CN109284504A

CN109284504A - It grinds to call the score using the security of deep learning model and analyses method and device

Info

Publication number: CN109284504A
Application number: CN201811228761.7A
Authority: CN
Inventors: 叶曙峰; 蒋逸文; 陈泽晖; 顾研
Original assignee: Ping An Technology Shenzhen Co Ltd
Current assignee: Ping An Technology Shenzhen Co Ltd
Priority date: 2018-10-22
Filing date: 2018-10-22
Publication date: 2019-01-29
Also published as: WO2020082609A1

Abstract

The present invention discloses a kind of security using deep learning model and grinds report analysis method, comprising: receives security to be analyzed and grinds report；Report progress subordinate sentence is ground to the security and obtains each subordinate sentence；Each subordinate sentence is scored to obtain the first scoring of each subordinate sentence and score based on first using the first deep learning model to determine viewpoint class subordinate sentence；Each viewpoint class subordinate sentence is scored to obtain the second scoring of each viewpoint class subordinate sentence and score based on second using the second deep learning model to determine industry viewpoint class subordinate sentence；Every profession and trade viewpoint class subordinate sentence is scored to obtain the third scoring of every profession and trade viewpoint class subordinate sentence using third deep learning model；The whole emotion trend of report is ground according to second, third determining security that score of every profession and trade viewpoint class subordinate sentence.The present invention realizes the analysis that security are ground with report in the way of the intelligent scoring of deep learning model progress, can solve the problems, such as that the prior art is ground for security and report the efficiency of analytical plan and accuracy rate lower, improve the efficiency and accuracy rate that security are ground with analysis of calling the score.

Description

It grinds to call the score using the security of deep learning model and analyses method and device

Technical field

The present invention relates to field of computer technology, and in particular to a kind of security using deep learning model grind the side of analysis of calling the score Method and device.

Background technique

Security research report can be also simply referred to as security and grind report, refer to related researcher's (such as research in securities broker company Personnel etc.) on the value of security and Related product or influences the factor of its market price and analyze, made research report It accuses.

Report is ground to security to analyze, and can be understood security in time and be ground in report about sides such as industry, policy, investment feasibilities Face situation, but main at present still ground by manual type to security reports progress reading analysis, to obtain the core views of author Equal useful informations, such mode need to expend a large amount of manpowers, and efficiency and accuracy rate are all lower.Pass through sentiment dictionary in addition, also having The scheme analyzed, for example by being compared with repertorie crucial in sentiment dictionary, carried out with the sentence ground to security in report Sentiment analysis, the opposite context for having isolated sentence of such mode are difficult to be compared sentence and improve accurately analysis, especially When for the sentence with incidence relations such as cause and effect, turnover classes, the accuracy rate of analysis is more undesirable for it.

It is directed to security in the related technology and grinds the efficiency of report analytical plan and the problem that accuracy rate is lower, not yet proposes have at present The solution of effect.

Summary of the invention

It grinds to call the score the purpose of the present invention is to provide a kind of security using deep learning model and analyses method, apparatus, calculating Machine equipment and readable storage medium storing program for executing, and then above-mentioned problems of the prior art are overcome to a certain extent, verification can be improved Certificate grinds the efficiency and accuracy rate for analysis of calling the score.

The present invention is to solve above-mentioned technical problem by following technical proposals:

According to an aspect of the invention, there is provided a kind of security using deep learning model grind report analysis method, packet Include following steps:

S01, the security to be analyzed for receiving input grind report；

S02 grinds report to the security to be analyzed and carries out subordinate sentence processing, obtains security to be analyzed and grinds each subordinate sentence in reporting；

S03 is analysed to each subordinate sentence that security are ground in report and is commented using preparatory trained first deep learning model Point, to obtain grinding the first scoring of each subordinate sentence in report for security to be analyzed, and ground based on the first scoring from security to be analyzed Viewpoint class subordinate sentence is determined in each subordinate sentence in report；

S04 is analysed to each viewpoint class subordinate sentence that security are ground in report and uses preparatory trained second deep learning model Score, with obtain for security to be analyzed grind report in each viewpoint class subordinate sentence second scoring, and based on second scoring from Security to be analyzed grind in each viewpoint class subordinate sentence in report and determine industry viewpoint class subordinate sentence；

S05 is analysed to the every profession and trade viewpoint class subordinate sentence that security are ground in report and uses preparatory trained third deep learning Model scores, to obtain grinding the third scoring of the every profession and trade viewpoint class subordinate sentence in report, the third for security to be analyzed It scores for determining that security to be analyzed grind the emotion trend of the every profession and trade viewpoint class subordinate sentence in report；

S06, according to security to be analyzed grind report in every profession and trade viewpoint class subordinate sentence second scoring and third scoring determine to Analysis security grind the whole emotion trend of report.

Further, S02 grinds report to security to be analyzed and carries out subordinate sentence processing, obtains security to be analyzed and grinds each point in reporting Sentence, comprising:

Report is ground to security to be analyzed according to the symbol of preset type and carries out subordinate sentence processing, security to be analyzed is obtained and grinds in report Each subordinate sentence；

Each subordinate sentence in report is ground to security to be analyzed and carries out word segmentation processing, security to be analyzed is obtained and grinds each participle in reporting；

Based on the preset dictionary including participle with the corresponding conversion relationship of numerical value, it is each in report to determine that security to be analyzed are ground Segment corresponding numerical value；

According to definitive result, it is analysed to security and grinds the subordinate sentence that each subordinate sentence in report is converted into numerical value vector format.

Further, the training process of the first deep learning model, includes the following steps:

Step 110, determine that the first data set, first data set include grinding plucking for report to the security of preset record in advance The multiple subordinate sentences obtained after subordinate sentence processing are partially carried out, wherein each subordinate sentence has the first kind label marked in advance, the One type label includes viewpoint class and non-viewpoint class；

Step 120, each subordinate sentence for being labeled with first kind label is subjected to word segmentation processing, obtains being labeled with first kind mark Each participle of label；

Step 130, each participle for being labeled with first kind label is converted to by corresponding first numerical value according to presetting rule, And by be labeled with first kind label respectively to segment that the first corresponding numerical value is stored in preset include participle and numerical value In the dictionary of corresponding conversion relationship；

Step 140, according to the dictionary, each subordinate sentence for being labeled with first kind label is converted into numerical value vector format Subordinate sentence forms the first numerical value vector subordinate sentence set；

Step 150, the subordinate sentence of the first preset quantity is chosen from the first numerical value vector subordinate sentence set as the first training number According to；

Step 160, the first training data is trained through deep learning model, to obtain the first deep learning model.

Further, the training process of the second deep learning model, includes the following steps:

Step 210, determine that the second data set, second data set include grinding plucking for report to the security of preset record in advance The multiple subordinate sentences obtained after subordinate sentence processing are partially carried out, wherein each subordinate sentence has the first kind label marked in advance, the One type label includes viewpoint class and non-viewpoint class, and is labeled with the subordinate sentence of viewpoint class label while having second marked in advance Type label, Second Type label include industry viewpoint class and non-industry viewpoint class；

Step 220, each subordinate sentence for being labeled with Second Type label is subjected to word segmentation processing, obtains being labeled with Second Type mark Each participle of label；

Step 230, each participle for being labeled with Second Type label is converted to by corresponding second value according to presetting rule, And by be labeled with Second Type label respectively to segment that corresponding second value is stored in preset include participle and numerical value In the dictionary of corresponding conversion relationship；

Step 240, according to the dictionary, each subordinate sentence for being labeled with Second Type label is converted into numerical value vector format Subordinate sentence forms second value vector subordinate sentence set；

Step 250, the subordinate sentence of the second preset quantity is chosen from second value vector subordinate sentence set as the second training number According to；

Step 260, the second training data is trained through deep learning model, to obtain the second deep learning model.

Further, the training process of third deep learning model, includes the following steps:

Step 310, determine that third data set, the third data set include grinding plucking for report to the security of preset record in advance The multiple subordinate sentences obtained after subordinate sentence processing are partially carried out, wherein each subordinate sentence has the first kind label marked in advance, the One type label includes viewpoint class and non-viewpoint class, and is labeled with the subordinate sentence of viewpoint class label while having second marked in advance Type label, Second Type label includes industry viewpoint class and non-industry viewpoint class, and is labeled with point of industry viewpoint class label Sentence has the third type label marked in advance simultaneously, and third type label includes be expected to rise class and class expected to fall；

Step 320, each subordinate sentence for being labeled with third type label is subjected to word segmentation processing, obtains being labeled with third type mark Each participle of label；

Step 330, each participle for being labeled with third type label is converted to by corresponding third value according to presetting rule, And by be labeled with third type label respectively to segment that corresponding third value is stored in preset include participle and numerical value In the dictionary of corresponding conversion relationship；

Step 340, according to the dictionary, each subordinate sentence for being labeled with third type label is converted into numerical value vector format Subordinate sentence forms third value vector subordinate sentence set；

Step 350, the subordinate sentence of third preset quantity is chosen from third value vector subordinate sentence set as third training number According to；

Step 360, third training data is trained through deep learning model, to obtain third deep learning model.

Further, the deep learning model is shot and long term memory network machine learning model.

Further, S06 grinds the second scoring of the every profession and trade viewpoint class subordinate sentence in report according to security to be analyzed and third is commented Divide the whole emotion trend for determining that security to be analyzed grind report, comprising:

It calculates security to be analyzed and grinds the product that the second scoring of each industry viewpoint class subordinate sentence in report is scored with third, as First product value of each industry viewpoint class subordinate sentence；

The sum for calculating the first product value of all industry viewpoint class subordinate sentences, as first and value；

The sum for calculating the second scoring of all industry viewpoint class subordinate sentences, as second and value；

By first and value divided by second and value, the whole emotion scoring that security to be analyzed grind report is obtained；

Judge that whether the security to be analyzed grind the whole emotion scoring of report higher than preset scoring threshold value；

If so, the whole emotion trend for determining that security to be analyzed grind report is to be expected to rise, if not, it is determined that security to be analyzed are ground The whole emotion trend of report is expected to fall.

Further, the method also includes:

The determining entirety that report is ground with security of third scoring of the every profession and trade viewpoint class subordinate sentence in report is ground according to security to be analyzed The consistent industry viewpoint class subordinate sentence of emotion trend；

The industry viewpoint class subordinate sentence that third scoring highest or minimum predetermined number are chosen from definitive result, as wait divide Analysis security grind core views and the output of report.

To achieve the goals above, the present invention also provides a kind of security using deep learning model to grind analysis apparatus of calling the score, Include:

Receiving module, security to be analyzed for receiving input grind report；

Subordinate sentence module carries out subordinate sentence processing for grinding report to security to be analyzed, obtains security to be analyzed and grinds each point in reporting Sentence；

First grading module grinds each subordinate sentence in reporting using preparatory trained first depth for being analysed to security Practise model score, with obtain for security to be analyzed grind report in each subordinate sentence first scoring, and based on first scoring from Security to be analyzed grind in each subordinate sentence in report and determine viewpoint class subordinate sentence；

Second grading module grinds each viewpoint class subordinate sentence in reporting using in advance trained second for being analysed to security Deep learning model scores, to obtain grinding the second scoring of each viewpoint class subordinate sentence in report, and base for security to be analyzed It is ground in each viewpoint class subordinate sentence in report in the second scoring from security to be analyzed and determines industry viewpoint class subordinate sentence；

Third grading module grinds the every profession and trade viewpoint class subordinate sentence in reporting using trained in advance for being analysed to security Third deep learning model scores, and is commented with obtaining grinding the third of the every profession and trade viewpoint class subordinate sentence in report for security to be analyzed Point, the third scoring is for determining that security to be analyzed grind the emotion trend of the every profession and trade viewpoint class subordinate sentence in report；

Emotion trend determining module, for grinding the second scoring of the every profession and trade viewpoint class subordinate sentence in report according to security to be analyzed And third scoring determines that security to be analyzed grind the whole emotion trend of report.

Further, the subordinate sentence module, comprising:

Clause unit grinds report to security to be analyzed for the symbol according to preset type and carries out subordinate sentence processing, obtains wait divide Analysis security grind each subordinate sentence in report；

Participle unit, each subordinate sentence for being ground in report to security to be analyzed carry out word segmentation processing, obtain security to be analyzed and grind Each participle in report；

Numerical value determination unit, for based on it is preset include participle with the corresponding conversion relationship of numerical value dictionary, determine to Analysis security grind the corresponding numerical value of each participle in report；

Converting unit, for being analysed to each subordinate sentence that security are ground in report and being converted into numerical value Vector Lattices according to definitive result The subordinate sentence of formula.

Further, emotion trend determining module, comprising:

First computing unit, for calculate security to be analyzed grind report in each industry viewpoint class subordinate sentence second scoring with The product of third scoring, the first product value as each industry viewpoint class subordinate sentence；

Second computing unit, the sum of the first product value for calculating all industry viewpoint class subordinate sentences, as first and value；

Third computing unit, the sum of the second scoring for calculating all industry viewpoint class subordinate sentences, as second and value；

4th computing unit, for first and value divided by second and value, to be obtained the whole emotion that security to be analyzed grind report Scoring；

Judging unit, for judging that whether security to be analyzed grind the whole emotion scoring of report higher than preset scoring threshold value；

Emotion trend determination unit is when being, to determine that security to be analyzed are ground for the judging result in the judging unit The whole emotion trend of report is to be expected to rise, and when the judging result of the judging unit is no, determines that security to be analyzed grind the whole of report Body emotion trend is expected to fall.

Further, described device, further includes:

Industry viewpoint class subordinate sentence determining module, for grinding the of the every profession and trade viewpoint class subordinate sentence in report according to security to be analyzed The determining consistent industry viewpoint class subordinate sentence of whole emotion trend that report is ground with security of three scorings；

Module is chosen, for choosing the industry viewpoint class of third scoring highest or minimum predetermined number from definitive result Subordinate sentence grinds core views and the output of report as security to be analyzed.

To achieve the goals above, the present invention also provides a kind of computer equipments, including memory, processor and storage On a memory and the computer program that can run on a processor, the processor realize the above method when executing described program The step of.

To achieve the goals above, the present invention also provides a kind of computer readable storage medium, it is stored thereon with computer Program, when described program is executed by processor the step of the realization above method.

Security provided by the invention using deep learning model, which grind to call the score, analyses method, apparatus, computer equipment and readable Storage medium can first be analysed to security and grind report progress subordinate sentence processing, then by each subordinate sentence using preparatory trained first depth Learning model scores to obtain the first of each subordinate sentence the scoring, and judges whether subordinate sentence is viewpoint class point according to the first scoring Then the viewpoint class subordinate sentence judged is used preparatory trained second deep learning model to score to obtain each sight by sentence Second scoring of point class subordinate sentence, and judge whether each viewpoint class subordinate sentence is industry viewpoint class subordinate sentence according to the second scoring, next Preparatory trained third deep learning model is used to score to obtain every profession and trade the industry viewpoint class subordinate sentence judged The third of viewpoint class subordinate sentence scores, and the emotion trend of every profession and trade viewpoint class subordinate sentence is judged according to third scoring, last according to each The second scoring and third scoring of industry viewpoint class subordinate sentence determine that the security to be analyzed grind the whole emotion trend of report.By above-mentioned Scheme can pick out viewpoint class subordinate sentence, industry viewpoint class based on preparatory trained deep learning model by way of scoring Subordinate sentence and the emotion trend for determining industry viewpoint class subordinate sentence, and determine that the security grind the entirety of report eventually by objective scoring Emotion trend, as a result, it is above-mentioned using deep learning model intelligent scoring and analyze process, can not only greatly save manpower, and And analysis efficiency can be improved and analyze the accuracy rate of result.

Detailed description of the invention

Fig. 1 is that the security according to an embodiment of the present invention using deep learning model grind one kind of report analysis method optionally Flow diagram；

Fig. 2 is that the security according to an embodiment of the present invention using deep learning model grind the one kind for analysis apparatus of calling the score optionally Program module schematic diagram；

Fig. 3 is that the security according to an embodiment of the present invention using deep learning model grind the another optional of analysis apparatus of calling the score Program module schematic diagram；

Fig. 4 be the security according to an embodiment of the present invention using deep learning model grind analysis apparatus of calling the score another is optional Program module schematic diagram；

Fig. 5 is a kind of optional hardware structure schematic diagram of computer equipment according to an embodiment of the present invention.

Specific embodiment

In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to the accompanying drawings and embodiments, right The present invention is further elaborated.It should be appreciated that described herein, specific examples are only used to explain the present invention, not For limiting the present invention.Based on the embodiments of the present invention, those of ordinary skill in the art are not before making creative work Every other embodiment obtained is put, shall fall within the protection scope of the present invention.

Embodiment one

Report analysis method is ground to the security provided by the invention using deep learning model with reference to the accompanying drawing to be illustrated.

Fig. 1 is a kind of optional flow diagram that the present invention grinds report analysis method using the security of deep learning model, As shown in Figure 1, this method may comprise steps of:

S01, the security to be analyzed for receiving input grind report.

The extraction of its core views is carried out, when receiving After one security to be analyzed grinds report, the format that first can grind report to this security judges.Specifically, can determine whether this security Whether the format for grinding report is text type, such as " .txt " format, " .doc " format etc..If judging result is no, that is to say, that The format that the security grind report is not text type, then the format that the security can be ground to report is converted to text type, such as the security The format for grinding report is portable document format (PDF format), and tools such as existing " PDFParser " may be used by PDF format The security that security grind report conversion txt format grind report.With this, the security to be analyzed that can uniformly receive grind the format of report, with more square Just and the more efficient content for grinding report to security to be analyzed is read out.

S02 grinds report to security to be analyzed and carries out subordinate sentence processing, obtains security to be analyzed and grinds each subordinate sentence in reporting.

In the present embodiment, report can be ground to security to be analyzed according to the symbol of preset type and carries out subordinate sentence processing, such as can According to comma, ", fullstop ".", point number " ", dash "-", bracket " [], [] ", branch "；" etc. symbols, to be analyzed Security grind report and carry out subordinate sentence processing, grind each subordinate sentence in reporting to obtain security to be analyzed.

Obtain security to be analyzed grind report in each subordinate sentence after, can further using jieba word segmentation module to each subordinate sentence into Row word segmentation processing grinds each participle in reporting to obtain security to be analyzed.

After obtaining each participle that security to be analyzed are ground in report, can based on it is preset include segmenting corresponding with numerical value turn The dictionary for changing relationship, determining that security to be analyzed grind the numerical value corresponding to respectively segmenting in report (is in the present embodiment integer type number Value).

Then, it grinds in report further according to determining as a result, being analysed to security for each corresponding numerical value of participle each by multiple The subordinate sentence of participle composition is converted into the subordinate sentence of numerical value vector format.It, can in the next steps, by point of numerical value vector format with this In sentence input deep learning model, so that deep learning model score to each subordinate sentence and can be according to the type to score to subordinate sentence Judged.

S03 is analysed to each subordinate sentence that security are ground in report and is commented using preparatory trained first deep learning model Point, to obtain grinding the first scoring of each subordinate sentence in report for security to be analyzed, and ground based on the first scoring from security to be analyzed Viewpoint class subordinate sentence is determined in each subordinate sentence in report.

Firstly, being first illustrated to the training process of the first deep learning model, which may include following steps:

Step 110, the first data set is determined.

For example, the security for selecting preset record (such as 300) in advance grind report, for example a certain proportion of viewpoint ratio can be selected It is more apparent, especially industry viewpoint is obvious, clearly security grind report for emotion trend comparison, can also select a certain proportion of sight Point is relatively fuzzyyer, especially industry viewpoint is relatively fuzzyyer, less clearly security grind report etc. to emotion trend, then can be to these cards The abstract part that certificate grinds report carries out subordinate sentence and handles to obtain multiple subordinate sentences, then carries out the first kind to each subordinate sentence manually The mark of type label, the first kind label may include viewpoint class and non-viewpoint class.That is, through first kind label for labelling Afterwards, the subordinate sentence in the first data set may include viewpoint class subordinate sentence and non-viewpoint class subordinate sentence.

Step 120, each subordinate sentence for being labeled with first kind label is subjected to word segmentation processing, obtains being labeled with first kind mark Each participle of label.

For example, can be used existing jieba word segmentation module to each subordinate sentence for being labeled with first kind label in the first data set Word segmentation processing is carried out, to obtain each participle for being labeled with first kind label.

Step 130, each participle for being labeled with first kind label is converted to by corresponding first numerical value according to presetting rule, And by be labeled with first kind label respectively to segment that the first corresponding numerical value is stored in preset include participle and numerical value In the dictionary of corresponding conversion relationship.

Specifically, can be according to presetting rule (such as random transition, or word can be preset according to actual needs to number The transformation rule etc. of value) each participle for being labeled with first kind label is converted to numerical value one by one and (in the present embodiment, can be described as First numerical value), wherein the first numerical value can such as be the numerical value of integer integer type, and can will be labeled with first kind label Respectively participle is stored in the dictionary of the preset corresponding conversion relationship including participle and numerical value with the first numerical value.

In the present embodiment, can by the corresponding conversion relationship for being labeled with each participle of first kind label and the first numerical value with The form of file or tables of data is stored in dictionary, and content can be as shown in table 1.

The corresponding conversion relationship of table 1 participle and numerical value

Participle	Numerical value
		Participle 1	2
Participle 2	175
		Participle 3	50
...	...

Step 140, according to dictionary, each subordinate sentence for being labeled with first kind label is converted into point of numerical value vector format Sentence forms the first numerical value vector subordinate sentence set.

That is, the first kind can be labeled with by each according to the above-mentioned dictionary including participle with the corresponding conversion relationship of numerical value Type label segments the subordinate sentence formed by multiple, is converted into the subordinate sentence of numerical value vector format, and form numerical value vector subordinate sentence set, In the present embodiment, it can be described as the first numerical value vector subordinate sentence set.

Step 150, the subordinate sentence of the first preset quantity is chosen from the first numerical value vector subordinate sentence set as the first training number According to.

Usually before carrying out deep learning model training, the survey for the training set of training and for test can be first determined Examination collection, in the present embodiment, can choose preset quantity from above-mentioned first numerical value vector subordinate sentence set and (in the present embodiment, may be used Referred to as the first preset quantity, such as accounting 90%) subordinate sentence as the first training data, the subordinate sentence of remaining (such as accounting 10%) It then can be used as the first test data.

In the present embodiment, which can be shot and long term memory network (LSTM, Long Short-Term Memory) machine learning model can be configured some key parameters, such as embeding layer size before training (embedding size), hidden layer size (hidden layer size), batch training size (batch size), bulk sample This cycle-index (num epochs), forgetting rate (dropout), activation primitive (activation, classifying type variable uses Sigmoid), the parameters such as loss function (loss, classifying type problem use binary-crossentropy).It then, can be by first Training data input LSTM machine learning model is trained, to obtain the first deep learning model after training.

In addition, also the first test data can be inputted the first depth after training finishes and obtains the first deep learning model Learning model is tested, to obtain the accuracy rate of the first deep learning model, through test of many times, our the first depth Practising accuracy rate of the model in the first test data can reach 85% or so.It is lower than preassigned (such as 70%) in accuracy rate When, can by adjusting parameter, be adjusted etc. modes to training set data and re-start training, to obtain meeting preassigned First deep learning model of accuracy rate, thus trained first depth model of utility can it is more accurate to subordinate sentence into Row scores and judges sentence type.

Can be analysed to each subordinate sentence that security are ground in report as a result, uses above-mentioned first deep learning model to score to obtain To the first scoring of each subordinate sentence ground for security to be analyzed in report, and can be ground in report based on the first scoring from security to be analyzed Viewpoint class subordinate sentence is determined in each subordinate sentence.

In the present embodiment, the scoring which can think for the model, the scoring can be considered probability (0 to 1 it Between), for example model thinks to compare the subordinate sentence for being partial to viewpoint class, then closer to 1 (such as 0.75), model is thought to compare for scoring It is partial to the subordinate sentence of non-viewpoint class, then scores closer to 0 (such as 0.22).

After obtaining the first scoring, it can be compared according to first scoring with pre-set first score threshold, and Viewpoint class subordinate sentence is judged according to comparison result, for example the subordinate sentence that the first scoring is greater than the first score threshold can be determined as viewpoint First scoring is determined as non-viewpoint subordinate sentence no more than the subordinate sentence of the first score threshold by class subordinate sentence.It in the present embodiment, can should First score threshold is set as 0.5, that is to say, that and the first subordinate sentence of the scoring no more than 0.5 can be identified as non-viewpoint class subordinate sentence, It is no longer participate in subsequent step；First subordinate sentence of the scoring greater than 0.5 can be identified as viewpoint class subordinate sentence, can continue to participate in subsequent step Suddenly.

Class of subordinate sentence is analyzed in a manner of intelligent scoring using preparatory trained first deep learning model with this Type (viewpoint class subordinate sentence or non-viewpoint class subordinate sentence), on the one hand, manpower can be greatlyd save, improve analysis efficiency；On the other hand, it can mention The accuracy rate of high analyte result.

S04 is analysed to each viewpoint class subordinate sentence that security are ground in report and uses preparatory trained second deep learning model Score, with obtain for security to be analyzed grind report in each viewpoint class subordinate sentence second scoring, and based on second scoring from Security to be analyzed grind in each viewpoint class subordinate sentence in report and determine industry viewpoint class subordinate sentence.

Firstly, being first illustrated to the training process of the second deep learning model, which may include following steps:

Step 210, the second data set is determined.

For example, selecting preset record (such as 300) security in advance grinds report, for example a certain proportion of viewpoint can be selected and compared Obviously, especially industry viewpoint is obvious, clearly security grind report for emotion trend comparison, can also select a certain proportion of viewpoint Compare that fuzzy, especially industry viewpoint is relatively fuzzy, less clearly security grind report etc. to emotion trend, then can security be ground with report Abstract part carry out subordinate sentence handle to obtain multiple subordinate sentences, then manually to each subordinate sentence carry out first kind label Mark, which may include viewpoint class and non-viewpoint class.That is, after first kind label for labelling, the Subordinate sentence in two data sets may include viewpoint class subordinate sentence and non-viewpoint class subordinate sentence.

Then, the mark of Second Type label, the Second Type mark then are manually carried out to each viewpoint class subordinate sentence Label may include industry viewpoint class and non-industry viewpoint class.That is, after Second Type label for labelling, in the second data set Subordinate sentence may include viewpoint class subordinate sentence and non-viewpoint class subordinate sentence, and further include industry viewpoint class subordinate sentence and non-row in viewpoint class subordinate sentence Industry viewpoint class subordinate sentence.

Step 220, each subordinate sentence for being labeled with Second Type label is subjected to word segmentation processing, obtains being labeled with Second Type mark Each participle of label.

That is, by each subordinate sentence for being labeled with Second Type label in the second data set (namely in the second data set Industry viewpoint class subordinate sentence and non-industry viewpoint class subordinate sentence) carry out word segmentation processing, for example, existing jieba word segmentation module can be used Word segmentation processing is carried out to above-mentioned each subordinate sentence, to obtain each participle for being labeled with Second Type label.

Step 230, each participle for being labeled with Second Type label is converted to by corresponding second value according to presetting rule, And by be labeled with Second Type label respectively to segment that corresponding second value is stored in preset include participle and numerical value In the dictionary of corresponding conversion relationship.

Specifically, can be according to presetting rule (such as random transition, or word can be preset according to actual needs to number The transformation rule etc. of value) each participle for being labeled with Second Type label is converted to numerical value one by one and (in the present embodiment, can be described as Second value), wherein the first numerical value can such as be the numerical value of integer integer type, and can will be labeled with Second Type label Respectively participle is stored in the dictionary of the preset corresponding conversion relationship including participle and numerical value with second value.

In the present embodiment, can by the corresponding conversion relationship for being labeled with each participle of Second Type label and second value with The form of file or tables of data is stored in dictionary, and content can be as shown in Table 1 above.

Step 240, according to dictionary, each subordinate sentence for being labeled with Second Type label is converted into point of numerical value vector format Sentence forms second value vector subordinate sentence set.

That is, second class can be labeled with by each according to the above-mentioned dictionary including participle with the corresponding conversion relationship of numerical value Type label segments the subordinate sentence formed by multiple, is converted into the subordinate sentence of numerical value vector format, and form numerical value vector subordinate sentence set, In the present embodiment, it can be described as second value vector subordinate sentence set.

Step 250, the subordinate sentence of the second preset quantity is chosen from second value vector subordinate sentence set as the second training number According to.

Usually before carrying out deep learning model training, the survey for the training set of training and for test can be first determined Examination collection, in the present embodiment, can choose preset quantity from above-mentioned second value vector subordinate sentence set and (in the present embodiment, may be used Referred to as the second preset quantity, such as accounting 90%) subordinate sentence as the second training data, the subordinate sentence of remaining (such as accounting 10%) It then can be used as the second test data.

In the present embodiment, which can be shot and long term memory network (LSTM, Long Short-Term Memory) machine learning model can be configured some key parameters, such as embeding layer size before training (embedding size), hidden layer size (hidden layer size), batch training size (batch size), bulk sample This cycle-index (num epochs), forgetting rate (dropout), activation primitive (activation, classifying type variable uses Sigmoid), the parameters such as loss function (loss, classifying type problem use binary-crossentropy).It then, can be by second Training data input LSTM machine learning model is trained, to obtain the second deep learning model after training.

In addition, also the second test data can be inputted the second depth after training finishes and obtains the second deep learning model Learning model is tested, to obtain the accuracy rate of the second deep learning model.Through test of many times, our the second depth The accuracy rate practised in the second test data can reach 85% or so.It, can when accuracy rate is lower than preassigned (such as 70%) It by adjusting parameter, it is adjusted etc. modes to training set data re-starts training, to obtain meeting the accurate of preassigned Second deep learning model of rate, so that trained second depth model of utility more accurate can comment subordinate sentence Divide and judges sentence type.

It as a result, can be by each viewpoint class subordinate sentence ground by the security to be analyzed determined in S03 in report using above-mentioned second depth Learning model scores, and to obtain grinding the second scoring of each viewpoint class subordinate sentence in report for security to be analyzed, and can be based on Second scoring grinds in the viewpoint class subordinate sentence in report from security to be analyzed and determines industry viewpoint class subordinate sentence.

In the present embodiment, the scoring which can think for the model, the scoring can be considered probability (0 to 1 it Between), for example model thinks to compare the subordinate sentence for being partial to industry viewpoint class, then closer to 1 (such as 0.85), model is thought for scoring Compare the subordinate sentence for being partial to non-industry viewpoint class, then scores closer to 0 (such as 0.3).It, can basis after obtaining the second scoring Second scoring is compared with pre-set second score threshold, and judges viewpoint class subordinate sentence according to comparison result, than The subordinate sentence that second scoring is greater than the second score threshold can be such as determined as to industry viewpoint class subordinate sentence, the second scoring is not more than second point The subordinate sentence of number threshold value is determined as non-industry viewpoint subordinate sentence.In the present embodiment, 0.5 can be set by second score threshold, That is the second subordinate sentence of the scoring no more than 0.5 can be identified as non-industry viewpoint class subordinate sentence, it is no longer participate in subsequent step；The Two subordinate sentences of the scoring greater than 0.5 can be identified as industry viewpoint class subordinate sentence, can continue to participate in subsequent step.

Viewpoint class point is analyzed in a manner of intelligent scoring using preparatory trained second deep learning model with this The concrete type (industry viewpoint class subordinate sentence or non-industry viewpoint class subordinate sentence) of sentence, on the one hand, manpower can be greatlyd save, improve analysis Efficiency；On the other hand, the accuracy rate of analysis result can be improved.

S05 is analysed to the every profession and trade viewpoint class subordinate sentence that security are ground in report and uses preparatory trained third deep learning Model scores, to obtain grinding the third scoring of the every profession and trade viewpoint class subordinate sentence in report, the third for security to be analyzed It scores for determining that security to be analyzed grind the emotion trend of the every profession and trade viewpoint class subordinate sentence in report.

Firstly, being first illustrated to the training process of third deep learning model, which may include following steps:

Step 310, third data set is determined.

For example, we can grind report by well-chosen preset record (such as 300) security in advance, for example certain proportion can be selected Viewpoint it is obvious, especially industry viewpoint is obvious, clearly security grind report for emotion trend comparison, can also select certain The viewpoint of ratio is relatively fuzzyyer, especially industry viewpoint is relatively fuzzyyer, less clearly security grind report etc. to emotion trend, then may be used Subordinate sentence is carried out to the abstract part that security grind report to handle to obtain multiple subordinate sentences, and the then is carried out to each subordinate sentence manually The mark of one type label, the first kind label may include viewpoint class and non-viewpoint class.That is, through first kind label After mark, the subordinate sentence in the first data set may include viewpoint class subordinate sentence and non-viewpoint class subordinate sentence.

Then, the mark of Second Type label, the Second Type mark then are manually carried out to each viewpoint class subordinate sentence Label may include industry viewpoint class and non-industry viewpoint class.That is, after Second Type label for labelling, in the second data set Subordinate sentence may include viewpoint class subordinate sentence and non-viewpoint class subordinate sentence, while further include industry viewpoint class subordinate sentence and non-row in viewpoint class subordinate sentence Industry viewpoint class subordinate sentence.

Next, carrying out the mark of third type label, the third to each industry viewpoint class subordinate sentence manually again Type label may include be expected to rise class and class expected to fall.That is, the subordinate sentence after third type label mark, in third data set It may include viewpoint class subordinate sentence and non-viewpoint class subordinate sentence, and further include industry viewpoint class subordinate sentence and non-industry viewpoint in viewpoint class subordinate sentence Class subordinate sentence, and further include be expected to rise class industry viewpoint class subordinate sentence and class industry viewpoint class subordinate sentence expected to fall in industry viewpoint class subordinate sentence.

Step 320, each subordinate sentence for being labeled with third type label is subjected to word segmentation processing, obtains being labeled with third type mark Each participle of label.

That is, each subordinate sentence (namely being expected to rise in third data set that third type will be labeled in third data set Class industry viewpoint class subordinate sentence and class industry viewpoint class subordinate sentence expected to fall) carry out word segmentation processing.For example, existing jieba can be used to segment Module carries out word segmentation processing to above-mentioned each subordinate sentence, to obtain each participle for being labeled with third type label.

Step 330, each participle for being labeled with third type label is converted to by corresponding third value according to presetting rule, And by be labeled with third type label respectively to segment that corresponding third value is stored in preset include participle and numerical value In the dictionary of corresponding conversion relationship.

Specifically, can be according to presetting rule (such as random transition, or word can be preset according to actual needs to number The transformation rule etc. of value) each participle for being labeled with third type label is converted to numerical value one by one and (in the present embodiment, can be described as Third value), wherein third value can such as be the numerical value of integer integer type, and can will be labeled with third type label Respectively participle is stored in the dictionary of the preset corresponding conversion relationship including participle and numerical value with third value.

In the present embodiment, can by the corresponding conversion relationship for being labeled with each participle of third type label and third value with The form of file or tables of data is stored in dictionary, and content can be as shown in Table 1 above.

Step 340, according to the dictionary, each subordinate sentence for being labeled with third type label is converted into numerical value vector format Subordinate sentence forms third value vector subordinate sentence set.

That is, third class can be labeled with by each according to the above-mentioned dictionary including participle with the corresponding conversion relationship of numerical value Type label segments the subordinate sentence formed by multiple, is converted into the subordinate sentence of numerical value vector format, and form numerical value vector subordinate sentence set, In the present embodiment, it can be described as third value vector subordinate sentence set.

Step 350, the subordinate sentence of third preset quantity is chosen from third value vector subordinate sentence set as third training number According to.

Usually before carrying out deep learning model training, the survey for the training set of training and for test can be first determined Examination collection, in the present embodiment, can choose preset quantity from above-mentioned third value vector subordinate sentence set and (in the present embodiment, may be used Referred to as third preset quantity, such as accounting 90%) subordinate sentence as third training data, the subordinate sentence of remaining (such as accounting 10%) It then can be used as third test data.

In the present embodiment, which can be shot and long term memory network (LSTM, Long Short-Term Memory) machine learning model can be configured some key parameters, such as embeding layer size before training (embedding size), hidden layer size (hidden layer size), batch training size (batch size), bulk sample This cycle-index (num epochs), forgetting rate (dropout), activation primitive (activation, classifying type variable uses Sigmoid), the parameters such as loss function (loss, classifying type problem use binary-crossentropy).It then, can be by third Training data input LSTM machine learning model is trained, to obtain third deep learning model after training.

In addition, also third test data can be inputted third depth after training finishes and obtains third deep learning model Learning model is tested, to obtain the accuracy rate of the third deep learning model.Through test of many times, our third depth The accuracy rate practised in third test data can reach 85% or so.It, can when accuracy rate is lower than preassigned (such as 70%) It by adjusting parameter, it is adjusted etc. modes to training set data re-starts training, to obtain meeting the accurate of preassigned The third deep learning model of rate, so that the trained third depth model of utility more accurate can comment subordinate sentence Divide and judges sentence type according to demand.

It as a result, can be by the every profession and trade viewpoint class subordinate sentence ground by the security to be analyzed determined in S04 in report using above-mentioned third Deep learning model scores, to obtain grinding the third scoring of the every profession and trade viewpoint class subordinate sentence in report for security to be analyzed, And the emotion trend for determining that security to be analyzed grind the industry viewpoint class subordinate sentence in report that can be scored based on third, namely class industry of being expected to rise Viewpoint class subordinate sentence or class industry viewpoint class subordinate sentence expected to fall.

In the present embodiment, the scoring that third scoring can be thought for the model, the scoring can be considered probability (0 to 1 it Between), for example model thinks the subordinate sentence for comparing class industry viewpoint class of being partial to be expected to rise, then scoring is closer to 1 (such as 0.85), mould Type thinks to compare the subordinate sentence for being partial to class industry viewpoint class expected to fall, then scores closer to 0 (such as 0.3).It is commented obtaining third After point, it can be scored according to the third and be compared with pre-set third score threshold, and further sentenced according to comparison result It is disconnected to be expected to rise or the degree of mood expected to fall, for example, third can be scored be greater than third score threshold subordinate sentence be judged to being expected to rise mood compared with Third scoring is judged to being expected to rise by the lower industry viewpoint class subordinate sentence of mood high or expected to fall no more than the subordinate sentence of third score threshold Mood is lower or the higher industry viewpoint class subordinate sentence of mood expected to fall.In the present embodiment, second score threshold can be set to Between 0.5-0.6, be preferably arranged to 0.6, that is to say, that subordinate sentence of the third scoring greater than 0.6 can be identified as being expected to rise mood compared with The lower industry viewpoint class subordinate sentence of mood high or expected to fall；Subordinate sentence of the third scoring no more than 0.6 can be identified as being expected to rise mood compared with The higher industry viewpoint class subordinate sentence of mood low or expected to fall.

With this, using preparatory trained third deep learning model, obtained in a manner of intelligent scoring each for determining The third class of the emotion trend of industry viewpoint class subordinate sentence scores, and is class industry sight of being expected to rise to be used to analyze every profession and trade viewpoint class subordinate sentence Point class subordinate sentence or class industry viewpoint class subordinate sentence expected to fall, and the degree for mood expected to fall of being expected to rise, on the one hand, people can be greatlyd save Power improves analysis efficiency；On the other hand, the accuracy rate of analysis result can be improved.

In the present embodiment, the second scoring of the every profession and trade viewpoint class subordinate sentence that security to be analyzed are ground in report is obtained in S04, And it after obtaining the third scoring that security to be analyzed grind every profession and trade viewpoint class subordinate sentence in report in S05, comments in combination with the two The whole emotion trend for grinding report to security to be analyzed is divided to assess.

In specific implementation, can first calculate security to be analyzed grind report in each industry viewpoint class subordinate sentence second scoring with The product of third scoring, the first product value as each industry viewpoint class subordinate sentence；The first of all industry viewpoint class subordinate sentences is calculated again The sum of product value, as first and value；Then the sum for calculating the second scoring of all industry viewpoint class subordinate sentences, as second and value； Next, by first and value divided by second and value, with obtain security to be analyzed grind report whole emotion scoring (it can be appreciated that Ups and downs scoring), since the second scoring and third scoring are all between 0 to 1, the security to be analyzed being calculated grind report Whole emotion score be also between 0 to 1.

After security to be analyzed are calculated and grind the whole emotion scoring of report, it can continue to judge that entirety emotion scoring is No to be higher than preset scoring threshold value, which can such as be set as between 0.5-0.6 based on practical experience, be preferably provided with It is 0.6.

If whole emotion scoring is higher than above-mentioned preset scoring threshold value, it can determine that the security to be analyzed grind the whole emotion of report Trend is to be expected to rise, if whole emotion scoring is not higher than above-mentioned preset scoring threshold value, can determine that the security to be analyzed grind the whole of report Body emotion trend is expected to fall.

It, can be by the way that the security to be analyzed that deep learning model is chosen be ground with the of the every profession and trade viewpoint class subordinate sentence in report with this Two scorings (can be used for determining whether for more apparent industry viewpoint) and third scoring (can be used for determining the feelings of every profession and trade viewpoint Sense trend) carry out COMPREHENSIVE CALCULATING mode, obtain security to be analyzed grind report whole emotion scoring, and pass through the security to be analyzed The comparison result for grinding whole the emotion scoring and desired indicator of report, obtains the whole emotion trend that security to be analyzed grind report, thus Available more objective and accurate emotion trend analysis result.

In addition, can also be ground to security to be analyzed in report after determining that security to be analyzed grind the whole emotion trend of report Core views extract.

In specific implementation, for example, can be analysed to security grind every profession and trade viewpoint class subordinate sentence in report according to its corresponding the Three are divided into two parts, and 0.6 part, and two parts are not more than including part of the third scoring greater than 0.6 and third scoring It can be ranked up with positive sequence (score value is from big to small).As a result, can according to the third of every profession and trade viewpoint class subordinate sentence score determine with Security grind the consistent industry viewpoint class subordinate sentence of whole emotion trend of report.

That is, third can be scored greater than 0.6 if the whole emotion trend that security to be analyzed grind report is to be expected to rise Industry viewpoint class subordinate sentence is determined as grinding the consistent industry viewpoint class subordinate sentence of whole emotion trend of report with security to be analyzed；If to The whole emotion trend that analysis security grind report is industry viewpoint class subordinate sentence expected to fall, then third can scoring no more than 0.6, is determined For the consistent industry viewpoint class subordinate sentence of whole emotion trend for grinding report with security to be analyzed.

After the consistent industry viewpoint class subordinate sentence of whole emotion trend for grinding report with security to be analyzed has been determined, if wait divide The whole emotion trend that analysis security grind report is to be expected to rise, and can choose third and score the industry viewpoint class subordinate sentence of highest predetermined number, Core views and the output of report are ground as security to be analyzed；If security to be analyzed grind report whole emotion trend be it is expected to fall, it is optional It takes third to score the industry viewpoint class subordinate sentence of minimum predetermined number, core views and the output of report is ground as security to be analyzed. Wherein the predetermined number can be configured according to actual needs, in the present embodiment, for example may be configured as 2.

That is, can score and be greater than in above-mentioned third if the whole emotion trend that security to be analyzed grind report is to be expected to rise The industry viewpoint class subordinate sentence that 2 thirds scoring highest (being ordered as first, second) is chosen in 0.6 part waits for point as this Analysis security grind core views and the output of report；If security to be analyzed grind report whole emotion trend be it is expected to fall, can be above-mentioned the Three scorings are no more than the industry viewpoint class for choosing 2 thirds scoring minimum (be ordered as last, second) in 0.6 part Subordinate sentence grinds core views and the output of report as the security to be analyzed.

With this, it can select that grind the whole emotion trend of report with security to be analyzed consistent and most according to objective appraisal result The subordinate sentence of core views is represented, to can guarantee the accuracy for choosing result, in order to which user accurately understands the security to be analyzed Grind the core views of report.

According to each embodiment of the present embodiment, it can first be analysed to security and grind report progress subordinate sentence processing, then by each point Sentence uses preparatory trained first deep learning model to score to obtain the first of each subordinate sentence the scoring, and comments according to first Divide and judge whether subordinate sentence is viewpoint class subordinate sentence, then by the viewpoint class subordinate sentence judged using preparatory trained second depth It practises model to score to obtain the second of each viewpoint class subordinate sentence the scoring, and whether each viewpoint class subordinate sentence is judged according to the second scoring For industry viewpoint class subordinate sentence, the industry viewpoint class subordinate sentence judged next is used into preparatory trained third deep learning mould Type scores to obtain the scoring of the third of every profession and trade viewpoint class subordinate sentence, and judges every profession and trade viewpoint class subordinate sentence according to third scoring Emotion trend, finally according to the second of every profession and trade viewpoint class subordinate sentence scoring and third score determine the security to be analyzed grind report Whole emotion trend.Through the above scheme, it can be picked out by way of scoring based on preparatory trained deep learning model Viewpoint class subordinate sentence, industry viewpoint class subordinate sentence and the emotion trend for determining industry viewpoint class subordinate sentence, and eventually by objectively commenting Point determine that the security grind the whole emotion trend of report, it is above-mentioned using deep learning model intelligent scoring and the process analyzed as a result, Manpower can be not only greatlyd save, and analysis efficiency can be improved and analyze the accuracy rate of result.

Through the above description of the embodiments, those skilled in the art can be understood that above-described embodiment side Method can be realized by means of software and necessary general hardware platform, naturally it is also possible to by hardware, but in many cases The former is more preferably embodiment.

Embodiment two

The security using deep learning model provided in one based on the above embodiment grind report analysis method, in the present embodiment A kind of security using deep learning model are provided and grind analysis apparatus of calling the score, specifically, Fig. 2 to 4 shows this and utilizes deep learning The security of model grind the optional structural block diagram for analysis apparatus of calling the score, which grinds analysis apparatus quilt of calling the score One or more program modules are divided into, one or more program module is stored in storage medium, and by one or more Performed by a processor, to complete the present invention.The so-called program module of the present invention is to refer to complete a series of of specific function Computer program instructions section, is more suitable for describing being ground analysis apparatus of calling the score using the security of deep learning model and is storing than program itself Implementation procedure in medium, the function of each program module of the present embodiment will specifically be introduced by being described below.

As shown in Fig. 2, the security using deep learning model grind analysis apparatus 20 of calling the score can include:

Receiving module 21, the security to be analyzed that can receive input grind report；

Subordinate sentence module 22 can be used for grinding security to be analyzed report and carry out subordinate sentence processing, obtains security to be analyzed and grinds in report Each subordinate sentence；

It is deep using in advance trained first to can be used for being analysed to each subordinate sentence that security are ground in report for first grading module 23 Degree learning model scores, and to obtain grinding the first scoring of each subordinate sentence in report for security to be analyzed, and comments based on first Divide in each subordinate sentence ground from security to be analyzed in report and determines viewpoint class subordinate sentence；

Second grading module 24 can be used for being analysed to security and grind each viewpoint class subordinate sentence in reporting using trained in advance Second deep learning model scores, to obtain grinding the second scoring of each viewpoint class subordinate sentence in report for security to be analyzed, And it is ground in each viewpoint class subordinate sentence in report based on the second scoring from security to be analyzed and determines industry viewpoint class subordinate sentence；

Third grading module 25 can be used for being analysed to the every profession and trade viewpoint class subordinate sentence that security are ground in report and be trained using preparatory Good third deep learning model scores, to obtain grinding the of the every profession and trade viewpoint class subordinate sentence in report for security to be analyzed Three scorings, the third scoring is for determining that security to be analyzed grind the emotion trend of the every profession and trade viewpoint class subordinate sentence in report；

Emotion trend determining module 26 can be used for grinding second of the every profession and trade viewpoint class subordinate sentence in report according to security to be analyzed Scoring and third scoring determine that security to be analyzed grind the whole emotion trend of report.

Further, shown in referring to Fig. 3, subordinate sentence module 22 be may particularly include:

Clause unit 221 can be used for grinding report progress subordinate sentence processing to security to be analyzed according to the symbol of preset type, obtain Security to be analyzed grind each subordinate sentence in report；

Participle unit 222 can be used for grinding security to be analyzed each subordinate sentence in report and carry out word segmentation processing, obtains card to be analyzed Certificate grinds each participle in report；

Numerical value determination unit 223 can be used for based on the preset dictionary including participle with the corresponding conversion relationship of numerical value, really Fixed security to be analyzed grind the corresponding numerical value of each participle in report；

Converting unit 224, can be used for being analysed to according to definitive result security grind report in each subordinate sentence be converted into numerical value to Measure the subordinate sentence of format.

In the present embodiment, the training process of the first deep learning model, it may include following steps:

Step 140, according to dictionary, each subordinate sentence for being labeled with first kind label is converted into point of numerical value vector format Sentence forms the first numerical value vector subordinate sentence set；

In the present embodiment, the training process of the second deep learning model, it may include following steps:

Step 210, the second data set is determined, wherein the second data set includes grinding report to the security of preset record in advance Abstract part carries out the multiple subordinate sentences obtained after subordinate sentence processing, wherein each subordinate sentence has the first kind label marked in advance, First kind label includes viewpoint class and non-viewpoint class, and is labeled with the subordinate sentence of viewpoint class label while having marked in advance Two type labels, Second Type label include industry viewpoint class and non-industry viewpoint class；

Step 240, according to dictionary, each subordinate sentence for being labeled with Second Type label is converted into point of numerical value vector format Sentence forms second value vector subordinate sentence set；

In the present embodiment, the training process of third deep learning model, it may include following steps:

Step 310, third data set is determined, wherein third data set includes grinding report to the security of preset record in advance Abstract part carries out the multiple subordinate sentences obtained after subordinate sentence processing, wherein each subordinate sentence has the first kind label marked in advance, First kind label includes viewpoint class and non-viewpoint class, and is labeled with the subordinate sentence of viewpoint class label while having marked in advance Two type labels, Second Type label includes industry viewpoint class and non-industry viewpoint class, and is labeled with industry viewpoint class label Subordinate sentence has the third type label marked in advance simultaneously, and third type label includes be expected to rise class and class expected to fall；

Step 340, according to dictionary, each subordinate sentence for being labeled with third type label is converted into point of numerical value vector format Sentence forms third value vector subordinate sentence set；

Further, deep learning model is shot and long term memory network machine learning model.

In addition, emotion trend determining module 26 may particularly include referring to shown in Fig. 4:

First computing unit 261 grinds second commenting for each industry viewpoint class subordinate sentence in report for calculating security to be analyzed Divide the product to score with third, the first product value as each industry viewpoint class subordinate sentence；

Second computing unit 262, the sum of the first product value for calculating all industry viewpoint class subordinate sentences, as the first He Value；

Third computing unit 263, the sum of the second scoring for calculating all industry viewpoint class subordinate sentences, as the second He Value；

4th computing unit 264, for first and value divided by second and value, to be obtained the whole feelings that security to be analyzed grind report Sense scoring；

Judging unit 265, for judging that whether security to be analyzed grind the whole emotion scoring of report higher than preset scoring threshold value；

Emotion trend determination unit 266 can be used for determining card to be analyzed when the judging result of judging unit 265 is to be The whole emotion trend that certificate grinds report is to be expected to rise, and when the judging result of judging unit 265 is no, determines that security to be analyzed grind report Whole emotion trend is expected to fall.

Further, which grinds the analytical equipment of report, may also include that

Module is chosen, for choosing the industry viewpoint class of third scoring highest or minimum predetermined number from definitive result Subordinate sentence grinds core views and the output of report as security.

About the device in above-described embodiment, wherein each unit, module execute the concrete mode of operation related It is described in detail in the embodiment of this method, no detailed explanation will be given here.

Each embodiment through this embodiment can first be analysed to security and grind report progress subordinate sentence processing, then by each point Sentence uses preparatory trained first deep learning model to score to obtain the first of each subordinate sentence the scoring, and comments according to first Divide and judge whether subordinate sentence is viewpoint class subordinate sentence, then by the viewpoint class subordinate sentence judged using preparatory trained second depth It practises model to score to obtain the second of each viewpoint class subordinate sentence the scoring, and whether each viewpoint class subordinate sentence is judged according to the second scoring For industry viewpoint class subordinate sentence, the industry viewpoint class subordinate sentence judged next is used into preparatory trained third deep learning mould Type scores to obtain the scoring of the third of every profession and trade viewpoint class subordinate sentence, and judges every profession and trade viewpoint class subordinate sentence according to third scoring Emotion trend, finally according to the second of every profession and trade viewpoint class subordinate sentence scoring and third score determine the security to be analyzed grind report Whole emotion trend.Through the above scheme, it can be picked out by way of scoring based on preparatory trained deep learning model Viewpoint class subordinate sentence, industry viewpoint class subordinate sentence and the emotion trend for determining industry viewpoint class subordinate sentence, and eventually by objectively commenting Point determine that the security grind the whole emotion trend of report, it is above-mentioned using deep learning model intelligent scoring and the process analyzed as a result, Manpower can be not only greatlyd save, and analysis efficiency can be improved and analyze the accuracy rate of result.

Embodiment three

The present embodiment also provides a kind of computer equipment, can such as execute the smart phone, tablet computer, notebook of program Computer, desktop computer, rack-mount server, blade server, tower server or Cabinet-type server are (including independent Server cluster composed by server or multiple servers) etc..As shown in figure 5, the computer equipment 50 of the present embodiment to It is few to include but is not limited to: memory 51, the processor 52 of connection can be in communication with each other by system bus, as shown in Figure 5.It needs to refer to Out, Fig. 5 illustrates only the computer equipment 50 with component 51-52, it should be understood that being not required for implementing all The component shown, the implementation that can be substituted is more or less component.

In the present embodiment, memory 51 (i.e. readable storage medium storing program for executing) includes flash memory, hard disk, multimedia card, card-type memory (for example, SD or DX memory etc.), random access storage device (RAM), static random-access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read only memory (PROM), magnetic storage, magnetic Disk, CD etc..In some embodiments, memory 51 can be the internal storage unit of computer equipment 50, such as the calculating The hard disk or memory of machine equipment 50.In further embodiments, memory 51 is also possible to the external storage of computer equipment 50 The plug-in type hard disk being equipped in equipment, such as the computer equipment 50, intelligent memory card (Smart Media Card, SMC), peace Digital (Secure Digital, SD) card, flash card (Flash Card) etc..Certainly, memory 51 can also both include meter The internal storage unit for calculating machine equipment 50 also includes its External memory equipment.In the present embodiment, memory 51 is commonly used in storage It is installed on the operating system and types of applications software of computer equipment 50, such as the card using deep learning model of embodiment two Certificate grinds the program code etc. for analysis apparatus of calling the score.In addition, memory 51 can be also used for temporarily storing and export or will The Various types of data of output.

Processor 52 can be in some embodiments central processing unit (Central Processing Unit, CPU), Controller, microcontroller, microprocessor or other data processing chips.The processor 52 is commonly used in control computer equipment 50 overall operation.In the present embodiment, program code or processing data of the processor 52 for being stored in run memory 51, Such as analysis apparatus of calling the score etc. is ground using the security of deep learning model.

Example IV

The present embodiment also provides a kind of computer readable storage medium, such as flash memory, hard disk, multimedia card, card-type memory (for example, SD or DX memory etc.), random access storage device (RAM), static random-access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read only memory (PROM), magnetic storage, magnetic Disk, CD, server, App are stored thereon with computer program, phase are realized when program is executed by processor using store etc. Answer function.The computer readable storage medium of the present embodiment is used to grind analysis apparatus of calling the score, quilt using the security of deep learning model Realize that the security using deep learning model of embodiment one grind report analysis method when processor executes.

The serial number of the above embodiments of the invention is only for description, does not represent the advantages or disadvantages of the embodiments.

The above is only a preferred embodiment of the present invention, is not intended to limit the scope of the invention, all to utilize this hair Equivalent structure or equivalent flow shift made by bright specification and accompanying drawing content is applied directly or indirectly in other relevant skills Art field, is included within the scope of the present invention.

Claims

1. a kind of security using deep learning model grind report analysis method, which comprises the steps of:

S01, the security to be analyzed for receiving input grind report；

S02 grinds report to security to be analyzed and carries out subordinate sentence processing, obtains security to be analyzed and grinds each subordinate sentence in reporting；

S03 is analysed to each subordinate sentence that security are ground in report and is scored using preparatory trained first deep learning model, with It obtains grinding the first scoring of each subordinate sentence in report for security to be analyzed, and is ground in report based on the first scoring from security to be analyzed Viewpoint class subordinate sentence is determined in each subordinate sentence；

S04 is analysed to each viewpoint class subordinate sentence that security are ground in report and is carried out using trained second deep learning model in advance Scoring, to obtain being directed to the second scoring of each viewpoint class subordinate sentence that security to be analyzed are ground in report, and based on the second scoring from wait divide Analysis security grind in each viewpoint class subordinate sentence in report and determine industry viewpoint class subordinate sentence；

S05 is analysed to the every profession and trade viewpoint class subordinate sentence that security are ground in report and uses preparatory trained third deep learning model It scores, to obtain grinding the third scoring of the every profession and trade viewpoint class subordinate sentence in report, the third scoring for security to be analyzed For determining that security to be analyzed grind the emotion trend of the every profession and trade viewpoint class subordinate sentence in report；

S06, according to security to be analyzed grind report in every profession and trade viewpoint class subordinate sentence second scoring and third scoring determination it is to be analyzed Security grind the whole emotion trend of report.

2. the security according to claim 1 using deep learning model grind report analysis method, which is characterized in that S02 pairs Security to be analyzed grind report and carry out subordinate sentence processing, obtain security to be analyzed and grind each subordinate sentence in reporting, comprising:

Report is ground to security to be analyzed according to the symbol of preset type and carries out subordinate sentence processing, security to be analyzed is obtained and grinds each point in reporting Sentence；

Based on the preset dictionary including participle with the corresponding conversion relationship of numerical value, determine that security to be analyzed grind each participle in report Corresponding numerical value；

3. the security according to claim 1 using deep learning model grind report analysis method, which is characterized in that first is deep The training process for spending learning model, includes the following steps:

Step 110, determine that the first data set, first data set include the abstract portion that in advance security of preset record are ground with report Divide the multiple subordinate sentences for carrying out obtaining after subordinate sentence processing, wherein each subordinate sentence has the first kind label marked in advance, the first kind Type label includes viewpoint class and non-viewpoint class；

Step 120, each subordinate sentence for being labeled with first kind label is subjected to word segmentation processing, obtains being labeled with first kind label Each participle；

Step 130, each participle for being labeled with first kind label is converted to by corresponding first numerical value according to presetting rule, and will Be labeled with first kind label respectively segment the first corresponding numerical value be stored in it is preset include segment it is corresponding with numerical value In the dictionary of transformational relation；

Step 140, according to the dictionary, each subordinate sentence for being labeled with first kind label is converted into point of numerical value vector format Sentence forms the first numerical value vector subordinate sentence set；

Step 150, the subordinate sentence of the first preset quantity is chosen as the first training data from the first numerical value vector subordinate sentence set；

4. the security according to claim 1 using deep learning model grind report analysis method, which is characterized in that second is deep The training process for spending learning model, includes the following steps:

Step 210, determine that the second data set, second data set include the abstract portion that in advance security of preset record are ground with report Divide the multiple subordinate sentences for carrying out obtaining after subordinate sentence processing, wherein each subordinate sentence has the first kind label marked in advance, the first kind Type label includes viewpoint class and non-viewpoint class, and is labeled with the subordinate sentence of viewpoint class label while having the Second Type marked in advance Label, Second Type label include industry viewpoint class and non-industry viewpoint class；

Step 220, each subordinate sentence for being labeled with Second Type label is subjected to word segmentation processing, obtains being labeled with Second Type label Each participle；

Step 230, each participle for being labeled with Second Type label is converted to by corresponding second value according to presetting rule, and will Be labeled with Second Type label respectively segment corresponding second value be stored in it is preset include segment it is corresponding with numerical value In the dictionary of transformational relation；

Step 240, according to the dictionary, each subordinate sentence for being labeled with Second Type label is converted into point of numerical value vector format Sentence forms second value vector subordinate sentence set；

Step 250, the subordinate sentence of the second preset quantity is chosen from second value vector subordinate sentence set as the second training data；

5. the security according to claim 1 using deep learning model grind report analysis method, which is characterized in that third is deep The training process for spending learning model, includes the following steps:

Step 310, determine that third data set, the third data set include the abstract portion that in advance security of preset record are ground with report Divide the multiple subordinate sentences for carrying out obtaining after subordinate sentence processing, wherein each subordinate sentence has the first kind label marked in advance, the first kind Type label includes viewpoint class and non-viewpoint class, and is labeled with the subordinate sentence of viewpoint class label while having the Second Type marked in advance Label, Second Type label includes industry viewpoint class and non-industry viewpoint class, and the subordinate sentence for being labeled with industry viewpoint class label is same When with the third type label marked in advance, third type label includes be expected to rise class and class expected to fall；

Step 320, each subordinate sentence for being labeled with third type label is subjected to word segmentation processing, obtains being labeled with third type label Each participle；

Step 330, each participle for being labeled with third type label is converted to by corresponding third value according to presetting rule, and will Be labeled with third type label respectively segment corresponding third value be stored in it is preset include segment it is corresponding with numerical value In the dictionary of transformational relation；

Step 340, according to the dictionary, each subordinate sentence for being labeled with third type label is converted into point of numerical value vector format Sentence forms third value vector subordinate sentence set；

Step 350, the subordinate sentence of third preset quantity is chosen from third value vector subordinate sentence set as third training data；

6. grinding report analysis method using the security of deep learning model according to claim 3 or 4 or 5, which is characterized in that The deep learning model is shot and long term memory network machine learning model.

7. the security according to claim 1 using deep learning model grind report analysis method, which is characterized in that S06 root The second scoring of the every profession and trade viewpoint class subordinate sentence in report is ground according to security to be analyzed and third scoring determines that security to be analyzed grind report Whole emotion trend, comprising:

It calculates security to be analyzed and grinds the product that the second scoring of each industry viewpoint class subordinate sentence in report is scored with third, as each First product value of industry viewpoint class subordinate sentence；

If so, the whole emotion trend for determining that security to be analyzed grind report is to be expected to rise, if not, it is determined that security to be analyzed grind report Whole emotion trend is expected to fall.

8. the security according to claim 1 using deep learning model grind report analysis method, which is characterized in that also wrap It includes:

The determining whole emotion that report is ground with security of third scoring of the every profession and trade viewpoint class subordinate sentence in report is ground according to security to be analyzed The consistent industry viewpoint class subordinate sentence of trend；

The industry viewpoint class subordinate sentence that third scoring highest or minimum predetermined number are chosen from definitive result, as card to be analyzed Certificate grinds the core views of report and output.

9. a kind of security using deep learning model grind analysis apparatus of calling the score characterized by comprising

Receiving module, security to be analyzed for receiving input grind report；

Subordinate sentence module carries out subordinate sentence processing for grinding report to the security to be analyzed, obtains security to be analyzed and grinds each point in reporting Sentence；

First grading module grinds each subordinate sentence in reporting using preparatory trained first deep learning mould for being analysed to security Type scores to obtain grinding the first scoring of each subordinate sentence in report for security to be analyzed, and scores based on first to be analyzed Security grind in each subordinate sentence in report and determine viewpoint class subordinate sentence；

Second grading module grinds each viewpoint class subordinate sentence in reporting using preparatory trained second depth for being analysed to security Learning model scores to obtain grinding the second scoring of each viewpoint class subordinate sentence in report for security to be analyzed, and is based on second It scores to grind from security to be analyzed and determines industry viewpoint class subordinate sentence in each viewpoint class subordinate sentence in reporting；

Third grading module grinds the every profession and trade viewpoint class subordinate sentence in reporting using preparatory trained third for being analysed to security Deep learning model scores to obtain grinding the third scoring of the every profession and trade viewpoint class subordinate sentence in report, institute for security to be analyzed Third scoring is stated for determining that security to be analyzed grind the emotion trend of the every profession and trade viewpoint class subordinate sentence in report；

Emotion trend determining module, for grinding the second scoring and the of the every profession and trade viewpoint class subordinate sentence in report according to security to be analyzed Three scorings determine that security to be analyzed grind the whole emotion trend of report.

10. a kind of computer equipment, the computer equipment include memory, processor and storage on a memory and can be The computer program run on processor, which is characterized in that the processor realizes claim 1 to 8 when executing described program The step of any one the method.

11. a kind of computer readable storage medium, is stored thereon with computer program, it is characterised in that: described program is processed The step of any one of claim 1 to 8 the method is realized when device executes.