A data processing method and device for labeling and cleaning in cooperation, and a medium

Abstract

The embodiment of the specification discloses a data processing method and device for labeling and cleaning cooperation and medium, relates to the technical field of data processing, and the method comprises the following steps: obtaining an original data set, extracting semantic features in the original data set, marking the original data set, and generating labeled guided thermal data; input the labeled guided thermal data into a labeling node, real-time collect the labeling behavior interaction data of the target user of the labeling node in the labeled guided thermal data, the labeling behavior interaction data includes operation data coordinate sequence and coordinate correction sequence; based on the labeling behavior interaction data, identify the semantic suspicious area, so as to detect the entity boundary of the semantic suspicious area through the cleaning node, and generate the corresponding semantic level cleaning suggestion; according to the semantic level cleaning suggestion, the labeling of the semantic suspicious area is corrected, the labeling information corresponding to the original data set is determined, and according to the semantic level cleaning suggestion, the data cleaning rule is optimized, so as to realize data cleaning.

Description

Technical Field

[0001] This specification relates to the field of data processing technology, and in particular to a data processing method, device and medium for coordinated annotation and cleaning. Background Technology

[0002] In current data processing workflows, annotation and cleaning are crucial preprocessing steps, but they are typically relatively independent. During annotation, the lack of standardized criteria and misunderstandings can lead to errors or non-standard results. Furthermore, raw data may contain significant noise and errors; labeling without cleaning increases workload and error rates. Traditional data cleaning methods often rely solely on statistical characteristics or simple rules, failing to fully utilize semantic information and struggle to accurately identify and handle complex errors. Moreover, the separate annotation and cleaning processes make data processing cumbersome, inefficient, and compromise data quality, failing to meet the urgent demand for high-quality data in fields like artificial intelligence and big data.

[0003] When labeling is prioritized, raw data containing potential quality defects must be fully labeled, resulting in a significant waste of labeling resources on samples that essentially require cleaning, leading to substantial redundant and ineffective labeling. Conversely, if cleaning is prioritized, processing mechanisms relying on statistical features or simple rules struggle to identify complex semantic inconsistencies, causing data with deep semantic value to be mistakenly cleaned due to normal surface statistical characteristics. Therefore, in existing data processing workflows, data labeling and data cleaning are executed sequentially and independently, resulting in data processing efficiency and quality failing to meet requirements. Summary of the Invention

[0004] This specification provides one or more embodiments of a data processing method, device, and medium that combines labeling and cleaning to solve the following technical problem: In existing data processing workflows, data labeling and data cleaning are executed sequentially and independently, resulting in data processing efficiency and quality failing to meet requirements.

[0005] One or more embodiments of this specification employ the following technical solutions:

[0006] This specification provides one or more embodiments of a data processing method for collaborative annotation and cleaning. The method includes: acquiring an original dataset; extracting semantic features from the original dataset; labeling the original dataset; generating annotation-guided heatmap data; inputting the annotation-guided heatmap data into annotation nodes; and collecting in real-time annotation behavior interaction data of target users in the annotation-guided heatmap data, wherein the annotation behavior interaction data includes an operation data coordinate sequence and a coordinate correction sequence; identifying semantically questionable regions based on the annotation behavior interaction data; performing entity boundary detection on the semantically questionable regions through cleaning nodes; generating corresponding semantic-level cleaning suggestions; correcting the annotations of the semantically questionable regions according to the semantic-level cleaning suggestions; determining the annotation information corresponding to the original dataset; and optimizing the data cleaning rules according to the semantic-level cleaning suggestions to achieve data cleaning.

[0007] This specification provides one or more embodiments of a data processing apparatus for coordinated labeling and cleaning, comprising:

[0008] At least one processor; and,

[0009] A memory communicatively connected to the at least one processor; wherein,

[0010] The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the above-described method.

[0011] This specification provides one or more embodiments of a non-volatile computer storage medium storing computer-executable instructions configured to perform the above-described method.

[0012] The above-mentioned technical solutions adopted in the embodiments of this specification can achieve the following beneficial effects: Through the technical solutions of the embodiments of this specification, the spatial semantic labeling method of the annotation-guided heat data can accurately identify high semantic value regions in the original dataset before the annotation starts, significantly reduce the scope of annotation work through pre-screening, and thus greatly reduce the investment of annotation resources for invalid data to be cleaned; Based on the dynamic semantic parsing of annotation behavior interaction data, a real-time transformation path from the operation data coordinate sequence and coordinate correction sequence to semantically questionable regions is constructed. Through the deep coupling of behavior patterns and spatiotemporal features, the traditional unquantifiable annotation confusion is transformed into locatable semantic anomaly coordinates, providing accurate cleaning targets for cleaning nodes; The annotation of semantically questionable regions is corrected according to semantic-level cleaning suggestions, improving the quality of annotation information; The semantic completion logic corresponding to the cleaning suggestions is compiled into incremental parameters of the rule base, enabling the cleaning process to dynamically adapt to complex semantic scenarios, fundamentally solving the problems of resource waste, semantic decay and rule rigidity caused by the separation of annotation and cleaning, and improving the efficiency of annotation and cleaning while ensuring data quality. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments or prior art of this specification, the drawings used in the description of the embodiments or prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0014] Figure 1 A flowchart illustrating a data processing method for coordinated labeling and cleaning provided in an embodiment of this specification;

[0015] Figure 2 This is a schematic diagram of a data processing device for collaborative labeling and cleaning, provided as an embodiment of this specification. Detailed Implementation

[0016] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0017] This specification provides a data processing method for collaborative labeling and cleaning. It should be noted that the execution entity in this specification embodiment can be a server or any device with data processing capabilities. Figure 1 This is a flowchart illustrating a data processing method for coordinated annotation and cleaning provided in an embodiment of this specification, as shown below. Figure 1 As shown, the main steps include the following:

[0018] Step S101: Obtain the original dataset, extract semantic features from the original dataset, label the original dataset, and generate labeled guided heatmap data.

[0019] Extracting semantic features from the original dataset and labeling the original dataset to generate labeled guidance heatmap data specifically includes: obtaining text data and image data from the original dataset; performing semantic extraction on the text data and image data respectively to determine the text highlight coordinate set and the image warning area coordinate set; and labeling the original dataset using the text highlight coordinate set and the image warning area coordinate set to generate the labeled guidance heatmap data. Semantic extraction is performed on the text data and the image data to determine the text highlight coordinate set and the image warning region coordinate set. Specifically, this includes: determining the word vector sequence corresponding to the text data; parsing the word vector sequence using a pre-trained language model to determine the polarity score corresponding to each word vector; filtering positive and negative word subsets from the word vector sequence based on the polarity score corresponding to each word vector; determining the first positive word and the first negative word as conflicting word pairs when the dependency distance between the first positive word in the positive word subset and the first negative word in the negative word subset is not greater than a preset distance threshold; determining the text highlight coordinate set based on the word pair position information of multiple conflicting word pairs; acquiring the image data from the original data; parsing the pixel gradient tensor of the image data using a preset edge detection model to determine gradient abrupt change regions; and determining the image warning region coordinate set based on the gradient abrupt change coordinates corresponding to the gradient abrupt change regions.

[0020] In traditional data preprocessing pipelines, annotation users need to process the entire dataset in the traditional annotation process. However, raw datasets often contain semantically mixed regions, such as contradictory descriptions in text and objects with blurred boundaries in images. These regions essentially require cleaning, filtering, or reconstruction. Without pre-labeling guidance, a lot of effort will be spent on low-value semantic regions. The role of annotation-guided heatmap data is to establish semantic spatial navigation. By pre-identifying the coordinates of sentiment conflicts in text and abnormal areas in image structure through algorithms, it is transformed into visual heatmaps, allowing annotators to focus on processing high-semantic-value regions and avoid wasting resources on dirty data that needs cleaning.

[0021] In one embodiment of this specification, the raw dataset to be processed is obtained. It should be noted that the raw dataset includes data of multiple modalities, such as text data and image data. Therefore, the generation of labeled and guided heatmap data includes a process for generating a text highlight coordinate set and a process for generating an image warning region coordinate set.

[0022] The text data in the original dataset is processed into word-level vectors. A pre-trained language model, such as a BERT-like model based on the Transformer architecture, is used to parse the word vector sequences. A fully connected neural network layer is added to the end of the language model to output the polarity score of each word vector in the range [-1, 1]. Positive values ​​indicate positive semantic tendency, and negative values ​​indicate negative semantic tendency. Two types of keywords are located by traversing the sequence: high-positive-score words, such as "extremely fast" and "stable," and high-negative-score words, such as "stuttering" and "fault." These high-positive-score and high-negative-score words can be defined using preset thresholds. For example, positive words with a polarity score > 0.8 are identified as a positive word subset; similarly, negative words with a polarity score < -0.8 are identified as a negative word subset. It should be noted that 0.8 and -0.8 are examples here, and the thresholds can be set according to the actual application scenario. The syntactic distance between words is calculated using a dependency parsing tree. When the length of the connection path between a pair of positive and negative words in the dependency tree is lower than a set threshold, such as the dependency distance between "smooth" and "heating" in "runs smoothly but overheats severely" being only 2, it is determined to be a conflicting word pair. This threshold can be set to 5, or it can be set according to requirements. Based on the word pair position information of multiple conflicting word pairs, the start and end positions of the character offsets of all conflicting word pairs in the text stream are extracted to determine the text highlight coordinate set, which is then highlighted in the annotation interface, such as mapped as a yellow semi-transparent highlight area.

[0023] In the process of generating the coordinate set of the image warning region, multi-scale gradient field calculations are performed on the image data in the original dataset. An improved edge detection operator, such as a hybrid convolution kernel that integrates Canny and Sobel characteristics, is used to generate a pixel-level gradient magnitude map. Based on the statistical analysis of local gradient distribution, a dynamic threshold is set. After Gaussian smoothing of the gradient magnitude matrix, the set of pixels with gradient values ​​significantly higher than the surrounding mean is extracted. The standard for "significantly higher" can be determined by a threshold difference between the gradient value and the surrounding mean. When the difference between the gradient value of a pixel and the surrounding mean is greater than a preset threshold, it is identified as a pixel set. A connected component analysis algorithm is applied to cluster discrete high-gradient pixels to generate the coordinates of the minimum bounding rectangle enclosing the abrupt change region. In the annotation interface, the rectangular overlay is rendered as a red transparent mask. The mask transparency is negatively correlated with the gradient peak within the region, i.e., the more significant the anomaly, the lower the transparency. A coordinate index table is generated simultaneously to record the mapping relationship between the region and the original data.

[0024] Through the above technical solutions, the algorithm predicts high-risk areas that require cleaning intervention, allowing users to avoid low-value data during operation and avoid resource consumption caused by full annotation in traditional processes. Conflict word pair identification based on dependency syntactic distance and abnormal region localization based on dynamic gradient establish accurate spatial semantic navigation for the annotation process, solving the semantic rework problem caused by blind annotation and effectively improving annotation efficiency.

[0025] Step S102: Input the annotation guidance heatmap data into the annotation node and collect the annotation behavior interaction data of the target user in the annotation guidance heatmap data in real time.

[0026] The annotation behavior interaction data includes an operation data coordinate sequence and a coordinate correction sequence;

[0027] In one embodiment of this specification, annotation guidance thermal data is input to the annotation node to guide the user to perform annotation operations at the annotation node.

[0028] The system collects real-time interaction data on the annotation behavior of the target user within the annotation guidance heatmap. Specifically, this includes: listening to user interface interaction events within the annotation node, determining the screen coordinates triggered by the target user on the annotation canvas, converting these screen coordinates into data space coordinates to determine the operation data coordinate sequence, which includes operation data space coordinates and operation timestamps; based on this operation data coordinate sequence, determining the region hover duration, filtering data regions where the hover duration exceeds a preset threshold, and determining a region hover duration mapping table, which includes hover region spatial coordinates and hover duration; acquiring the real-time annotation event log corresponding to the annotation tool, extracting label modification events within each hover region spatial coordinate based on the hover region spatial coordinates in the region hover duration mapping table and the real-time annotation event log; and determining the coordinate correction sequence corresponding to each hover region spatial coordinate based on the operation timestamp within the operation data coordinate sequence, where the coordinate correction sequence represents the label modification event flow within the hover region spatial coordinate.

[0029] In semantic collaborative processing, traditional methods often use annotation behavior only for post-event statistical analysis, failing to influence cleaning decisions in real time. The cognitive load of annotation users in semantically ambiguous areas, such as repeated boundary modifications, cannot be immediately translated into cleaning instructions, wasting the semantic correction intent. Secondly, the cleaning process can only respond to static rules, failing to perceive the semantic ambiguity inherent in dynamic behaviors, such as prolonged hesitation regarding contradictory descriptions. By establishing a real-time conversion channel between behavior and semantics, spatial coordinate sequences and temporal patterns are captured as annotation operations occur, quantifying cognitive states into computable interactive features.

[0030] In one embodiment of this specification, a user interface listener is deployed when the annotation node is started to monitor user interface interaction events in the annotation node and capture mouse movement and click event streams in real time. When the cursor moves on the annotation canvas, mouse movement events on the annotation canvas are monitored, and screen pixel coordinates are sampled at fixed time intervals (e.g., 20 frames per second). The screen coordinate system is mapped to the data space coordinate system through a preset affine transformation matrix. Text data is converted into character offsets, such as characters 15-20 in a sentence, and image data is converted into pixel bounding boxes, such as the coordinates of the top left corner (120, 80) - the coordinates of the bottom right corner (150, 110). The data space coordinates and timestamps of each sampling point are encapsulated into an operation data coordinate sequence to form the original behavior trajectory containing spatiotemporal information.

[0031] Perform sliding window analysis on the data space coordinate sequence. When multiple consecutive sampling points are located in the same data area (meaning ±5 characters of text or a 10×10 pixel area of ​​an image), and the displacement is less than the tolerance threshold, start a hover timer. Continuously monitor whether the cursor leaves the boundary of this area. If it remains stationary, accumulate the duration; if it exceeds the range, stop the timer and record the total hover duration. Filter the coordinates of areas exceeding the baseline hover duration (e.g., 3 seconds) to generate an area hover duration mapping table. The key is the data space area identifier, and the value is the duration accurate to milliseconds.

[0032] Synchronously monitor label modification events from the annotation tool, recording the semantic label type, modification timestamp, and affected data coordinate range for each operation. For example, changing the semantic label type from "negative" to "suspicious." Align each key region in the region hover duration mapping table with the annotation event log in both time and space, filtering out all modification events occurring within that region whose timestamps fall within the hover duration. Arrange the label modification actions within that region in chronological order, forming a coordinate correction sequence. Key attributes include operation type, target label, and operation timestamp. The operation type here includes any one or more of the following: create, delete, and modify operations.

[0033] Through the above technical solution, the precise mapping from screen coordinates to data space coordinates restores the continuous process of the user in the text semantic conflict area or image structural ambiguity area during the annotation process, forming a two-dimensional quantification of spatial positioning and temporal continuity; the dynamic correlation analysis between hovering duration and modification events can distinguish between unconscious operations and deep decision-making behaviors, such as repeated corrections after long pauses, and obtain semantic doubt areas that affect annotation quality, thus solving the problem of the disconnect between traditional cleaning rules and dynamic annotation semantics.

[0034] Step S103: Based on the labeled behavioral interaction data, identify semantically questionable regions, and perform entity boundary detection on the semantically questionable regions through the cleaning nodes to generate corresponding semantic-level cleaning suggestions.

[0035] Based on the annotation behavior interaction data, semantically questionable regions are identified, specifically including: analyzing the label modification time stream according to the coordinate correction sequence to determine at least one corresponding label modification event and counting the number of label modifications; determining the adjacent modification time interval of the label modification event based on the operation data coordinate sequence; and identifying the semantically questionable region by the number of label modifications and the adjacent modification time interval.

[0036] In complex data annotation scenarios, traditional methods rely on manual sampling or post-hoc statistical analysis to identify problem areas, resulting in time lags. Frequent modification behaviors during the annotation process, such as repeatedly adjusting sentiment labels on the same text fragment, reveal semantic uncertainties that cannot be captured in real time. Furthermore, fine-tuning of bounding boxes in image annotation often implies discrepancies in structural understanding, but conventional methods only consider the final result and ignore the decision-making process. Repeated modifications by annotation users to certain areas may be due to operational habits rather than genuine semantic uncertainty. Hovering duration exceeding a threshold reflects decision-making hesitation, indicating that annotation users are weighing multiple interpretations. Combined with frequent modification behaviors, this helps eliminate misjudgments; for example, rapid operations might be due to interface lag. By deeply deconstructing the behavioral data flow, coupling the operational density in the temporal dimension with coordinate corrections in the spatial dimension, we can achieve in-depth analysis from behavioral appearances to semantic contradictions.

[0037] In one embodiment of this specification, within a pre-screened hovering area's spatial coordinate range, the original operation records corresponding to the coordinate correction sequence are parsed. All label modification events occurring within this area are extracted chronologically, with each event containing the operation type, the modified label value, and a timestamp accurate to milliseconds. The total number of modification events within a unit time window is counted and recorded as the label modification count; this unit time window can be set to a 5-second sliding window. The time difference between adjacent modification events is calculated based on the operation timestamps, generating a time interval sequence. The sliding window method is used to analyze the interval distribution pattern. When the standard deviation of the time interval within the window is significantly lower than the average interval (e.g., less than 1 / 3 of the mean), it indicates highly concentrated continuous modification behavior. The interval decay trend is simultaneously detected; if adjacent intervals continuously shorten, such as from 2.1 seconds to 1.3 seconds and then to 0.8 seconds, it reflects an escalation of decision-making anxiety among the labeling user. It should be noted that this process needs to exclude invalid operation interference and filter instantaneous events with a duration lower than a preset minimum threshold (e.g., 0.3 seconds) to avoid mistakenly identifying accidental interface touches as valid modifications.

[0038] When the number of modifications exceeds a threshold and the time interval shows a decreasing trend, the region is considered semantically questionable. This threshold can be set to four times. For text data, lexical analysis is used to verify semantic conflicts, such as whether the modification event occurs in a region of contradictory sentiment words. For image data, edge gradient detection results are correlated, such as whether high-frequency modification regions overlap with gradient abrupt change regions. Finally, the spatial coordinates and confidence score of the semantically questionable region are output. The confidence score is calculated by weighting the number of modifications, the interval decay slope, and the intensity of semantic conflict.

[0039] By identifying semantically questionable regions, a real-time transformation path from behavioral data to cleaning instructions is constructed. Spatial coordinates are generated based on high-frequency correction patterns and temporal decay features in labeled behavioral interaction data, so that computationally intensive operations such as dependency parsing and sub-pixel reconstruction are applied only to key regions, effectively improving operational efficiency compared to the full data processing mode. Furthermore, the correction results of semantically questionable regions can be transformed into rule parameters, such as adding sentiment analysis scenario constraints and adjusting edge detection sensitivity, forming a closed-loop optimization chain of annotation feedback, rule iteration, and cleaning upgrade, which has the ability to continuously adapt to complex semantic scenarios.

[0040] The system performs entity boundary detection on the semantically questionable region through cleaning nodes, generating corresponding semantic-level cleaning suggestions. Specifically, this includes: performing structural topology analysis on the semantically questionable region to label semantic anomaly types, where each semantic anomaly type includes one or more of the following: missing components, broken paths, and conflicting modifications; dynamically generating the semantic-level cleaning suggestion based on the semantic anomaly type by querying a pre-defined cleaning rule base. Specifically, the structural topology analysis on the semantically questionable region to label semantic anomaly types includes: performing structural topology analysis on text-type semantically questionable regions, parsing syntactic core components and modification relationships to generate text logical dependency information; performing structural topology analysis on image-type semantically questionable regions, reconstructing continuous boundary paths to generate image geometric topology information; and labeling the semantic anomaly type corresponding to the semantically questionable region based on the text logical dependency information and / or the image geometric topology information. Based on the textual logical dependency information and / or the image geometric topology information, the semantic anomaly type corresponding to the semantically questionable region is marked, specifically including: when the subject node or object node in the textual logical dependency information lacks a referent, the anomaly type is determined to be component missing; when the boundary path length of the image geometric topology information is lower than the effective connectivity threshold, the anomaly type is determined to be path break; when there are opposing semantic modifiers bound to the same node in the textual logical dependency information, the anomaly type is conflict modification.

[0041] In traditional data processing workflows, cleaning operations are typically independent of annotation and semantic understanding. Cleaning rules based on statistical features cannot identify deep structural anomalies, such as semantic ambiguity caused by missing subjects in text. They can only handle superficial errors like spelling corrections. Furthermore, simple interpolation repair of blurred image boundaries distorts the original semantics, such as smoothing a lobulated tumor outline into an ellipse. By using structural topology analysis, annotation cognition is transformed into machine-executable semantic completion instructions. Semantically questionable regions only indicate the location of the problem, such as character offsets in a sentence, without revealing the type of structural defect. Entity boundary detection transforms spatial coordinates into a quantifiable translation layer of structural anomalies. Cleaning suggestions can then address semantic contradictions, improving cleaning efficiency.

[0042] In one embodiment of this specification, dependency syntax tree parsing is performed on semantically ambiguous regions of the text to separate the core components and subordinate modification relations. First, the core predicate nodes in the syntax tree are located, tracing the subject noun phrases and object components they govern. Second, the modification binding state is detected, identifying the positions of degree adverbs and negation words. Finally, structured logical dependency information is generated, where nodes represent the core elements (subject, predicate, object) and edges represent modification relations. Key processes include tracing the origin of dangling modifiers, such as marking "fast power consumption" as a detached modifier when there is no subject, and verifying the binding state of contradictory words, such as "high efficiency" and "high power consumption" simultaneously modifying the same predicate.

[0043] Image sub-regions are extracted based on the bounding box coordinates of semantically questionable regions and input into a super-resolution model for pixel-level reconstruction. A four-fold upsampled image is generated through multi-scale feature fusion, and non-maximum suppression is performed along the object's edge trajectory to extract sub-pixel-level keypoint sets. A path tracing algorithm is used to connect adjacent keypoints, generating a geometric topology map, which is a closed polygon. Double verification is performed on the reconstructed path: the total path length is measured, and if the tumor contour path is less than a predetermined proportion of the expected perimeter, it is considered a break; and the local curvature change rate is calculated, and if the abrupt change in the angle between adjacent line segments exceeds a tolerance threshold, it is marked as a structural malformation region.

[0044] The text's logical dependency information and the image's geometric topology information are matched against a pre-defined anomaly pattern library. When a subject / object node is detected to have no valid lexical binding (e.g., no subject before the predicate), it is marked as a missing component anomaly. When a boundary path has discontinuous nodes or a length below a threshold (e.g., the reconstructed blood vessel length is less than the original detection value), it is marked as a path break. When the same core node is attached with semantically conflicting modifiers, it is marked as a conflict modifier. Based on the anomaly type query, the cleaning rule library is used. Missing components trigger suggestions to supplement missing subject classes, path breaks trigger reconnection coordinate instructions, and conflict modifiers trigger guidance based on application scenarios. Finally, the cleaning operation information is output in natural language format.

[0045] Through the aforementioned technical solutions, structural topology analysis delves into the semantic layer, unifying syntactic structural defects in text and geometric topological breaks in images into quantifiable anomaly types, thus identifying the core contradictions affecting annotation quality. In text scenarios, it accurately identifies deep semantic defects such as missing subjects and modification conflicts; in the image domain, it precisely diagnoses visual defects such as boundary breaks and structural distortions. Problems in the annotation process are transformed into clear anomaly classification labels. Based on precise anomaly type matching and cleaning suggestions, customized repair solutions for structural defects are provided, achieving semantic-level error correction rather than data deletion. This allows the cleaning process to directly inherit the semantic understanding capabilities of the annotation user, repairing data while preserving its original value, and avoiding the problem of erroneous deletion of valid data in complex semantic scenarios during traditional cleaning processes.

[0046] Step S104: Correct the annotations of semantically questionable regions according to semantic-level cleaning suggestions, determine the annotation information corresponding to the original dataset, and optimize the data cleaning rules according to semantic-level cleaning suggestions to achieve data cleaning.

[0047] In one embodiment of this specification, an interactive layer displaying semantic-level cleaning suggestions is shown on the annotation interface, such as a semi-transparent annotation box covering text or flashing boundary points in an image. User actions are recorded, and when the user clicks the "Accept" button, the corrected key semantic elements are extracted. Newly added scene-qualifying words and completed subject components are captured in the text data; the adjusted key point coordinate sequence is recorded in the image data. Simultaneously, semantic markers actively added by the user are identified, such as checking the "Scene Verification Required" checkbox, generating a snapshot of the correction result with spatial location markers. The semantic supplementary information in the correction snapshot is parsed and associated with the spatial coordinates of the original data. The physical location of semantically questionable areas in the original dataset is located, and existing rule entries covering the spatial location are retrieved from the cleaning rule base. Parameters are dynamically compiled and optimized according to the correction type. In component completion type corrections, the added subject / object words are extracted, and subject constraints are added to the rule condition field; in the case of scene refinement type corrections, scene-qualifying phrases are extracted, and exemption conditions are added to the negative sentiment rules, such as NOT scene = "during gameplay". For structural optimization corrections, the redrawn boundary coordinate set is fitted as a vector path, a boundary smoothness threshold is calculated, and the image structure rules are updated. Additionally, a rule weight iterator can be set to increase the priority of new rules when they reach a preset accuracy on the validation set, otherwise triggering a manual review process.

[0048] The spatial semantic labeling method for guiding thermal data through the embodiments of this specification can accurately identify high semantic value regions in the original dataset before labeling begins. Pre-screening significantly reduces the scope of labeling work, thereby greatly reducing the investment of labeling resources on invalid data to be cleaned. Based on dynamic semantic parsing of labeling behavior interaction data, a real-time transformation path from operational data coordinate sequences and coordinate correction sequences to semantically questionable regions is constructed. Through deep coupling of behavioral patterns and spatiotemporal features, traditionally unquantifiable labeling confusion is transformed into locatable semantic anomaly coordinates, providing precise cleaning targets for cleaning nodes. Labeling of semantically questionable regions is corrected according to semantic-level cleaning suggestions, improving the quality of labeling information. The semantic completion logic corresponding to the cleaning suggestions is compiled into incremental parameters of the rule base, enabling the cleaning process to dynamically adapt to complex semantic scenarios. This fundamentally solves the problems of resource waste, semantic decay, and rule rigidity caused by the separation of labeling and cleaning, improving labeling and cleaning efficiency while ensuring data quality.

[0049] It should be noted that the embodiments in this specification apply to a data processing system. This system provides diverse annotation tools. For text data, it supports manual annotation of keywords, entity names, sentiment categories, and other information. Annotators can select text segments and add corresponding tags through a visual interface. For image data, it supports rectangular boxes, polygonal boxes, point annotations, and other methods to annotate the category and location of target objects in the image. Simultaneously, the system sets up an annotation rule library to standardize the annotation process, such as specifying label naming conventions and annotation granularity, to ensure the consistency and accuracy of the annotation results. Furthermore, to improve annotation efficiency, the system has an automatic annotation assistance function. Based on existing labeled data and machine learning algorithms, it performs preliminary annotation on new data. Annotators only need to review and correct the preliminary annotation results. In other words, the methods in the embodiments of this specification are for scenarios where manual annotation is involved. Establish a mapping relationship between annotation results and data cleaning rules. For example, if multiple contradictory sentiment tags appear in text data annotations, or if the labeled target objects in image data are significantly inconsistent with the actual content, the data is judged as erroneous and deleted. For data with missing annotations, if the missing parts have a minor impact on the overall data value, they can be supplemented by filling in default values ​​or interpolating based on similar data; if the impact is significant, they are deleted. Simultaneously, use annotation information to perform consistency checks on the data, such as checking whether the annotations of the same entity in text are consistent, and whether there are discrepancies in the annotations of similar target objects in images, correcting inconsistent data.

[0050] For text data, the annotated text data is converted according to a unified encoding format and character set, and the text length is normalized. For image data, the images are uniformly adjusted to the specified size, resolution, and color mode, and the coordinate information of the annotations is converted according to a unified coordinate system. In addition, the annotation labels are standardized, with labels with similar meanings merged, and the naming and encoding methods of the labels are unified, so that the processed data has consistency and standardization in both format and semantics, which facilitates subsequent data use.

[0051] This specification also provides an embodiment of a data processing device that combines labeling and cleaning, such as... Figure 2 As shown, the device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the above-described method.

[0052] This specification also provides a non-volatile computer storage medium storing computer-executable instructions configured to perform the above-described method.

[0053] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0054] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0055] The devices, media, and methods provided in the embodiments of this specification are one-to-one correspondences. Therefore, the devices and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.

[0056] Those skilled in the art will understand that embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, this specification may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this specification may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0057] This specification is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this specification. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0058] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0059] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0060] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0061] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0062] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0063] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Without further limitation, the phrase "comprising a…" … ” The definition of a specific element does not preclude the presence of other identical elements in the process, method, product, or apparatus that includes the element.

[0064] The above description is merely one or more embodiments of this specification and is not intended to limit this specification. Various modifications and variations can be made to the one or more embodiments of this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of one or more embodiments of this specification should be included within the scope of the claims of this specification.

Claims

1. A data processing method that combines labeling and cleaning, characterized in that, The method includes: Obtain the original dataset, extract semantic features from the original dataset, label the original dataset, and generate labeled guided heatmap data; The annotation guidance heatmap data is input into the annotation node, and the annotation behavior interaction data of the target user in the annotation guidance heatmap data is collected in real time. The annotation behavior interaction data includes operation data coordinate sequence and coordinate correction sequence. Based on the labeled behavioral interaction data, semantically questionable regions are identified, and entity boundary detection is performed on the semantically questionable regions through the cleaning node to generate corresponding semantic-level cleaning suggestions. The annotations of the semantically questionable regions are corrected according to the semantic-level cleaning suggestions, the annotation information corresponding to the original dataset is determined, and the data cleaning rules are optimized according to the semantic-level cleaning suggestions to achieve data cleaning; Extracting semantic features from the original dataset, labeling the original dataset, and generating labeled guided heatmap data, specifically including: Obtain the text data and image data from the original dataset; Semantic extraction is performed on the text data and the image data respectively to determine the text highlight coordinate set and the image warning area coordinate set; The original dataset is labeled using the text highlight coordinate set and the image warning region coordinate set to generate the labeled guidance heatmap data. The semantically questionable region is subjected to entity boundary detection by the cleaning node, and corresponding semantic-level cleaning suggestions are generated, including: Structural topology analysis is performed on the semantically questionable region to mark semantic anomaly types, wherein the semantic anomaly types include any one or more of component missing, path breakage and conflict modification; Based on the semantic anomaly type, a preset cleaning rule base is queried to dynamically generate the semantic-level cleaning suggestions; Structural topology analysis is performed on the semantically questionable regions to identify semantic anomaly types, specifically including: Structural topological analysis is performed on semantically ambiguous regions of text types to parse syntactic core components and modification relations in order to generate textual logical dependency information; Structural topological analysis is performed on semantically ambiguous regions of an image type to reconstruct continuous boundary paths and generate image geometric topological information; Based on the textual logical dependency information and / or the image geometric topology information, mark the semantic anomaly type corresponding to the semantically questionable region.

2. The data processing method for collaborative labeling and cleaning according to claim 1, characterized in that, Semantic extraction is performed on the text data and the image data respectively to determine the text highlight coordinate set and the image warning region coordinate set, specifically including: The word vector sequence corresponding to the text data is determined, and the word vector sequence is parsed using a pre-trained language model to determine the polarity score corresponding to each word vector; Based on the polarity score corresponding to each word vector, a positive word subset and a negative word subset are selected from the word vector sequence; When the dependency distance between the first positive word in the positive word subset and the first negative word in the negative word subset is not greater than a preset distance threshold, the first positive word and the first negative word are determined to be a conflicting word pair. Based on the word pair position information of multiple conflicting word pairs, a text highlight coordinate set is determined. Image data is obtained from the original data, and the pixel gradient tensor of the image data is analyzed using a preset edge detection model to determine the gradient abrupt change region. The coordinate set of the image warning area is determined based on the gradient change coordinates corresponding to the gradient change region.

3. The data processing method for collaborative annotation and cleaning according to claim 1, characterized in that, Real-time collection of annotation behavior interaction data of target users in the annotation guidance heatmap, specifically including: Listen to the user interface interaction events in the annotation node, determine the screen coordinates triggered by the target user in the annotation canvas, convert the screen coordinates into data space coordinates to determine the operation data coordinate sequence, wherein the operation data coordinate sequence includes operation data space coordinates and operation timestamp; Based on the operation data coordinate sequence, the hovering duration of the region is determined, and the data regions whose hovering duration exceeds the preset duration threshold are filtered to determine the region hovering duration mapping table, wherein the region hovering duration mapping table includes the spatial coordinates of the hovering region and the hovering duration. Obtain the real-time annotation event log corresponding to the annotation tool, and extract the label modification event within each of the hovering area spatial coordinates based on the hovering duration mapping table and the real-time annotation event log; Based on the operation timestamp of each hovering area spatial coordinate within the operation data coordinate sequence, the coordinate correction sequence corresponding to each hovering area spatial coordinate is determined, wherein the coordinate correction sequence is used to represent the label modification event stream within the hovering area spatial coordinate.

4. The data processing method for collaborative labeling and cleaning according to claim 1, characterized in that, Based on the labeled behavioral interaction data, semantically questionable regions are identified, specifically including: Based on the coordinate correction sequence, the label modification time stream is analyzed to determine the corresponding at least one label modification event and to count the number of label modifications. Based on the operation data coordinate sequence, determine the adjacent modification time interval of the label modification event; The semantically questionable region is determined by the number of times the label is modified and the time interval between adjacent modifications.

5. The data processing method for collaborative annotation and cleaning according to claim 1, characterized in that, Based on the textual logical dependency information and / or the image geometric topology information, the semantic anomaly type corresponding to the semantically questionable region is marked, specifically including: When the subject node or object node in the text logical dependency information lacks a referent, the abnormal type is determined to be component missing. When the boundary path length of the image geometric topology information is lower than the effective connectivity threshold, the anomaly type is determined to be path breakage; When there are opposing semantic modifiers that bind the same node in the text logical dependency information, the anomaly type is determined to be conflict modification.

6. A data processing device for coordinated labeling and cleaning, characterized in that, The device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method as described in any one of claims 1-5.

7. A non-volatile computer storage medium storing computer-executable instructions, characterized in that, The computer-executable instructions are configured to perform the method as described in any one of claims 1-5.

Images