processing. Through the above preprocessing, the characters included in the data to be processed can be removed, so that text data can be obtained, and further, the text data can be split to obtain a character sequence.

[0078] Optionally, after obtaining the data to be processed, the data of the banking system may be preprocessed to obtain a character sequence. The data to be processed may include some bank customer data information. The preprocessing may include data cleaning, word segmentation and data labeling. Data cleaning is performed on the acquired data to be processed to remove meaningless values ​​that may appear in the data to be processed. The meaningless values ​​may include Hypertext Markup Language (HTML) tags, non-text data, null values, etc. Through the preprocessing, text data in the data to be processed may be extracted. Furthermore, the text data may be split to obtain a character sequence, which may include at least one word or element.

[0079] Optionally, after obtaining the data to be processed, plain text information (i.e., text data) that can be used for text analysis can be extracted from the data to be processed. In the banking field, the data to be processed may come from various channels, including but not limited to customer service conversation records, transaction details, loan application forms, financial reports, etc. These data may contain non-text elements such as HTML tags, pictures, tables, etc. Therefore, extracting text data is the first step to convert these messy data into an analyzable format. At the same time, considering that Chinese entities may consist of a single character or multiple consecutive characters, and word boundaries are not always obvious, the text data can be split to obtain character sequences. Splitting into character sequences enables the model to analyze character by character to ensure accurate identification of entity boundaries.

[0080] In this step, irrelevant noise can be removed through data cleaning and extraction to ensure that the model processes pure text information, thereby improving the accuracy and efficiency of entity recognition. In addition, this helps the training and inference of the model in subsequent steps and reduces misrecognition caused by data format issues. At the same time, character-level splitting can capture the fine-grained information of the entity, especially for long entities or entities containing complex naming patterns, and can effectively identify the boundaries of the entity, which is crucial to the accuracy of entity recognition. And the character-level representation also makes it easier for the model to utilize character-level features, such as character shape, strokes, etc., which is very important for identifying entities with similar or partially similar text.

[0081] As an optional implementation, step S203, obtaining associated data having an association degree with a character sequence greater than an association degree threshold, includes: annotating the character sequence to obtain an annotated character sequence; converting the annotated character sequence into a word embedding vector; and using dictionary information to determine associated data having an association degree with the word embedding vector greater than an association degree threshold, wherein the dictionary information is used to provide additional semantics and / or characters to the word embedding vector.

[0082] In this embodiment, the character sequence can be annotated to obtain an annotated character sequence, and the annotated annotation information is the initial annotation. The annotated character sequence can be converted into a word embedding vector, and the dictionary information can be used to determine the associated data in the dictionary information whose association with the word embedding vector is greater than the association threshold. The above-mentioned dictionary information can be used to provide additional semantics and / or characters to the word embedding vector.

[0083] Optionally, this embodiment can use the BIO annotation method to annotate the character sequence to obtain an annotated character sequence. It should be noted that this is only an example, and the annotation method is not specifically limited. Through this annotation method, B, I, and O can be identified to annotate the text data respectively, and different letter annotations represent different meanings. Table 1 is a annotation information table. As shown in Table 1, a character can be annotated as BX, indicating that the character is at the beginning of the text and the character belongs to type X. If the character is marked as IX, it means that the character is in the middle of the text and the character belongs to type X. If the character is marked as O, it means that the character does not belong to any type. In this application, the three categories of asset name, organization name and personal name are used as the main entity recognition types. It should be noted that this is only an example, and the above letters are not specifically limited.

[0084] Table 1 Labeling information table

[0085]

[0086] Optionally, after entity labeling, a labeling character sequence can be obtained, and the labeling character sequence can be used for vectorization processing of subsequent models to obtain a word embedding vector. Using the word embedding vector can contribute to a better word segmentation effect.

[0087] In this embodiment, a shallow neural network model (e.g., Word2vec model) can be used to transform the labeled character sequence to obtain a word embedding vector. The output layer of the above-mentioned Word2vec model can be a probability distribution regression method (e.g., Softmax regression method), which can map all values ​​to the region (0,1), and the output conclusion can be a vector, that is, the probability value of obtaining the corresponding label sequence. It can be used to convert words in natural language into feature vectors as input to the next layer of model, and its learning goal can be to maximize the log-likelihood function (e.g., L=∑ w∈c log p (w|context(w))).

[0088] Optionally, the above Word2vec model can convert the preprocessed words into a fixed-length vector representation to achieve the purpose of converting the labeled character sequence into a word embedding vector. Among them, the above Word2vec model can capture the semantics and contextual relationship of words and provide rich feature representation for subsequent models.

[0089] In this embodiment, after the word embedding vector is obtained, Lattice can be used to compare the dictionary information with the word embedding vector to obtain related data.

[0090] Optionally, based on word segmentation, the Lattice structure can further consider the character information within each word, as well as possible polysemous word information in the dictionary. For example, "Bank of China" may have "Bank" and "China" as subwords in the dictionary. This information will be used by the model to select the most relevant features for the current context through the gating mechanism, so as to obtain the most relevant data.

[0091] Optionally, the Lattice structure is used as an enhanced word embedding and feature extraction method in natural language processing (NLP) tasks, especially in entity recognition and word polysemy processing, and its input may include two parts: word embedding vector and dictionary information. The dictionary information can be a dictionary containing additional word information, such as multiple interpretations of polysemous words, internal components of words, word parts of speech, etc. When processing "Bank of China", the dictionary information may contain information about "China" and "Bank" as subwords, as well as multiple possible semantic interpretations of "Bank of China". It should be noted that this is only an example, and there is no specific restriction on the content of the dictionary information.

[0092] Optionally, in the Lattice structure, not only the basic word sequence is considered, but also a lattice is constructed through the dictionary information, in which each node can represent a word or a part of a word (such as a character in a word). This lattice contains all possible word and character combinations, providing a rich source of information for subsequent feature extraction. Furthermore, by inputting the word embedding vector into the Lattice structure, at least one enhanced feature representation can be obtained, which can include additional features extracted from the input word sequence and dictionary information.

[0093] In this step, the character sequence is labeled to obtain the labeled character sequence, which is then converted into a word embedding vector and combined with the dictionary information to filter the highly correlated related data. This can enhance the model's understanding ability, accurately locate entities, and improve the recognition rate. This is especially effective when dealing with long sentences and polysemy problems, which optimizes the accuracy and efficiency of knowledge extraction in the banking field.

[0094] As an optional implementation, context information of a character sequence in data to be processed is determined based on associated data, including: integrating the associated data and word embedding vectors to obtain integrated data; and obtaining context data of the integrated data.

[0095] In this embodiment, the associated data and the word embedding vector can be integrated to obtain integrated data, and the integrated data can be transmitted to the BiLSTM layer, so that the context information of the character sequence in the data to be processed can be determined. The above integrated data can be a vector set including the word embedding vector and the associated data, and can be a word embedding sequence.

[0096] Optionally, the input of the BiLSTM layer may be integrated data obtained by integrating word embeddings and associated data. The integrated data may be a word embedding sequence, which may include semantic information of the word and possible additional features, such as word frequency, part of speech, etc.

[0097] For example, for an input text sequence, such as "Bank of China and China Merchants Bank conducted a transaction of 100 million", each word (or character) will be converted into a vector to form a vector sequence, which can contain at least one word embedding vector, such as [vec(China), vec(Bank), vec(with), vec(China Merchants Bank), vec(conduct)...] At the same time, the Lattice structure can be used to obtain the associated data corresponding to each word or character, which can come from dictionary information, such as the frequency of occurrence and word meaning in the dictionary. The above information can be integrated into the word vector, and once the integrated data is obtained, the integrated data can be used as an additional input of BiLSTM.

[0098] Optionally, the output of the BiLSTM layer can be a serialized vector, where each vector corresponds to a word or character in the input sequence. Each output vector contains the semantic and contextual information of the word in the context, which is crucial for subsequent sequence labeling tasks (for example, the CRF layer in named entity recognition). Specifically:

[0099] Optionally, the BiLSTM layer can generate two vectors for each word in the input sequence, representing the flow of information from left to right (forward LSTM) and from right to left (backward LSTM). These two vectors are typically merged into a fixed-length vector to form a complete contextual representation of the word. Sequence labeling: Although the direct output of the BiLSTM layer is a serialized vector, these vectors are further processed by subsequent layers (such as the CRF layer) to generate sequence labels for each word. In named entity recognition, this means assigning a label to each word, such as B-Person (beginning of a person entity), I-Organization (inside an organization entity), O (non-entity), etc.

[0100] For example, assuming that the input text sequence is "Wang moved from Guangzhou to Shenzhen", each word is converted into a word embedding vector: Input: [vec(Wang), vec(from), vec(Guangzhou), vec(moving), vec(to), vec(Shenzhen)]. BiLSTM output: After being processed by the BiLSTM layer, the representation of each word will contain the contextual information of the previous and next words, forming a serialized vector output, such as [vec'(Wang), vec'(from), vec'(Guangzhou), vec'(moving), vec'(to), vec'(Shenzhen)], where vec' can represent a vector containing contextual information. Further, for the named entity recognition task, each output vector of the BiLSTM layer can be used by the subsequent CRF layer or Softmax layer to generate the corresponding entity label: Output: [B-Person, O, B-Location, O, O, I-Location]. It should be noted that the above data processing process is only for illustrative purposes and is not specifically limited here.

[0101] Optionally, the Lattice structure layer of word feature information is combined with the dictionary information and matched with all the character sequence information obtained to obtain related data. Further, the BiLSTM neural network layer is used to process the context information for modeling and training. The input sentence T can be expressed as T = {b1, b2, b3...b n}, where b i It can be used to represent the i-th character in a sentence, with a total of n character vectors. i,j It can be used to represent words, i can be used to represent the starting position of a word, and j can be used to represent the ending position of a word. The calculation formula of the BiLSTM layer word sequence can be:

[0102]

[0103]

[0104] in, and It can be the forget gate set input into the BiLSTM neural network. wt Yes w The model parameters of . σ represents the normalization function (sigmoid function), and tanh is the output activation function.

[0105] Optionally, by calculating the gate information of the BiLSTM model, the weight of the current character and vocabulary in the current sentence text can be obtained according to the annotation features of the current sequence. The vocabulary weight calculation formula is as follows:

[0106]

[0107] Finally, update the state of the character vector at index position j. The state update formula is as follows:

[0108]

[0109] In this step, the integration of associated data and word embedding vectors enhances the model's understanding of entity semantics, and the acquisition of contextual data further improves the accuracy of the model's entity recognition and context perception. This not only solves the problem of fuzzy entity boundaries and polysemy, but also improves the model's performance in processing long sentences and complex contexts, thereby optimizing the efficiency and quality of knowledge extraction.

[0110] As an optional implementation, step S204, based on the associated data and context information, determines the label information of the character sequence, including: based on the associated data and context information, identifying the character sequence to obtain multiple initial label information of the character sequence; and determining the label information from the multiple initial label information.

[0111] In this embodiment, the character sequence can be identified based on the associated data and context information to obtain multiple initial label information of the character sequence. The final label information can be determined from the multiple initial label information, that is, the text label can be output for information sequence annotation and segmentation, and the CRF layer that globally normalizes the data information features can obtain the optimal solution for the label sequence.

[0112] Optionally, the character sequence can be identified based on the associated data and context information to obtain multiple initial label information, and label information among the multiple initial label information can be determined based on the probability of the initial label information, and the label information can be the initial label information with the highest probability.

[0113] Optionally, the input of the CRF layer can be a sequence output by the previous neural network layer (such as a BiLSTM layer), and each sequence element can be a vector. The sequence can contain associated data and context information, which can characterize the feature representation of the word or character at the corresponding position in the context. The associated data and context are processed by the CRF layer to obtain a serialized label, which corresponds to each word or character in the input sequence. These label information can be selected from a predefined label set and can be used to identify the category of the word or character. For example, in named entity recognition, the label can be B-Person (beginning of a person's name), I-Person (inside a person's name), O (non-entity), etc.

[0114] Optionally, the CRF layer determines the optimal label configuration for the entire sequence through global optimization to ensure that the label sequence is not only locally reasonable but also consistent throughout the sequence, which is particularly important for handling the boundary recognition of entities in sequence data.

[0115] For example, assume we have an input sequence, which is the output from the BiLSTM layer and represents the word vector features in the sentence "Li works in Guangzhou": Input: \(\mathbf{h} = [\mathbf{h}_{Li Ming}, \mathbf{h}_{in Guangzhou}, \mathbf{h}_{works}]\), where each \(\mathbf{h}\) vector contains the semantic and context information of the word. The output of the CRF layer can assign a label to each word, and this label can indicate whether it belongs to an entity and the type of the entity: Output: \(\mathbf{y} = [B - Person, B - Location, O]\).

[0116] Optionally, in the above - mentioned way, the CRF layer can output an optimal label sequence based on the overall context information of the sequence to identify entities and other structural information in the input text, thus playing a key role in text understanding and information extraction tasks.

[0117] In this embodiment, information sequence annotation and segmentation can be performed on the output text labels. The CRF layer that globally normalizes the data information features obtains the optimal solution of the label sequence. For a given sentence \(x\), the probability calculation formula for its label sequence \(y\) is as follows:

[0118]

[0119] where, \(\sum\) y′ \(\exp(\text{score}(x, y'))\) represents the sum of the exponents of the scores for each token sequence. Assuming the total length of the sequence is \(i\) and the total number of tokens is \(k\), the total number of sequences can be expressed as \(k\) i ; where \(y\) represents the true label value.

[0120] Among them, the score function calculation formula is as follows:

[0121]

[0122] Among them, \(p(i, j)\) can be used to represent the emission probability that the LSTM model output assigns \(x\) i to the label \(y\) i ; \(q(y\) i-1 , \(y\) i ) can be used to represent the transition probability of the transition matrix of the CRF. \(y_0\) and \(y\) nIndicates the starting and ending character positions of a sentence.

[0123] Alternatively, BiLSTM-CRF can be trained using the maximum likelihood method, and the corresponding loss function can be expressed as follows:

[0124] -log(p(y|x))=-score(x,y)+logZ(x)

[0125] Where F = log Z (x) = log ∑ y exp(score(x,y)).

[0126] As an optional implementation, determining label information from multiple initial label information includes: retrieving a conditional random field model to determine evaluation information of the initial label information; and determining label information from multiple initial label information based on the evaluation information.

[0127] In this embodiment, the conditional random field model can also be called to determine the evaluation information of the initial label information. The evaluation information can be the probability corresponding to the initial label information. Based on the evaluation information, the final label information can be determined from multiple initial label information.

[0128] Optionally, this embodiment can collect text data related to the bank, including customer information, transaction records, internal documents, etc., which may contain entities that need to be identified, such as asset names, organization names, and personal names. The acquired data to be processed can be cleaned to remove meaningless or interfering information such as HTML tags, non-text data, and null values. Word segmentation: Use a word segmentation tool suitable for bank corpus to segment the text and divide the long text into words or phrases. Data annotation: Perform BIO annotation on the data after word segmentation, where B indicates the start of the entity, I indicates the inside of the entity, and O indicates a non-entity. For example, for the name "Wang Mou", it is labeled as B-Dname, I-Dname, I-Dname, indicating that this is the beginning and continuation of a name entity. Furthermore, the Word2Vec model can be used to vectorize the words and convert them into feature vectors that can be input into a neural network. Lattice structure: Construct a Lattice structure to combine characters and dictionary information to form a network structure, in which each node can be a word, a character, or a sub-word in the dictionary. BiLSTM layer: uses a bidirectional long short-term memory network to process the sequence output by the Lattice structure. It can process information from two directions (forward and backward) and capture features in the context. CRF layer: The conditional random field layer is used for sequence labeling to ensure that the label configuration of the entire sequence conforms to the rules and context of entity recognition and select the optimal label sequence.

[0129] Optionally, a sequence embedded with Word2Vec words and character and sub-word information contained in the Lattice structure is input. Lattice-BiLSTM layer processing: The Lattice layer obtains character and dictionary information, selects the most relevant features to the context through a gating mechanism for integration, and obtains integrated data. The BiLSTM layer processes the integrated sequence (i.e., integrated data), learns context information and long-term dependencies, and obtains context information. CRF layer: The output of the BiLSTM layer is used as input, and the optimal label sequence is selected through global sequence optimization to perform entity recognition, so that the final entity extraction result can be output, which can include the identified entity type and boundary. The label sequence output by the model can also be normalized to ensure that the output entity information format is consistent, which is convenient for subsequent analysis and application. The extracted entity information can be used for subsequent data analysis, customer information management, or risk assessment and other internal bank applications.

[0130] For example, for entity recognition of loan application documents, the following data to be processed can be obtained from the loan application documents: "Wang Mou applied for a mortgage of 1 million yuan at the bank's Guangzhou branch." Data preprocessing is performed on the data to be processed, which may include data cleaning, word segmentation and data annotation. Among them, data cleaning may include: removing HTML tags, non-text characters, etc. in the text to ensure that only pure text content is retained. The text is segmented into words to obtain the sequence after word segmentation: "Wang Mou", "at", "bank", "Guangzhou branch", "apply", "a", "a", "1 million yuan", "of", "mortgage". For BIO annotation, the sequence after word segmentation is annotated for the purpose of entity recognition. For example, "Wang Mou" is annotated as B-Dname, I-Dname, I-Dname, indicating that this is the beginning and continuation of a person name entity; "Bank" can be annotated as B-Org, I-Org, indicating that this is the beginning and continuation of an organization entity; "1 million yuan" is annotated as B-Asset, I-Asset, indicating that this is the beginning and continuation of an asset entity.

[0131] Furthermore, the word sequence after word segmentation is converted into a word embedding vector. At the same time, the additional word or sub-word information in the Lattice structure will also be converted into the corresponding vector representation. The above word vector is input into the BiLSTM layer for processing. The BiLSTM layer can select the words and character information most relevant to the context through the gating mechanism. For example, when processing "Wang Mou", the Lattice structure may consider "Wang" and "Mou" as character information at the same time, and "Wang Mou" as a complete name entity. BiLSTM will learn the association between this information and the context, such as the relationship between "Wang Mou" and "Bank", to identify "Wang Mou" as a name entity. The CRF layer can be used for the final sequence labeling to ensure that the entity recognition result of the entire sentence is globally optimal and avoid the trap of local optimality. For example, ensure that "Wang Mou" and "Bank" are correctly identified as name and organization name entities, rather than splitting them into discontinuous entities. The identified entity information is further processed and normalized. For example, an entity such as "Bank Guangzhou Branch" that includes subordinate units is unified into "Bank" or "Bank Guangzhou Branch" in a specific format to facilitate subsequent business analysis or data integration. Finally, the output results can be obtained, which may include: Name entity: "Wang Mou" Organization entity: "Bank", "Guangzhou Branch" Asset entity: "1 million yuan".

[0132] In this step, the above steps fully utilize the advantages of neural networks and statistical models, combined with the characteristics of bank data, and effectively improve the accuracy and efficiency of entity extraction through the construction and training of multi-layer models. From data cleaning to model training, to sequence labeling and normalization processing, each step is aimed at optimizing the performance of entity recognition and ensuring that the final output entity information is accurate and meets actual needs.

[0133] In an embodiment of the present application, a configured interface protocol and an interface protocol to be configured are obtained, difference information between the two is determined, and the difference information is converted using the configuration data of the target interface to obtain the configuration data. The configured interface protocol can be updated according to the configuration data to obtain the interface protocol to be configured, thereby achieving the technical effect of improving the accuracy of text parsing and solving the technical problem of low accuracy in text parsing.

[0134] Another optional specific implementation is described in detail below.

[0135] At present, each operation of the rule-based entity extraction method is that the system performs semantic analysis on the text according to the established rules. However, these rule-based solutions have great limitations, require huge manpower and material resources, and are not scalable enough.

[0136] In related technologies, one difficulty in knowledge extraction is that the feature information of sequence-annotated characters and words is difficult to use effectively, and there are difficulties in solving and identifying the problem of polysemy. It is impossible to make dynamic adjustments according to the specific context, and the flexibility is insufficient. It is difficult to extract all the semantic information of a long sentence, and there is a possibility of semantic loss.

[0137] With the application and development of neural network technology in artificial intelligence technology, most of the neural network models that can be widely used in banking systems are static models. Although they are relatively versatile to a certain extent, they cannot be dynamically adjusted according to the specific context, lack flexibility, and have difficulties in solving and identifying polysemy problems. It is difficult to extract all the semantic information of a sentence from a long sentence, and there is a situation of semantic loss. In addition, the data of the banking system is significantly different from the laboratory data and has more obvious characteristics.

[0138] To solve the above problems, in this embodiment, a method for entity extraction based on the Lattice dictionary is proposed. The method proposes a neural network knowledge extraction model that can fully utilize the characteristics of bank data. Through this model, when dealing with knowledge extraction of data with obvious characteristics such as bank data, better recognition rate and extraction effect can be achieved.

[0139] Optionally, the model can use the Lattice word model in the knowledge extraction stage, making full use of character and subsequence information, and obtaining related characters and words through the selection of gating structures, so as to obtain better entity information extraction results. The output data is then used as the input of the BiLSTM-CRF model for sequence identification and normalization. The excellent word information collection capability of the Lattice structure makes the model effect much better than the traditional extraction model.

[0140] Next, the entity extraction method based on the Lattice dictionary in this application is further introduced.

[0141] In this embodiment, an entity extraction model is formed by integrating the Lattice structure based on the bank data dictionary with a bidirectional long short-term memory model whose output layer is a conditional random field. A bidirectional long short-term memory entity extraction model based on the Lattice structure (Word2vec-Lattice-BiLSTM-CRF, referred to as WLBC) is proposed. Through bank data, the Lattice structure based on the dictionary obtains more word information in the input layer, and then uses the output data as the input of the BiLSTM-CRF model for sequence identification and normalization. The excellent word information collection capability of the Lattice structure and the contextual processing capability of the BiLSTM model that takes into account data information make the model effect much better than the traditional knowledge extraction model.

[0142] Optionally, the method can be used to process data from a banking system, where the data obtained consists of some bank customer data information. Data preprocessing includes data cleaning, word segmentation and data annotation. Data cleaning is performed in the acquired data to remove possible meaningless values, including HTML tags, non-text data, null values, etc. The data annotation method used in this application is BIO annotation, which is widely used and can transform the annotation problem into the simplest problem. The characters are annotated with B, I, and O, respectively, and different letter annotations represent different meanings. A character is annotated with BX, indicating that the character is at the beginning of the text and that the character belongs to type X. If a character is annotated with IX, it means that the character is in the middle of the text and that the character belongs to type X. If a character is annotated with O, it means that the character does not belong to any type. In this application, the three major categories of asset names, organization names, and personal names are used as the main entity recognition types.

[0143] Optionally, the text data after entity annotation can be used for vectorization processing of subsequent models, which helps to achieve better word segmentation effects.

[0144] Figure 3 is a flow chart of an entity extraction method according to an embodiment of the present application. Figure 3 As shown, the method may include the following steps:

[0145] Step S301, obtaining data to be processed.

[0146] In this embodiment, bank-related text data is collected to obtain data to be processed, which may include customer information, transaction records, internal documents, etc. The above-mentioned data to be processed may contain entities that need to be identified, such as asset names, organization names, and personal names.

[0147] Step S302: pre-process the data to be processed.

[0148] In this embodiment, the acquired data to be processed can be cleaned to remove meaningless or interfering information such as HTML tags, non-text data, and null values. Furthermore, a word segmentation tool suitable for bank corpus can be used to segment the text to divide the long text into words or phrases. The segmented data is annotated with BIO data.

[0149] Step S303, processing the preprocessed data to obtain output data.

[0150] In this embodiment, the Word2Vec model is used to vectorize words and convert them into feature vectors that can be input into a neural network.

[0151] Optionally, constructing the Lattice structure may include: combining characters and dictionary information to form a network structure, wherein each node may be a word, a character, or a sub-word in the dictionary.

[0152] Optionally, the BiLSTM layer can process the sequence output by the Lattice structure using a bidirectional long short-term memory network, which can process information from two directions (forward and backward) to capture features in the context.

[0153] Optionally, the CRF layer can be used for sequence labeling to ensure that the label configuration of the entire sequence conforms to the rules and context of entity recognition and select the optimal label sequence.

[0154] In this embodiment, a sequence embedded with Word2Vec words is input, as well as character and sub-word information contained in the Lattice structure. Lattice-BiLSTM layer processing may include: the Lattice layer obtains character and dictionary information, and selects the most relevant features to the context for integration through a gating mechanism. The BiLSTM layer processes the integrated sequence to learn context information and long-term dependencies. The CRF layer can use the output of the BiLSTM layer as input, select the optimal label sequence through global sequence optimization, and perform entity recognition, so that the final entity extraction result can be output, including the recognized entity type and boundary.

[0155] Step S304, normalizing the output data.

[0156] In this embodiment, the label sequence output by the model can be normalized to ensure that the output entity information format is consistent, which is convenient for subsequent analysis and application.

[0157] Optionally, the entity extraction process is completed and the extracted entity information is used for subsequent internal bank applications such as data analysis, customer information management or risk assessment.

[0158] In this embodiment, the above method fully utilizes the advantages of neural networks and statistical models, combines the characteristics of bank data, and effectively improves the accuracy and efficiency of entity extraction through the construction and training of multi-layer models. From data cleaning to model training, to sequence labeling and normalization processing, each step is aimed at optimizing the performance of entity recognition and ensuring that the final output entity information is accurate and meets actual needs.

[0159] Figure 4 is a schematic diagram of an entity extraction model according to an embodiment of the present application, such as Figure 4As shown, the entity extraction model may be a Lattice-BiLSTM-CRF model, which may include: a Lattice dictionary model 401, a BiLSTM model 402, a CRF model 403, an input layer 404 and a Softmax layer 405.

[0160] In this embodiment, the input layer (Input Layer) may be the bottom layer of the model, which receives preprocessed text data and may include character sequences and word segmentation information.

[0161] In this embodiment, the above-mentioned Lattice dictionary model 401 can be deployed in the Lattice dictionary model layer (Lattice Dictionary Layer), where the input layer data is fused with the pre-built Lattice dictionary information, and the Lattice structure is matched through the dictionary information to obtain richer information between characters and words (that is, association information), which may include possible entity boundaries and types. The output of this layer is a word embedding vector that integrates the dictionary information, which is used to capture the semantics and contextual associations of the word.

[0162] In this embodiment, after receiving the output of the Lattice dictionary model layer, the BiLSTM layer can use its bidirectional characteristics to comprehensively consider the previous and next contexts of the input sequence and further extract the contextual features of each word or character, which helps to solve the problems of polysemy and long sentence understanding.

[0163] In this embodiment, CRF is used as the top layer of the model. The CRF layer can be used to make final label predictions for the sequence output by BiLSTM to ensure the consistency and coherence of the entire sequence. The CRF layer selects the most likely entity label sequence through global optimization, thereby avoiding the trap of local optimal solutions and improving the accuracy of entity recognition.

[0164] In this embodiment, the Softmax layer can be used to normalize the data.

[0165] In this embodiment, new technologies such as knowledge graphs are fully utilized, and the advantages of bank data foundation are combined. The bank's characteristic data is used in Lattice word structure, which can effectively obtain character and word sequence information, and obtain related characters and words through the selection of gated structures, which also improves the knowledge extraction effect of the model. By combining a bidirectional long short-term memory network, the effect of connecting context is better, thereby obtaining more information features, and more effectively utilizing context and word information, which improves the entity recognition effect of the model.

[0166] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0167] Example 2

[0168] The embodiment of the present application also provides a device for determining tag information. It should be noted that the device for determining tag information in the embodiment of the present application can be used to execute the method for determining tag information provided in the embodiment of the present application. The device for determining tag information provided in the embodiment of the present application is introduced below.

[0169] According to an embodiment of the present application, a device for implementing the above-mentioned method for determining tag information is also provided. Figure 5 is a schematic diagram of a device for determining tag information according to an embodiment of the present application, such as Figure 5 As shown, the device 500 for determining tag information includes: a first acquiring unit 501 , a splitting unit 502 , a second acquiring unit 503 , a determining unit 504 and an identifying unit 505 .

[0170] The first acquisition unit 501 is used to acquire data to be processed.

[0171] The splitting unit 502 is used to split the data to be processed into character sequences.

[0172] The second acquisition unit 503 is used to acquire the associated data whose association degree with the character sequence is greater than the association degree threshold.

[0173] The determining unit 504 is configured to determine context information of the character sequence in the data to be processed based on the associated data.

[0174] The identification unit 505 is used to determine the label information of the character sequence based on the associated data and the context information.

[0175] The apparatus for determining label information provided in the embodiment of the present application obtains data to be processed through a first acquisition unit; splits the data to be processed into character sequences through a splitting unit; obtains associated data having an association degree with the character sequence greater than an association degree threshold through a second acquisition unit; determines context information of the character sequence in the data to be processed based on the associated data through a determination unit; and determines label information of the character sequence based on the associated data and the context information through an identification unit, thereby achieving the technical effect of improving the accuracy of text parsing and solving the technical problem of low accuracy of text parsing.

[0176] It should be noted that the first acquisition unit 501, the splitting unit 502, the second acquisition unit 503, the determination unit 504 and the identification unit 505 correspond to steps S201 to S205 in Example 1, and the five modules are the same as the examples and application scenarios implemented by the corresponding steps, but are not limited to the contents disclosed in the above-mentioned Example 1. It should be noted that the above-mentioned modules or units can be hardware components or software components stored in a memory (e.g., memory 104) and processed by one or more processors (e.g., processors 102a, 102b, ..., 102n), and the above-mentioned modules can also be part of the device and can be run in the computer terminal 10 provided in Example 1.

[0177] Example 3

[0178] An embodiment of the present application may provide an electronic device, Figure 6 is a structural block diagram of an electronic device according to an embodiment of the present application. Figure 6 As shown, the electronic device may include: one or more ( Figure 6 (only one is shown) processor 602, memory 604, storage controller, and peripheral interface, wherein the peripheral interface is connected to the radio frequency module, audio module and display.

[0179] Among them, the memory can be used to store software programs and modules, such as program instructions / modules corresponding to the methods and devices in the embodiments of the present application, and the processor executes various functional applications and data processing by running the software programs and modules stored in the memory, that is, realizing the above-mentioned method. The memory may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include a memory remotely arranged relative to the processor, and these remote memories may be connected to the terminal via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0180] The processor can call the information and application stored in the memory through the transmission device to perform the following steps: obtain the data to be processed; split the data to be processed into character sequences; obtain associated data whose association with the character sequence is greater than an association threshold; based on the associated data, determine the context information of the character sequence in the data to be processed; based on the associated data and the context information, determine the label information of the character sequence.

[0181] The processor can also call the information and application programs stored in the memory through the transmission device to perform the following steps: extracting text data from the data to be processed; splitting the text data to obtain a character sequence.

[0182] The processor can also call the information and application stored in the memory through the transmission device to perform the following steps: labeling the character sequence to obtain a labeled character sequence; converting the labeled character sequence into a word embedding vector; using dictionary information to determine associated data whose association with the word embedding vector is greater than an association threshold, wherein the dictionary information is used to provide additional semantics and / or characters to the word embedding vector.

[0183] The processor can also call the information and application programs stored in the memory through the transmission device to perform the following steps: integrate the associated data and the word embedding vector to obtain integrated data; and obtain context data of the integrated data.

[0184] The processor can also call the information and application stored in the memory through the transmission device to perform the following steps: based on the associated data and context information, identify the character sequence to obtain multiple initial label information of the character sequence; determine the label information from the multiple initial label information.

[0185] The processor can also call the information and application programs stored in the memory through the transmission device to perform the following steps: call the conditional random field model to determine the evaluation information of the initial label information; based on the evaluation information, determine the label information from multiple initial label information.

[0186] By adopting the embodiment of the present application, a method for determining label information is provided, by obtaining data to be processed, splitting the data to be processed, and obtaining multiple character sequences, associated data with a correlation degree greater than an association degree threshold value with the character sequence can be obtained, and based on the associated data, the context information of the character sequence in the data to be processed can be accurately determined, and based on the associated data and the context information, the meaning of the character sequence can be determined, so that the label information corresponding to the character sequence can be accurately determined, thereby achieving the technical effect of improving the accuracy of text parsing and solving the technical problem of low accuracy in text parsing.

[0187] It can be understood by those skilled in the art that Figure 6 The structure shown is for illustration only, and the electronic device may also be a terminal device such as a smart phone (such as an Android phone, an iOS phone, etc.), a tablet computer, a PDA, a mobile Internet device (Mobile Internet Devices, abbreviated as MID), a personal access device (Personal Access Device, abbreviated as PAD), etc. Figure 6 The structure of the electronic device is not limited. Figure 6 More or fewer components (such as network interfaces, display devices, etc.) shown in, or having Figure 6 Different configurations are shown.

[0188] A person of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing the hardware related to the terminal device through a program, and the program can be stored in a computer-readable storage medium, and the storage medium may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk, etc.

[0189] Example 4

[0190] The embodiment of the present application further provides a storage medium. Optionally, in this embodiment, the storage medium can be used to store the program code executed by the tag information determination method provided in the first embodiment.

[0191] Optionally, in this embodiment, the above storage medium may be located in any computer terminal in a computer terminal group in a computer network, or in any mobile terminal in a mobile terminal group.

[0192] The present application also provides a computer program product, which, when executed on a data processing device, is suitable for executing the steps of the method for determining tag information.

[0193] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0194] In the above embodiments of the present application, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0195] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic, for example, the division of units is only a logical function division, and there may be other division methods in actual implementation, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0196] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0197] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0198] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art or all or part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, server or network device, etc.) to perform all or part of the steps of each embodiment method of the present application. The aforementioned storage medium includes: U disk, read-only memory (Read-Only Memory, referred to as ROM), random access memory (Random Access Memory, referred to as RAM), mobile hard disk, disk or optical disk and other media that can store program codes.

[0199] The above are only preferred implementations of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A method for determining tag information, characterized in that: include: Get the data to be processed; Splitting the data to be processed into character sequences; Acquire association data whose association degree with the character sequence is greater than an association degree threshold; Based on the associated data, determining context information of the character sequence in the data to be processed; Based on the association data and the context information, label information of the character sequence is determined.

2. The method according to claim 1, characterized in that The step of splitting the data to be processed into character sequences comprises: Extracting text data from the data to be processed; The text data is split to obtain the character sequence.

3. The method according to claim 1, characterized in that: The acquiring of the associated data having an association degree with the character sequence greater than an association degree threshold comprises: Marking the character sequence to obtain a marked character sequence; Convert the labeled character sequence into a word embedding vector; The associated data having an association degree with the word embedding vector greater than the association degree threshold is determined by using the dictionary information, wherein the dictionary information is used to provide additional semantics and / or characters to the word embedding vector.

4. The method according to claim 3, characterized in that The determining, based on the associated data, context information of the character sequence in the data to be processed includes: Integrate the associated data and the word embedding vector to obtain integrated data; The context data of the integrated data is obtained.

5. The method according to claim 1, characterized in that The determining the label information of the character sequence based on the associated data and the context information includes: Based on the associated data and the context information, the character sequence is identified to obtain a plurality of initial label information of the character sequence; The tag information is determined from a plurality of initial tag information.

6. The method according to claim 5, characterized in that The step of determining the label information from the plurality of initial label information includes: Retrieving a conditional random field model to determine evaluation information of the initial label information; Based on the evaluation information, the label information is determined from a plurality of initial label information.

7. A device for determining tag information, characterized in that: include: A first acquisition unit, used to acquire data to be processed; A splitting unit, used for splitting the data to be processed into character sequences; A second acquisition unit, configured to acquire associated data having an association degree greater than an association degree threshold with the character sequence; A determination unit, configured to determine context information of the character sequence in the data to be processed based on the associated data; The recognition unit is used to determine the label information of the character sequence based on the associated data and the context information.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored executable program, wherein when the executable program is executed, the device where the computer-readable storage medium is located is controlled to execute the method according to any one of claims 1 to 6.

9. An electronic device, characterized in that: include: A memory storing an executable program; A processor, configured to run the program, wherein the program executes the method according to any one of claims 1 to 6 when running.

10. A computer program product comprising computer instructions, characterized in that: When the computer instructions are executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images