Bond Information Parsing Method, Device and Electronic Device Based on Deep Learning Algorithm

Abstract

The present invention discloses a bond information parsing method, device and electronic device based on a deep learning algorithm. The method includes: pre-constructing a deep learning algorithm model; obtaining a target domain data set and constructing input features according to a preset annotation guide; sequentially performing coarse-grained pre-training and fine-grained pre-training on the encoder in the deep learning algorithm model according to the input features to generate a target encoder; setting the decoder of the deep learning algorithm to generate a target decoder; generating a target deep learning algorithm model according to the target encoder and the target decoder; training the target deep learning algorithm model according to the input features to generate a bond parsing model; inputting the bond information to be parsed into the bond parsing model and outputting a parsing result. The embodiments of the present invention realize named entity recognition for financial bond transactions, improve the overall accuracy in tasks rich in numerical data, effectively solve the problem of entity nesting in named entity recognition, and improve the parsing efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular, to a method, device, and electronic device for parsing bond information based on a deep learning algorithm. Background Art

[0002] In text processing, a common requirement is to extract valuable information from a piece of text. For example, in the requirement for booking a hotel, key information such as location and time needs to be extracted from unstructured text information. The same requirement also exists in the field of financial bonds, that is, to extract valuable information from unstructured text information.

[0003] In the field of natural language processing, named entity recognition is a relatively mature sequence labeling task, which is a process of predicting entities with specific meanings in an input sentence, such as bond names, bond codes, and institution names. In existing deep learning algorithms, relatively classic techniques include sequence labeling models such as LSTM+CRF, Bert+CRF, and Bert+BiLSTM+CRF.

[0004] Existing pre-trained models are built on constructing language models, and the goal is to learn general text representations. The disadvantage is that they ignore the rich knowledge in named entity recognition. For example, the pre-training process of the Bert model is to randomly mask some words in a sentence and then predict these words. This method cannot well integrate entity information and boundary information in named entities into the model, resulting in poor parsing effects.

[0005] Therefore, the existing technology still needs to be improved and developed. Summary of the Invention

[0006] In view of the above deficiencies of the existing technology, the present invention provides a method, device, and electronic device for parsing bond information based on a deep learning algorithm, aiming to solve the problem that the pre-trained model of the bond parsing method based on the deep learning algorithm in the existing technology cannot integrate entity information and boundary information in named entities into the model, resulting in poor parsing effects.

[0007] The technical solution of the present invention is as follows:

[0008] The first embodiment of the present invention provides a method for parsing bond information based on a deep learning algorithm, the method comprising:

[0009] Pre-construct a deep learning algorithm model; the deep learning algorithm model includes an encoder and a decoder;

[0010] Obtain a target domain data set, and construct input features corresponding to the target domain data set according to a preset annotation guide;

[0011] Perform coarse-grained pre-training and fine-grained pre-training on the encoder in the deep learning algorithm model in sequence according to the input features to generate a target encoder;

[0012] Set the decoder of the deep learning algorithm to generate a target decoder, and the target decoder uses a pointer network;

[0013] Generate a target deep learning algorithm model according to the target encoder and the target decoder;

[0014] Train the target deep learning algorithm model according to the input features to generate a bond parsing model;

[0015] Obtain the bond information to be parsed, input the bond information into the bond parsing model, and output the parsing result. Further, the deep learning algorithm model is pre-constructed; the deep learning algorithm model includes an encoder and a decoder, including:

[0016] Pre-construct the encoder of the deep learning algorithm model, and the encoder is a BERT model;

[0017] Pre-construct a decoder, and the decoder is used to decode the encoding;

[0018] Generate a deep learning algorithm model according to the encoder and the decoder.

[0019] Further, the target domain dataset is obtained, and according to the preset annotation guidelines, the input features corresponding to the target domain dataset are constructed, including:

[0020] Obtain the annotation guidelines corresponding to the bond data;

[0021] Obtain the target domain dataset, and generate input features composed of annotation guidelines and input data according to the preset annotation guidelines and the target domain dataset.

[0022] Further, the coarse-grained pre-training of the encoder in the deep learning algorithm model according to the input features includes:

[0023] Obtain a large number of public datasets to generate weakly supervised data;

[0024] Pre-train the encoder of the deep learning algorithm according to the input features, and supervise the encoder with the weakly supervised data to generate a first encoder, and the first encoder is used to judge the entities and non-entities in the input sentence.

[0025] Further, the fine-grained pre-training of the encoder in the deep learning algorithm model according to the input features includes:

[0026] Obtain a named entity dictionary in the financial field, scan the input features according to the named entity dictionary, and automatically annotate according to the maximum matching algorithm;

[0027] Through an automatic screening strategy, after each round of training, the trained model is used to re-predict the dataset, and the labels with high confidence are re-annotated;

[0028] Use the remotely supervised data for the next round of data training. After the training is completed, the target encoder is output.

[0029] Furthermore, the decoder of the deep learning algorithm is set to generate a target decoder. The target decoder adopts a pointer network and includes:

[0030] The decoder of the deep learning algorithm is set, and two softmax multi-classifiers are used for classification, which are respectively denoted as the first softmax multi-classifier and the second softmax multi-classifier. The first softmax multi-classifier is used to label the start of the entity, and the second softmax multi-classifier is used to label the end of the entity;

[0031] The loss function is defined as the sum of the cross-entropies of the two softmax multi-classifiers to generate the target decoder.

[0032] Furthermore, the bond information to be parsed is obtained, and the bond information is input into the bond parsing model, and the parsing result is output, including:

[0033] Obtain the bond information to be parsed, and input the bond information into the encoder of the bond parsing model;

[0034] The encoder performs text feature encoding on the bond information to be parsed to generate a character vector with a fixed dimension;

[0035] The character vector is input into the decoder of the bond parsing model, and the decoded parsing result is output.

[0036] Another embodiment of the present invention provides a bond information parsing device based on a deep learning algorithm. The device includes:

[0037] A model construction module for pre-constructing a deep learning algorithm model; the deep learning algorithm model includes an encoder and a decoder;

[0038] A data processing module for obtaining a target domain dataset and constructing input features corresponding to the target domain dataset according to a preset annotation guide;

[0039] A pre-training module for sequentially performing coarse-grained pre-training and fine-grained pre-training on the encoder in the deep learning algorithm model according to the input features to generate a target encoder;

[0040] A decoder setting module for setting a decoder of a deep learning algorithm to generate a target decoder, where the target decoder uses a pointer network;

[0041] A target model generation module for generating a target deep learning algorithm model according to the target encoder and the target decoder;

[0042] A training module for training the target deep learning algorithm model according to input features to generate a bond parsing model;

[0043] An analysis module for obtaining bond information to be analyzed, inputting the bond information into the bond parsing model, and outputting an analysis result.

[0044] Another embodiment of the present invention provides an electronic device, which includes at least one processor; and,

[0045] A memory communicatively connected to the at least one processor; wherein,

[0046] The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above-mentioned bond information parsing method based on a deep learning algorithm.

[0047] Another embodiment of the present invention further provides a non-volatile computer-readable storage medium, which stores computer-executable instructions, and when the computer-executable instructions are executed by one or more processors, the one or more processors can execute the above-mentioned bond information parsing method based on a deep learning algorithm.

[0048] Advantageous effects: The embodiments of the present invention can implement named entity recognition for financial bond transactions, improve the overall accuracy in tasks rich in numerical data; effectively solve the problem of entity nesting in named entity recognition, and improve the parsing efficiency. Description of the Drawings

[0049] The present invention will be further described below in conjunction with the drawings and embodiments. In the drawings:

[0050] Figure 1 Is a flowchart of a preferred embodiment of a bond information parsing method based on a deep learning algorithm of the present invention;

[0051] Figure 2 Is a schematic diagram of the functional modules of a preferred embodiment of a bond information parsing device based on a deep learning algorithm of the present invention;

[0052] Figure 3 Is a schematic diagram of the hardware structure of a preferred embodiment of an electronic device of the present invention. Detailed Embodiments

[0053] To make the objectives, technical solutions, and effects of the present invention clearer and more explicit, the present invention will be further described in detail below. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0054] The embodiments of the present invention will be introduced below with reference to the accompanying drawings.

[0055] In the prior art, in the decoder part, the currently popular CRF module does not take into account the meaning of each entity category itself. Especially in the vertical field of finance, each entity type in the bond parsing module has a relatively detailed annotation guide. Traditional named entity recognition devices such as the BERT+CRF structure cannot utilize the annotation documents of bond data. The annotation documents contain rich text knowledge, such as explanations of entities, which can guide the model to better learn the underlying features of entities; the CRF decoder device is difficult to process entity nesting. It is difficult to accurately identify ambiguous entities.

[0056] In view of the above problems, the embodiments of the present invention provide a bond information parsing method based on a deep learning algorithm. Please refer to Figure 1 , Figure 1 which is a flowchart of a preferred embodiment of a bond information parsing method based on a deep learning algorithm of the present invention. As Figure 1 shown, it includes:

[0057] Step S100: Pre-build a deep learning algorithm model; the deep learning algorithm model includes an encoder and a decoder;

[0058] Step S200: Obtain a target domain data set, and construct input features corresponding to the target domain data set according to a preset annotation guide;

[0059] Step S300: Perform coarse-grained pre-training and fine-grained pre-training on the encoder in the deep learning algorithm model in sequence according to the input features to generate a target encoder;

[0060] Step S400: Set the decoder of the deep learning algorithm to generate a target decoder, and the target decoder uses a pointer network;

[0061] Step S500: Generate a target deep learning algorithm model according to the target encoder and the target decoder;

[0062] Step S600: Train the target deep learning algorithm model according to the input features to generate a bond parsing model;

[0063] Step S700: Obtain the bond information to be parsed, input the bond information into the bond parsing model, and output the parsing result.

[0064] In specific implementation, the model in the embodiments of the present invention is used to identify named entities. Named entity recognition is an encoder-decoder model in terms of the model. Specifically, the encoder learns semantic representations, and the decoder learns downstream tasks such as classification. Based on the data characteristics in the financial field on the basis of the basic end-to-end deep learning framework, this solution introduces a reading comprehension mechanism of <annotation guide, input text>, pre-trains the data in bond transactions on a large-scale corpus to enhance semantic representations, and then inputs the prior knowledge (i.e., the annotation guide) into the encoder part of the model to guide the neural network to learn deeper structural relationships. Finally, a pointer network is added to the decoding layer to solve the problem of entity nesting. The whole solution maintains an end-to-end learning form and is an algorithm model solution with strong interpretability and generalization ability.

[0065] The goal of pre-training is to train the model on a public-domain dataset to obtain good semantic representations for better learning of private-domain data. Since it is difficult to obtain relevant datasets for financial bonds and pre-trained models have high requirements for the amount of data, the common traditional method of directly pre-training based on a large amount of text is not feasible. To improve the above deficiencies and better integrate entity information into the pre-trained model, this solution adopts a two-stage pre-trained model of "coarse-grained" and "fine-grained" optimized for named entity recognition.

[0066] By learning a large amount of data, a better semantic representation H is obtained. In this step, we will fine-tune on business data to obtain the final bond parsing model. Construct the features of <question sentence, input text> according to the annotation document as the input of the model; the encoder part adopts the encoder part of the pre-trained model; the decoder part adopts a pointer network, that is, 2 Softmax classifiers are used to predict the start and end of the entity respectively.

[0067] The embodiments of the present invention are mainly used for financial bond data and can also be used for semi-structured text information in other fields.

[0068] In one embodiment, a deep learning algorithm model is pre-constructed; the deep learning algorithm model includes an encoder and a decoder, including:

[0069] Pre-construct the encoder of the deep learning algorithm model, and the encoder is a BERT model;

[0070] Pre-construct a decoder, and the decoder is used to decode the encoding;

[0071] Generate a deep learning algorithm model according to the encoder and the decoder.

[0072] In specific implementation, an encoder of a deep learning algorithm model is pre-constructed, and the encoder is a BERT model; a decoder is pre-constructed, and the decoder is used to decode the encoding; a deep learning algorithm model is generated according to the encoder and the decoder. BERT is an acronym for "Bidirectional Encoder Representations from Transformers". As a whole, it is an autoencoder language model (Autoencoder LM), and two tasks are designed to pre-train the model. The first task is to train the language model in the way of MaskLM. Generally speaking, when inputting a sentence, some words to be predicted are randomly selected and then replaced with a special symbol [MASK]. Then, the model is allowed to learn the words to be filled in these places according to the given labels. The second task adds an additional sentence-level continuity prediction task on the basis of the bidirectional language model, that is, predicting whether two texts input to BERT are continuous texts. Introducing this task can better enable the model to learn the relationship between continuous text segments.

[0073] In some other embodiments, other Bert variant models can be adopted, or LSTM and CNN encoders can be adopted for speed requirements.

[0074] In one embodiment, a target domain data set is obtained, and according to a preset annotation guide, input features corresponding to the target domain data set are constructed, including:

[0075] An annotation guide corresponding to the bond data is obtained;

[0076] A target domain data set is obtained, and input features composed of the annotation guide and the input data are generated according to the preset annotation guide and the target domain data set.

[0077] In specific implementation, the annotation guide refers to the instruction document used by annotators for data annotation. For example, for the entity "price", the corresponding annotation guide is that "it is commonly a digital string, and the integer part is commonly two or three digits, and the decimal part is commonly two, three, or four digits. In order to make full use of the entity description information in the annotation guide, the input layer is constructed by splicing the questions constructed from the annotation guide and the input sentence, that is, " <cls>Question sentence <sep>Input sentence "。"

[0078] In one embodiment, a coarse-grained pre-training is performed on an encoder in a deep learning algorithm model according to input features, including:

[0079] Obtaining a large number of publicly available datasets to generate weakly supervised data;

[0080] Pre-training the encoder of the deep learning algorithm according to the input features, and using the weakly supervised data to supervise the encoder to generate a first encoder, where the first encoder is used to determine entities and non-entities in the input sentence.

[0081] In specific implementation, first, a coarse-grained training is performed on the model. The goal of this stage is to enable the model to better learn entity boundaries. Since the Wikipedia document dataset is large and there is a natural weak correspondence relationship (anchor text) such as <category, entity>, a large number of publicly available datasets are collected in this stage to generate weakly supervised data, and then a reading comprehension parsing device is used to warm up the model. The goal of prediction is to find entity boundaries and enable the model to learn to determine entities and non-entities in the input sentence, so the construction of the question is simplified to finding entities. The construction of the input of the model is as follows: <cls>Find entity <sep>Input sentence.

[0082] In one embodiment, fine-grained pre-training of the encoder in the deep learning algorithm model according to the input features includes:

[0083] Obtain the named entity dictionary in the financial field, scan the input features according to the named entity dictionary, and automatically annotate according to the maximum matching algorithm;

[0084] Through the automatic screening strategy, after each round of training, use the trained model to re-predict the data set and re-annotate the labels with high confidence;

[0085] Use the remotely supervised data for the next round of data training. After the training is completed, output the target encoder.

[0086] In specific implementation, the purpose of fine-grained named entity pre-training is to enable the model to better learn the underlying features related to entities.

[0087] This part mainly uses the knowledge of the named entity dictionary related to finance to generate label data and conduct training. On the basis of the previous step, we scan the entities in the data set according to the named entity dictionary and automatically annotate according to the maximum matching algorithm. The specific working principle is as follows: Assume that the longest word in the dictionary has i Chinese characters, then use the first i characters in the current string of the processed document as the matching field to search the dictionary. If such an i-word exists in the dictionary, the matching is successful, and the matching field is segmented as a word. If such an i-word cannot be found in the dictionary, the matching fails, remove the last character from the matching field, and re-match the remaining string. Proceed in this way until the matching is successful, that is, a word is segmented or the length of the remaining string is zero. In this way, one round of matching is completed, and then the next i-character string is taken for matching processing until the document is scanned.

[0088] However, the data set generated in this way will have a large amount of noise, such as missing labels and boundary errors. To solve such problems, this solution proposes a self-picking (automatic screening) strategy. After each round of training, use the trained model to re-predict the data set and re-annotate the labels with high confidence. Specifically, define a threshold σ. When p start , p end are both greater than σ, label text[start:end] as the corresponding entity, and use this new data set for the next round of training. Then use the remotely supervised data for the next round of training. For the vertical field of financial spot bonds, construct questions according to the annotation guidelines for each entity type, and use the reading comprehension parsing device for training.

[0089] In one embodiment, the decoder of the deep learning algorithm is set to generate a target decoder. The target decoder uses a pointer network and includes:

[0090] The decoder of the deep learning algorithm is set. Two softmax multi-classifiers are used for classification, denoted as the first softmax multi-classifier and the second softmax multi-classifier respectively. The first softmax multi-classifier is used to label the start of the entity, and the second softmax multi-classifier is used to label the end of the entity.

[0091] The loss function is defined as the sum of the cross-entropies of the two softmax multi-classifiers to generate the target decoder.

[0092] In specific implementation, bond texts have characteristics such as high complexity and strong ambiguity. Different from the common CRF decoder module, in order to effectively solve the entity nesting problem, this solution uses two softmax multi-classifiers, one to label the start of the entity and one to label the end of the entity. The part between the start and end positions is the entity:

[0093]

[0094] The final loss function is defined as the sum of the cross-entropies of the two softmax multi-classifiers:

[0095]

[0096] The target decoder is generated according to the definition result.

[0097] In one embodiment, the bond information to be parsed is obtained, and the bond information is input into the bond parsing model to output the parsing result, including:

[0098] The bond information to be parsed is obtained, and the bond information is input into the encoder of the bond parsing model.

[0099] The encoder performs text feature encoding on the bond information to be parsed to generate a character vector with a fixed dimension.

[0100] The character vector is input into the decoder of the bond parsing model to output the decoded parsing result.

[0101] In specific implementation, the bond information to be parsed is obtained. The bond information can be bond-related text in a chat tool. The bond information is input into the encoder of the bond parsing model based on the Bert model. The encoder based on the Bert model inputs the character vector into the decoder of the bond parsing model to output the decoded parsing result in bond format.

[0102] In the embodiments of the present invention, a reading comprehension framework is integrated into a named entity resolution model for learning; multi-granularity pre-training is performed through a large amount of data, and then the obtained model is migrated to specific services; named entity recognition for financial bond transactions is achieved, and the overall accuracy rate in tasks rich in numerical data is increased by more than 2%-5%; the problem of entity nesting in named entity recognition can also be effectively solved, and the parsing efficiency is improved.

[0103] It should be noted that there is not necessarily a certain order among the above steps. Those of ordinary skill in the art can understand according to the description of the embodiments of the present invention that in different embodiments, the above steps can have different execution orders, that is, they can be executed in parallel or exchanged, etc.

[0104] Another embodiment of the present invention provides a bond information parsing device based on a deep learning algorithm, as Figure 2 shown. The device 1 includes:

[0105] A model construction module 11 for pre-constructing a deep learning algorithm model; the deep learning algorithm model includes an encoder and a decoder;

[0106] A data processing module 12 for obtaining a target domain data set and constructing input features corresponding to the target domain data set according to a preset annotation guide;

[0107] A pre-training module 13 for sequentially performing coarse-grained pre-training and fine-grained pre-training on the encoder in the deep learning algorithm model according to the input features to generate a target encoder;

[0108] A decoder setting module 14 for setting the decoder of the deep learning algorithm to generate a target decoder, and the target decoder adopts a pointer network;

[0109] A target model generation module 15 for generating a target deep learning algorithm model according to the target encoder and the target decoder;

[0110] A training module 16 for training the target deep learning algorithm model according to the input features to generate a bond parsing model;

[0111] An analysis module 17 for obtaining the bond information to be analyzed, inputting the bond information into the bond parsing model, and outputting an analysis result.

[0112] For the specific implementation manner, refer to the method embodiments, which will not be elaborated here.

[0113] Another embodiment of the present invention provides an electronic device, as Figure 3 shown. The electronic device 10 includes:

[0114] One or more processors 110 and a memory 120, Figure 3 Taking a processor 110 as an example for introduction, the processor 110 and the memory 120 can be connected through a bus or other means. Figure 3 Taking the connection through a bus as an example.

[0115] The processor 110 is used to complete various control logics of the electronic device 10. It can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), a single-chip microcomputer, an ARM (Acorn RISC Machine), or other programmable logic devices, discrete gate or transistor logic, discrete hardware controls, or any combination of these components. Additionally, the processor 110 can also be any conventional processor, microprocessor, or state machine. The processor 110 can also be implemented as a combination of computing devices, for example, a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors combined with a DSP core, or any other such configuration.

[0116] The memory 120, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the program instructions corresponding to the bond information parsing method based on the deep learning algorithm in the embodiments of the present invention. The processor 110 executes various functional applications and data processing of the device 10 by running the non-volatile software programs, instructions, and units stored in the memory 120, that is, implementing the bond information parsing method based on the deep learning algorithm in the above method embodiments.

[0117] The memory 120 can include a program storage area and a data storage area. Among them, the program storage area can store an operating device and application programs required for at least one function; the data storage area can store data created according to the use of the device 10, etc. In addition, the memory 120 can include high-speed random access memory, and can also include non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other non-volatile solid-state storage devices. In some embodiments, the memory 120 can optionally include a memory remotely set relative to the processor 110, and these remote memories can be connected to the device 10 through a network. Examples of the above network include but are not limited to the Internet, an enterprise intranet, a local area network, a mobile communication network, and their combinations.

[0118] One or more units are stored in the memory 120 and, when executed by one or more processors 110, execute the bond information parsing method based on the deep learning algorithm in any of the above method embodiments. For example, execute the Figure 1 method steps S100 to step S700 described above.

[0119] An embodiment of the present invention provides a non-volatile computer-readable storage medium. The computer-readable storage medium stores computer-executable instructions, which are executed by one or more processors. For example, the method steps S100 to S700 described above are executed. Figure 1 in the above.

[0120] As an example, the non-volatile storage medium can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable ROM (EEPROM), or flash memory. The volatile memory can include random access memory (RAM) as an external cache memory. By way of illustration and not limitation, RAM can be obtained in many forms such as synchronous RAM (SRAM), dynamic RAM, (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), Synchlink DRAM (SLDRAM), and direct Rambus (Rambus) RAM (DRRAM). The disclosed memory controls or memories of the operating environment described herein are intended to include one or more of these and / or any other suitable types of memories.

[0121] Another embodiment of the present invention provides a computer program product. The computer program product includes a computer program stored on a non-volatile computer-readable storage medium. The computer program includes program instructions that, when executed by a processor, cause the processor to execute the bond information parsing method based on a deep learning algorithm in the above method embodiment. For example, the method steps S100 to S700 described above are executed. Figure 1 in the above.

[0122] The embodiments described above are merely illustrative. 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 to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0123] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a general hardware platform, and of course can also be implemented by hardware. Based on such an understanding, the above technical solution, in essence, or the part that contributes to the related technology can be embodied in the form of a software product, and this computer software product can exist in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of each embodiment or some parts of the embodiments.

[0124] Among other things, conditional language such as "able to", "can", "may", or "could", unless specifically stated otherwise or otherwise understood within the context in which it is used, generally is intended to convey that a particular embodiment can include (while other embodiments do not include) a particular feature, element, and / or operation. Thus, such conditional language generally is also intended to imply that the feature, element, and / or operation is in some way required for one or more embodiments or that one or more embodiments must include logic for deciding whether or not the feature, element, and / or operation is included or will be performed in any particular embodiment, with or without input or prompting.

[0125] What has been described in this specification and the drawings herein includes examples of a bond information parsing method and apparatus capable of providing a deep learning algorithm. Of course, it is not possible to describe every conceivable combination of elements and / or methods for the purpose of describing the various features of the present disclosure, but it can be recognized that many additional combinations and permutations of the disclosed features are possible. Thus, it is apparent that various modifications can be made to the present disclosure without departing from the scope or spirit thereof. Additionally, or in the alternative, other embodiments of the present disclosure may be apparent from consideration of the specification and drawings and from practice of the present disclosure as presented herein. The intention is that the examples presented in this specification and the drawings be considered illustrative in all respects and not restrictive. Although specific terms are employed herein, they are used in a generic and descriptive sense and not for purposes of limitation.< / sep> < / cls> < / sep> < / cls>

Claims

1. A method for parsing bond information based on deep learning algorithms, characterized in that , the method includes: Pre - construct a deep learning algorithm model; the deep learning algorithm model includes an encoder and a decoder; Obtain a target domain data set, and construct input features corresponding to the target domain data set according to a preset annotation guide; Perform coarse - grained pre - training and fine - grained pre - training on the encoder in the deep learning algorithm model in sequence according to the input features to generate a target encoder; Set the decoder of the deep learning algorithm to generate a target decoder, and the target decoder uses a pointer network; Generate a target deep learning algorithm model according to the target encoder and the target decoder; Train the target deep learning algorithm model according to the input features to generate a bond parsing model; Obtain the bond information to be parsed, input the bond information into the bond parsing model, and output the parsing result; The coarse - grained pre - training of the encoder in the deep learning algorithm model according to the input features includes: Obtain a large number of public data sets to generate weakly - supervised data; Pre - train the encoder of the deep learning algorithm according to the input features, and use the weakly - supervised data to supervise the encoder to generate a first encoder, and the first encoder is used to judge entities and non - entities in the input sentence; The fine - grained pre - training of the encoder in the deep learning algorithm model according to the input features includes: Obtain a named - entity dictionary in the financial field, scan the input features according to the named - entity dictionary, and automatically annotate according to the maximum - matching algorithm; Through an automatic screening strategy, use the trained model to re - predict the data set after each round of training, and re - annotate the labels with high confidence; Use remotely - supervised data for the next round of data training, and after training is completed, output the target encoder.

2. The method according to claim 1, characterized in that, The pre - construction of the deep learning algorithm model; the deep learning algorithm model includes an encoder and a decoder, includes: Pre - construct the encoder of the deep learning algorithm model, and the encoder is a BERT model; Pre - construct a decoder, and the decoder is used to decode the encoding; Generate a deep learning algorithm model according to the encoder and the decoder.

3. The method according to claim 2, characterized in that, The obtaining of the target domain data set and the construction of the input features corresponding to the target domain data set according to the preset annotation guide includes: Obtain the annotation guide corresponding to the bond data; Obtain the target domain data set, and generate input features composed of the annotation guide and input data according to the preset annotation guide and the target domain data set.

4. The method according to claim 1, characterized in that, The setting of the decoder of the deep learning algorithm to generate a target decoder, and the target decoder uses a pointer network, includes: Set the decoder of the deep learning algorithm, and use two softmax multi - classifiers for classification, which are respectively denoted as the first softmax multi - classifier and the second softmax multi - classifier. The first softmax multi - classifier is used to label the start of the entity, and the second softmax multi - classifier is used to label the end of the entity; Define the loss function as the sum of the cross-entropies of two softmax multi-classifiers, and generate a target decoder.

5. The method according to claim 1, wherein, the obtaining of the bond information to be parsed, inputting the bond information into the bond parsing model, and outputting the parsing result includes: obtaining the bond information to be parsed, and inputting the bond information into the encoder of the bond parsing model; the encoder performs text feature encoding on the bond information to be parsed, and generates a character vector of a fixed dimension; inputting the character vector into the decoder of the bond parsing model, and outputting the decoded parsing result.

6. A bond information parsing device based on a deep learning algorithm, wherein, the device includes: a model construction module for pre-constructing a deep learning algorithm model; the deep learning algorithm model includes an encoder and a decoder; a data processing module for obtaining a target domain data set and constructing input features corresponding to the target domain data set according to a preset annotation guide; a pre-training module for sequentially performing coarse-grained pre-training and fine-grained pre-training on the encoder in the deep learning algorithm model according to the input features, and generating a target encoder; a decoder setting module for setting the decoder of the deep learning algorithm to generate a target decoder, and the target decoder adopts a pointer network; a target model generation module for generating a target deep learning algorithm model according to the target encoder and the target decoder; a training module for training the target deep learning algorithm model according to the input features to generate a bond parsing model; a parsing module for obtaining the bond information to be parsed, inputting the bond information into the bond parsing model, and outputting the parsing result; the performing of the coarse-grained pre-training on the encoder in the deep learning algorithm model according to the input features includes: obtaining a large number of public data sets to generate weakly supervised data; performing pre-training on the encoder of the deep learning algorithm according to the input features, and using the weakly supervised data to supervise the encoder to generate a first encoder, and the first encoder is used to judge entities and non-entities in the input sentence; the performing of the fine-grained pre-training on the encoder in the deep learning algorithm model according to the input features includes: obtaining a named entity dictionary in the financial field, scanning the input features according to the named entity dictionary, and automatically annotating according to the maximum matching algorithm; through an automatic screening strategy, using the trained model to re-predict the data set after each round of training, and re-annotating the labels with high confidence; using the remotely supervised data for the next round of data training, and after the training is completed, outputting the target encoder.

7. An electronic device, wherein, the electronic device includes at least one processor; and, a memory communicatively connected to the at least one processor; wherein, the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the bond information parsing method based on the deep learning algorithm according to any one of claims 1-5.

8. A non-volatile computer-readable storage medium, wherein, The non-volatile computer-readable storage medium stores computer-executable instructions that, when executed by one or more processors, enable the one or more processors to execute the bond information parsing method based on a deep learning algorithm according to any one of claims 1-5.

Images