Investment research report processing method and device based on deep learning and electronic equipment

Abstract

The present application relates to the technical field of data processing, and more particularly to a kind of based on deep learning's investment research report processing method, device and electronic equipment, the method comprises: obtaining investment research report set, report is text structured analysis, record source attribute and release time;Extract structured text in structural unit, according to the information density constraint of source attribute corresponding screening key structural unit, determine structure distribution characteristics and content attribute characteristics;Release time and feature value construct time series, verify the distribution dispersion and mutation constraint of curve segment, set the initial screening label to abnormal period report;To initial screening report extraction semantic representation vector, construct semantic correlation graph and identify abnormal connected component, set depth processing label;Determine multiple verification time windows to depth processing report, extract the time domain fluctuation parameter of content attribute characteristics, compared with the preset fluctuation constraint range to set key annotation label.The present application effectively improves the screening efficiency of investment research report.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method, apparatus, and electronic device for processing investment research reports based on deep learning. Background Technology

[0002] Investment research reports are crucial for financial investment decisions, encompassing multi-dimensional information such as macroeconomic analysis, industry trend assessment, corporate financial evaluation, and risk warnings. With the increasing frequency and volume of report releases, efficiently sifting through this vast sea of ​​reports to identify key reports worthy of in-depth analysis has become a critical technical challenge for financial institutions and investment research firms.

[0003] Existing investment research report processing solutions typically employ the following methods: First, report screening is based on keyword matching or rule templates. This method relies on manually set keyword libraries or rule sets, making it difficult to adapt to the diversity of report structures and content expressions, and easily overlooking reports with abnormal information density but lacking prominent keywords. Second, a fixed threshold statistical analysis method is used to uniformly process reports from different sources, ignoring the inherent differences in structural distribution and content attributes between different publishing institutions and different types of reports, resulting in a high misjudgment rate. Third, only individual reports are analyzed independently, without considering the continuous change patterns of structural distribution and content attributes of reports from the same source over time, making it impossible to capture abnormal trends caused by fluctuations in report quality or information changes.

[0004] The aforementioned technical deficiencies make it difficult for existing solutions to quickly screen a large number of reports during periods of intensive report release, and also make it impossible to identify potential risks in advance through temporal correlation changes when there are no obvious abnormalities in the report content. This makes it difficult to meet the comprehensive requirements of efficient processing and accurate labeling of investment research reports. Summary of the Invention

[0005] The purpose of this invention is to provide a method, apparatus, and electronic device for processing investment research reports based on deep learning, so as to solve at least one of the problems existing in the prior art.

[0006] To achieve the above objectives, according to one aspect of this application, the present invention provides a deep learning-based investment research report processing method, comprising:

[0007] Obtain the set of investment research reports to be processed, perform text structured parsing on each investment research report in the set, determine the structured text data corresponding to each investment research report, and record the source attributes and publication time of each investment research report;

[0008] Extract several structural units from each of the structured text data, screen key structural units, and then determine the structural distribution characteristics of the key structural units, while simultaneously extracting the content attribute characteristics of each structured text data.

[0009] Time-series correlation analysis was performed based on the structured text data of each investment research report to obtain the time-series correlation analysis results;

[0010] The system acquires real-time resource status parameters of the processing system, dynamically adjusts the time series correlation analysis results based on the real-time resource status parameters, and verifies the distribution discrete constraints and distribution mutation constraints of each curve segment based on the time series correlation analysis results, and then sets initial screening labels for each investment research report.

[0011] In response to the initial screening label being set on the investment research report, semantic feature extraction is performed on the investment research report to determine the semantic representation vector of the investment research report;

[0012] Determine the semantic association strength between semantic representation vectors, construct an association graph based on the semantic association strength, identify abnormal connected components in the association graph, and set the corresponding deep processing labels for the investment research report based on the distribution characteristics of the abnormal connected components.

[0013] In response to the investment research report being tagged with in-depth processing, several verification time windows are determined based on the structured text data of the investment research report. Within each verification time window, the time-domain fluctuation parameters of the content attribute features are extracted. Based on the time-domain fluctuation parameters and the predetermined fluctuation constraint range, key annotation tags are set for the investment research report.

[0014] As a further improvement of the present invention, several structural units are extracted from each of the structured text data, and key structural units are selected from the several structural units based on the attribute information of each structural unit.

[0015] Based on the distribution of key structural units in the investment research report, the structural distribution characteristics are determined, and the content attribute characteristics of each structured text data are extracted simultaneously.

[0016] As a further improvement of the present invention, the process of screening key structural units is as follows:

[0017] Determine the unit attributes of several structural units in structured text data, wherein the unit attributes include unit length and information density within the unit;

[0018] Based on historical samples, determine the information density constraint range corresponding to the source attributes of the investment research report:

[0019] Obtain several historical investment research reports with the same source attributes as samples, and annotate the structural units corresponding to key content;

[0020] The upper limit of the information density constraint range is set to 1.3 times the sample mean, and the lower limit is set to 0.7 times the sample mean. The information density of each labeled structural unit is calculated, and the sample mean of the information density of all labeled structural units is solved.

[0021] The information density constraint range is a closed interval formed by the upper and lower limits of the constraint.

[0022] As a further improvement of the present invention, the process of determining the structural distribution characteristics is as follows:

[0023] If the unit attributes of a structural unit meet the structural screening conditions, then the structural unit is determined to be a key structural unit; the structural screening conditions are that the unit length is greater than a predetermined length threshold and the information density within the unit is within the corresponding information density constraint range.

[0024] Determine the shortest distance between each key structural unit and its nearest key structural unit, and calculate the mean of the shortest distance as the structural distribution feature;

[0025] If the number of key structural units in the investment research report is less than two, the average shortest interval cannot be calculated, and the structural distribution characteristics of the investment research report are marked as null.

[0026] As a further improvement to this invention, the process of performing time-series correlation analysis based on the structured text data of each investment research report to obtain the time-series correlation analysis results is as follows:

[0027] Using the publication time of each investment research report in the collection as a timestamp, and the structural distribution characteristics and content attribute characteristics as feature values, a time series dataset is established.

[0028] Using the publication time as the horizontal axis and the feature value as the vertical axis, the cubic spline interpolation method is used to fit curves to the effective data points in the time series dataset, forming a continuous time series curve of structural distribution features and a time series curve of content attribute features. If the time interval between two consecutive effective data points in the time series dataset exceeds the preset maximum interpolation interval, curve breakpoint processing is performed at that interval to divide the entire curve into independent curve segments.

[0029] The continuous and unbroken curve segments are further divided into several initial curve segments according to the preset basic analysis window duration. Each initial curve segment corresponds to a continuous time sub-interval, and each initial curve segment contains at least two valid peaks. If the number of valid peaks in an initial curve segment after division is less than two, the initial curve segment is merged with the next adjacent initial curve segment.

[0030] As a further improvement of the present invention, real-time resource status parameters of the processing system are obtained;

[0031] Calculate the overall system load index, and maintain the basic analysis window duration as the effective window duration when the overall system load index is less than the preset load safety threshold; otherwise, dynamically compress the basic analysis window duration according to the following formula:

[0032] Determine the effective window duration, and then re-divide the initial curve segment based on the effective window duration.

[0033] As a further improvement of the present invention, the verification process of the distributed discrete constraint and the distributed abrupt change constraint is as follows:

[0034] Extract the amplitude variance of the peaks in each curve segment of the time-series change curve of structural distribution features and the time-series change curve of content attribute features to verify the distribution discrete constraint: if the amplitude variance of the peak is within the corresponding variance constraint range, then the curve segment is determined to meet the distribution discrete constraint.

[0035] Extract the variation peaks in each curve segment of the temporal change curve of structural distribution features and the temporal change curve of content attribute features, and determine the proportion of variation peaks to verify the distribution mutation constraint: if the proportion of variation peaks is less than the predetermined mutation threshold, the curve segment is determined to meet the distribution mutation constraint.

[0036] The variation peak is a peak in a curve segment whose peak amplitude differs from the average amplitude of the curve segment by a ratio greater than the corresponding difference ratio threshold.

[0037] The difference ratio is defined as: |peak amplitude - average amplitude of curve segment| / average amplitude of curve segment;

[0038] If any curve segment does not meet the discrete distribution constraint or the abrupt change distribution constraint, then trace back the time domain segment corresponding to that curve segment; set initial screening tags for all investment research reports corresponding to the time domain segment.

[0039] As a further improvement of the present invention, the process of constructing the association map includes:

[0040] Calculate the cosine similarity between the semantic representation vectors corresponding to any two reference dimensions, and use the result as the semantic association strength;

[0041] Using each reference dimension as a node and the semantic association strength between reference dimensions as edge weights, an undirected weighted association graph is constructed. Then, connected component analysis is performed on the association graph to determine several connected components.

[0042] The process of identifying anomalous connected components in the correlation graph includes:

[0043] Determine the average internal edge weight of each connected component and the maximum edge weight between each connected component and an external node; wherein, for a connected component containing only one node, its average internal edge weight is defined as zero.

[0044] If the ratio of the average edge weight to the maximum edge weight of a connected component is less than a predetermined connectivity anomaly threshold, the connected component is determined to be an abnormal connected component; if the connected component is an isolated node, it is directly determined to be an abnormal connected component.

[0045] The ratio of the number of nodes in the abnormal connected component to the total number of nodes is used as the abnormal coverage ratio. If the abnormal coverage ratio is greater than a predetermined coverage ratio threshold, a depth processing label is set for the research report.

[0046] According to another aspect of this application, a deep learning-based investment research report processing apparatus is provided, comprising:

[0047] The data acquisition and structuring module is used to acquire a set of investment research reports to be processed, perform text structuring parsing on each investment research report in the set, determine the structured text data corresponding to each investment research report, and record the source attributes and publication time of each investment research report;

[0048] The feature extraction module is used to extract several structural units from each of the structured text data, screen key structural units, and then determine the structural distribution features of the key structural units, while simultaneously extracting the content attribute features of each structured text data.

[0049] The time series analysis module is used to perform time series correlation analysis based on the structured text data of each investment research report to obtain the time series correlation analysis results;

[0050] The initial screening label setting module is used to obtain the real-time resource status parameters of the processing system, dynamically adjust the time series correlation analysis results based on the real-time resource status parameters, and verify the distribution discrete constraints and distribution mutation constraints of each curve segment based on the time series correlation analysis results, and then set initial screening labels for each investment research report.

[0051] The semantic representation module is used to extract semantic features from the investment research report in response to the initial screening label being set, and to determine the semantic representation vector of the investment research report.

[0052] The deep processing label setting module is used to determine the semantic association strength between semantic representation vectors, construct an association graph based on the semantic association strength, identify abnormal connected components in the association graph, and set deep processing labels for the corresponding investment research reports based on the distribution characteristics of the abnormal connected components.

[0053] The key annotation tag setting module is used to respond to the investment research report being set with in-depth processing tags. Based on the structured text data of the investment research report, several verification time windows are determined. Within each verification time window, the time-domain fluctuation parameters of the content attribute features are extracted. Based on the time-domain fluctuation parameters and the predetermined fluctuation constraint range, the key annotation tags of the investment research report are set.

[0054] According to another aspect of this application, an electronic device is provided, comprising:

[0055] One or more processors;

[0056] Storage device for storing one or more programs;

[0057] When the one or more programs are executed by the one or more processors, the one or more processors implement the deep learning-based investment research report processing method.

[0058] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0059] By performing structured analysis and extracting structural units from investment research reports, and adaptively determining the information density constraint range based on source attributes, key structural units are screened and structural distribution characteristics are identified. This effectively overcomes the problem that fixed threshold strategies are ill-suited to the structural differences in reports from different sources. Furthermore, a time-series correlation analysis is performed on the publication time of multiple reports with their structural distribution characteristics and content attribute characteristics, forming a continuous change curve. By validating the distribution discreteness and abrupt change constraints of the curve segments, abnormal fluctuations in the report structure and content attributes can be captured from a time-series perspective, avoiding omissions caused by independent analysis of single reports. Further, semantic feature extraction is performed on the initially screened reports, and a correlation graph is constructed. By identifying abnormal connected components, reports with semantically abnormal correlations are located, achieving in-depth screening. Finally, by validating the time-domain fluctuation parameters within the time window, key annotation labels are precisely set, thus meeting the comprehensive requirements for efficient initial screening, in-depth analysis, and key annotation of massive amounts of investment research reports without increasing manual review costs. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 This is a flowchart illustrating the deep learning-based investment research report processing method in this embodiment.

[0062] Figure 2This is a schematic diagram of the structure of the deep learning-based investment research report processing device provided in this embodiment.

[0063] Figure 3 This is a schematic diagram of the structure of the electronic device provided in this embodiment. Detailed Implementation

[0064] To more clearly illustrate the present invention, the following description, in conjunction with preferred embodiments and accompanying drawings, further explains the invention. Similar components in the drawings are indicated by the same reference numerals. Those skilled in the art should understand that the specific description below is illustrative rather than restrictive and should not be construed as limiting the scope of protection of the present invention.

[0065] It should be noted that although the terms first, second, third, etc., may be used in the embodiments of this application for description, these descriptions should not be limited to these terms. These terms are only used to distinguish the descriptions. For example, without departing from the scope of the embodiments of this application, first can also be referred to as second, and similarly, second can also be referred to as first.

[0066] The acquisition, storage, use, and processing of data in this application all comply with the relevant provisions of national laws and regulations.

[0067] It should first be noted that the deep learning-based investment research report processing method, device, and electronic equipment provided in this application are applied to the processing and analysis of financial investment research reports. By acquiring multi-source heterogeneous investment research report data and performing structured parsing and time-series correlation analysis, it achieves efficient identification and highlighting of potential anomalies in massive amounts of investment research reports. This application is particularly suitable for batch processing scenarios during periods of intensive investment research report release, where a large number of reports need to be quickly screened to identify those worthy of in-depth analysis.

[0068] For data monitoring in the aforementioned application scenarios, the bottlenecks are jointly affected by the diversity of investment research report structures, fluctuations in content attributes, and semantic drift between reports. As the frequency of report releases increases and the number of reports accumulates, fixed threshold analysis strategies are difficult to adapt to the structural differences of reports from different sources. Furthermore, traditional single-article analysis ignores the continuous change patterns of structural distribution and content attributes of reports from the same source in the temporal dimension, leading to reports with abnormal information density being easily missed or misjudged.

[0069] In response, this embodiment provides an adaptive processing method that can jointly identify the structural distribution characteristics and content attribute characteristics of investment research reports, and incorporate time-series discrete analysis, semantic association graph construction and fluctuation verification into a unified framework. This method meets the comprehensive processing requirements for efficient initial screening and key annotation of investment research reports without increasing the cost of manual review.

[0070] Example 1:

[0071] To apply the above-mentioned application scenarios, this embodiment provides a deep learning-based investment research report processing method, the flowchart of which is shown below. Figure 1 As shown, it includes:

[0072] Step S1: Obtain the set of investment research reports to be processed, perform text structured parsing on each investment research report in the set, determine the structured text data corresponding to each investment research report, and record the source attributes and publication time of each investment research report;

[0073] Step S2: Extract several structural units from each of the structured text data, screen key structural units, and then determine the structural distribution characteristics of the key structural units, while simultaneously extracting the content attribute characteristics of each structured text data.

[0074] Step S3: Perform time-series correlation analysis on the structured text data of each investment research report to obtain the time-series correlation analysis results. The time-series correlation analysis results include: time-series change curves of structural distribution characteristics and time-series change curves of content attribute characteristics.

[0075] Step S4: Obtain the real-time resource status parameters of the processing system, dynamically adjust the time series correlation analysis results based on the real-time resource status parameters, and verify the distribution discrete constraints and distribution mutation constraints of each curve segment based on the time series correlation analysis results, and then set preliminary screening labels for each investment research report.

[0076] Step S5: In response to the initial screening label being set on the investment research report, perform semantic feature extraction on the investment research report to determine the semantic representation vector of the investment research report;

[0077] Step S6: Determine the semantic association strength between semantic representation vectors, construct an association graph based on the semantic association strength, identify abnormal connected components in the association graph, and set the corresponding deep processing labels for the investment research report based on the distribution characteristics of the abnormal connected components.

[0078] Step S7: In response to the investment research report being set with in-depth processing tags, several verification time windows are determined based on the structured text data of the investment research report. Within each verification time window, the time-domain fluctuation parameters of the content attribute features are extracted. Based on the time-domain fluctuation parameters and the predetermined fluctuation constraint range, the key annotation tags of the investment research report are set.

[0079] It should be noted that the investment research reports mentioned in this embodiment include at least one of the following: research reports from publicly available information release platforms, industry analysis reports, and periodically disclosed corporate documents. Each investment research report is accessed by the processing system in electronic document form via a data interface. This embodiment, while standardizing the format and extracting the text of the investment research reports, parses and records their source attributes from the report metadata or the header area. The source attributes include at least one of the following: the type of publishing institution, the report type, and the field covered by the report. The publication time is the official publication date recorded in the original document of each investment research report, distinguished from the time the report is accessed by the processing system.

[0080] Specifically, the implementation process of step S2 includes:

[0081] Extract several structural units from each of the structured text data, and select key structural units from the several structural units based on the attribute information of each structural unit.

[0082] Based on the distribution of key structural units in the investment research report, the structural distribution characteristics are determined, and the content attribute characteristics of each structured text data are extracted simultaneously.

[0083] The process of selecting key structural units is as follows:

[0084] Determine the unit attributes of several structural units in structured text data, wherein the unit attributes include unit length and information density within the unit;

[0085] Based on historical samples, determine the information density constraint range corresponding to the source attributes of the investment research report:

[0086] Obtain several historical investment research reports with the same source attributes as samples, and annotate the structural units corresponding to key content;

[0087] Calculate the information density of each labeled structural unit and solve for the sample mean of the information density of all labeled structural units;

[0088] The upper limit of the information density constraint range is set to 1.3 times the sample mean, and the lower limit is set to 0.7 times the sample mean. The information density constraint range is a closed interval formed by the upper and lower limits of the constraint.

[0089] Specifically, the process of determining the structural distribution characteristics is as follows:

[0090] If the unit attributes of a structural unit meet the structural screening conditions, then the structural unit is determined to be a key structural unit; the structural screening conditions are that the unit length is greater than a predetermined length threshold and the information density within the unit is within the corresponding information density constraint range.

[0091] Determine the shortest distance between each key structural unit and its nearest key structural unit, and calculate the mean of the shortest distance as the structural distribution feature;

[0092] If the number of key structural units in the investment research report is less than two, the average shortest interval cannot be calculated, and the structural distribution characteristics of the investment research report are marked as null.

[0093] It should be noted that, in this embodiment, the information density within a unit is the ratio of the number of predefined effective information characters within the structural unit to the total number of characters; the content attribute feature is the mean of the content features corresponding to each structural unit in the structured text data, and the content feature is one of semantic richness and terminology density, wherein the terminology density is the proportion of professional domain-specific words in the structured text data to the total number of words in the structured text data, and the semantic richness in this embodiment is the type-tag ratio (TTR);

[0094] Meanwhile, the predetermined length threshold is set to avoid misjudging short sentences, ensure that the unit carries sufficient information, and conform to the average length of most investment research report paragraphs. In this embodiment, it is set to 500 characters.

[0095] Specifically, in step S3, the process of performing time-series correlation analysis based on the structured text data of each investment research report to obtain the time-series correlation analysis results is as follows:

[0096] Using the publication time of each investment research report in the collection as a timestamp, and the structural distribution characteristics and content attribute characteristics as feature values, a time series dataset is established.

[0097] Using the publication time as the horizontal axis and the feature value as the vertical axis, the cubic spline interpolation method is used to fit curves to the effective data points in the time series dataset, forming a continuous time series curve of structural distribution features and a time series curve of content attribute features. If the time interval between two consecutive effective data points in the time series dataset exceeds the preset maximum interpolation interval, curve breakpoint processing is performed at that interval to divide the entire curve into independent curve segments.

[0098] The continuous and unbroken curve segments are further divided into several initial curve segments according to the preset basic analysis window duration. Each initial curve segment corresponds to a continuous time sub-interval, and each initial curve segment contains at least two valid peaks. If the number of valid peaks in an initial curve segment after division is less than two, the initial curve segment is merged with the next adjacent initial curve segment.

[0099] In an exemplary embodiment, the duration of the basic analysis window is set to 30 calendar days, which allows the monthly window to effectively smooth out weekly noise and achieve a balance between statistical stability and analytical timeliness; the maximum interpolation interval is set to 60 calendar days, which is used when there are no reports for more than two months, indicating that the source may be interrupted, and independent segments can avoid forming false continuous trends.

[0100] It is worth noting that in this embodiment, cases where the curve is shorter than the basic analysis window duration are not classified to avoid algorithm interruption.

[0101] Specifically, the process of dynamically adjusting the time-series correlation analysis results in step S4 is as follows:

[0102] Obtain real-time resource status parameters of the processing system; the real-time resource status parameters include: memory occupancy rate and CPU utilization rate of the processing system.

[0103] The comprehensive load index of the computing system is L = α × Rm + β × Rc, where Rm and Rc are the memory usage rate and CPU utilization rate of the processing system, respectively, and α and β are weighting coefficients, satisfying α + β = 1.

[0104] The system load comprehensive index L is compared with the preset load safety threshold Lh: if L is not greater than Lh, the basic analysis window duration Wb is maintained as the effective window duration We; if L is greater than Lh, the basic analysis window duration is dynamically compressed according to the following formula:

[0105] Set We = Wb × [1 - γ × (L - Lh)], where γ is the compression gain coefficient. When We is less than the minimum window duration Wmin, we take We = Wmin.

[0106] The initial curve segments are re-divided according to the effective window duration We: if We is less than Wb, each original initial curve segment is further divided according to the effective window duration We; if We = Wb, the initial curve segments remain unchanged; if the comprehensive system load index L continues to decrease to L≤Lh and remains for more than 10 seconds, We is restored to Wb.

[0107] In an exemplary embodiment, α=0.65 and β=0.35 are set to highlight the impact of memory pressure, given that the memory usage of text feature data is often a processing bottleneck. Lh is selected within the range of [0.70, 0.85], and in this embodiment, it is preferably 0.75 to intervene in advance when the resource utilization rate reaches 75%, reserving a 25% buffer to cope with subsequent bursts and avoid instantaneous overload. The compression gain coefficient ranges from (0, 2], and in this embodiment, it is preferably γ=1.2 to moderately increase the gain, so that the window shrinks rapidly as the load increases, ensuring processing timeliness and not falling below the minimum effective length. The minimum window duration is set to 7 calendar days, with a basic report release cycle of one week, to ensure that the window contains at least several report points and maintain the statistical validity of the curve segment.

[0108] It should be noted that when acquiring the real-time resource status parameters of the processing system, the data is collected at a fixed period. In this embodiment, the fixed period is set to 5 seconds.

[0109] Specifically, the verification process for the distributed discrete constraints and the distributed abrupt change constraints is as follows:

[0110] Extract the amplitude variance of the peaks in each curve segment of the time-series change curve of structural distribution features and the time-series change curve of content attribute features to verify the distribution discrete constraint: if the amplitude variance of the peak is within the corresponding variance constraint range, then the curve segment is determined to meet the distribution discrete constraint.

[0111] Extract the variation peaks in each curve segment of the temporal change curve of structural distribution features and the temporal change curve of content attribute features, and determine the proportion of variation peaks to verify the distribution mutation constraint: if the proportion of variation peaks is less than the predetermined mutation threshold, the curve segment is determined to meet the distribution mutation constraint.

[0112] The variation peak is a peak in a curve segment whose peak amplitude differs from the average amplitude of the curve segment by a ratio greater than the corresponding difference ratio threshold.

[0113] The difference ratio is defined as: |peak amplitude - average amplitude of curve segment| / average amplitude of curve segment;

[0114] If any curve segment does not meet the discrete distribution constraint or the abrupt change distribution constraint, then trace back the time domain segment corresponding to that curve segment; set initial screening tags for all investment research reports corresponding to the time domain segment.

[0115] In an exemplary embodiment, the mutation threshold is set to 0.2, allowing 20% ​​of the peaks to be abnormal peaks, thus capturing true anomalies in the sequence while providing tolerance for random local perturbations. Simultaneously, the difference ratio threshold for structural distribution features is set to 0.5, and the difference ratio threshold for content distribution features is set to 0.4. The lower limit of the variance constraint range for structural distribution features is set to 0.7 times the mean variance of the normal reported peak amplitude, and the upper limit is set to 1.3 times. The lower limit of the variance constraint range for content distribution features is also set to 0.7 times the mean variance of the normal reported peak amplitude, and the upper limit is set to 1.3 times.

[0116] The historical normal investment research reports mentioned are those that have been archived by staff.

[0117] Specifically, the process of determining the semantic representation vector of the investment research report is as follows:

[0118] Determine the source of the investment research report;

[0119] Based on a predefined reference dimension mapping table that matches the source attribute, several reference dimensions associated with the source attribute are determined.

[0120] The pre-trained semantic encoding model is invoked to perform semantic feature extraction on the structured text data across each reference dimension:

[0121] The structured text data under each reference dimension is segmented according to a predetermined word segmentation strategy to determine the word sequence corresponding to each reference dimension;

[0122] The word sequences corresponding to each reference dimension are input into a pre-trained semantic coding model. After processing by the coding layer of the semantic coding model, the semantic representation vectors corresponding to each reference dimension are output.

[0123] In one exemplary embodiment, the semantic encoding model is a pre-trained language representation model based on the Transformer architecture, and the dimension of the semantic representation vector is set to 768 dimensions;

[0124] Meanwhile, the predetermined word segmentation strategy can be based on a word segmentation table, and the specific process is a technical means well known to those skilled in the art, and is not limited in this application.

[0125] It should be noted that the reference dimension mapping table is pre-built and records the matching relationship between each source attribute and the corresponding analysis chapter in the report. It can be understood that if the source attribute does not have a matching record in the reference dimension mapping table, a default reference dimension set is used, which includes the report summary dimension, the main text analysis dimension, and the risk warning dimension.

[0126] Specifically, the process of constructing the association graph includes:

[0127] Calculate the cosine similarity between the semantic representation vectors corresponding to any two reference dimensions, and use the result as the semantic association strength;

[0128] Using each reference dimension as a node and the semantic association strength between the reference dimensions as the edge weight, an undirected weighted association graph is constructed. Then, connected component analysis is performed on the association graph to determine several connected components.

[0129] Specifically, the process of identifying anomalous connected components in an association graph includes:

[0130] Determine the average internal edge weight of each connected component and the maximum edge weight between each connected component and an external node; wherein, for a connected component containing only one node, its average internal edge weight is defined as zero.

[0131] If the ratio of the average edge weight to the maximum edge weight of a connected component is less than a predetermined connectivity anomaly threshold, the connected component is determined to be an abnormal connected component; if the connected component is an isolated node, it is directly determined to be an abnormal connected component.

[0132] The ratio of the number of nodes in the abnormal connected component to the total number of nodes is used as the abnormal coverage ratio. If the abnormal coverage ratio is greater than a predetermined coverage ratio threshold, a depth processing label is set for the research report.

[0133] It is worth noting that, in this embodiment, the isolated node is defined as a node whose internal average edge weight and maximum edge weight are both zero; meanwhile, in an exemplary embodiment, the connectivity anomaly threshold is selected within the interval [0.3, 0.5], preferably 0.4, and the coverage ratio threshold is selected within the interval [0.25, 0.45], preferably 0.35.

[0134] Specifically, the implementation process of step S7 includes:

[0135] The publication time of the investment research report is determined, and several historical time nodes are determined by tracing back from the publication time as the starting point according to a predetermined time interval;

[0136] Each verification time window is determined based on a predetermined window duration, with each historical time point serving as the starting point.

[0137] Extract the content attribute feature values ​​corresponding to each investment research report contained within each verification time window;

[0138] Calculate the ratio of the standard deviation to the mean of the content attribute feature values ​​within each verification time window, and use the calculation result as the time domain fluctuation parameter. If the number of valid investment research reports contained in a certain verification time window is less than three, the time domain fluctuation parameter of that window is marked as invalid and will not participate in subsequent comparisons.

[0139] If the time-domain fluctuation parameter of any valid verification time window exceeds the fluctuation constraint range, a key label will be set for the investment research report; if the time-domain fluctuation parameters of all valid verification time windows are within the fluctuation constraint range, no key label will be set for the investment research report.

[0140] It should be noted that the fluctuation constraint range described in this embodiment is determined in advance based on the time-domain fluctuation parameters of continuously published reports in the historical normal investment research report set.

[0141] Example 2:

[0142] Please see Figure 2 As shown, this is a schematic diagram of the structure of the deep learning-based investment research report processing device provided in this embodiment, including:

[0143] The data acquisition and structuring module is used to acquire a set of investment research reports to be processed, perform text structuring parsing on each investment research report in the set, determine the structured text data corresponding to each investment research report, and record the source attributes and publication time of each investment research report;

[0144] The feature extraction module is used to extract several structural units from each of the structured text data, screen key structural units, and then determine the structural distribution features of the key structural units, while simultaneously extracting the content attribute features of each structured text data.

[0145] The time series analysis module is used to perform time series correlation analysis based on the structured text data of each investment research report to obtain the time series correlation analysis results;

[0146] The initial screening label setting module is used to obtain the real-time resource status parameters of the processing system, dynamically adjust the time series correlation analysis results based on the real-time resource status parameters, and verify the distribution discrete constraints and distribution mutation constraints of each curve segment based on the time series correlation analysis results, and then set initial screening labels for each investment research report.

[0147] The semantic representation module is used to extract semantic features from the investment research report in response to the initial screening label being set, and to determine the semantic representation vector of the investment research report.

[0148] The deep processing label setting module is used to determine the semantic association strength between semantic representation vectors, construct an association graph based on the semantic association strength, identify abnormal connected components in the association graph, and set deep processing labels for the corresponding investment research reports based on the distribution characteristics of the abnormal connected components.

[0149] The key annotation tag setting module is used to respond to the investment research report being set with in-depth processing tags. Based on the structured text data of the investment research report, several verification time windows are determined. Within each verification time window, the time-domain fluctuation parameters of the content attribute features are extracted. Based on the time-domain fluctuation parameters and the predetermined fluctuation constraint range, the key annotation tags of the investment research report are set.

[0150] The deep learning-based investment research report processing device provided in this application can execute the deep learning-based investment research report processing method provided in any embodiment of this application, and has the corresponding functional modules and beneficial effects of the method execution.

[0151] Example 3:

[0152] From a hardware perspective, to realize the deep learning-based investment research report processing method in a computer, this application also provides an electronic device, which can be a server, and its internal structure diagram can be as follows. Figure 3 As shown:

[0153] The computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores server data. The network interface communicates with external terminals via a network connection. When executed by the processor, the computer program implements a deep learning-based investment research report processing method.

[0154] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0155] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A deep learning-based investment research report processing method, characterized in that, include: Obtain the set of investment research reports to be processed, perform text structured parsing on each investment research report in the set, determine the structured text data corresponding to each investment research report, and record the source attributes and publication time of each investment research report; Extract several structural units from each of the structured text data, screen key structural units, and then determine the structural distribution characteristics of the key structural units, while simultaneously extracting the content attribute characteristics of each structured text data. Time-series correlation analysis was performed based on the structured text data of each investment research report to obtain the time-series correlation analysis results; The system acquires real-time resource status parameters of the processing system, dynamically adjusts the time series correlation analysis results based on the real-time resource status parameters, and verifies the distribution discrete constraints and distribution mutation constraints of each curve segment based on the time series correlation analysis results, and then sets initial screening labels for each investment research report. In response to the initial screening label being set on the investment research report, semantic feature extraction is performed on the investment research report to determine the semantic representation vector of the investment research report; Determine the semantic association strength between semantic representation vectors, construct an association graph based on the semantic association strength, identify abnormal connected components in the association graph, and set the corresponding deep processing labels for the investment research report based on the distribution characteristics of the abnormal connected components. In response to the investment research report being tagged with in-depth processing, several verification time windows are determined based on the structured text data of the investment research report. Within each verification time window, the time-domain fluctuation parameters of the content attribute features are extracted. Based on the time-domain fluctuation parameters and the predetermined fluctuation constraint range, key annotation tags are set for the investment research report.

2. The deep learning-based investment research report processing method according to claim 1, characterized in that, Extract several structural units from each of the structured text data, and select key structural units from the several structural units based on the attribute information of each structural unit. Based on the distribution of key structural units in the investment research report, the structural distribution characteristics are determined, and the content attribute characteristics of each structured text data are extracted simultaneously.

3. The deep learning-based investment research report processing method according to claim 2, characterized in that, The process of screening key structural units is as follows: Determine the unit attributes of several structural units in structured text data, wherein the unit attributes include unit length and information density within the unit; Based on historical samples, determine the information density constraint range corresponding to the source attributes of the investment research report: Obtain several historical investment research reports with the same source attributes as samples, and annotate the structural units corresponding to key content; The upper limit of the information density constraint range is set to 1.3 times the sample mean, and the lower limit is set to 0.7 times the sample mean. The information density of each labeled structural unit is calculated, and the sample mean of the information density of all labeled structural units is solved. The information density constraint range is a closed interval formed by the upper and lower limits of the constraint.

4. The deep learning-based investment research report processing method according to claim 3, characterized in that, The process for determining the structural distribution characteristics is as follows: If the unit attributes of a structural unit meet the structural screening conditions, then the structural unit is determined to be a key structural unit; the structural screening conditions are that the unit length is greater than a predetermined length threshold and the information density within the unit is within the corresponding information density constraint range. Determine the shortest distance between each key structural unit and its nearest key structural unit, and calculate the mean of the shortest distance as the structural distribution feature; If the number of key structural units in the investment research report is less than two, the average shortest interval cannot be calculated, and the structural distribution characteristics of the investment research report are marked as null.

5. The deep learning-based investment research report processing method according to claim 4, characterized in that, The process of performing time-series correlation analysis on structured text data from various investment research reports to obtain the results is as follows: Using the publication time of each investment research report in the collection as a timestamp, and the structural distribution characteristics and content attribute characteristics as feature values, a time series dataset is established. Using the publication time as the horizontal axis and the feature value as the vertical axis, the cubic spline interpolation method is used to fit curves to the effective data points in the time series dataset, forming a continuous time series curve of structural distribution features and a time series curve of content attribute features. If the time interval between two consecutive effective data points in the time series dataset exceeds the preset maximum interpolation interval, curve breakpoint processing is performed at that interval to divide the entire curve into independent curve segments. The continuous and unbroken curve segments are further divided into several initial curve segments according to the preset basic analysis window duration. Each initial curve segment corresponds to a continuous time sub-interval, and each initial curve segment contains at least two valid peaks. If the number of valid peaks in an initial curve segment after division is less than two, the initial curve segment is merged with the next adjacent initial curve segment.

6. The deep learning-based investment research report processing method according to claim 5, characterized in that, Obtain real-time resource status parameters of the processing system; Calculate the overall system load index, and maintain the basic analysis window duration as the effective window duration when the overall system load index is less than the preset load safety threshold; otherwise, dynamically compress the basic analysis window duration according to the following formula: Determine the effective window duration, and then re-divide the initial curve segment based on the effective window duration.

7. The deep learning-based investment research report processing method according to claim 6, characterized in that, The verification process for the distributed discrete constraints and the distributed abrupt change constraints is as follows: Extract the amplitude variance of the peaks in each curve segment of the time-series change curve of structural distribution features and the time-series change curve of content attribute features to verify the distribution discrete constraint: if the amplitude variance of the peak is within the corresponding variance constraint range, then the curve segment is determined to meet the distribution discrete constraint. Extract the variation peaks in each curve segment of the temporal change curve of structural distribution features and the temporal change curve of content attribute features, and determine the proportion of variation peaks to verify the distribution mutation constraint: if the proportion of variation peaks is less than the predetermined mutation threshold, the curve segment is determined to meet the distribution mutation constraint. The variation peak is a peak in a curve segment whose peak amplitude differs from the average amplitude of the curve segment by a ratio greater than the corresponding difference ratio threshold. The difference ratio is defined as: |peak amplitude - average amplitude of curve segment| / average amplitude of curve segment; If any curve segment does not meet the discrete distribution constraint or the abrupt change distribution constraint, then trace back the time domain segment corresponding to that curve segment; set initial screening tags for all investment research reports corresponding to the time domain segment.

8. The deep learning-based investment research report processing method according to claim 7, characterized in that, The process of constructing the association graph includes: Calculate the cosine similarity between the semantic representation vectors corresponding to any two reference dimensions, and use the result as the semantic association strength; Using each reference dimension as a node and the semantic association strength between reference dimensions as edge weights, an undirected weighted association graph is constructed. Then, connected component analysis is performed on the association graph to determine several connected components. The process of identifying anomalous connected components in the correlation graph includes: Determine the average internal edge weight of each connected component and the maximum edge weight between each connected component and an external node; wherein, for a connected component containing only one node, its average internal edge weight is defined as zero. If the ratio of the average edge weight to the maximum edge weight of a connected component is less than a predetermined connectivity anomaly threshold, the connected component is determined to be an abnormal connected component; if the connected component is an isolated node, it is directly determined to be an abnormal connected component. The ratio of the number of nodes in the abnormal connected component to the total number of nodes is used as the abnormal coverage ratio. If the abnormal coverage ratio is greater than a predetermined coverage ratio threshold, a depth processing label is set for the research report.

9. A deep learning-based investment research report processing apparatus, applied to the deep learning-based investment research report processing method as described in any one of claims 1-8, characterized in that, include: The data acquisition and structuring module is used to acquire a set of investment research reports to be processed, perform text structuring parsing on each investment research report in the set, determine the structured text data corresponding to each investment research report, and record the source attributes and publication time of each investment research report; The feature extraction module is used to extract several structural units from each of the structured text data, screen key structural units, and then determine the structural distribution features of the key structural units, while simultaneously extracting the content attribute features of each structured text data. The time series analysis module is used to perform time series correlation analysis based on the structured text data of each investment research report to obtain the time series correlation analysis results; The initial screening label setting module is used to obtain the real-time resource status parameters of the processing system, dynamically adjust the time series correlation analysis results based on the real-time resource status parameters, and verify the distribution discrete constraints and distribution mutation constraints of each curve segment based on the time series correlation analysis results, and then set initial screening labels for each investment research report. The semantic representation module is used to extract semantic features from the investment research report in response to the initial screening label being set, and to determine the semantic representation vector of the investment research report. The deep processing label setting module is used to determine the semantic association strength between semantic representation vectors, construct an association graph based on the semantic association strength, identify abnormal connected components in the association graph, and set deep processing labels for the corresponding investment research reports based on the distribution characteristics of the abnormal connected components. The key annotation tag setting module is used to respond to the investment research report being set with in-depth processing tags. Based on the structured text data of the investment research report, several verification time windows are determined. Within each verification time window, the time-domain fluctuation parameters of the content attribute features are extracted. Based on the time-domain fluctuation parameters and the predetermined fluctuation constraint range, the key annotation tags of the investment research report are set.

10. An electronic device, characterized in that, The electronic device includes: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the deep learning-based investment research report processing method according to any one of claims 1-8.

Images