AI-based data classification method, platform and electronic device

Abstract

The application relates to the technical field of natural language processing, in particular to an AI-based data classification method, platform and electronic equipment, which comprises the following steps: collecting target data; performing standardization processing on the target text data to obtain standard text, and constructing a negative sentiment feature and a sentiment turning index according to the standard text, and determining a sentiment feature based on the negative sentiment feature and the sentiment turning index; extracting an acoustic feature according to the standard text and the target voice data, and constructing an emergency acoustic index according to the acoustic feature; classifying the target data according to the sentiment feature and the emergency acoustic index; constructing a target behavior feature according to the standard text, and updating the classification result of the target data according to the target behavior feature. The application effectively improves the accuracy of e-commerce complaint work order data classification and reduces the missed detection rate of emergency complaints.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of natural language processing, and particularly relates to an AI-based data classification method and platform and electronic equipment. BACKGROUND

[0002] The complaint ticket classification of a traditional e-commerce platform mainly relies on manual review or keyword matching based on rules, which has significant limitations: on the one hand, manual processing is inefficient and subjective, and it is difficult to cope with the real-time grading needs of a large number of tickets; on the other hand, pure text analysis cannot capture the emotional fluctuations in speech, leading to misjudgment of the true urgency of users.

[0003] In addition, existing methods often ignore dialectical expression, emotional turning points and behavioral intentions, and simply rely on negative word density for classification, resulting in missed high-priority complaints. These defects lead to misallocation of customer service resources, prolong the response cycle of major issues, and increase the risk of platform operation. SUMMARY

[0004] The present application aims to provide an AI-based data classification method, platform and electronic equipment to solve at least one of the problems in the prior art.

[0005] To achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0006] An AI-based data classification method, comprising:

[0007] Collecting target data;

[0008] Standardizing the target text data to obtain a standard text, and constructing a negative sentiment feature and an emotional turning point index according to the standard text, and determining a sentiment feature based on the negative sentiment feature and the emotional turning point index;

[0009] Extracting acoustic features from the standard text and the target speech data, and constructing an urgent acoustic index according to the acoustic features;

[0010] Classifying the target data according to the sentiment feature and the urgent acoustic index;

[0011] Constructing a target behavior feature according to the standard text, and updating the classification result of the target data according to the target behavior feature.

[0012] Further, the target text data is standardized to obtain a standard text, and a negative sentiment feature Qs is constructed according to the standard text;

[0013] An emotional turning point index is constructed according to the standard text, and a sentiment feature is determined according to the negative sentiment feature and the emotional turning point index of the standard text;

[0014] segment the standard text, and extract the sentiment value of each segment of the standard text, and the sentiment value of the ith segment of the standard text is denoted as Si;

[0015] Construct a sentiment turning index according to the sentiment value of each segment of the standard text, and the expression of the sentiment turning index is:

[0016] ; In the formula, I is the number of segments of the standard text, zz is the sentiment turning index, and S(i-1) is the sentiment value of the (i-1)th segment of the standard text.

[0017] Further, determine the sentiment feature QT according to the negative sentiment feature of the standard text and the sentiment turning index, QT=0.6×Qs+0.4×zz.

[0018] Further, the standard text of the standard speech transcription text after standardization processing is taken as a standard speech transcription text, and the total number of words of the standard speech transcription text is extracted, and the ratio of the total number of words of the standard speech transcription text to the duration of the target speech data is taken as the target speech speed MY.

[0019] Further, one fundamental frequency point is extracted every 50ms of the target speech data to obtain a fundamental frequency sequence (F1, F2,..., Fn), F1 is the fundamental frequency of the first fundamental frequency point, F2 is the fundamental frequency of the second fundamental frequency point, Fn is the fundamental frequency of the nth fundamental frequency point, and n is the number of fundamental frequency points;

[0020] The variance of each fundamental frequency point in the fundamental frequency sequence is calculated, and the calculation result is taken as the pitch variance FG;

[0021] Frame the target speech data according to a preset frame length to obtain an audio frame sequence, and obtain the energy of each audio frame, denoted as Ni, i is the audio frame number, and when Ni is less than or equal to the energy threshold, the audio frame is determined as a silent frame;

[0022] Merge the continuous silent frames into a silent segment, and when the silent segment is less than or equal to the preset duration, the silent segment is determined as an invalid pause, and when the silent segment is greater than the preset duration, the silent segment is determined as a valid pause;

[0023] Statistically count the number of valid pauses in the target speech data as Y, and the ratio of Y to the duration of the target speech data is taken as the pause frequency TD;

[0024] Construct an emergency acoustic index SSR based on the target speech speed MY, the pitch variance FG and the pause frequency TD.

[0025] Further, when the sentiment feature weight×QT+acoustic weight×SSR is less than or equal to the classification threshold, the target data is classified as normal data, and when the sentiment feature weight×QT+acoustic weight×SSR is greater than the classification threshold, the target data is classified as emergency data.

[0026] Further, the standard text is matched with the preset right protection keywords to determine a right protection tendency index;

[0027] The standard text is matched with the preset platform keywords to determine a platform responsibility index;

[0028] The standard text is matched with the preset group keywords to determine a group characteristic index.

[0029] Further, a target behavior characteristic is determined according to the right protection tendency index, the platform responsibility index and the group characteristic index, and a classification result of the target data is updated according to the target behavior characteristic;

[0030] The expression of the target behavior characteristic is MT = 0.5 x right protection tendency index + 0.2 x platform responsibility index + 0.3 x group characteristic index, and MT is the target behavior characteristic;

[0031] When the target behavior characteristic MT is less than a behavior characteristic threshold, the classification result of the target data is not updated, and when the target behavior characteristic MT is greater than or equal to the behavior characteristic threshold, if the classification result of the target data is ordinary data, the classification result of the target data is updated to urgent data, and if the classification result of the target data is urgent data, the classification result of the target data is not updated.

[0032] According to another aspect of the present application, an AI-based data classification platform is provided, comprising:

[0033] An acquisition unit is configured to collect target data;

[0034] An emotion characteristic construction unit is configured to perform standardization processing on the target text data to obtain a standard text, and construct a negative emotion characteristic and an emotion turning index according to the standard text, and determine an emotion characteristic based on the negative emotion characteristic and the emotion turning index;

[0035] An acoustic characteristic construction unit is configured to extract an acoustic characteristic according to the standard text and the target voice data, and construct an urgent acoustic index according to the acoustic characteristic;

[0036] A classification unit is configured to classify the target data according to the emotion characteristic and the urgent acoustic index;

[0037] A behavior characteristic construction unit is configured to construct a target behavior characteristic according to the standard text, and update a classification result of the target data according to the target behavior characteristic.

[0038] According to still another aspect of the present application, an electronic device is provided, comprising:

[0039] One or more processors;

[0040] A storage device configured to store one or more programs;

[0041] When the one or more programs are executed by the one or more processors, the one or more processors implement the AI-based data classification method.

[0042] The beneficial effects of the present application are as follows: through multi-modal AI fusion and three-level verification mechanism, the accuracy and timeliness of complaint work order classification are systematically improved, and text and acoustic features are integrated: dialect standardization ensures analysis fairness, emotional turning point index identifies implicit dissatisfaction, acoustic parameter quantifies emotional intensity, solves one-sidedness of single mode, introduces behavior characteristics to dynamically correct classification: legal risk warning of rights protection keywords, platform responsibility marking accelerates internal coordination, group signal captures public opinion fermentation clues, forms a comprehensive decision-making framework of "emotion + behavior + responsibility attribution", and finally realizes intelligent grading of complaint work orders, ensures resource tilt to high-value customers and high-risk events, shortens the response time of emergency work orders, reduces the complaint rate, at the same time provides emotion and intention for customer service personnel Stereoscopic insight, optimize user experience and enterprise risk control efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0044] Figure 1 The flowchart of the AI-based data classification method of the present embodiment.

[0045] Figure 2 The flowchart of the emotion feature determination method of the present embodiment.

[0046] Figure 3 The flowchart of the target data classification result updating method of the present embodiment.

[0047] Figure 4 The structural diagram of the AI-based data classification platform of the present embodiment.

[0048] Figure 5 The structural diagram of the electronic device of the present embodiment. DETAILED DESCRIPTION

[0049] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described, obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the protection scope of the present application.

[0050] In the present application, the term "comprising", "containing" or any other variant thereof is intended to cover the non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes the elements inherent to such process, method, article or equipment. Without more limitation, the element defined by the sentence "including a…" does not exclude the presence of other identical elements in the process, method, article or equipment including the element.

[0051] Specifically, the AI-based data classification method described in the embodiment is applied to the classification of complaint work orders containing text and voice on an e-commerce platform.

[0052] Please refer to Figure 1 As shown in the figure, it is a flowchart of the AI-based data classification method of the embodiment, which includes:

[0053] Step S101, collecting target data, the target data including target text data and target voice data, the target text data being original text of a user complaint work order and voice transcription text of the user complaint work order, and the target voice data being voice audio of the user complaint work order.

[0054] For example, in the embodiment, pure text in UTF-8 format can be directly obtained through the database API of the e-commerce platform, voice transcription text can be obtained using the Ali Cloud intelligent voice interaction service, and target voice data can be obtained through APP voice messages; the collection method of the target data is not specifically limited in the embodiment, and can be freely set by those skilled in the art according to the needs.

[0055] Specifically, by automatically collecting the original text, voice transcription text and voice audio of user complaints, it is ensured that the subsequent analysis can cover the full-dimensional information expressed by the user. This solves the problem that voice information is ignored or not fully transcribed in traditional work order processing, and provides a complete data basis for subsequent analysis.

[0056] Please continue to refer to Figure 1 As shown in the figure, the AI-based data classification method further includes:

[0057] In step S102, the target text data is standardized to obtain a standard text, and a negative sentiment feature and a sentiment turning index are constructed according to the standard text, and the sentiment feature is determined based on the negative sentiment feature and the sentiment turning index.

[0058] Referring to Figure 2 The determination method of the sentiment feature includes:

[0059] In step S201, the target text data is standardized to obtain a standard text, and a negative sentiment feature is constructed according to the standard text.

[0060] Specifically, the process of standardizing the target text data is as follows:

[0061] The pre-trained dialect BERT is used to detect the dialect area of the target text data, and the dialect words are recognized and replaced with Mandarin synonyms, and the standardized target text data is used as the standard text;

[0062] The process of constructing the negative sentiment feature according to the standard text is as follows:

[0063] The number of negative sentiment words in the standard text is m1, the total number of words in the standard text is m0, and the negative sentiment feature is constructed according to m0 and m1, and the expression of the negative sentiment feature is:

[0064] Qs=m1 / m0+0.1×m0 / M, M is a text word quantity threshold.

[0065] Specifically, the text word quantity threshold in the embodiment is a critical word number for triggering text length correction, which avoids diluting the negative sentiment density of long text, and the value range is [80, 120], and the value of M in the embodiment is 100.

[0066] Specifically, the number of negative sentiment words is extracted by multi-level matching combined with context analysis: first, the negative sentiment words are matched based on the customized sentiment word library of e-commerce, then the real sentiment polarity of the words in the specific context is analyzed by the BERT model, and the interference of negative sentences (such as "not bad") and the recognition of unregistered words are excluded; finally, the total number of words confirmed as negative sentiment is counted, the general negative words of the Hownet of the Chinese Knowledge Grid and the sentiment word library of Tsinghua University Li Jun are fused as the basic word library, and the high-frequency negative words are extracted from the e-commerce platform review data to construct the word library; the method of obtaining the number of negative sentiment words is not limited in the embodiment, and the person skilled in the art can freely set it according to the needs.

[0067] Specifically, the total number of words of the text can be collected by using a word segmentation tool in the embodiment.

[0068] Specifically, the process of not standardizing the target text in the embodiment is not specifically limited, and a person skilled in the art can freely set it according to the needs.

[0069] Specifically, the dialect vocabulary is converted into a Mandarin synonym, the regional expression difference is eliminated to interfere with the sentiment analysis, and it is ensured that complaints of users in different regions can be fairly evaluated. Meanwhile, by calculating the negative word density and introducing the text length correction, the recognition of real negative emotions is avoided due to the long text and the content redundancy, and the core dissatisfaction of the user is more accurately captured.

[0070] Please continue to refer to Figure 2 As shown, the sentiment feature extraction method further includes:

[0071] In step S202, the sentiment turning index is constructed according to the standard text, and the sentiment feature is determined according to the negative sentiment feature of the standard text and the sentiment turning index.

[0072] Specifically, the standard text is segmented, and the sentiment value of each segment of the standard text is extracted, and the sentiment value of the i-th segment of the standard text is denoted as Si.

[0073] The sentiment turning index is constructed according to the sentiment value of each segment of the standard text, and the expression of the sentiment turning index is:

[0074]

[0075] In the formula, I is the number of segments of the standard text, zz is the sentiment turning index, S(i-1) is the sentiment value of the (i-1)th segment of the standard text.

[0076] The sentiment feature QT is determined according to the negative sentiment feature of the standard text and the sentiment turning index, and QT=0.6×Qs+0.4×zz.

[0077] Specifically, in the embodiment, when the standard text is segmented, the Chinese sentence end punctuation (such as. ;,!?, etc.) is used as the reference to cut, the sentence containing transition conjunctions (such as but / only / although, etc.) is protected (temporary markers are inserted to prevent mis-cutting), and the long sentence without punctuation is disassembled by a Chinese word segmentation tool, then recombined according to 2-6 word semantic units, and finally, the adjacent segments of the super short are merged and the empty segments are filtered to ensure that each segment carries independent emotional semantics; the segmentation process of the standard text in the embodiment is not specifically limited, and a person skilled in the art can freely set it according to the needs.

[0078] ​Exemplarily, in the embodiment, a RoBERTa-wwm-ext Chinese pre-training model can be used as a basic architecture, the characteristics of e-commerce complaint text are adapted through domain fine-tuning, and the sentiment value (range [-1, 1]) of each segment of standard text is obtained by fine-tuning the model using the complaint data with artificially labeled sentiment intensity. In the embodiment, the above settings are not specifically limited, and a person skilled in the art can freely set them according to the needs.

[0079] Specifically, the text sentiment fluctuation is analyzed by segment, the integrity of transition sentences such as "but" and "however" is specially protected, and the dramatic change of user emotion is quantified by combining the sentiment jump degree between paragraphs (such as suddenly changing from praise to criticism), so as to identify hidden strong dissatisfaction or potential escalation risks.

[0080] Please continue to refer to Figure 1 As shown, the AI-based data classification method further includes:

[0081] In step S103, acoustic features are extracted from the standard text and the target voice data, and an emergency acoustic index is constructed according to the acoustic features.

[0082] Specifically, the standard text of the standardized voice transcription text is taken as a standard voice transcription text, the total number of words of the standard voice transcription text is extracted, and the ratio of the total number of words of the standard voice transcription text to the duration of the target voice data is taken as the target speech speed MY.

[0083] A fundamental frequency point is extracted every 50 ms for the target voice data to obtain a fundamental frequency sequence (F1, F2,..., Fn), F1 is the fundamental frequency of the first fundamental frequency point, F2 is the fundamental frequency of the second fundamental frequency point, Fn is the fundamental frequency of the nth fundamental frequency point, and n is the number of fundamental frequency points.

[0084] The variance of each fundamental frequency point in the fundamental frequency sequence is calculated, and the calculation result is taken as the pitch variance FG.

[0085] The target voice data is divided into frames according to a preset frame length to obtain an audio frame sequence, and the energy of each audio frame is obtained. When the energy of the audio frame is less than or equal to an energy threshold, the audio frame is determined as a silent frame, otherwise, the audio frame is determined as a normal frame.

[0086] Continuous silent frames are merged into a silent section. When the silent section is less than or equal to a preset duration, the silent section is determined as an invalid pause. When the silent section is greater than the preset duration, the silent section is determined as a valid pause.

[0087] The number of valid pauses in the target voice data is counted as Y, and the ratio of Y to the duration of the target voice data is taken as the pause frequency TD.

[0088] An emergency acoustic index is constructed based on the target speech speed MY, the pitch variance FG, and the pause frequency TD, and the expression of the emergency acoustic index is:

[0089] SSR=0.5×MY / MZ+0.3×tanh(FG / FZ)+0.2×TD;

[0090] In the formula, MZ is a speech speed threshold, and FZ is a pitch variance threshold.

[0091] Specifically, in the embodiment, the preset frame length is 25 ms, and the preset time length is 0.5 seconds.

[0092] Specifically, the energy threshold is a critical energy value for distinguishing whether a speech frame is silent, and the energy threshold has a value range of -60 dB to -45 dB, and in the embodiment, the energy threshold is -50 dB.

[0093] Specifically, in the embodiment, the audio frame energy can be obtained through a Python audio processing library, and the fundamental frequency sequence can be obtained through a pyworld library of Python; the above setting is not specifically limited in the embodiment, and a person skilled in the art can freely set according to the needs.

[0094] Specifically, the speech speed threshold is an upper limit of normal speech speed (words per second), and the speech speed threshold has a value range of 4.0-6.0 words per second, and in the embodiment, the value of the speech speed threshold is 5 words per second, and the pitch variance threshold is an upper limit of the pitch variance of speech, and the pitch variance threshold has a value range of 120-150 Hz², and in the embodiment, the pitch variance threshold is 140 Hz².

[0095] Specifically, in the embodiment, the unit of the time length of the target speech data is second, and in the construction process of the emergency acoustic index, the unit of the pause frequency is not considered, and only the value of the pause frequency is analyzed.

[0096] Specifically, the speech speed, the pitch fluctuation, and the abnormal pause frequency are extracted from the speech, and the sound characteristics are converted into quantifiable emergency signals. For example, too fast speech speed may reflect excited emotion, voice tremor (high pitch variance) implies anxiety, and long pause may be caused by emotional choking. These physical indicators are complementary to the text emotion, and are especially suitable for identifying the hidden state of the user who is strongly suppressing anger or is on the verge of collapse.

[0097] Please continue to refer to Figure 1 As shown in the figure, the AI-based data classification method further includes:

[0098] In step S104, the target data is classified according to the emotional features and the emergency acoustic index.

[0099] Specifically, when the emotional feature weight x QT + the acoustic weight x SSR is less than or equal to the classification threshold, the target data is classified as normal data, and when the emotional feature weight x QT + the acoustic weight x SSR is greater than the classification threshold, the target data is classified as emergency data.

[0100] Specifically, in the embodiment, the sum of the emotional feature weight and the acoustic weight is 1, the emotional feature weight is 0.7, the acoustic weight is 0.3, and the classification threshold is an emotional and acoustic weighted score threshold, and the value range of the classification threshold is [0.7, 0.8], and in the embodiment, the value of the classification threshold is 0.75.

[0101] Specifically, the text emotional intensity and the voice emergency degree are fused to avoid misjudgment of a single mode, for example, the user's text is calm but the voice is trembling, or the tone is smooth but the right word is frequently mentioned, and the system can recognize the real emergency degree through weighted comprehensive scoring. This ensures that high-priority complaints are handled first, optimizing customer service resource allocation.

[0102] Please continue to refer to Figure 1 As shown in the figure, the AI-based data classification method further includes:

[0103] Step S105, constructing a target behavior feature according to a standard text, and updating a classification result of the target data according to the target behavior feature.

[0104] Please refer to Figure 3 As shown in the figure, the target data classification result updating method includes:

[0105] Step S301, matching the standard text with a preset right word keyword to determine a right word tendency index.

[0106] Specifically, the standard text is matched with the preset right word keyword, and the number of successful matches is recorded as W, if W = 0, the right word tendency index is set to 0, if W is greater than 0 and less than 3, the right word tendency index is set to 0.5, and if W is greater than or equal to 3, the right word tendency index is set to 1.

[0107] Please continue to refer to Figure 3 As shown in the figure, the target data classification method further includes:

[0108] Step S302, matching the standard text with a preset platform keyword to determine a platform responsibility index.

[0109] Specifically, the standard text is matched with the preset platform keyword, if the matching is successful, the platform responsibility index is set to 1, and if the matching is not successful, the platform responsibility index is set to 0.

[0110] Specifically, identify whether the problem directly points to the platform itself, quickly distinguish between platform responsibility and third-party seller responsibility, ensure that key problems are handled by a dedicated team, and avoid responsibility shirking.

[0111] Please continue to refer to Figure 3 As shown, the target data classification method further includes:

[0112] Step S303, the standard text is matched with the preset group keyword to determine the group characteristic index.

[0113] Specifically, the standard text is matched with the preset group keyword, if the matching is successful, the group characteristic index is set to 1, and if the matching is not successful, the group characteristic index is set to 0.

[0114] Please continue to refer to Figure 3 As shown, the target data classification method further includes:

[0115] Step S304, determining the target behavior characteristic according to the right protection tendency index, the platform responsibility index and the group characteristic index, and updating the classification result of the target data according to the target behavior characteristic.

[0116] Specifically, the expression of the target behavior characteristic is: MT=0.5×right protection tendency index+0.2×platform responsibility index+0.3×group characteristic index, MT is the target behavior characteristic;

[0117] When the target behavior characteristic MT is less than the behavior characteristic threshold, the classification result of the target data is not updated, and when the target behavior characteristic MT is greater than or equal to the behavior characteristic threshold, if the classification result of the target data is ordinary data, the classification result of the target data is updated to urgent data, and if the classification result of the target data is urgent data, the classification result of the target data is not updated.

[0118] Specifically, the behavior characteristic threshold is a multi-feature comprehensive score threshold, and in this embodiment, the behavior characteristic threshold is 0.5.

[0119] Specifically, the three behavior signals of right protection, responsibility attribution and group diffusion are comprehensively considered to perform secondary calibration on the original classification result to prevent underestimating the actual risk due to the surface calm emotion.

[0120] Specifically, the matching success in this embodiment refers to that one preset keyword appears in the standard text.

[0121] Please refer to Figure 4 As shown, the AI-based data classification platform includes:

[0122] The acquisition unit is configured to collect target data.

[0123] The emotion feature construction unit is configured to perform standardization processing on the target text data to obtain a standard text, and construct a negative emotion feature and an emotion turning index according to the standard text, and determine the emotion feature based on the negative emotion feature and the emotion turning index;

[0124] The acoustic feature construction unit is configured to extract an acoustic feature according to the standard text and the target voice data, and construct an emergency acoustic index according to the acoustic feature;

[0125] The classification unit is configured to classify the target data according to the emotion feature and the emergency acoustic index;

[0126] The behavior feature construction unit is configured to construct a target behavior feature according to the standard text, and update the classification result of the target data according to the target behavior feature.

[0127] The AI-based data classification platform provided by the embodiments of the present application can execute the AI-based data classification method provided by any of the embodiments of the present application, and has the corresponding function modules and beneficial effects of the execution method.

[0128] Please refer to Figure 5 Fig. 1 shows a structural schematic diagram of an electronic device in the embodiment of the present application. The electronic device in the embodiment of the present application can include but is not limited to mobile terminals such as mobile phones, notebook computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Tablet Personal Computers), PMPs (Portable Multimedia Players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), wearable electronic devices, and the like, and fixed terminals such as digital TVs, desktop computers, smart home devices, and the like. Figure 5 The electronic device shown is only an example, and should not impose any limitation on the functions and use range of the embodiments of the present application.

[0129] As shown in Fig. 2, the electronic device includes the following components: Figure 5

[0130] The processor 501 includes at least one of a central processing unit (CPU), a graphics processing unit (GPU), or a field-programmable gate array (FPGA). Its function is to invoke the computer program and data stored in the memory 502, and generate control instructions accordingly;

[0131] The memory 502 includes a random access memory (RAM) and / or a non-volatile memory (NVM). The NVM can be a flash memory, a solid state drive (SSD), or a combination thereof. The memory is used to store computer programs, intermediate data generated during processing, and a historical data set;

[0132] ​The communication interface 503 comprises a wired communication module and a wireless communication module; the wired communication module supports Ethernet or RS-485 protocol and is used for connecting a sensor network; the wireless communication module supports LoRa, 5G or satellite communication protocol and is used for transmitting processing results to a remote server.

[0133] The system bus 504 adopts a PCI Express or AXI (AXI) bus architecture and is used for realizing high-speed data interaction and clock synchronization among the processor 501, the memory 502 and the communication interface 503.

[0134] The embodiment further provides a computer readable storage medium, which physically stores computer executable instructions, when the instructions are transmitted to a processing unit via an integrated circuit substrate, the instructions are encapsulated and processed through a data channel of a bus system and then solidified to a non-volatile storage area of a storage module, and the executable instructions are configured to realize the complete technical solution of the AI-based data classification method when a processor executes.

[0135] Obviously, the above embodiments of the present application are merely examples for clearly illustrating the present application, and are not intended to limit the implementation manners of the present application. For those skilled in the art, on the basis of the above description, other different forms of changes or variations can also be made, and it is impossible to enumerate all the implementation manners here. Any changes or variations derived from the technical solution of the present application still fall within the protection scope of the present application.

Claims

1. An AI-based data classification method, characterized by, The method comprises the following steps: Collect target data; Standardize the target text data to obtain a standard text, and construct a negative sentiment feature and a sentiment turning index according to the standard text, and determine a sentiment feature based on the negative sentiment feature and the sentiment turning index; Extract acoustic features from the standard text and the target voice data, and construct an emergency acoustic index according to the acoustic features; Classify the target data according to the sentiment feature and the emergency acoustic index; Construct a target behavior feature according to the standard text, and update the classification result of the target data according to the target behavior feature; Standardize the target text data to obtain a standard text, and construct a negative sentiment feature Qs according to the standard text; Construct a sentiment turning index according to the standard text, and determine a sentiment feature according to the negative sentiment feature and the sentiment turning index of the standard text; Segment the standard text, and extract the sentiment value of each segment of the standard text, denoted as Si; Construct a sentiment turning index according to the sentiment value of each segment of the standard text, and the expression of the sentiment turning index is: ; wherein I is the number of segments of the standard text, zz is the sentiment shift index, and S(i-1) is the sentiment value of the (i-1)th segment of the standard text. Extract a fundamental frequency point every 50 ms from the target voice data to obtain a fundamental frequency sequence (F1, F2,..., Fn), F1 is the fundamental frequency of the first fundamental frequency point, F2 is the fundamental frequency of the second fundamental frequency point, Fn is the fundamental frequency of the nth fundamental frequency point, and n is the number of fundamental frequency points; Calculate the variance of each fundamental frequency point in the fundamental frequency sequence, and take the calculation result as the pitch variance FG; Frame the target voice data according to a preset frame length to obtain an audio frame sequence, and obtain the energy of each audio frame, denoted as Ni, i is the audio frame number, and when Ni is less than or equal to an energy threshold, the audio frame is determined as a silent frame; Merge consecutive silent frames into a silent segment, and when the silent segment is less than or equal to a preset time length, the silent segment is determined as an invalid pause, and when the silent segment is greater than the preset time length, the silent segment is determined as a valid pause; Count the number of valid pauses in the target voice data as Y, and take the ratio of Y to the time length of the target voice data as the pause frequency TD; Construct an emergency acoustic index SSR based on the target speed MY, the pitch variance FG and the pause frequency TD; When the sentiment feature weight × QT + acoustic weight × SSR is less than or equal to a classification threshold, the target data is classified as normal data, and when the sentiment feature weight × QT + acoustic weight × SSR is greater than the classification threshold, the target data is classified as emergency data. 2.The AI-based data classification method of claim 1, wherein, Determine the sentiment feature QT according to the negative sentiment feature and the sentiment turning index of the standard text, QT = 0.6 × Qs + 0.4 × zz. 3.The AI-based data classification method of claim 2, wherein, Take the standard text of the standardized voice transcription text as a standard voice transcription text, extract the total number of words of the standard voice transcription text, and take the ratio of the total number of words of the standard voice transcription text to the time length of the target voice data as the target speed MY. 4.The AI-based data classification method of claim 3, wherein, Match the standard text with a preset right protection keyword to determine a right protection tendency index; Match the standard text with a preset platform keyword to determine a platform responsibility index; Match the standard text with a preset group keyword to determine a group feature index. 5.The AI-based data classification method of claim 4, wherein, Determine the target behavior characteristic according to the rights protection tendency index, the platform responsibility index and the group characteristic index, and update the classification result of the target data according to the target behavior characteristic; The expression of the target behavior characteristic is MT=0.5*rights protection tendency index+0.2*platform responsibility index+0.3*group characteristic index, and MT is the target behavior characteristic; When the target behavior characteristic MT is less than the behavior characteristic threshold, the classification result of the target data is not updated, and when the target behavior characteristic MT is greater than or equal to the behavior characteristic threshold, if the classification result of the target data is ordinary data, the classification result of the target data is updated to emergency data, and if the classification result of the target data is emergency data, the classification result of the target data is not updated.

6. An AI-based data classification platform, characterized by, It comprises: An acquisition unit is configured to collect target data; An emotional feature construction unit is configured to perform standardization processing on the target text data to obtain standard text, construct a negative emotional feature and an emotional turning point index according to the standard text, and determine an emotional feature based on the negative emotional feature and the emotional turning point index; An acoustic feature construction unit is configured to extract an acoustic feature according to the standard text and the target voice data, and construct an emergency acoustic index according to the acoustic feature; A classification unit is configured to classify the target data according to the emotional feature and the emergency acoustic index; A behavior feature construction unit is configured to construct a target behavior characteristic according to the standard text, and update the classification result of the target data according to the target behavior characteristic; The target text data is standardized to obtain standard text, and a negative emotional feature Qs is constructed according to the standard text; An emotional turning point index is constructed according to the standard text, and an emotional feature is determined according to the negative emotional feature and the emotional turning point index of the standard text; The standard text is segmented, and the emotional value of each segment of the standard text is extracted, and the emotional value of the ith segment of the standard text is denoted as Si; An emotional turning point index is constructed according to the emotional value of each segment of the standard text, and the expression of the emotional turning point index is: ; wherein I is the number of segments of the standard text, zz is the sentiment shift index, and S(i-1) is the sentiment value of the (i-1)th segment of the standard text. A fundamental frequency point is extracted every 50 ms of the target voice data to obtain a fundamental frequency sequence (F1, F2,..., Fn), F1 is the fundamental frequency of the first fundamental frequency point, F2 is the fundamental frequency of the second fundamental frequency point, Fn is the fundamental frequency of the nth fundamental frequency point, and n is the number of fundamental frequency points; The variance of each fundamental frequency point in the fundamental frequency sequence is calculated, and the calculation result is taken as the pitch variance FG; The target voice data is segmented according to a preset frame length to obtain an audio frame sequence, and the energy of each audio frame is obtained, denoted as Ni, i is the audio frame number, and when Ni is less than or equal to an energy threshold, the audio frame is determined as a silent frame; Continuous silent frames are merged into a silent segment, and when the silent segment is less than or equal to a preset time length, the silent segment is determined as an invalid pause, and when the silent segment is greater than the preset time length, the silent segment is determined as a valid pause; The number of valid pauses in the target voice data is counted as Y, and the ratio of Y to the time length of the target voice data is taken as the pause frequency TD; An emergency acoustic index SSR is constructed based on the target speech rate MY, the pitch variance FG and the pause frequency TD; When the emotional feature weight x QT + the acoustic weight x SSR is less than or equal to a classification threshold, the target data is classified as normal data, and when the emotional feature weight x QT + the acoustic weight x SSR is greater than the classification threshold, the target data is classified as emergency data.

7. An electronic device, comprising: The electronic device includes: one or more processors; a 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 AI-based data classification method as claimed in any one of claims 1-5.

Images