A policy text interpretation method and device

Abstract

The present application relates to the technical field of deep learning, and provides a policy text interpretation method and device, which comprises the following steps: extracting a core sentence to be analyzed from a policy text to be analyzed; inputting the core sentence to be analyzed and a historical core sentence of each historical policy text as a set of reasoning comparison data into a pre-trained core sentence comparison model to predict the classification of the reasoning comparison data; and calculating a quantitative evaluation value of the core sentence to be analyzed according to the classification of the reasoning comparison data and the quantitative evaluation value of the historical core sentence in the reasoning comparison data; wherein the core sentence comparison model is trained by using historical comparison data composed of historical core sentences of each two historical policy texts and classification labels of the historical comparison data to obtain a comparison deep learning network model. The present application can improve the accuracy and efficiency of quantitative evaluation of the policy text to be analyzed, and further improve the accuracy of the prediction of the tightness degree of the policy text to be analyzed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, in particular to a policy text interpretation method and device. BACKGROUND

[0002] In the prior art, the tightness of the policy text for subsequent operation has a strong guiding significance. For example, the financial policy text (including meeting and speech content) issued by the financial regulatory agency has a strong forward-looking significance for subsequent financial operation. More specifically, the monetary policy issued by the central bank of each country has a strong forward-looking guiding effect on whether the subsequent monetary policy of other countries is to raise interest rates or lower interest rates. There are two popular methods in the academic field: the first method is to screen the keywords in the policy text, and then quantize by cumulative scoring mode; the second method is to use machine learning algorithm to split the policy file by sentence, and then label each sentence as positive, negative or neutral.

[0003] Taking the interpretation of the monetary policy file as an example, after classifying each sentence, the quantization score of this monetary policy report is: quantization score = (# number of positive sentences - # number of negative sentences) / # total number of sentences.

[0004] This score is between [-1, 1], if the quantization score is 0, it means that the central bank's attitude towards tightness is neutral; if the quantization score is 1, it means that the central bank's attitude towards tightness is very tight (positive classification), and it may raise interest rates significantly; if the score is -1, it means that the central bank's attitude towards tightness is very loose (negative), and it may significantly lower interest rates.

[0005] The existing technical solution gives the same weight to the hundreds of sentences in the report, which contains a considerable amount of text on policy operations that are not important. This "rough reading" of all sentences in the report and then three-classification operation has a low upper limit of quantization score accuracy, and is easily affected by the distribution of sentences in the report. In addition, this quantization method does not combine the actual action after the policy is released, and there is a problem that the quantization score does not match the actual situation.

[0006] In addition, due to the small amount of text in some policy texts, it is difficult to construct a reasonable modeling method for quantization interpretation on a small sample. The accuracy of the existing keyword scoring mode is not high enough, and deep learning requires large amounts of training data, which is a labor-intensive and difficult to annotate work. SUMMARY

[0007] The present application is used to solve the problems of low accuracy and low efficiency in the quantization interpretation of existing policy texts.

[0008] To solve the above technical problems, the first aspect of the present application provides a policy text interpretation method, comprising:

[0009] extracting a core sentence to be analyzed from a policy text to be analyzed;

[0010] inputting the core sentence to be analyzed and the historical core sentence of each historical policy text as a set of reasoning comparison data into a pre-trained core sentence comparison model, and predicting the classification of the reasoning comparison data;

[0011] According to the classification of the reasoning comparison data and the quantitative evaluation value of the historical core sentence in the reasoning comparison data, the quantitative evaluation value of the core sentence to be analyzed is calculated;

[0012] The core sentence comparison model is trained by using the historical comparison data composed of two historical core sentences of each historical policy text and the classification label of the historical comparison data; the classification label of the historical comparison data is determined according to the difference between the quantitative evaluation values of the two historical core sentences in the historical comparison data, and the quantitative evaluation value of the core sentence reflects the tightness of the policy adjustment action.

[0013] As a further embodiment of the present application, the core sentence to be analyzed is extracted from the policy text to be analyzed, comprising:

[0014] splitting the policy text to be analyzed to obtain a plurality of sentences;

[0015] inputting each sentence into a pre-trained binary classification model to obtain the classification result of each sentence; wherein the binary classification model is trained by using each sentence in the historical policy text and the labeled label of each sentence, and the labeled label includes a core sentence label and a non-core sentence label;

[0016] According to the sentence with the core sentence label in the classification result, the core sentence to be analyzed is determined.

[0017] As a further embodiment of the present application, the core sentence comparison model establishment process comprises:

[0018] extracting historical core sentences from each historical policy text;

[0019] collecting action indicator values after each historical policy text is published;

[0020] According to the action indicator values after each historical policy text is published, the quantitative evaluation value of each historical core sentence is calculated;

[0021] Each two historical core sentences is taken as a historical comparison data, and the classification label of each set of historical comparison data is determined according to the difference between the quantitative evaluation values of the two historical core sentences in each set of historical comparison data;

[0022] According to historical comparison data and classification labels of the historical comparison data, the contrast deep learning network model is trained to obtain a core sentence contrast model.

[0023] As a further embodiment of the present application, the deep learning network model comprises a Bert pre-training model and a classifier.

[0024] The Bert pre-training model is used to extract semantic vectors in the core sentence.

[0025] The classifier is used to classify according to the semantic vectors of the core sentence.

[0026] As a further embodiment of the present application, the quantitative evaluation value of each historical core sentence is calculated according to the action indicator value after each historical policy text is published, comprising:

[0027] According to all the action indicator values after the publication of each historical policy text, the normalized values of each action indicator after the publication of each historical policy text are calculated.

[0028] For each historical policy text, the normalized values of all action indicators of the historical policy text are averaged to obtain the quantitative evaluation value of the historical core sentence of the historical policy text.

[0029] As a further embodiment of the present application, the normalized values of each action indicator after the publication of each historical policy text are calculated according to all the action indicator values after the publication of each historical policy text, comprising:

[0030] The normalized values of each action indicator after the publication of each historical policy text are calculated by the following formula:

[0031]

[0032] v· is the normalized action indicator value, v is the action indicator value after the publication of the historical policy text, is the median of the positive action indicator values in all historical policy texts, is the median of the negative action indicator values in all historical policy texts.

[0033] As a further embodiment of the present application, the classification label of each set of historical comparison data is determined according to the difference between the quantitative evaluation values of the two historical core sentences in each set of historical comparison data, comprising:

[0034] The classification label of the historical comparison data is determined by the following formula:

[0035]

[0036] Wherein, (a, b) is historical comparison data, a and b are historical core sentences, (a, b)Label is the label of historical comparison data (a, b), and ε is a tolerance value.

[0037] As a further embodiment of the present application, the quantitative evaluation value of the core sentence to be analyzed is calculated according to the classification of the inference comparison data and the quantitative evaluation value of the historical core sentence in the inference comparison data, comprising:

[0038] From all the inference comparison data, the inference comparison data of a preset classification is screened out, and the preset classification represents that the tightness degree corresponding to the core sentence in the inference comparison data is similar;

[0039] The quantitative evaluation values of the historical core sentences in the screened inference comparison data are weighted and averaged to obtain the quantitative evaluation value of the core sentence of the policy text to be analyzed.

[0040] As a further embodiment of the present application, the policy text interpretation method further comprises:

[0041] According to the quantitative evaluation value of the core sentence to be analyzed, the estimated adjustment range of the policy text to be analyzed is calculated by reverse normalization.

[0042] The second aspect of the present application provides a policy text interpretation device, comprising:

[0043] The core sentence extraction unit is configured to extract the core sentence to be analyzed from the policy text to be analyzed;

[0044] The comparison prediction unit is configured to input the core sentence to be analyzed and the historical core sentence of each historical policy text as a set of inference comparison data into a pre-trained core sentence comparison model to predict the classification of the inference comparison data.

[0045] The evaluation unit is configured to calculate the quantitative evaluation value of the core sentence to be analyzed according to the classification of the inference comparison data and the quantitative evaluation value of the historical core sentence in the inference comparison data.

[0046] Wherein, the core sentence comparison model is trained by using the historical comparison data composed of the historical core sentences of each two historical policy texts and the classification label of the historical comparison data to obtain a comparison deep learning network model; the classification label of the historical comparison data is determined according to the difference between the quantitative evaluation values of the two historical core sentences in the historical comparison data, and the quantitative evaluation value of the core sentence reflects the tightness degree of the policy adjustment action.

[0047] The third aspect of the present application provides a computer device, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to realize the method of any one of the preceding embodiments.

[0048] The fourth aspect of the present application provides a computer readable storage medium, the computer readable storage medium stores a computer program, the computer program is executed by the processor of the computer device to realize the method of any one of the preceding embodiments.

[0049] The fifth aspect of the present application provides a computer program product, the computer program product comprises a computer program, the computer program is executed by the processor of the computer device to realize the method of any one of the preceding embodiments.

[0050] The policy text interpretation method and device provided by the present application considers that the core sentence in the policy text is used to express the subsequent position attitude in practice, and then affects the policy adjustment action. Based on this, the core sentence to be analyzed is extracted from the policy text to be analyzed, and the policy text is interpreted based on the core sentence. This can simplify the processing amount of the text to be analyzed and improve the processing efficiency. By forming historical comparison data from every two historical core sentences of historical policy texts, determining the classification label of the historical comparison data according to the difference of the quantitative evaluation values of the two historical core sentences in the historical comparison data, the sample data enhancement can be realized, and then the comparison deep learning training model is implemented on the small sample. Based on the historical comparison data and the classification label thereof, the core sentence comparison model is trained, the change attitude of the policy influence can be more accurately grasped, the internal correlation between the two core sentences and the difference of the quantitative evaluation values of the two core sentences is mined, the accuracy and efficiency of the quantitative evaluation of the policy text to be analyzed are improved, and then the accuracy of the prediction of the looseness degree of the policy text to be analyzed is improved.

[0051] In summary, the present application can overcome the problem of low precision of the core sentence comparison model caused by small sample by the core sentence extraction + comparison mode quantitative evaluation method, and can accurately capture the difference between the core sentences by the self-control mechanism based on interaction of two core sentences, thereby improving the accuracy of the analysis of the policy text to be analyzed.

[0052] In order to make the above and other objects, features and advantages of the present application more apparent, the following preferred embodiments are described in detail below, and the accompanying drawings are described as follows. BRIEF DESCRIPTION OF DRAWINGS

[0053] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0054] Figure 1 The flow chart of the core sentence comparison model establishment process of the embodiment of the present application is shown;

[0055] Figure 2 A structural diagram of a deep learning network model compared with an embodiment of the application is shown;

[0056] Figure 3 An allocation diagram of a training set and a test set of an embodiment of the application is shown;

[0057] Figure 4 A first flowchart of a policy text interpretation method of an embodiment of the application is shown;

[0058] Figure 5 A flowchart of a core sentence extraction process of an embodiment of the application is shown;

[0059] Figure 6 A flowchart of a quantitative evaluation value calculation process of a core sentence to be analyzed of an embodiment of the application is shown;

[0060] Figure 7 A second flowchart of a policy text interpretation method of an embodiment of the application is shown;

[0061] Figure 8 A structural diagram of a policy text interpretation device of an embodiment of the application is shown;

[0062] Figure 9 A structural diagram of a computer device of an embodiment of the application is shown.

[0063] Explanation of the drawing symbols:

[0064] 201, a Bert pre-training model;

[0065] 202, a classifier;

[0066] 801, a core sentence extraction unit;

[0067] 802, a comparison prediction unit;

[0068] 803, an evaluation unit;

[0069] 804, a reverse normalization unit;

[0070] 902, a computer device;

[0071] 904, a processor;

[0072] 906, a memory;

[0073] 908, a driving mechanism;

[0074] 910, an input / output module;

[0075] 912, an input device;

[0076] 914, an output device;

[0077] 916, presentation device;

[0078] 918, graphical user interface;

[0079] 920, network interface;

[0080] 922, communication link;

[0081] 924, communication bus. DETAILED DESCRIPTION

[0082] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all the other embodiments obtained by those skilled in the art without creative work are within the protection scope of the present application.

[0083] It should be noted that the terms "first", "second" and the like in the description and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, device, product or apparatus that includes a series of steps or units does not necessarily have to include only those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to the process, method, product or apparatus.

[0084] The present specification provides method operation steps as described in the embodiments or flowcharts, but more or fewer operation steps can be included based on routine or non-creative work. The order of steps listed in the embodiments is only one of the many execution orders of the steps, and does not represent the only execution order. In actual system or device product execution, the method order shown in the embodiments or drawings can be executed in sequence or in parallel.

[0085] It should be noted that the policy text interpretation method and device of the present application can be used for financial policy texts (such as monetary policy texts) published in the financial field, and can also be used for policy documents and meeting minutes published by any government regulatory agency with obvious core paragraphs (i.e. expressing attitude and position) other than the financial field. The specific field of the policy text of the present application is not limited.

[0086] It should be noted that the data involved in the present application (including but not limited to data for analysis, stored data, displayed data, etc.) are all information and data authorized by the user or fully authorized by all parties, and the acquisition, transmission, storage, use and processing of related data comply with relevant laws, regulations and standards of relevant countries and regions.

[0087] It should be noted that in the embodiments of the present application, some existing industry solutions such as software, components, models, etc. may be mentioned, which should be considered as exemplary, and the purpose is only to illustrate the feasibility of the implementation of the technical solutions of the present application, but it does not mean that the applicant has or will necessarily use the solution.

[0088] In the prior art, the interpretation of policy documents is mainly realized by keyword matching quantization and sentence-by-sentence interpretation, and the existing solutions have the problems of low accuracy and low efficiency.

[0089] In addition, the present inventors have found that some policy texts have descriptions that can express core policy positions (referred to as core sentences), and after the policy texts are published, these core sentences will affect subsequent action indicators, which comprehensively reflect the tightness of the policy, for example, the action indicators corresponding to the financial policy text include the reserve ratio of large banks and small and medium-sized banks, the one-year medium-term lending facility (MLF) interest rate, and the 7-day and 14-day reverse repurchase interest rate. These indicators can reflect whether the central bank's monetary policy is relaxed or tightened.

[0090] And considering that the existing policy text is a small sample event, for example, a financial policy text, if the core sentence of the policy text and the tightness of the policy text are directly used to establish a network model, the model accuracy will be affected.

[0091] In order to solve the problems of low accuracy and low efficiency of the existing policy text solution and the above-mentioned technical problems found by the present inventors in the research and development process, the present application first establishes a core sentence comparison model, and then performs policy text interpretation based on the core sentence comparison model.

[0092] As shown in FIG. 1, Figure 1 the core sentence comparison model establishment process includes:

[0093] Step 101, extracting historical core sentences from each historical policy text.

[0094] In this step, the historical policy text is a policy document of various formats for a certain matter, such as a monetary policy document in pdf, word, etc. format, which can be obtained from the central bank website. The specific matter is determined according to the actual application scene, and the present application does not limit the specific content of the policy text.

[0095] In implementation, the core sentence generally mentions the future position and trend, and the historical core sentence can be extracted manually from the historical policy text. In some real-time manners, the historical core sentence includes multiple continuous sentences.

[0096] For example, the core sentence in the historical monetary policy text in the second quarter of 2023 is “The prudent monetary policy should be accurate and powerful, and better play the dual functions of the total amount and structure of monetary policy tools to support the recovery and development of the real economy. A variety of monetary policy tools should be used comprehensively to maintain reasonable liquidity and keep the growth rate of money supply and social financing scale basically matching the nominal economic growth rate.”

[0097] Step 102, collect the action indicator values after the publication of each historical policy text.

[0098] In this step, the action indicator is an indicator that affects the degree of tightness, and is obtained based on data statistics after the publication of the policy text. For example, for the monetary and financial policy text, the corresponding action indicators include the reserve ratio of large banks and small and medium-sized banks, the one-year medium-term lending facility (MLF) interest rate, and the 7-day and 14-day reverse repurchase interest rates.

[0099] In detail, the action indicator values collected in this step are the action indicator values in a predetermined time period after the publication of the historical policy text, for example, the action indicator values in three months after the publication of the policy text.

[0100] Step 103, calculate the quantitative evaluation value of each historical core sentence according to the action indicator values after the publication of each historical policy text.

[0101] In the implementation of this step, first, the normalized values of each action indicator after the publication of each historical policy text are calculated according to the action indicator values after the publication of each historical policy text. Then, for each historical policy text, the normalized values of all action indicators of the historical policy text are averaged to obtain the quantitative evaluation value of the historical core sentence of the historical policy text.

[0102] Specifically, the normalized values of each action indicator after the publication of each historical policy text can be calculated by the following formula:

[0103]

[0104] where v· is the normalized action indicator value, v is the action indicator value after the publication of the historical policy text, is the median of the positive action indicator values in all historical policy texts, is the median of the negative action indicator values in all historical policy texts.

[0105] By the above formula, each action indicator can be normalized to [-1, 1]. For the example in step 103, if the maximum value of the historical interest rate reduction indicator is 75 BP, the median value of the interest rate reduction interval is 50 BP, and the interest rate reduction amplitude in the current time interval is 25 BP, then the normalized score is max(-1, -0.25), that is, -0.25.

[0106] For the example in step 103, for the above five indicators, each indicator has a positive action median value and a negative action median value. After normalizing each action indicator value, the average value of the normalized values of the five indicators in the same time interval is obtained, and the average value is taken as the quantitative evaluation value of the historical core sentence.

[0107] The quantitative evaluation value of the historical core sentence comprehensively reflects the comprehensive influence of the core sentence on each action indicator after the historical policy text is released. For example, the monetary policy document expresses the subsequent position attitude through the core sentence, and the quantitative evaluation value of the core sentence in the monetary policy document is determined according to the actual operation action of the central bank, which can reflect the position of the regulatory agency and to a large extent indicates what action will be taken next.

[0108] In step 104, each two historical core sentences are taken as a historical comparison data, and the classification label of each group of historical comparison data is determined according to the difference between the quantitative evaluation values of the two historical core sentences in each group of historical comparison data.

[0109] Suppose that 87 historical core sentences are obtained, then 87*87, that is, 7569 historical comparison data will be obtained through this step, which correspond to 87*87 classification labels, for example, the classification label of the historical comparison data can be represented as C i,j , wherein i and j are the labels of the historical comparison data, and the classification labels of all historical comparison data can be represented as:

[0110]

[0111] , wherein N is the total number of historical core sentences.

[0112] In some embodiments, the classification label of each group of historical comparison data is determined according to the difference between the quantitative evaluation values of the two historical core sentences in each group of historical comparison data, including:

[0113] The classification label of the historical comparison data is determined by the following formula:

[0114]

[0115] Wherein, (a, b) is historical comparison data, a and b are historical core sentences, (a, b)Label is the label of historical comparison data (a, b), and epsilon is a tolerance value. In some embodiments, the value of epsilon can be set to 0.1. In actual implementation, the value of epsilon can be set according to actual needs, and the specific value of epsilon is not limited in the present application.

[0116] When the label of historical comparison data (a, b) is 1, it indicates that the core sentence a (causing policy action) is more tightening than the core sentence b (causing policy action). When the label is -1, it indicates that the core sentence a (causing policy action) is more relaxing than the core sentence b (causing policy action). When the label is 0, it indicates that the tightening degree of the core sentence a (causing policy action) is similar to that of the core sentence b (causing policy action).

[0117] In step 105, the comparison deep learning network model is trained according to the historical comparison data and the classification label of the historical comparison data, and a core sentence comparison model is obtained.

[0118] The core sentence comparison model established in the embodiment can more accurately grasp the change attitude of policy influence, mine the internal correlation between the two core sentences and the quantitative evaluation value difference of the two core sentences, improve the accuracy of quantitative evaluation of the policy text to be analyzed, and further improve the accuracy of the tightening degree prediction caused by the policy text to be analyzed.

[0119] In an embodiment of the present application, the step 101 is executed before the above-mentioned step 101, which further includes parsing and sentence splitting of historical policy texts. Specifically, all historical policy texts are synchronously acquired, and the acquired texts are parsed. According to the structure of the texts, a structured chapter, section and sentence splitting is performed by using a policy expression processing scheme, and special styles such as columns, abstracts and icons are standardized.

[0120] Taking a monetary policy document as an example, the document can include chapters, abstracts, charts, columns and the like. Since the formats of policy text reports published by different government regulatory agencies are different, the embodiment develops a targeted parsing algorithm for the format of a specific report, which can improve the parsing efficiency of the policy text.

[0121] In an embodiment of the present application, as shown in Figure 2 The comparison deep learning network model includes a Bert pre-training model 201 and a classifier 202.

[0122] The Bert pre-training model 201 is used to extract semantic vectors in the core sentences. The classifier 202 is used to classify according to the semantic vectors of the core sentences. Specifically, the classifier 202 is implemented by using an attention mechanism.

[0123] The comparative deep learning network model of the embodiment adopts an open source sentence-transformer, which is an algorithm framework specially fine-tuned for semantic vectors of sentences based on the open source Transformer and Bert algorithms of Google. Compared with the semantic representation of the sentence vector based on the representation, the interaction-based training framework with higher accuracy is adopted. This framework can simultaneously receive two sentences separated by SEP, and uses a self-attention mechanism to accurately capture the subtle differences between the two sentences.

[0124] In an embodiment of the present application, in order to improve the extraction efficiency of subsequent core sentences, after determining the historical core sentences of the historical policy text in the above step 102, a two-classification model for core sentence recognition is established. Specifically, the two-classification model is obtained by training using each sentence and the annotation label of each sentence in the historical policy text, and the annotation label includes a core sentence label and a non-core sentence label.

[0125] In a specific example of the present application, from 2001 to the first quarter of 2023, there are 89 monetary policy reports, and based on this, 89 core sentences are extracted. We mark all non-core sentences in the 89 reports as "non-core sentences", and mark the 89 core sentences as "core sentences", and then train a two-classification model for core sentences by fine-tuning the Bert model. Based on the two-classification model, the core sentences of future monetary policy reports can be extracted.

[0126] Based on the manual extraction of core sentences, the two-classification model for core sentence recognition is trained, which can improve the efficiency of extracting core sentences in subsequent policy documents.

[0127] In a specific embodiment of the present application, the above Figure 1When the method is applied to monetary policy texts, a core sentence comparison model between the core sentences of the monetary policy texts and the central bank's real monetary policy operations (such as interest rate hikes and reductions) after the release of the monetary policy texts can be established, based on which the quantitative evaluation of future monetary policy texts and the future adjustment direction and amplitude can be predicted. The release interval of the monetary policy report is generally about three months. For example, the central bank generally releases the monetary policy report of the fourth quarter of the previous year in February and the monetary policy report of the first quarter of the current year in May. In this embodiment, the central bank's actions in this time interval are taken as the main observation indicators, such as a 0.5 percentage point reduction in the reserve ratio and a 25BP interest rate reduction. In specific implementation, the following five indicators of the reserve ratio of large and small banks, the one-year medium-term lending facility (MLF) interest rate, and the 7-day and 14-day reverse repurchase interest rates are mainly detected to judge whether the central bank's monetary policy (operation) is relaxed or tightened. The selection basis of the five indicators is that the reserve ratio controls the bank's lending ability and affects the long-term liquidity of the market; the medium-term lending facility adjusts the medium-term liquidity of the bank and provides a relatively flexible financing method; and the reverse repurchase interest rate affects the short-term market interest rate and is suitable for quickly adjusting the short-term funding supply. The central bank generally adjusts the long-term, medium-term, and short-term policy stance through these three policy tools.

[0128] Since there is no standard method for combining multiple indicators into one indicator, in this embodiment, the change amplitude of the central bank's action in the three-month time range before the release of the next report after the release of each monetary policy report is taken as the initial variable (such as a 25BP interest rate reduction). Then, the above normalization formula is used to normalize each action indicator change amplitude, so that the value of each action indicator change amplitude can be normalized to [-1, 1]. Finally, the average of the normalized values of the five indicators in the current time interval is taken as the final quantitative evaluation value of the current monetary policy core sentence.

[0129] From 2001 to the first quarter of 2023, a total of 89 monetary policy texts were sampled, excluding 2 special scenes in 2008. Each two core sentences were taken as a historical comparison data, and the core sentences in the historical comparison data were cross-labeled with each other to form a 87*87 label matrix, that is, 7569 training samples, each of which includes two core sentences and the corresponding label.

[0130] The label matrix is shown below:

[0131]

[0132] The label of the comparison pair (sentence a, sentence b) composed of two core sentences is calculated by the following formula

[0133]

[0134] ε is a tolerance value, we set it to 0.1. Where Label is 1, indicating that sentence a is more tightening than sentence b, when Label is -1, indicating that sentence a is more relaxed than sentence b, when Label is 0, then the two sentences are similar in tightness.

[0135] The labeling formula can refer to the foregoing embodiments.

[0136] For the split of the training set and the validation set, in order to ensure that the core sentence comparison model does not see the core sentence appearing in the test set at all during training, we process it as shown in Figure 3 The first 78 core sentences are compared with each other to form the training set. The 79th to 87th core sentences and other sentences are compared to form the test set.

[0137] Finally, our training and test data are shown in Table 1:

[0138] Table 1

[0139]

[0140] Using the training set to train Figure 2 The comparison neural network model shown in the core sentence comparison model. Using the test set to test the accuracy of the core sentence comparison model, if the accuracy is not met, adjust the structure of the comparison neural network model and retrain.

[0141] On the basis of establishing the core sentence comparison model, the interpretation of the policy text can be executed. Specifically, as shown in Figure 4 The policy text interpretation method includes:

[0142] Step 401, extracting the core sentence to be analyzed from the policy text to be analyzed.

[0143] Step 402, input the core sentence to be analyzed and the historical core sentence of each historical policy text as a set of reasoning comparison data into the pre-trained core sentence comparison model, and predict the classification of the reasoning comparison data.

[0144] Taking the monetary policy text as an example, there are 87 historical core sentences, and 87 reasoning comparison data and 87 reasoning comparison data classifications are obtained.

[0145] The classification of the reasoning comparison data includes three categories, represented by -1, 1 and 0, -1 represents that the action after the core sentence of the policy text to be analyzed is released is more relaxed than the action after the core sentence of the historical policy text is released (referred to as more relaxed), 1 represents that the action after the core sentence of the policy text to be analyzed is released is more tightened than the action after the core sentence of the historical policy text is released (referred to as more tightened), and 0 represents that the action after the core sentence of the policy text to be analyzed is released is similar to the action after the core sentence of the historical policy text is released (referred to as similar in tightness).

[0146] In step 403, the quantitative evaluation value of the core sentence to be analyzed is calculated according to the classification of the reasoning comparison data and the quantitative evaluation value of the historical core sentence in the reasoning comparison data.

[0147] The core sentence comparison model is trained by using historical comparison data composed of two historical core sentences of each historical policy text and a classification label of the historical comparison data, and the classification label of the historical comparison data is determined according to the difference between the quantitative evaluation values of the two historical core sentences in the historical comparison data, and the quantitative evaluation value of the core sentence reflects the tightness of the policy adjustment action.

[0148] Based on the core sentence comparison model, the reading efficiency and accuracy of the policy text to be analyzed can be improved.

[0149] As shown in FIG. 5, the step 401 of extracting the core sentence to be analyzed from the policy text to be analyzed includes the following steps. Figure 5

[0150] In step 501, the policy text to be analyzed is split to obtain a plurality of sentences.

[0151] In this step, the structured chapter segmentation and sentence splitting can be performed by using a regular expression processing method according to the structure of the policy text.

[0152] In step 502, each sentence is input into a pre-trained binary classification model to obtain a classification result of each sentence.

[0153] The binary classification model is trained by using each sentence in the historical policy text and a labeled label of each sentence, and the labeled label includes a core sentence label and a non-core sentence label. The training process of the binary classification model can refer to the foregoing embodiments, which will not be described in detail here.

[0154] In step 503, the core sentence to be analyzed is determined according to the sentence with a core sentence label.

[0155] In this step, a plurality of adjacent sentences with core sentence labels are determined as the core sentence to be analyzed. If there is a sentence with a non-core sentence label between two sentences with core sentence labels, the sentence with the non-core sentence label is also set as a core sentence label. ​

[0156] The embodiment can improve the efficiency of core sentence recognition of the policy text to be analyzed through the setting of the binary classification model.

[0157] In an embodiment of the present application, as shown in Figure 6 The step 403 calculates the quantitative evaluation value of the core sentence to be analyzed according to the classification of the inference comparison data and the quantitative evaluation value of the historical core sentence in the inference comparison data, and includes:

[0158] In step 601, the inference comparison data of a preset classification is screened out from all the inference comparison data, and the preset classification represents that the core sentences in the inference comparison data are similar in tightness. For example, the inference comparison data with a label of 0 in the foregoing embodiment.

[0159] In step 602, the quantitative evaluation values of the historical core sentences in the screened inference comparison data are weighted and averaged to obtain the quantitative evaluation value of the core sentence of the policy text to be analyzed.

[0160] In some embodiments, the quantitative evaluation score of the core sentence in the policy text to be analyzed can also be determined in the following manner:

[0161] Suppose that the quantitative evaluation score of the core sentence in the policy text to be analyzed is x, and the value range of x is [-1, 1]. All possible values in the range [-1, 1] are enumerated, such as -1, -0.99, -0.98, …, and 1. The enumeration value that is most consistent with the strong-weak comparison of the score of the historical core sentence and the label of -1, 0, and 1 predicted by the algorithm (the same proportion is the highest) is taken as the final score.

[0162] In an embodiment of the present application, as shown in Figure 7 In addition to the steps 401 to 403, the policy text interpretation method also includes:

[0163] In step 404, the estimated adjustment range of the policy text to be analyzed is reversely normalized according to the quantitative evaluation value of the core sentence to be analyzed.

[0164] The estimated adjustment range is calculated according to the following formula:

[0165] Estimated adjustment range = median of historical action indicators * 2 * quantitative evaluation value of core sentence to be analyzed.

[0166] The quantitative evaluation value of the core sentence to be analyzed can be understood as a degree ratio. If it is 0.5, the value is the median in history, and if it is 0.9, it is 1.8 times the median in history.

[0167] For example, the quantitative evaluation value of the core sentence to be analyzed is -0.3, which means that there is a possibility of interest rate reduction. According to the algorithm mentioned in the third step of the training phase, the median of the interest rate reduction interval is 50BP, so the future interest rate reduction range is calculated in reverse from the single indicator of interest rate reduction, that is, the interest rate will be reduced by 30BP.

[0168] When the application is applied to the monetary policy text scene, the F1 of the deep learning algorithm on the training set is about 0.8, and considering the complexity of predicting the central bank's tightening action in the next 90 days, this indicator has good application effect. In the third quarter of 2023, the algorithm will be deployed to the production environment, and in the subsequent three quarters of the central bank's tightening prediction scene, it has reached 100% accuracy in the direction of the central bank's tightening action, and about 70% accuracy in the prediction of the tightening range. The application can provide the accuracy of policy text classification, quantitative analysis and adjustment range prediction.

[0169] Based on the same inventive concept, the application also provides a policy text interpretation device, as described in the following embodiments. Since the principle of solving problems of the policy text interpretation device is similar to that of the policy text interpretation method, the implementation of the device can be referred to the policy text interpretation method, and the repeated parts will not be described again.

[0170] Specifically, as shown in Figure 8 The policy text interpretation device includes:

[0171] The core sentence extraction unit 801 is configured to extract a core sentence to be analyzed from a policy text to be analyzed.

[0172] The comparison prediction unit 802 is configured to input the core sentence to be analyzed and a historical core sentence of each historical policy text as a set of reasoning comparison data into a pre-trained core sentence comparison model, and predict a classification of the reasoning comparison data.

[0173] The evaluation unit 803 is configured to calculate a quantitative evaluation value of the core sentence to be analyzed according to the classification of the reasoning comparison data and a quantitative evaluation value of the historical core sentence in the reasoning comparison data.

[0174] The core sentence comparison model is obtained by training a comparison deep learning network model using historical comparison data composed of two historical core sentences of each historical policy text and classification labels of the historical comparison data, and the classification labels of the historical comparison data are determined according to the difference between the quantitative evaluation values of the two historical core sentences in the historical comparison data, and the quantitative evaluation value of the core sentence reflects the tightening degree of the policy adjustment action.

[0175] Further, the method further comprises: a reverse normalization unit 804 configured to calculate an estimated adjustment range of the policy text to be analyzed according to the quantitative evaluation value of the core sentence to be analyzed.

[0176] The embodiment can more accurately grasp the changing attitude of policy influence, mine the internal correlation between the two core sentences and the quantitative evaluation value difference of the two core sentences, improve the accuracy of quantitative evaluation of the policy text to be analyzed, and further improve the accuracy of the tightness degree prediction of the policy text to be analyzed.

[0177] The policy text interpretation method and device can be applied to a financial policy text scene, and the application adopts core sentence extraction + comparison mode quantitative scoring to interpret and quantify the financial policy.

[0178] Because the speech text of the financial regulatory agency is generally a small amount of word adjustment on the template of the past core sentence, the expression mode of the policy core sentence is very close, and the difference of some core words is very significant.

[0179] In summary, the prior art generally concentrates on treating all sentences of the policy equally without distinguishing the core policy semantic sentences.

[0180] And in the modeling mode, only the frequency of the key words or the occurrence times of the sentence classification can be quantified, and the upper limit of the accuracy is not high.

[0181] In the embodiment of the application, a computer device is further provided. Figure 9As shown, computer device 902 can include one or more processors 904, such as one or more central processing units (CPUs), each of which can implement one or more hardware threads. Computer device 902 can also include any memory 906 for storing any kind of information, such as code, settings, data, and the like. Without limitation, for example, memory 906 can include any one or combination of: any type of RAM, any type of ROM, a flash memory device, a hard disk, an optical disk, and the like. More generally, any memory can use any technology for storing information. Further, any memory can provide volatile or non-volatile retention of information. Further, any memory can represent a fixed or removable component of computer device 902. In one case, computer device 902 can perform any operation of associated instructions stored in any memory or combination of memories when processor 904 executes the associated instructions. Computer device 902 can also include one or more drive mechanisms 908, such as a hard disk drive mechanism, an optical disk drive mechanism, and the like, for interacting with any memory.

[0182] Computer device 902 can also include an input / output module 910 (I / O) for receiving various inputs (via input devices 912) and for providing various outputs (via output devices 914). One particular output mechanism can include a presentation device 916 and associated graphical user interface 918 (GUI). In other embodiments, input / output module 910 (I / O), input devices 912, and output devices 914 can also not be included, merely as a computer device in a network. Computer device 902 can also include one or more network interfaces 920 for exchanging data with other devices via one or more communication links 922. One or more communication buses 924 couple the above-described components together.

[0183] Communication links 922 can be implemented in any manner, such as through a local area network, a wide area network (e.g., the Internet), a point-to-point connection, and the like, or any combination thereof. Communication links 922 can include any combination of hardwired links, wireless links, routers, gateway functionality, name servers, and the like governed by any protocol or combination of protocols.

[0184] The embodiments of the present application further provide a computer readable storage medium, having stored thereon a computer program, which, when executed by a processor, performs the steps of the above method.

[0185] The embodiments of the present application further provide a computer readable instruction, wherein when a processor executes the instruction, the program therein causes the processor to perform the method of any of the preceding embodiments.

[0186] It should be understood that the size of the sequence number of the above processes does not mean the order of execution in various embodiments of the present application, and the execution order of the processes should be determined according to its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0187] It should also be understood that in the embodiments of the present application, the term "and / or" is only a description of the association relationship of the associated objects, which means that there can be three relationships. For example, A and / or B can represent three cases of A alone, A and B together, and B alone. In addition, the character " / " in the present application generally represents that the front and rear associated objects are in an "or" relationship.

[0188] Those skilled in the art can realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed in the present application can be realized in electronic hardware, computer software or a combination of both. In order to clearly illustrate the interchangeability of hardware and software, the components and steps of the examples have been described in the above description in general terms. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. A person skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0189] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described system, device and unit can refer to the corresponding processes in the foregoing method embodiments, which will not be described here.

[0190] In several embodiments provided by the present application, it should be understood that the disclosed system, device and method can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division, and there can be another division manner in actual implementation. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interfaces, devices or units, and can also be electrical, mechanical or other forms of connection.

[0191] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e. they can be located in one place, or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiments of the present application.

[0192] In addition, each function unit in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software function unit.

[0193] When the integrated unit is realized in the form of a software function unit and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the part of the prior art that contributes to the present application, or all or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, including a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various program code storage media.

[0194] The principles and implementation manners of the present application are described in the specific embodiments in the present application. The above embodiment description is only used to help understand the method of the present application and its core idea; meanwhile, for those skilled in the art, according to the idea of the present application, the specific implementation manner and application range will be changed; in summary, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A method for interpreting a policy text, characterized by, The application comprises the following steps: extracting a core sentence to be analyzed from a policy text to be analyzed; inputting the core sentence to be analyzed and a historical core sentence of each historical policy text into a pre-trained core sentence comparison model as a set of reasoning comparison data, and predicting a classification of the reasoning comparison data; calculating a quantitative evaluation value of the core sentence to be analyzed according to the classification of the reasoning comparison data and the quantitative evaluation value of the historical core sentence in the reasoning comparison data, comprising: screening the reasoning comparison data of a preset classification from all the reasoning comparison data, wherein the preset classification represents that the core sentences in the reasoning comparison data are similar in tightness; and performing weighted average processing on the quantitative evaluation values of the historical core sentences in the screened reasoning comparison data to obtain the quantitative evaluation value of the core sentence of the policy text to be analyzed; wherein the core sentence comparison model is obtained by training a comparison deep learning network model using historical comparison data composed of historical core sentences of each two historical policy texts and classification labels of the historical comparison data; the classification labels of the historical comparison data are determined according to the difference between the quantitative evaluation values of the two historical core sentences in the historical comparison data, and the quantitative evaluation value of the core sentence reflects the tightness of the policy adjustment action; wherein the core sentence comparison model is established by the following steps: extracting historical core sentences from each historical policy text; collecting action indicator values after each historical policy text is published, wherein the action indicator is an indicator that affects the tightness, and the action indicator values are obtained based on data statistics after the policy document is published; calculating normalized values of each action indicator after each historical policy text is published according to the action indicator values after each historical policy text is published; for each historical policy text, performing average processing on the normalized values of all the action indicators of the historical policy text to obtain the quantitative evaluation value of the historical core sentence of the historical policy text; determining a classification label of each set of historical comparison data according to the difference between the quantitative evaluation values of the two historical core sentences in each set of historical comparison data; training a comparison deep learning network model according to the historical comparison data and the classification labels of the historical comparison data to obtain the core sentence comparison model; wherein the deep learning network model comprises a Bert pre-training model and a classifier; the Bert pre-training model is used to extract semantic vectors in the core sentence; the classifier is used to classify according to the semantic vectors of the core sentence.

2. The method of claim 1, wherein, extracting a core sentence to be analyzed from a policy text to be analyzed, comprising: splitting the policy text to be analyzed to obtain multiple sentences; inputting each sentence into a pre-trained binary classification model to obtain classification results of each sentence; wherein the binary classification model is trained using sentences in historical policy texts and labeled labels of the sentences, and the labeled labels comprise core sentence labels and non-core sentence labels; determining the core sentence to be analyzed according to the sentences with core sentence labels in the classification results.

3. The method of claim 1, wherein, calculating normalized values of each action indicator after each historical policy text is published according to the action indicator values after each historical policy text is published, comprising: calculating the normalized values of each action indicator after each historical policy text is published by using the following formula: ； wherein, is the normalized action indicator value, v is the action indicator value after the historical policy text is released, is the median of the positive action indicator values in all historical policy texts, is the median of the negative action indicator values in all historical policy texts.

4. The method of claim 1, wherein, According to the difference between the quantitative evaluation values of the two historical core sentences in each set of historical comparison data, the classification label of each set of historical comparison data is determined, including: The classification label of the historical comparison data is determined by using the following formula: ； Wherein, (a, b) is historical comparison data, a and b are historical core sentences, (a, b) Label is a label of historical comparison data (a, b), is a tolerance value.

5. The method of claim 1, wherein, Also includes: According to the quantitative evaluation value of the to-be-analyzed core sentence, the estimated adjustment range of the to-be-analyzed policy text is calculated by reverse normalization.

6. A policy text interpretation device, characterized by, Including: The core sentence extraction unit is used to extract the to-be-analyzed core sentence from the to-be-analyzed policy text; The comparison prediction unit is used to input the to-be-analyzed core sentence and the historical core sentence of each historical policy text as a set of reasoning comparison data into the pre-trained core sentence comparison model, and predict the classification of the reasoning comparison data; The evaluation unit is used to calculate the quantitative evaluation value of the to-be-analyzed core sentence according to the classification of the reasoning comparison data and the quantitative evaluation value of the historical core sentence in the reasoning comparison data, including: from all reasoning comparison data, screening reasoning comparison data of a preset classification, the preset classification represents that the core sentence in the reasoning comparison data is similar in tightness; the quantitative evaluation values of the historical core sentences in the screened reasoning comparison data are weighted and averaged to obtain the quantitative evaluation value of the core sentence of the to-be-analyzed policy text; Wherein, the core sentence comparison model is obtained by training a comparison deep learning network model using historical comparison data composed of historical core sentences of each two historical policy texts and classification labels of historical comparison data; the classification label of the historical comparison data is determined according to the difference between the quantitative evaluation values of the two historical core sentences in the historical comparison data, and the quantitative evaluation value of the core sentence reflects the tightness of the policy adjustment action; Wherein, the core sentence comparison model establishment process includes: Extracting historical core sentences from each historical policy text; Collecting action indicator values after each historical policy text is published, wherein the action indicator is an indicator that affects the tightness, and is obtained based on data statistics after the policy document is published; According to the normalized values of each action indicator after all historical policy texts are published, the normalized values of each action indicator after each historical policy text is published are calculated; For each historical policy text, the normalized values of all action indicators of the historical policy text are averaged to obtain the quantitative evaluation value of the historical core sentence of the historical policy text; Each two historical core sentences are taken as a historical comparison data, and the classification label of each set of historical comparison data is determined according to the difference between the quantitative evaluation values of the two historical core sentences in each set of historical comparison data; According to the historical comparison data and the classification label of the historical comparison data, a comparison deep learning network model is trained to obtain a core sentence comparison model; Wherein, the deep learning network model includes: a Bert pre-training model and a classifier; The Bert pre-training model is used to extract the semantic vector in the core sentence; The classifier is used for classification according to the semantic vector of the core sentence.

7. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to realize the method of any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, the computer-readable storage medium comprising: The computer program is executed by the processor of the computer device to realize the method of any one of claims 1 to 5.

9. A computer program product comprising a computer program, characterized in that, The computer program, when executed by a processor of a computer device, implements the method of any one of claims 1 to 5.

Images