WO2020082613A1 - 利用深度学习模型的证券研报核心观点提取方法及装置 - Google Patents

利用深度学习模型的证券研报核心观点提取方法及装置 Download PDF

Info

Publication number
WO2020082613A1
WO2020082613A1 PCT/CN2019/070292 CN2019070292W WO2020082613A1 WO 2020082613 A1 WO2020082613 A1 WO 2020082613A1 CN 2019070292 W CN2019070292 W CN 2019070292W WO 2020082613 A1 WO2020082613 A1 WO 2020082613A1
Authority
WO
WIPO (PCT)
Prior art keywords
clause
research report
analyzed
securities research
opinion
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2019/070292
Other languages
English (en)
French (fr)
Inventor
叶曙峰
梁桂贤
吴冶成
黄鸿顺
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Ping An Technology Shenzhen Co Ltd
Original Assignee
Ping An Technology Shenzhen Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Ping An Technology Shenzhen Co Ltd filed Critical Ping An Technology Shenzhen Co Ltd
Publication of WO2020082613A1 publication Critical patent/WO2020082613A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00Handling natural language data
    • G06F40/20Natural language analysis
    • G06F40/279Recognition of textual entities
    • G06F40/289Phrasal analysis, e.g. finite state techniques or chunking
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00Handling natural language data
    • G06F40/30Semantic analysis
    • G06F40/35Discourse or dialogue representation
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/049Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods

Definitions

  • This application relates to the field of computer technology, and in particular to a method and device for extracting core opinions of a securities research report using a deep learning model.
  • Securities research report also referred to as securities research report, refers to the research report made by relevant researchers (such as researchers in a securities company, etc.) to analyze the value of securities and related products, or factors that affect their market prices. .
  • the purpose of this application is to provide a method, device, computer equipment and readable storage medium for extracting the core opinions of a securities research report using a deep learning model, so as to overcome the problems in the prior art to a certain extent, which can be efficient and accurate Extracts the core opinions in the securities research report.
  • a method for extracting core opinions of a securities research report using a deep learning model including the following steps:
  • the present application also provides a device for extracting core opinions of a securities research report using a deep learning model, including:
  • the receiving module is used to receive the input securities research report to be analyzed
  • Sentence module which is used to perform sentence processing on the securities research report to be analyzed to obtain each clause in the securities research report to be analyzed;
  • the first scoring module is used to score each clause in the securities research report to be analyzed using a pre-trained first deep learning model to obtain a first score for each clause in the securities research report to be analyzed, and Determine opinion clauses from each clause in the securities research report to be analyzed based on the first score;
  • the second scoring module is used to score each opinion clause in the securities research report to be analyzed using a pre-trained second deep learning model to obtain the first sentence for each opinion category clause in the securities research report to be analyzed Second score, and based on the second score, determine the industry opinion clause from each opinion clause in the securities research report to be analyzed;
  • the extraction module is used to extract the second industry opinion clause with the highest score as the core opinion of the securities research report to be analyzed and output it.
  • the present application also provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor.
  • a deep learning model is used. The following steps of the method for extracting the core opinions of the Securities Research Report:
  • the present application also provides a computer-readable storage medium on which a computer program is stored, and when the program is executed by a processor, the following steps of a method for extracting core opinions of a securities research report using a deep learning model are implemented:
  • the method, device, computer equipment and readable storage medium for extracting the core opinions of the securities research report using the deep learning model provided in this application can first perform the sentence processing of the securities research report to be analyzed to obtain each clause, and then adopt each clause
  • the pre-trained first deep learning model scores to obtain the first score of each clause, and judges whether the clause is an opinion clause according to the first score, and then adopts the pre-trained opinion clause
  • the second deep learning model scores to obtain the second score of each opinion clause, and determines whether each opinion clause is an industry opinion clause according to the second score. After the industry opinion clause is judged, it can be extracted
  • the second industry opinion clause with the highest score is output as the core opinion in the securities research report to be analyzed.
  • a pre-trained deep learning model can be used to select opinion clauses and industry opinion clauses through intelligent scoring, and finally extract objective representative scores from industry opinion clauses that can represent the whole The clause of the core opinion of the securities research report. Therefore, the above process of intelligently scoring and analyzing using the deep learning model can not only greatly save manpower, but also improve the operation efficiency and the accuracy of extracting results.
  • FIG. 1 is a schematic flowchart of an alternative method for extracting core opinions of a securities research report using a deep learning model according to an embodiment of the present application;
  • FIG. 2 is a schematic diagram of an optional program module of an apparatus for extracting core opinions of a securities research report using a deep learning model according to an embodiment of the present application;
  • FIG. 3 is a schematic diagram of another optional program module of an apparatus for extracting core opinions of a securities research report using a deep learning model according to an embodiment of the present application;
  • FIG. 4 is a schematic diagram of an optional hardware architecture of a computer device according to an embodiment of the present application.
  • FIG. 1 is a schematic flowchart of an alternative method for extracting core opinions of a securities research report using a deep learning model, as shown in FIG. 1, the method may include the following steps:
  • a securities research report can be used as a unit to extract its core opinions.
  • the format of the securities research report can be judged first. Specifically, it can be determined whether the format of the securities research report is a text type, such as ".txt" format, ".doc” format, etc. If the judgment result is no, that is, the format of the securities research report is not a text type, the format of the securities research report can be converted to a text type, for example, the format of the securities research report is a portable file format (PDF format), then You can use the existing "PDFParser" and other tools to convert the PDF format of the securities research report into a txt format of the securities research report. In this way, the format of the received securities research report can be unified to read the content of the securities research report to be analyzed more conveniently and efficiently.
  • PDF format portable file format
  • the securities research report to be analyzed can be segmented according to a preset type of symbol, such as comma ",”, period “.”, Period “.”, Dash “-”, square bracket “ [], [] “, Semicolon”; "and other symbols, to deal with the analysis of the securities research report clauses, in order to get the analysis of the securities research report clauses.
  • the jieba word segmentation module may be further used to perform word segmentation processing on each clause to obtain each word segment in the securities research report to be analyzed.
  • the value corresponding to each word segment in the securities research report to be analyzed can be determined based on a preset dictionary that includes the corresponding conversion relationship between the word segment and the value (in this embodiment: Integer value).
  • each clause composed of multiple participles in the securities research report to be analyzed is converted into a clause in a numeric vector format.
  • the clauses in the numerical vector format can be input into the deep learning model, so that the deep learning model can score each clause and can judge the type of the clause according to the score.
  • the training process may include the following steps:
  • Step 110 Determine the first data set.
  • the first type tag may include opinion class and non-opinion class. That is to say, after being labeled by the first type label, the clauses in the first data set may include opinion clauses and non-opinion clauses.
  • Step 120 Perform segmentation processing on each clause labeled with the first type label to obtain each segmentation labeled with the first type label.
  • the existing jieba word segmentation module may be used to perform word segmentation processing on each clause labeled with the first type label in the first data set to obtain each word segment labeled with the first type label.
  • Step 130 Convert each participle marked with the first type label to a corresponding first value according to a preset rule, and store each participle marked with the first type label and its corresponding first value in a preset including participle and The dictionary of the corresponding conversion relationship of numerical values.
  • each participle marked with a label of the first type can be converted into a numeric value one by one according to preset rules (such as random conversion, or the conversion rule of words to numeric values can be preset according to actual needs, etc.) (in this embodiment , May be referred to as a first value), and may store each participle labeled with a label of the first type and the first value in a preset dictionary including the corresponding conversion relationship between the participle and the value.
  • preset rules such as random conversion, or the conversion rule of words to numeric values can be preset according to actual needs, etc.
  • each participle marked with the first type label and the first value can be stored in the dictionary in the form of a file or a data table, and the content can be shown in Table 1.
  • each clause labeled with the label of the first type is converted into a clause in a numeric vector format to form a first numeric vector clause set.
  • each clause composed of a plurality of word segments labeled with the first type label can be converted into a clause in a numeric vector format, and a numeric vector segment can be formed
  • the sentence set in this embodiment, may be referred to as a first numerical vector clause set.
  • Step 150 Select a first preset number of clauses from the first numerical vector clause set as the first training data.
  • the training set for training and the test set for testing can be determined.
  • a preset number such as Clauses with a ratio of more than 90%
  • the remaining clauses such as 10%
  • Step 160 Train the first training data through the deep learning model to obtain the first deep learning model.
  • the deep learning model may be a long-short-term memory network (LSTM, Long Short-Term Memory) machine learning model.
  • LSTM Long Short-Term Memory
  • some key parameters may be set, such as embedding size (embedding size), Hidden layer size (hidden layer size), batch training size (batch size), full sample number of cycles (num epochs), forgetting rate (dropout), activation function (activation, using sigmoid for variable types), loss function (loss, Use binary-crossentropy) and other parameters for different types of problems.
  • the first training data can be input into the LSTM machine learning model for training to obtain the first deep learning model after the training is completed.
  • the first test data can also be input into the first deep learning model for testing to obtain the accuracy of the first deep learning model.
  • our first The accuracy of the deep learning model on the first test data can reach about 85%.
  • the accuracy rate is lower than the predetermined standard (such as 70%)
  • the trained first depth model can score the sentence more accurately and determine the sentence type.
  • each clause in the securities research report to be analyzed can be scored using the above-mentioned first deep learning model to obtain a first score for each clause in the securities research report to be analyzed, and based on the first score Analyze the opinion clauses in each clause in the Securities Research Report.
  • the first score may be a score considered by the model, and the score may be regarded as a probability (between 0 and 1). For example, if the model thinks that the clause that is more biased towards the opinion category, the score is closer to 1 (such as 0.75), the model believes that the clauses that are more biased towards non-opinion categories, the score is closer to 0 (such as 0.22).
  • the first score can be compared with the preset first score threshold, and the opinion clause can be judged according to the comparison result, for example, the sentence with the first score greater than the first score threshold can be judged
  • the first score threshold can be set to 0.5, that is to say, clauses with a first score not greater than 0.5 can be determined as non-opinion clauses, and no longer participate in subsequent steps; the first score is greater than Clauses of 0.5 can be determined as opinion clauses, and you can continue to participate in subsequent steps.
  • the pre-trained first deep learning model is used to determine the type of clauses (sentence clauses or non-opinion clauses) in an intelligent scoring manner. On the one hand, it can greatly save manpower and improve judgment efficiency ; On the other hand, it can improve the accuracy of the judgment results.
  • the training process may include the following steps:
  • Step 210 Determine the second data set.
  • the abstract part of the securities research report can be sentence-processed to obtain multiple clauses, and then each clause can be manually labeled with the first type of label, the first Type tags may include opinion and non-opinion categories. That is to say, after being labeled by the first type label, the clauses in the second data set may include opinion clauses and non-opinion clauses.
  • each opinion class clause with a second type label, which may include an industry opinion class and a non-industry opinion class. That is to say, after being labeled by the second type of label, the clauses in the second data set may include opinion clauses and non-opinion clauses, and the opinion clauses also include industry opinion clauses and non-industry opinions Category clause.
  • Step 220 Perform segmentation processing on each clause labeled with the second type label to obtain each segmentation labeled with the second type label.
  • the clauses with the second type label in the second data set are used for word segmentation processing, for example, the existing The jieba word segmentation module performs word segmentation processing on the above clauses to obtain each word segment marked with a second type label.
  • Step 230 Convert each participle marked with the second type label to a corresponding second value according to a preset rule, and store each participle marked with the second type label and its corresponding second value in the preset including participle and The dictionary of the corresponding conversion relationship of numerical values.
  • each participle marked with a second type label can be converted into a numeric value one by one according to preset rules (such as random conversion, or a word-to-value conversion rule can be preset according to actual needs) (in this embodiment , May be referred to as a second value), and the conversion result of each participle labeled with the second type label and the second value may be stored in a preset dictionary including the corresponding conversion relationship between the participle and the value.
  • preset rules such as random conversion, or a word-to-value conversion rule can be preset according to actual needs
  • each participle marked with the second type label and the second value can be stored in the dictionary in the form of a file or a data table, and the content can be as shown in Table 1 above.
  • each clause labeled with the label of the second type is converted into a clause in a numeric vector format to form a second numeric vector clause set.
  • each clause composed of a plurality of word segments labeled with a second type label can be converted into a clause in a numeric vector format, and a numeric vector segment can be formed
  • the sentence set in this embodiment, may be referred to as a second numerical vector clause set.
  • Step 250 Select a second preset number of clauses from the second numerical vector clause set as the second training data.
  • the training set for training and the test set for testing can be determined before training the deep learning model.
  • a preset number such as Clauses with a ratio of more than 90%
  • the remaining clauses such as 10%
  • Step 260 Train the second training data through the deep learning model to obtain the second deep learning model.
  • the deep learning model may be a long-short-term memory network (LSTM, Long Short-Term Memory) machine learning model.
  • LSTM Long Short-Term Memory
  • some key parameters may be set, such as embedding size (embedding size), Hidden layer size (hidden layer size), batch training size (batch size), full sample number of cycles (num epochs), forgetting rate (dropout), activation function (activation, using sigmoid for variable types), loss function (loss, Use binary-crossentropy) and other parameters for different types of problems.
  • the second training data can be input to the LSTM machine learning model for training to obtain the second deep learning model after the training is completed.
  • the second test data may also be input into the second deep learning model for testing to obtain the accuracy rate of the second deep learning model.
  • the accuracy of our second deep learning on the second test data can reach about 85%.
  • the accuracy rate is lower than the predetermined standard (such as 70%)
  • the trained second depth model can score the sentence more accurately and determine the sentence type.
  • the opinion clauses in the securities research report to be analyzed determined in S03 can be scored using the above-mentioned second deep learning model to obtain a second score for each opinion category clause in the securities research report to be analyzed, Based on the second score, the industry opinion clause in the securities research report to be analyzed can be determined from the opinion clause.
  • the second score may be a score considered by the model, and the score may be regarded as a probability (between 0 and 1). For example, if the model considers a clause that is more biased towards industry opinions, the score is closer to At 1 (such as 0.85), the model believes that the clauses that are more biased towards non-industry opinion categories, the score is closer to 0 (such as 0.3).
  • the second score can be compared with a preset second score threshold, and the opinion clause can be determined according to the comparison result.
  • a clause with a second score greater than the second score threshold can be determined as
  • clauses whose second score is not greater than the second score threshold are determined as non-industry opinion clauses.
  • the second score threshold can be set to 0.5, that is to say, clauses with a second score not greater than 0.5 can be determined as non-industry opinion clauses, and no longer participate in subsequent steps; second score Clauses greater than 0.5 can be determined as industry opinion clauses and can continue to participate in subsequent steps.
  • the second score is a score given by the second deep learning model according to whether the clause is biased towards the industry opinion category or the non-industry opinion category, and the more biased it is towards the industry opinion category
  • the higher the second score of the clause ie, the closer it is to 1
  • the industry with the highest second score The opinion clause is regarded as the most important and core industry opinion.
  • the second score may be sorted (for example, sorted in a positive order), and the industry opinion category clause with the second highest score (for example, sorted first) may be used as the securities research to be analyzed.
  • the core opinions of the newspaper are output.
  • the securities research report to be analyzed can be processed into clauses to obtain each clause, and then each clause can be scored using a pre-trained first deep learning model to obtain the first clause of each clause.
  • the industry opinion category clause with the second highest score can be extracted as the securities research report to be analyzed
  • the core point of view is output.
  • a pre-trained deep learning model can be used to select opinion clauses and industry opinion clauses through intelligent scoring, and finally extract objective representative scores from industry opinion clauses that can represent the whole The clause of the core opinion of the securities research report. Therefore, the above process of intelligently scoring and analyzing using the deep learning model can not only greatly save manpower, but also improve the operation efficiency and the accuracy of extracting results.
  • FIGS. 2 to 3 An optional structural block diagram of the device for extracting core opinions of a securities research report using a deep learning model is divided into one or more program modules, and one or more program modules It is stored in a storage medium and executed by one or more processors to complete the application.
  • the program module referred to in this application refers to a series of computer program instruction segments that can perform specific functions. It is more suitable than the program itself to describe the execution process of the securities research report core viewpoint extraction device using the deep learning model in the storage medium. The following description will The functions of the program modules in this embodiment are specifically described.
  • the device 20 for extracting core opinions of a securities research report using a deep learning model may include:
  • the receiving module 21 can be used to receive the input securities research report to be analyzed
  • Sentence module 22 can be used to perform sentence processing on the securities research report to be analyzed to obtain each clause in the securities research report to be analyzed;
  • the first scoring module 23 can be used to score each clause in the securities research report to be analyzed using a pre-trained first deep learning model to obtain a first score for each clause in the securities research report to be analyzed, Based on the first score, the opinion clause is determined from the clauses in the securities research report to be analyzed;
  • the second scoring module 24 can be used to score each viewpoint class in the securities research report to be analyzed using a pre-trained second deep learning model to obtain a score for each opinion class in the securities research report to be analyzed The second score, and based on the second score, determine the industry opinion clause from each opinion clause in the securities research report to be analyzed;
  • the extraction module 25 can be used to extract the second industry opinion clause with the highest score as the core opinion of the securities research report to be analyzed for output.
  • the receiving module 21 may be specifically used for:
  • the format of the securities research report to be analyzed is converted to a text type.
  • clause module 22 may specifically include:
  • Sentence unit 221 is used to perform sentence processing on the securities research report to be analyzed according to preset symbols, to obtain each clause in the securities research report to be analyzed;
  • the word segmentation unit 222 is used to perform word segmentation processing on each clause in the securities research report to be analyzed to obtain each word segment in the securities research report to be analyzed;
  • the numerical value determining unit 223 is configured to determine the numerical value corresponding to each partial word in the securities research report to be analyzed based on a preset dictionary including the corresponding conversion relationship between the partial word and the numerical value;
  • the conversion unit 224 is configured to convert each clause in the securities research report to be analyzed into a clause in a numeric vector format according to the determination result.
  • the training process of the first deep learning model includes the following steps:
  • Step 110 Determine a first data set, where the first data set includes a plurality of clauses obtained by pre-sentencing the abstract section of the securities research report of a preset number of articles, each of which has a pre-marked
  • the first type label, the first type label includes opinion class and non-opinion class;
  • Step 120 Perform segmentation processing on each clause labeled with the first type label to obtain each segmentation labeled with the first type label;
  • Step 130 Convert each participle marked with the first type label to a corresponding first value according to a preset rule, and store each participle marked with the first type label and its corresponding first value in a preset including participle and The dictionary of the corresponding conversion relationship of numerical values;
  • Step 140 According to the dictionary, convert each clause marked with the first type label into a clause in a numeric vector format to form a first numeric vector clause set;
  • Step 150 Select a first preset number of clauses from the first numerical vector clause set as the first training data
  • Step 160 Train the first training data through the deep learning model to obtain the first deep learning model.
  • the training process of the second deep learning model includes the following steps:
  • Step 210 Determine a second data set, where the second data set includes a plurality of clauses obtained by pre-sentencing the abstract section of the securities research report of a preset number of articles, each of which has a pre-marked
  • the first type of label, the first type of label includes opinion class and non-opinion class, and the clause marked with the opinion class tag also has the pre-labeled second type tag, and the second type tag includes industry opinion class and non-industry opinion class;
  • Step 220 Perform segmentation processing on each clause labeled with the second type label to obtain each segmentation labeled with the second type label;
  • Step 230 Convert each participle marked with the second type label to a corresponding second value according to a preset rule, and store each participle marked with the second type label and its corresponding second value in the preset including participle and The dictionary of the corresponding conversion relationship of numerical values;
  • Step 240 According to the dictionary, convert each clause labeled with the second type label into a clause in a numeric vector format to form a second numeric vector clause set;
  • Step 250 Select a second preset number of clauses from the second numerical vector clause set as the second training data
  • Step 260 Train the second training data through the deep learning model to obtain the second deep learning model.
  • the deep learning model is a machine learning model for long- and short-term memory networks.
  • jieba word segmentation module is used to perform word segmentation processing on each clause.
  • the securities research report to be analyzed can be processed into clauses to obtain each clause, and then each clause can be scored using a pre-trained first deep learning model to obtain the first clause of each clause.
  • the industry opinion category clause with the second highest score can be extracted as the securities research report to be analyzed
  • the core point of view is output.
  • a pre-trained deep learning model can be used to select opinion clauses and industry opinion clauses through intelligent scoring, and finally extract objective representative scores from industry opinion clauses that can represent the whole The clause of the core opinion of the securities research report. Therefore, the above process of intelligently scoring and analyzing using the deep learning model can not only greatly save manpower, but also improve the operation efficiency and the accuracy of extracting results.
  • This embodiment also provides a computer device, such as a smartphone, a tablet computer, a notebook computer, a desktop computer, a rack server, a blade server, a tower server, or a rack server (including an independent server, or A server cluster composed of multiple servers), etc.
  • the computer device 40 of this embodiment includes at least but not limited to: a memory 41 and a processor 42 that can be connected to each other through a system bus, as shown in FIG. 4.
  • FIG. 4 only shows the computer device 40 having the components 41-42, but it should be understood that it is not required to implement all the components shown, and more or fewer components may be implemented instead.
  • the memory 41 (ie, readable storage medium) includes flash memory, hard disk, multimedia card, card-type memory (for example, SD or DX memory, etc.), random access memory (RAM), static random access memory (SRAM), Read only memory (ROM), electrically erasable programmable read only memory (EEPROM), programmable read only memory (PROM), magnetic memory, magnetic disk, optical disk, etc.
  • the memory 41 may be an internal storage unit of the computer device 40, such as a hard disk or a memory of the computer device 40.
  • the memory 41 may also be an external storage device of the computer device 40, for example, a plug-in hard disk equipped on the computer device 40, a smart memory card (Smart Media, Card, SMC), and secure digital (Secure Digital, SD) card, flash card (Flash Card), etc.
  • the memory 41 may also include both the internal storage unit of the computer device 40 and its external storage device.
  • the memory 41 is generally used to store the operating system and various application software installed in the computer device 40, for example, the program code of the device for extracting the core viewpoint of the securities research report using the deep learning model in Embodiment 2.
  • the memory 41 can also be used to temporarily store various types of data that have been output or are to be output.
  • the processor 42 may be a central processing unit (CPU), controller, microcontroller, microprocessor, or other data processing chip in some embodiments.
  • the processor 42 is generally used to control the overall operation of the computer device 40.
  • the processor 42 is used to run the program code or process data stored in the memory 41, for example, a device for extracting core opinions of a securities research report using a deep learning model.
  • This embodiment also provides a computer-readable storage medium, such as flash memory, hard disk, multimedia card, card-type memory (for example, SD or DX memory, etc.), random access memory (RAM), static random access memory (SRAM), only Read memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, server, App store, etc., on which computer programs are stored, When the program is executed by the processor, the corresponding function is realized.
  • the computer-readable storage medium of this embodiment is used for a securities research report core viewpoint extraction device using a deep learning model, and when executed by a processor, implements the securities research report core viewpoint extraction method of the first embodiment using a deep learning model.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • General Engineering & Computer Science (AREA)
  • Biomedical Technology (AREA)
  • Evolutionary Computation (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Machine Translation (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

一种利用深度学习模型的证券研报核心观点提取方法,包括:接收待分析证券研报;对该证券研报进行分句得到各分句;将各分句采用第一深度学习模型进行评分得到各分句的第一评分并基于第一评分确定观点类分句;将各观点类分句采用第二深度学习模型进行评分得到各观点类分句的第二评分并基于第二评分确定行业观点类分句;提取第二评分最高的行业观点类分句作为待分析证券研报的核心观点进行输出。利用深度学习模型进行智能化评分的方式实现对证券研报中核心观点的提取,可解决针对证券研报分析方案的效率及准确率较低的问题,可高效且准确的提取出证券研报中的核心观点。

Description

利用深度学习模型的证券研报核心观点提取方法及装置
相关申请的交叉引用
本申请申明享有2018年10月22日递交的申请号为CN2018112284905、名称为“利用深度学习模型的证券研报核心观点提取方法及装置”的中国专利申请的优先权,该中国专利申请的整体内容以参考的方式结合在本申请中。
技术领域
本申请涉及计算机技术领域,具体涉及一种利用深度学习模型的证券研报核心观点提取方法及装置。
背景技术
证券研究报告,也可简称为证券研报,是指相关研究人员(比如证券公司内的研究人员等)对证券及相关产品的价值、或者影响其市场价格的因素进行分析,所作出的研究报告。
对证券研报进行分析,可以及时了解证券研报中关于行业、政策、投资可行性等方面情况,但目前主要还是通过人工方式对证券研报进行阅读分析,以获得作者的核心观点等有用信息,此种方式需要耗费大量人力,效率及准确率都较低。
相关技术中针对证券研报分析方案的效率及准确率较低的问题,目前尚未提出有效的解决方案。
发明内容
本申请的目的在于提供一种利用深度学习模型的证券研报核心观点提取方法、装置、计算机设备及可读存储介质,进而在一定程度上克服上述现有技术中存在的问题,可高效且准确的提取出证券研报中的核心观点。
本申请是通过下述技术方案来解决上述技术问题:
根据本申请的一个方面,提供了一种利用深度学习模型的证券研报核心观点提取方法,包括如下步骤:
S01,接收输入的待分析证券研报;
S02,对待分析证券研报进行分句处理,得到待分析证券研报中的各分句;
S03,将待分析证券研报中的各分句采用预先训练好的第一深度学习模型进行评分,以得到针对待分析证券研报中的各分句的第一评分,并基于第一评分从待分析证券研报中的各分句中确定观点类分句;
S04,将待分析证券研报中的各观点类分句采用预先训练好的第二深度学习模型进行评分,以得到针对待分析证券研报中的各观点类分句的第二评分,并基于第二评分从待分析证券研报中的各观点类分句中确定行业观点类分句;
S05,提取第二评分最高的行业观点类分句作为待分析证券研报的核心观点进行输出。
为了实现上述目的,本申请还提供一种利用深度学习模型的证券研报核心观点提取装置,包括:
接收模块,用于接收输入的待分析证券研报;
分句模块,用于对待分析证券研报进行分句处理,得到待分析证券研报中的各分句;
第一评分模块,用于将待分析证券研报中的各分句采用预先训练好的第一深度学习模 型进行评分,以得到针对待分析证券研报中的各分句的第一评分,并基于第一评分从待分析证券研报中的各分句中确定观点类分句;
第二评分模块,用于将待分析证券研报中的各观点类分句采用预先训练好的第二深度学习模型进行评分,以得到针对待分析证券研报中的各观点类分句的第二评分,并基于第二评分从待分析证券研报中的各观点类分句中确定行业观点类分句;
提取模块,用于提取第二评分最高的行业观点类分句作为待分析证券研报的核心观点进行输出。
为了实现上述目的,本申请还提供一种计算机设备,包括存储器、处理器以及存储在存储器上并可在处理器上运行的计算机程序,所述处理器执行所述程序时实现利用深度学习模型的证券研报核心观点提取方法的以下步骤:
S01,接收输入的待分析证券研报;
S02,对待分析证券研报进行分句处理,得到待分析证券研报中的各分句;
S03,将待分析证券研报中的各分句采用预先训练好的第一深度学习模型进行评分,以得到针对待分析证券研报中的各分句的第一评分,并基于第一评分从待分析证券研报中的各分句中确定观点类分句;
S04,将待分析证券研报中的各观点类分句采用预先训练好的第二深度学习模型进行评分,以得到针对待分析证券研报中的各观点类分句的第二评分,并基于第二评分从待分析证券研报中的各观点类分句中确定行业观点类分句;
S05,提取第二评分最高的行业观点类分句作为待分析证券研报的核心观点进行输出。
为了实现上述目的,本申请还提供一种计算机可读存储介质,其上存储有计算机程序,所述程序被处理器执行时实现利用深度学习模型的证券研报核心观点提取方法的以下步骤:
S01,接收输入的待分析证券研报;
S02,对待分析证券研报进行分句处理,得到待分析证券研报中的各分句;
S03,将待分析证券研报中的各分句采用预先训练好的第一深度学习模型进行评分,以得到针对待分析证券研报中的各分句的第一评分,并基于第一评分从待分析证券研报中的各分句中确定观点类分句;
S04,将待分析证券研报中的各观点类分句采用预先训练好的第二深度学习模型进行评分,以得到针对待分析证券研报中的各观点类分句的第二评分,并基于第二评分从待分析证券研报中的各观点类分句中确定行业观点类分句;
S05,提取第二评分最高的行业观点类分句作为待分析证券研报的核心观点进行输出。
本申请提供的利用深度学习模型的证券研报核心观点提取方法、装置、计算机设备及可读存储介质,可先将待分析证券研报进行分句处理得到各分句,再将各分句采用预先训练好的第一深度学习模型进行评分以得到各分句的第一评分,并根据第一评分判断分句是否为观点类分句,然后将判断出的观点类分句采用预先训练好的第二深度学习模型进行评分以得到各观点类分句的第二评分,并根据第二评分判断各观点类分句是否为行业观点类分句,在判断出行业观点类分句后,可提取第二评分最高的行业观点类分句作为该待分析证券研报中的核心观点进行输出。通过上述方案,可利用预先训练好的深度学习模型通过智能化评分的方式挑选出观点类分句及行业观点类分句,并最终通过客观的评分从行业观点类分句中提取出能够代表整篇证券研报的核心观点的分句,由此,上述利用深度学习模型智能评分并分析的过程,不但可大大节省人力,而且可提高操作效率及提取结果的准确率。
附图说明
图1是根据本申请实施例的利用深度学习模型的证券研报核心观点提取方法的一种可 选的流程示意图;
图2是根据本申请实施例的利用深度学习模型的证券研报核心观点提取装置的一种可选的程序模块示意图;
图3是根据本申请实施例的利用深度学习模型的证券研报核心观点提取装置的另一种可选的程序模块示意图;
图4是根据本申请实施例的计算机设备的一种可选的硬件架构示意图。
具体实施方式
为了使本申请的目的、技术方案及优点更加清楚明白,以下结合附图及实施例,对本申请进行进一步详细说明。应当理解,此处所描述的具体实施例仅用以解释本申请,并不用于限定本申请。基于本申请中的实施例,本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例,都属于本申请保护的范围。
实施例1
下面结合附图对本申请提供的利用深度学习模型的证券研报核心观点提取方法进行说明。
图1为本申请利用深度学习模型的证券研报核心观点提取方法的一种可选的流程示意图,如图1所示,该方法可以包括以下步骤:
S01,接收输入的待分析证券研报。
在本实施例中,比如可以一篇证券研报为单位进行其核心观点的提取,当接收到一篇待分析的证券研报后,可先对该篇证券研报的格式进行判断。具体的,可判断该篇证券研报的格式是否为文本类型,比如“.txt”格式、“.doc”格式等。若判断结果为否,也就是说该证券研报的格式不是文本类型,则可将该证券研报的格式转换为文本类型,比如该证券研报的格式为便携文件格式(PDF格式),则可使用现有的“PDFParser”等工具将PDF格式的证券研报转换txt格式的证券研报。以此,可统一接收到的待分析证券研报的格式,以更方便且更高效的对待分析证券研报的内容进行读取。
S02,对待分析证券研报进行分句处理,得到待分析证券研报中的各分句。
在本实施例中,可根据预置类型的符号对待分析证券研报进行分句处理,比如可根据逗号“,”、句号“。”、点号“.”、破折号“—”、中括号“[]、【】”、分号“;”等符号,对待分析证券研报进行分句处理,以得到待分析证券研报中的各分句。
在得到待分析证券研报中的各分句后,可进一步采用jieba分词模块对各分句进行分词处理,以得到待分析证券研报中的各分词。
在得到待分析证券研报中的各分词后,可基于在预置的包括分词与数值的对应转换关系的字典,确定待分析证券研报中各分词所对应的数值(在本实施例中为整数型数值)。
然后,再根据各分词与其对应数值的确定结果,将待分析证券研报中每个由多个分词组成的分句转换成数值向量格式的分句。
以此,可在后续步骤中,将数值向量格式的分句输入深度学习模型中,以便深度学习模型对各分句进行评分并可根据评分对分句的类型进行判断。
S03,将待分析证券研报中的各分句采用预先训练好的第一深度学习模型进行评分,以得到针对待分析证券研报中的各分句的第一评分,并基于第一评分从待分析证券研报中的各分句中确定观点类分句。
首先,先对第一深度学习模型的训练过程进行说明,该训练过程可包括如下步骤:
步骤110,确定第一数据集。
比如,我们可预先精心挑选预置篇数(比如500篇)证券研报,比如可挑选一定比例的观点比较明显、尤其是行业观点比较明显的证券研报,还可挑选一定比例的观点比较模糊、尤其是行业观点比较模糊的证券研报等,然后可对这些证券研报的摘要部分进行分句 处理以得到多个分句,然后以人工方式对每个分句进行第一类型标签的标注,该第一类型标签可包括观点类和非观点类。也就是说,经第一类型标签标注后,在第一数据集中的分句可包括观点类分句和非观点类分句。
步骤120,将标注有第一类型标签的各分句进行分词处理,得到标注有第一类型标签的各分词。
比如,可使用现有的jieba分词模块对第一数据集中的标注有第一类型标签的各分句进行分词处理,以得到标注有第一类型标签的各分词。
步骤130,根据预置规则将标注有第一类型标签的各分词转换为对应的第一数值,并将标注有第一类型标签的各分词与其对应的第一数值存储于预置的包括分词与数值的对应转换关系的字典中。
具体的,可根据预置规则(比如随机转换,或者可根据实际需求预先设定词语到数值的转换规则等)将标注有第一类型标签的各分词一一转换为数值(在本实施例中,可称为第一数值),并可将标注有第一类型标签的各分词与第一数值存储于预置的包括分词与数值的对应转换关系的字典中。
在本实施例中,可将标注有第一类型标签的各分词与第一数值的对应转换关系以文件或者数据表的形式保存于字典中,内容可如表1所示。
表1分词与数值的对应转换关系
分词 数值
分词1 2
分词2 175
分词3 50
... ...
步骤140,根据字典,将标注有第一类型标签的各分句转换成数值向量格式的分句,形成第一数值向量分句集合。
也即,可根据上述包括分词与数值的对应转换关系的字典,将每个标注有第一类型标签的由多个分词组成的分句,转换成数值向量格式的分句,并形成数值向量分句集合,在本实施例中,可称为第一数值向量分句集合。
步骤150,从第一数值向量分句集合中选取第一预置数量的分句作为第一训练数据。
通常在进行深度学习模型训练之前,可先确定用于训练的训练集和用于测试的测试集,在本实施例中,可从上述第一数值向量分句集合中选取预置数量(比如占比90%)的分句作为第一训练数据,剩余(比如占比10%)的分句则可作为第一测试数据。
步骤160,将第一训练数据经深度学习模型进行训练,以获得第一深度学习模型。
在本实施例中,该深度学习模型可为长短期记忆网络(LSTM,Long Short-Term Memory)机器学习模型,在训练之前,可对一些关键参数进行设置,比如嵌入层大小(embedding size)、隐藏层大小(hidden layer size)、批次训练大小(batch size)、全样本循环次数(num epochs)、遗忘率(dropout)、激活函数(activation,分类型变量使用sigmoid)、损失函数(loss,分类型问题使用binary-crossentropy)等参数。然后,可将第一训练数据输入LSTM机器学习模型进行训练,以在训练完毕后获得第一深度学习模型。
此外,在训练完毕获得第一深度学习模型后,还可将第一测试数据输入第一深度学习模型进行测试,以得到该第一深度学习模型的准确率,经多次试验,我们的第一深度学习模型在第一测试数据上的准确率可达到85%左右。在准确率低于预定标准(比如70%)时,可通过调整参数、对训练集数据进行调整等方式重新进行训练,以得到符合预定标准的准确率的第一深度学习模型,从而可保证利用训练好的第一深度模型可更为准确的对分句进行评分并判断出句子类型。
由此,可将待分析证券研报中的各分句采用上述第一深度学习模型进行评分以得到针对待分析证券研报中的各分句的第一评分,并可基于第一评分从待分析证券研报中的各分 句中确定观点类分句。
在本实施例中,该第一评分可为该模型认为的评分,该评分可视为概率(在0至1之间),比如模型认为比较偏向于观点类的分句,则评分更接近于1(比如0.75),模型认为比较偏向于非观点类的分句,则评分更接近于0(比如0.22)。
在得到第一评分后,可根据该第一评分与预先设置的第一分数阈值进行比较,并根据比较结果判断出观点类分句,比如可将第一评分大于第一分数阈值的分句判定为观点类分句,将第一评分不大于第一分数阈值的分句判定为非观点分句。在本实施例中,可将该第一分数阈值设置为0.5,也就是说,第一评分不大于0.5的分句均可确定为非观点类分句,不再参与后续步骤;第一评分大于0.5的分句均可确定为观点类分句,可以继续参与后续步骤。
以此,采用预先训练好的第一深度学习模型,以智能化评分的方式来判断分句的类型(观点类分句或非观点类分句),一方面,可大大节省人力,提高判断效率;另一方面,可提高判断结果的准确率。
S04,将待分析证券研报中的各观点类分句采用预先训练好的第二深度学习模型进行评分,以得到针对待分析证券研报中的各观点类分句的第二评分,并基于第二评分从待分析证券研报中的各观点类分句中确定行业观点类分句。
首先,先对第二深度学习模型的训练过程进行说明,该训练过程可包括如下步骤:
步骤210,确定第二数据集。
比如,预先挑选预置篇数(比如500篇)证券研报,比如可挑选一定比例的观点比较明显、尤其是行业观点比较明显的证券研报,还可挑选一定比例的观点比较模糊、尤其是行业观点比较模糊的证券研报等,然后可对证券研报的摘要部分进行分句处理以得到多个分句,然后以人工方式对每个分句进行第一类型标签的标注,该第一类型标签可包括观点类和非观点类。也就是说,经第一类型标签标注后,第二数据集中的分句可包括观点类分句和非观点类分句。然后,再以人工方式对每个观点类分句进行第二类型标签的标注,该第二类型标签可包括行业观点类和非行业观点类。也就是说,经第二类型标签标注后,在第二数据集中的分句可包括观点类分句和非观点类分句,且观点类分句中还包括行业观点类分句和非行业观点类分句。
步骤220,将标注有第二类型标签的各分句进行分词处理,得到标注有第二类型标签的各分词。
也就是说,将第二数据集中标准有第二类型标签的各分句(也即第二数据集中的行业观点类分句和非行业观点类分句进行分词处理,比如,可使用现有的jieba分词模块对上述各分句进行分词处理,以得到标注有第二类型标签的各分词。
步骤230,根据预置规则将标注有第二类型标签的各分词转换为对应的第二数值,并将标注有第二类型标签的各分词与其对应的第二数值存储于预置的包括分词与数值的对应转换关系的字典中。
具体的,可根据预置规则(比如随机转换,或者可根据实际需求预先设定词语到数值的转换规则等)将标注有第二类型标签的各分词一一转换为数值(在本实施例中,可称为第二数值),并可将标注有第二类型标签的各分词与第二数值的转换结果存储于预置的包括分词与数值的对应转换关系的字典中。
在本实施例中,可将标注有第二类型标签的各分词与第二数值的对应转换关系以文件或者数据表的形式保存于字典中,内容可如上述表1所示。
步骤240,根据字典,将标注有第二类型标签的各分句转换成数值向量格式的分句,形成第二数值向量分句集合。
也即,可根据上述包括分词与数值的对应转换关系的字典,将每个标注有第二类型标签的由多个分词组成的分句,转换成数值向量格式的分句,并形成数值向量分句集合,在本实施例中,可称为第二数值向量分句集合。
步骤250,从第二数值向量分句集合中选取第二预置数量的分句作为第二训练数据。
通常在进行深度学习模型训练之前,可先确定用于训练的训练集和用于测试的测试集,在本实施例中,可从上述第二数值向量分句集合中选取预置数量(比如占比90%)的分句作为第二训练数据,剩余(比如占比10%)的分句则可作为第二测试数据。
步骤260,将第二训练数据经深度学习模型进行训练,以获得第二深度学习模型。
在本实施例中,该深度学习模型可为长短期记忆网络(LSTM,Long Short-Term Memory)机器学习模型,在训练之前,可对一些关键参数进行设置,比如嵌入层大小(embedding size)、隐藏层大小(hidden layer size)、批次训练大小(batch size)、全样本循环次数(num epochs)、遗忘率(dropout)、激活函数(activation,分类型变量使用sigmoid)、损失函数(loss,分类型问题使用binary-crossentropy)等参数。然后,可将第二训练数据输入LSTM机器学习模型进行训练,以在训练完毕后获得第二深度学习模型。
此外,在训练完毕获得第二深度学习模型后,还可将第二测试数据输入第二深度学习模型进行测试,以得到该第二深度学习模型的准确率。经多次试验,我们的第二深度学习在第二测试数据上的准确率可达到85%左右。在准确率低于预定标准(比如70%)时,可通过调整参数、对训练集数据进行调整等方式重新进行训练,以得到符合预定标准的准确率的第二深度学习模型,从而可保证利用训练好的第二深度模型可更为准确的对分句进行评分并判断出句子类型。
由此,可将由S03中判定的待分析证券研报中的各观点类分句采用上述第二深度学习模型进行评分以得到针对待分析证券研报中的各观点类分句的第二评分,并可基于第二评分从观点类分句中确定待分析证券研报中的行业观点类分句。
在本实施例中,该第二评分可为该模型认为的评分,该评分可视为概率(在0至1之间),比如模型认为比较偏向于行业观点类的分句,则评分更接近于1(比如0.85),模型认为比较偏向于非行业观点类的分句,则评分更接近于0(比如0.3)。
在得到第二评分后,可根据该第二评分与预先设置的第二分数阈值进行比较,并根据比较结果判断出观点类分句,比如可将第二评分大于二分数阈值的分句判定为行业观点类分句,第二评分不大于第二分数阈值的分句判定为非行业观点分句。在本实施例中,可将该第二分数阈值设置为0.5,也就是说,第二评分不大于0.5的分句均可确定为非行业观点类分句,不再参与后续步骤;第二评分大于0.5的分句均可确定为行业观点类分句,可以继续参与后续步骤。
以此,采用预先训练好的第二深度学习模型,以智能化评分的方式来判断观点类分句的具体类型(行业观点类分句或非行业观点类分句),一方面,可大大节省人力,提高判断效率;另一方面,可提高判断结果的准确率。
S05,提取第二评分最高的行业观点类分句作为待分析证券研报的核心观点进行输出。
在本实施例中,一方面,由于第二评分是由第二深度学习模型根据分句是偏向于行业观点类还是偏向于非行业观点类而给出的分数,且越偏向于行业观点类的分句的第二评分越高(即越接近1),另一方面,由于在通常情况下,整篇证券研报的核心观点必然包含于行业观点中,因此,可将第二评分最高的行业观点类分句看作是最为重要且核心的行业观点。
在具体实现时,可在得到第二评分后,将第二评分进行排序(比如以正序排序),并将第二评分最高(比如排序第一)的行业观点类分句作为待分析证券研报的核心观点进行输出。
根据本实施例的各个实施方式,可先将待分析证券研报进行分句处理得到各分句,再将各分句采用预先训练好的第一深度学习模型进行评分以得到各分句的第一评分,并根据第一评分判断分句是否为观点类分句,然后将判断出的观点类分句采用预先训练好的第二深度学习模型进行评分以得到各观点类分句的第二评分,并根据第二评分判断各观点类分句是否为行业观点类分句,在判断出行业观点类分句后,可提取第二评分最高的行业观点 类分句作为该待分析证券研报中的核心观点进行输出。通过上述方案,可利用预先训练好的深度学习模型通过智能化评分的方式挑选出观点类分句及行业观点类分句,并最终通过客观的评分从行业观点类分句中提取出能够代表整篇证券研报的核心观点的分句,由此,上述利用深度学习模型智能评分并分析的过程,不但可大大节省人力,而且可提高操作效率及提取结果的准确率。
通过以上的实施方式的描述,本领域的技术人员可以清楚地了解到上述实施例方法可借助软件加必需的通用硬件平台的方式来实现,当然也可以通过硬件,但很多情况下前者是更佳的实施方式。
实施例2
基于上述实施例1中提供的利用深度学习模型的证券研报核心观点提取方法,本实施例中提供一种利用深度学习模型的证券研报核心观点提取装置,具体地,图2至3示出了该利用深度学习模型的证券研报核心观点提取装置的可选的结构框图,该利用深度学习模型的证券研报核心观点提取装置被分割成一个或多个程序模块,一个或者多个程序模块被存储于存储介质中,并由一个或多个处理器所执行,以完成本申请。本申请所称的程序模块是指能够完成特定功能的一系列计算机程序指令段,比程序本身更适合描述利用深度学习模型的证券研报核心观点提取装置在存储介质中的执行过程,以下描述将具体介绍本实施例各程序模块的功能。
如图2所示,该利用深度学习模型的证券研报核心观点提取装置20可包括:
接收模块21,可用于接收输入的待分析证券研报;
分句模块22,可用于对待分析证券研报进行分句处理,得到待分析证券研报中的各分句;
第一评分模块23,可用于将待分析证券研报中的各分句采用预先训练好的第一深度学习模型进行评分,以得到针对待分析证券研报中的各分句的第一评分,并基于第一评分从待分析证券研报中的各分句中确定观点类分句;
第二评分模块24,可用于将待分析证券研报中的各观点类分句采用预先训练好的第二深度学习模型进行评分,以得到针对待分析证券研报中的各观点类分句的第二评分,并基于第二评分从待分析证券研报中的各观点类分句中确定行业观点类分句;
提取模块25,可用于提取第二评分最高的行业观点类分句作为待分析证券研报的核心观点进行输出。
在具体实现时,接收模块21,可具体用于:
判断接收到的待分析证券研报的格式是否为文本类型;
若否,则将所述待分析证券研报的格式转换为文本类型。
进一步的,参看图3所示,分句模块22,可具体包括:
分句单元221,用于根据预置类型的符号对待分析证券研报进行分句处理,得到待分析证券研报中的各分句;
分词单元222,用于对待分析证券研报中的各分句进行分词处理,得到待分析证券研报中的各分词;
数值确定单元223,用于基于预置的包括分词与数值的对应转换关系的字典,确定待分析证券研报中的各分词对应的数值;
转换单元224,用于根据确定结果,将待分析证券研报中的各分句转换成数值向量格式的分句。
在本实施例中,第一深度学习模型的训练过程,包括如下步骤:
步骤110,确定第一数据集,其中,第一数据集包括预先对预置篇数的证券研报的摘要部分进行分句处理后得到的多个分句,其中每个分句具有预先标注的第一类型标签,第一类型标签包括观点类和非观点类;
步骤120,将标注有第一类型标签的各分句进行分词处理,得到标注有第一类型标签的各分词;
步骤130,根据预置规则将标注有第一类型标签的各分词转换为对应的第一数值,并将标注有第一类型标签的各分词与其对应的第一数值存储于预置的包括分词与数值的对应转换关系的字典中;
步骤140,根据词典,将标注有第一类型标签的各分句转换成数值向量格式的分句,形成第一数值向量分句集合;
步骤150,从第一数值向量分句集合中选取第一预置数量的分句作为第一训练数据;
步骤160,将第一训练数据经深度学习模型进行训练,以获得第一深度学习模型。
在本实施例中,第二深度学习模型的训练过程,包括如下步骤:
步骤210,确定第二数据集,其中,第二数据集包括预先对预置篇数的证券研报的摘要部分进行分句处理后得到的多个分句,其中每个分句具有预先标注的第一类型标签,第一类型标签包括观点类和非观点类,且标注有观点类标签的分句同时具有预先标注的第二类型标签,第二类型标签包括行业观点类和非行业观点类;
步骤220,将标注有第二类型标签的各分句进行分词处理,得到标注有第二类型标签的各分词;
步骤230,根据预置规则将标注有第二类型标签的各分词转换为对应的第二数值,并将标注有第二类型标签的各分词与其对应的第二数值存储于预置的包括分词与数值的对应转换关系的字典中;
步骤240,根据字典,将标注有第二类型标签的各分句转换成数值向量格式的分句,形成第二数值向量分句集合;
步骤250,从第二数值向量分句集合中选取第二预置数量的分句作为第二训练数据;
步骤260,将第二训练数据经深度学习模型进行训练,以获得第二深度学习模型。
进一步的,深度学习模型为长短期记忆网络机器学习模型。
进一步的,采用jieba分词模块对各分句进行分词处理。
关于上述实施例中的装置,其中各个单元、模块执行操作的具体方式已经在有关该方法的实施例中进行了详细描述,此处将不做详细阐述说明。
通过本实施例的各个实施方式,可先将待分析证券研报进行分句处理得到各分句,再将各分句采用预先训练好的第一深度学习模型进行评分以得到各分句的第一评分,并根据第一评分判断分句是否为观点类分句,然后将判断出的观点类分句采用预先训练好的第二深度学习模型进行评分以得到各观点类分句的第二评分,并根据第二评分判断各观点类分句是否为行业观点类分句,在判断出行业观点类分句后,可提取第二评分最高的行业观点类分句作为该待分析证券研报中的核心观点进行输出。通过上述方案,可利用预先训练好的深度学习模型通过智能化评分的方式挑选出观点类分句及行业观点类分句,并最终通过客观的评分从行业观点类分句中提取出能够代表整篇证券研报的核心观点的分句,由此,上述利用深度学习模型智能评分并分析的过程,不但可大大节省人力,而且可提高操作效率及提取结果的准确率。
实施例3
本实施例还提供一种计算机设备,如可以执行程序的智能手机、平板电脑、笔记本电脑、台式计算机、机架式服务器、刀片式服务器、塔式服务器或机柜式服务器(包括独立的服务器,或者多个服务器所组成的服务器集群)等。如图4所示,本实施例的计算机设备40至少包括但不限于:可通过系统总线相互通信连接的存储器41、处理器42,如图4所示。需要指出的是,图4仅示出了具有组件41-42的计算机设备40,但是应理解的是,并不要求实施所有示出的组件,可以替代的实施更多或者更少的组件。
本实施例中,存储器41(即可读存储介质)包括闪存、硬盘、多媒体卡、卡型存储器 (例如,SD或DX存储器等)、随机访问存储器(RAM)、静态随机访问存储器(SRAM)、只读存储器(ROM)、电可擦除可编程只读存储器(EEPROM)、可编程只读存储器(PROM)、磁性存储器、磁盘、光盘等。在一些实施例中,存储器41可以是计算机设备40的内部存储单元,例如该计算机设备40的硬盘或内存。在另一些实施例中,存储器41也可以是计算机设备40的外部存储设备,例如该计算机设备40上配备的插接式硬盘,智能存储卡(Smart Media Card,SMC),安全数字(Secure Digital,SD)卡,闪存卡(Flash Card)等。当然,存储器41还可以既包括计算机设备40的内部存储单元也包括其外部存储设备。本实施例中,存储器41通常用于存储安装于计算机设备40的操作系统和各类应用软件,例如实施例2的利用深度学习模型的证券研报核心观点提取装置的程序代码等。此外,存储器41还可以用于暂时地存储已经输出或者将要输出的各类数据。
处理器42在一些实施例中可以是中央处理器(Central Processing Unit,CPU)、控制器、微控制器、微处理器、或其他数据处理芯片。该处理器42通常用于控制计算机设备40的总体操作。本实施例中,处理器42用于运行存储器41中存储的程序代码或者处理数据,例如利用深度学习模型的证券研报核心观点提取装置等。
实施例4
本实施例还提供一种计算机可读存储介质,如闪存、硬盘、多媒体卡、卡型存储器(例如,SD或DX存储器等)、随机访问存储器(RAM)、静态随机访问存储器(SRAM)、只读存储器(ROM)、电可擦除可编程只读存储器(EEPROM)、可编程只读存储器(PROM)、磁性存储器、磁盘、光盘、服务器、App应用商城等等,其上存储有计算机程序,程序被处理器执行时实现相应功能。本实施例的计算机可读存储介质用于利用深度学习模型的证券研报核心观点提取装置,被处理器执行时实现实施例1的利用深度学习模型的证券研报核心观点提取方法。
上述本申请实施例序号仅仅为了描述,不代表实施例的优劣。
通过以上的实施方式的描述,本领域的技术人员可以清楚地了解到上述实施例方法可借助软件加必需的通用硬件平台的方式来实现,当然也可以通过硬件,但很多情况下前者是更佳的实施方式。
以上仅为本申请的优选实施例,并非因此限制本申请的专利范围,凡是利用本申请说明书及附图内容所作的等效结构或等效流程变换,或直接或间接运用在其他相关的技术领域,均同理包括在本申请的专利保护范围内。

Claims (20)

  1. 一种利用深度学习模型的证券研报核心观点提取方法,其特征在于,包括如下步骤:
    S01,接收输入的待分析证券研报;
    S02,对待分析证券研报进行分句处理,得到待分析证券研报中的各分句;
    S03,将待分析证券研报中的各分句采用预先训练好的第一深度学习模型进行评分,以得到针对待分析证券研报中的各分句的第一评分,并基于第一评分从待分析证券研报中的各分句中确定观点类分句;
    S04,将待分析证券研报中的各观点类分句采用预先训练好的第二深度学习模型进行评分,以得到针对待分析证券研报中的各观点类分句的第二评分,并基于第二评分从待分析证券研报中的各观点类分句中确定行业观点类分句;
    S05,提取第二评分最高的行业观点类分句作为待分析证券研报的核心观点进行输出。
  2. 根据权利要求1所述的利用深度学习模型的证券研报核心观点提取方法,其特征在于,S01接收输入的待分析证券研报,包括:
    判断接收到的待分析证券研报的格式是否为文本类型;
    若否,则将所述待分析证券研报的格式转换为文本类型。
  3. 根据权利要求1所述的利用深度学习模型的证券研报核心观点提取方法,其特征在于,S02对待分析证券研报进行分句处理,得到待分析证券研报中的各分句,包括:
    根据预置类型的符号对待分析证券研报进行分句处理,得到待分析证券研报中的各分句;
    对待分析证券研报中的各分句进行分词处理,得到待分析证券研报中的各分词;
    基于预置的包括分词与数值的对应转换关系的字典,确定待分析证券研报中的各分词对应的数值;
    根据确定结果,将待分析证券研报中的各分句转换成数值向量格式的分句。
  4. 根据权利要求1所述的利用深度学习模型的证券研报核心观点提取方法,其特征在于,第一深度学习模型的训练过程,包括如下步骤:
    步骤110,确定第一数据集,所述第一数据集包括预先对预置篇数的证券研报的摘要部分进行分句处理后得到的多个分句,其中每个分句具有预先标注的第一类型标签,第一类型标签包括观点类和非观点类;
    步骤120,将标注有第一类型标签的各分句进行分词处理,得到标注有第一类型标签的各分词;
    步骤130,根据预置规则将标注有第一类型标签的各分词转换为对应的第一数值,并将标注有第一类型标签的各分词与其对应的第一数值存储于预置的包括分词与数值的对应转换关系的字典中;
    步骤140,根据所述词典,将标注有第一类型标签的各分句转换成数值向量格式的分句,形成第一数值向量分句集合;
    步骤150,从第一数值向量分句集合中选取第一预置数量的分句作为第一训练数据;
    步骤160,将第一训练数据经深度学习模型进行训练,以获得第一深度学习模型。
  5. 根据权利要求1所述的利用深度学习模型的证券研报核心观点提取方法,其特征在于,第二深度学习模型的训练过程,包括如下步骤:
    步骤210,确定第二数据集,所述第二数据集包括预先对预置篇数的证券研报的摘要部分进行分句处理后得到的多个分句,其中每个分句具有预先标注的第一类型标签,第一类型标签包括观点类和非观点类,且标注有观点类标签的分句同时具有预先标注的第二类型标签,第二类型标签包括行业观点类和非行业观点类;
    步骤220,将标注有第二类型标签的各分句进行分词处理,得到标注有第二类型标签 的各分词;
    步骤230,根据预置规则将标注有第二类型标签的各分词转换为对应的第二数值,并将标注有第二类型标签的各分词与其对应的第二数值存储于预置的包括分词与数值的对应转换关系的字典中;
    步骤240,根据所述字典,将标注有第二类型标签的各分句转换成数值向量格式的分句,形成第二数值向量分句集合;
    步骤250,从第二数值向量分句集合中选取第二预置数量的分句作为第二训练数据;
    步骤260,将第二训练数据经深度学习模型进行训练,以获得第二深度学习模型。
  6. 根据权利要求4或5所述的利用深度学习模型的证券研报核心观点提取方法,其特征在于,所述深度学习模型为长短期记忆网络机器学习模型。
  7. 根据权利要求3至5任一项所述的利用深度学习模型的证券研报核心观点提取方法,其特征在于,采用jieba分词模块对各分句进行分词处理。
  8. 一种利用深度学习模型的证券研报核心观点提取装置,其特征在于,包括:
    接收模块,用于接收输入的待分析证券研报;
    分句模块,用于对待分析证券研报进行分句处理,得到待分析证券研报中的各分句;
    第一评分模块,用于将待分析证券研报中的各分句采用预先训练好的第一深度学习模型进行评分,以得到针对待分析证券研报中的各分句的第一评分,并基于第一评分从待分析证券研报中的各分句中确定观点类分句;
    第二评分模块,用于将待分析证券研报中的各观点类分句采用预先训练好的第二深度学习模型进行评分,以得到针对待分析证券研报中的各观点类分句的第二评分,并基于第二评分从待分析证券研报中的各观点类分句中确定行业观点类分句;
    提取模块,用于提取第二评分最高的行业观点类分句作为待分析证券研报的核心观点进行输出。
  9. 一种计算机设备,所述计算机设备包括存储器、处理器以及存储在存储器上并可在处理器上运行的计算机程序,其特征在于,所述处理器执行所述程序时实现利用深度学习模型的证券研报核心观点提取方法的以下步骤:
    S01,接收输入的待分析证券研报;
    S02,对待分析证券研报进行分句处理,得到待分析证券研报中的各分句;
    S03,将待分析证券研报中的各分句采用预先训练好的第一深度学习模型进行评分,以得到针对待分析证券研报中的各分句的第一评分,并基于第一评分从待分析证券研报中的各分句中确定观点类分句;
    S04,将待分析证券研报中的各观点类分句采用预先训练好的第二深度学习模型进行评分,以得到针对待分析证券研报中的各观点类分句的第二评分,并基于第二评分从待分析证券研报中的各观点类分句中确定行业观点类分句;
    S05,提取第二评分最高的行业观点类分句作为待分析证券研报的核心观点进行输出。
  10. 根据权利要求9所述的计算机设备,其特征在于,S01接收输入的待分析证券研报,包括:
    判断接收到的待分析证券研报的格式是否为文本类型;
    若否,则将所述待分析证券研报的格式转换为文本类型。
  11. 根据权利要求9所述的计算机设备,其特征在于,S02对待分析证券研报进行分句处理,得到待分析证券研报中的各分句,包括:
    根据预置类型的符号对待分析证券研报进行分句处理,得到待分析证券研报中的各分句;
    对待分析证券研报中的各分句进行分词处理,得到待分析证券研报中的各分词;
    基于预置的包括分词与数值的对应转换关系的字典,确定待分析证券研报中的各分词对应的数值;
    根据确定结果,将待分析证券研报中的各分句转换成数值向量格式的分句。
  12. 根据权利要求9所述的计算机设备,其特征在于,第一深度学习模型的训练过程,包括如下步骤:
    步骤110,确定第一数据集,所述第一数据集包括预先对预置篇数的证券研报的摘要部分进行分句处理后得到的多个分句,其中每个分句具有预先标注的第一类型标签,第一类型标签包括观点类和非观点类;
    步骤120,将标注有第一类型标签的各分句进行分词处理,得到标注有第一类型标签的各分词;
    步骤130,根据预置规则将标注有第一类型标签的各分词转换为对应的第一数值,并将标注有第一类型标签的各分词与其对应的第一数值存储于预置的包括分词与数值的对应转换关系的字典中;
    步骤140,根据所述词典,将标注有第一类型标签的各分句转换成数值向量格式的分句,形成第一数值向量分句集合;
    步骤150,从第一数值向量分句集合中选取第一预置数量的分句作为第一训练数据;
    步骤160,将第一训练数据经深度学习模型进行训练,以获得第一深度学习模型。
  13. 根据权利要求9所述的计算机设备,其特征在于,第二深度学习模型的训练过程,包括如下步骤:
    步骤210,确定第二数据集,所述第二数据集包括预先对预置篇数的证券研报的摘要部分进行分句处理后得到的多个分句,其中每个分句具有预先标注的第一类型标签,第一类型标签包括观点类和非观点类,且标注有观点类标签的分句同时具有预先标注的第二类型标签,第二类型标签包括行业观点类和非行业观点类;
    步骤220,将标注有第二类型标签的各分句进行分词处理,得到标注有第二类型标签的各分词;
    步骤230,根据预置规则将标注有第二类型标签的各分词转换为对应的第二数值,并将标注有第二类型标签的各分词与其对应的第二数值存储于预置的包括分词与数值的对应转换关系的字典中;
    步骤240,根据所述字典,将标注有第二类型标签的各分句转换成数值向量格式的分句,形成第二数值向量分句集合;
    步骤250,从第二数值向量分句集合中选取第二预置数量的分句作为第二训练数据;
    步骤260,将第二训练数据经深度学习模型进行训练,以获得第二深度学习模型。
  14. 根据权利要求12或13所述的计算机设备,其特征在于,所述深度学习模型为长短期记忆网络机器学习模型。
  15. 一种计算机可读存储介质,其上存储有计算机程序,其特征在于:所述程序被处理器执行时实现利用深度学习模型的证券研报核心观点提取方法的以下步骤:
    S01,接收输入的待分析证券研报;
    S02,对待分析证券研报进行分句处理,得到待分析证券研报中的各分句;
    S03,将待分析证券研报中的各分句采用预先训练好的第一深度学习模型进行评分,以得到针对待分析证券研报中的各分句的第一评分,并基于第一评分从待分析证券研报中的各分句中确定观点类分句;
    S04,将待分析证券研报中的各观点类分句采用预先训练好的第二深度学习模型进行评分,以得到针对待分析证券研报中的各观点类分句的第二评分,并基于第二评分从待分析证券研报中的各观点类分句中确定行业观点类分句;
    S05,提取第二评分最高的行业观点类分句作为待分析证券研报的核心观点进行输出。
  16. 根据权利要求15所述的计算机可读存储介质,其特征在于,S01接收输入的待分析证券研报,包括:
    判断接收到的待分析证券研报的格式是否为文本类型;
    若否,则将所述待分析证券研报的格式转换为文本类型。
  17. 根据权利要求15所述的计算机可读存储介质,其特征在于,S02对待分析证券研报进行分句处理,得到待分析证券研报中的各分句,包括:
    根据预置类型的符号对待分析证券研报进行分句处理,得到待分析证券研报中的各分句;
    对待分析证券研报中的各分句进行分词处理,得到待分析证券研报中的各分词;
    基于预置的包括分词与数值的对应转换关系的字典,确定待分析证券研报中的各分词对应的数值;
    根据确定结果,将待分析证券研报中的各分句转换成数值向量格式的分句。
  18. 根据权利要求15所述的计算机可读存储介质,其特征在于,第一深度学习模型的训练过程,包括如下步骤:
    步骤110,确定第一数据集,所述第一数据集包括预先对预置篇数的证券研报的摘要部分进行分句处理后得到的多个分句,其中每个分句具有预先标注的第一类型标签,第一类型标签包括观点类和非观点类;
    步骤120,将标注有第一类型标签的各分句进行分词处理,得到标注有第一类型标签的各分词;
    步骤130,根据预置规则将标注有第一类型标签的各分词转换为对应的第一数值,并将标注有第一类型标签的各分词与其对应的第一数值存储于预置的包括分词与数值的对应转换关系的字典中;
    步骤140,根据所述词典,将标注有第一类型标签的各分句转换成数值向量格式的分句,形成第一数值向量分句集合;
    步骤150,从第一数值向量分句集合中选取第一预置数量的分句作为第一训练数据;
    步骤160,将第一训练数据经深度学习模型进行训练,以获得第一深度学习模型。
  19. 根据权利要求15所述的计算机可读存储介质,其特征在于,第二深度学习模型的训练过程,包括如下步骤:
    步骤210,确定第二数据集,所述第二数据集包括预先对预置篇数的证券研报的摘要部分进行分句处理后得到的多个分句,其中每个分句具有预先标注的第一类型标签,第一类型标签包括观点类和非观点类,且标注有观点类标签的分句同时具有预先标注的第二类型标签,第二类型标签包括行业观点类和非行业观点类;
    步骤220,将标注有第二类型标签的各分句进行分词处理,得到标注有第二类型标签的各分词;
    步骤230,根据预置规则将标注有第二类型标签的各分词转换为对应的第二数值,并将标注有第二类型标签的各分词与其对应的第二数值存储于预置的包括分词与数值的对应转换关系的字典中;
    步骤240,根据所述字典,将标注有第二类型标签的各分句转换成数值向量格式的分句,形成第二数值向量分句集合;
    步骤250,从第二数值向量分句集合中选取第二预置数量的分句作为第二训练数据;
    步骤260,将第二训练数据经深度学习模型进行训练,以获得第二深度学习模型。
  20. 根据权利要求18或19所述的计算机可读存储介质,其特征在于,所述深度学习模型为长短期记忆网络机器学习模型。
PCT/CN2019/070292 2018-10-22 2019-01-03 利用深度学习模型的证券研报核心观点提取方法及装置 Ceased WO2020082613A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201811228490.5A CN109388804A (zh) 2018-10-22 2018-10-22 利用深度学习模型的证券研报核心观点提取方法及装置
CN201811228490.5 2018-10-22

Publications (1)

Publication Number Publication Date
WO2020082613A1 true WO2020082613A1 (zh) 2020-04-30

Family

ID=65427490

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/070292 Ceased WO2020082613A1 (zh) 2018-10-22 2019-01-03 利用深度学习模型的证券研报核心观点提取方法及装置

Country Status (2)

Country Link
CN (1) CN109388804A (zh)
WO (1) WO2020082613A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115659023A (zh) * 2022-10-24 2023-01-31 天元大数据信用管理有限公司 一种基于标签自动标注的智能政策匹配方法及系统

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110110295B (zh) * 2019-04-04 2023-10-20 平安科技(深圳)有限公司 大样本研报信息提取方法、装置、设备及存储介质
CN110889291B (zh) * 2019-11-21 2023-06-27 山东浪潮科学研究院有限公司 一种基于深度学习和区块链的研报评价方法
CN113127595B (zh) * 2021-04-26 2022-08-16 数库(上海)科技有限公司 研报摘要的观点详情提取方法、装置、设备和存储介质

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160283583A1 (en) * 2014-03-14 2016-09-29 Tencent Technology (Shenzhen) Company Limited Method, apparatus, and storage medium for text information processing
CN106776551A (zh) * 2016-12-06 2017-05-31 桂林电子科技大学 一种英语作文情感观点的分析方法
CN108491406A (zh) * 2018-01-23 2018-09-04 深圳市阿西莫夫科技有限公司 信息分类方法、装置、计算机设备和存储介质

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101667194A (zh) * 2009-09-29 2010-03-10 北京大学 基于用户评论文本特征的自动摘要方法及其自动摘要系统
CN102841940B (zh) * 2012-08-17 2015-01-28 浙江大学 一种基于数据重构的文档摘要提取方法
US20180025268A1 (en) * 2016-07-21 2018-01-25 Tessera Advanced Technologies, Inc. Configurable machine learning assemblies for autonomous operation in personal devices
CN106227722B (zh) * 2016-09-12 2019-07-05 中山大学 一种基于上市公司公告摘要的自动提取方法
CN107797990A (zh) * 2017-10-18 2018-03-13 渡鸦科技(北京)有限责任公司 用于确定文本核心语句的方法和装置
CN108182247A (zh) * 2017-12-28 2018-06-19 东软集团股份有限公司 文摘生成方法和装置

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160283583A1 (en) * 2014-03-14 2016-09-29 Tencent Technology (Shenzhen) Company Limited Method, apparatus, and storage medium for text information processing
CN106776551A (zh) * 2016-12-06 2017-05-31 桂林电子科技大学 一种英语作文情感观点的分析方法
CN108491406A (zh) * 2018-01-23 2018-09-04 深圳市阿西莫夫科技有限公司 信息分类方法、装置、计算机设备和存储介质

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115659023A (zh) * 2022-10-24 2023-01-31 天元大数据信用管理有限公司 一种基于标签自动标注的智能政策匹配方法及系统

Also Published As

Publication number Publication date
CN109388804A (zh) 2019-02-26

Similar Documents

Publication Publication Date Title
CN110598203B (zh) 一种结合词典的军事想定文书实体信息抽取方法及装置
CN113051356A (zh) 开放关系抽取方法、装置、电子设备及存储介质
CN104050256B (zh) 基于主动学习的问答方法及采用该方法的问答系统
CN111368043A (zh) 基于人工智能的事件问答方法、装置、设备及存储介质
CN112052682A (zh) 事件实体联合抽取方法、装置、计算机设备及存储介质
WO2019218514A1 (zh) 网页目标信息的提取方法、装置及存储介质
CN111309910A (zh) 文本信息挖掘方法及装置
CN109933796A (zh) 一种公告文本关键信息提取方法及设备
WO2020082613A1 (zh) 利用深度学习模型的证券研报核心观点提取方法及装置
WO2022048363A1 (zh) 网站分类方法、装置、计算机设备及存储介质
CN112163424A (zh) 数据的标注方法、装置、设备和介质
CN109472022B (zh) 基于机器学习的新词识别方法及终端设备
CN107357765B (zh) Word文档碎片化方法及装置
CN109359296B (zh) 舆情情感识别方法、装置及计算机可读存储介质
CN106897262A (zh) 一种文本分类方法和装置以及处理方法和装置
CN109389418A (zh) 基于lda模型的供电服务客户诉求识别方法
WO2020082609A1 (zh) 利用深度学习模型的证券研报分析方法及装置
CN114461665A (zh) 用于生成语句转换模型的方法、装置及计算机程序产品
CN115081441A (zh) 运维指令意图识别方法、装置、电子设备和存储介质
CN109857957B (zh) 建立标签库的方法、电子设备及计算机存储介质
CN119739838A (zh) 多标签生成匹配的rag智能问答方法、装置、设备及介质
CN114297388A (zh) 一种文本关键词提取方法
CN118298449A (zh) 文档结构切分方法、装置、设备和介质
CN112115229A (zh) 文本意图识别方法、装置、系统以及文本分类系统
WO2020082612A1 (zh) 利用大数据的证券研报情感分析方法、装置及计算机设备

Legal Events

Date Code Title Description
NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205 A DATED 01/09/2021)

122 Ep: pct application non-entry in european phase

Ref document number: 19875089

Country of ref document: EP

Kind code of ref document: A1