38results about How to "Reduce manual labeling" patented technology

The invention relates to a device and method for acquiring road information from map data. The method includes: acquiring point cloud map data and semantic map data; extracting road information in a semantic map from the semantic map data; acquiring a primary road boundary from the point cloud map data; matching a road boundary in the road information in the semantic map with the point cloud map data to obtain road boundary information of a to-be-generated semantic map; matching lane line information in the road information in the semantic map with the point cloud map data to obtain lane lineinformation of the to-be-generated semantic map; obtaining road information of the to-be-generated semantic map according to the road boundary information and the lane line information of the to-be-generated semantic map. By means of matching of data in a traditional data map with point cloud in high-precision map data, manual annotation is greatly saved, and the map drawing speed is increased.

The embodiment of the invention provides a video playing progress control method and a terminal. The method comprises the following steps: displaying at least one character marker belonging to a video on a playing interface of the video; if a selection operation of a target character marker in the at least one character marker is detected, obtaining facial feature data of a target character corresponding to the target character marker; searching a plurality of target key frames with facial feature data matched with the facial feature data of the target character in the video, and displaying frame images of the plurality of target key frames on the playing interface in the form of thumbnails; and if a selection operation of a target thumbnail in a plurality of thumbnails is detected, determining a time point of the target key frame corresponding to the target thumbnail; and adjusting the playing progress of the video to the frame image corresponding to the time point. By adoption of the video playing progress control method provided by the invention, artificial markers can be reduced, a user can automatically adjust the video playing progress, and the user experience is improved.

The invention relates to artificial intelligence, can be used for medical image analysis auxiliary scenes, and provides an instance segmentation model sample screening method. The method comprises thesteps: reading an original data set, selecting a first to-be-labeled sample of which the information amount is greater than that of remaining samples from an unlabeled set based on an active learningmode, obtaining a first annotation set in a mode of manually annotating a plurality of first to-be-annotated samples; selecting a second to-be-labeled sample of which the confidence is higher than aset value from all the remaining samples based on a semi-supervised learning mode, obtaining a second labeling set in a mode of pseudo labeling of the second to-be-labeled sample, and taking the firstlabeling set, the second labeling set and the labeled set as a training set together. According to the method, a large number of samples used for training the image instance segmentation model can beobtained while the manual annotation amount of the samples is reduced, and then the more ideal instance segmentation model accuracy can be achieved. In addition, the invention also relates to a blockchain technology, and both the original data set and the training set can be stored in the blockchain.

The invention discloses a photo background similarity clustering method based on a convolutional neural network. The method comprises the following steps of preprocessing an original image based on aconvolutional neural network algorithm so as to correct a direction of an identification target in the original image; carrying out instance segmentation on foreground image features and background image features of the recognition target contained in the original image, and carrying out background extraction; performing background separation on the image subjected to instance segmentation; performing feature extraction on the separated background image to obtain a high-dimensional spatial feature map; and performing similarity clustering processing on the high-dimensional spatial feature map.The invention further provides a computer program system for implementing the method. According to the method, based on a pixel-level instance segmentation algorithm, foreground areas (portraits andidentity cards) in a real application scene are detected and removed, similarity comparison is carried out through background areas, and meanwhile, the recognition accuracy can be greatly improved byutilizing the convolutional neural network obtained through migration training.

The invention discloses a short text sentiment analysis method based on a CNN (Convolutional Neural Network) bidirectional GRU attention mechanism. The method comprises the following steps: preprocessing a short text, denoising, segmenting words, labeling part of voice, and removing stop words; expressing sentences into a word sequence through negative sampling training by taking words as units through a continuous bag-of-words model (CBOW), and mapping the word sequence into a multi-dimensional vector to construct a word vector set; calculating the occurrence frequency of sentiment words in different data set documents, calculating sentiment scores, and converting the sentiment scores into a sentiment feature vector matrix; word embedding and feature embedding topology serves as input ofa convolutional neural network, and sentence representation is obtained through convolution and pooling; through a bidirectional GRU recurrent neural network, negative word turning words are set as parameter query items of an attention mechanism to obtain representation; and combining the two representation topologies to serve as full connection layer input, and outputting an emotion analysis result.

The invention discloses a method and a device for training a supervised machine learning model. The method includes the steps: generating a plurality of man-made images comprising movement states of same target objects at different time points in one or more time periods; recording annotation data related to movement of the target objects in one or more time periods in the generating process of the man-made images; generating multimedia streaming comprising movement based on the man-made images; using a plurality of frame data of the multimedia streaming as a plurality of input data of the model to execute operation in the model and acquire derivation data related to movement; comparing the derivation data with the annotation data to determine whether to adjust parameters of the model or not. By the method, a large quantity of manual annotation needed in the training process of the model can be omitted.

The invention provides a method and a system for searching whether-class problem key sentences in a reading understanding task. The method comprises the steps of selecting existing reading understanding question and answer data, preprocessing the question and answer data to obtain a data set, and then mining semantic information of questions and sentences in chapter paragraphs in the data set based on a coding layer network to obtain word embedding representation of each word; constructing an algorithm model, and calculating the questions and the text paragraphs mined through the coding layernetwork by using a neural network model and a TFIDF to obtain key sentences of whether problems are classified or not; and inputting to-be-read understanding question and answer data into the trainedalgorithm model, and predicting whether key sentences of questions are classified or not. According to the method, more key sentence supports can be provided, the weight of the key sentence is calculated through the combination of the bidirectional gating loop network and the TF-IDF, and the efficiency and accuracy of answering the whether-class problems are improved.

A method and a device for training a supervised machine learning model are disclosed. The method includes the following steps: generating an artificial image containing a target object; recording annotation data related to the target object in the process of generating the artificial image; using the artificial image as the input data of a model to perform the operation in the model in order to obtain derivation data related to the target object; and comparing the derivation data with the annotation data to determine whether or not to adjust the parameters of the model. Through the method, a lot of manual annotation required in the training process of the model can be omitted.

The invention discloses a labeling model training method and device, and relates to the field of computers, in particular to the field of data processing. The specific implementation scheme is as follows: training an initial labeling model by utilizing a first sample with a reference label to obtain a first labeling model; labeling the second sample by utilizing the first labeling model to obtaina second sample with model labeling; obtaining a third sample with a reference label, wherein the third sample is a part of the second sample; and optimizing the first annotation model according to the second sample with the model annotation and the third sample with the reference annotation. The data volume required by model training can be reduced, so that the manual operation amount and the time cost are saved.

The invention discloses a premade barcode and patient information matching method. The method comprises the following steps of S1, performing mode and parameter setting in a blood collection tube distribution machine, and adding a premade barcode blood collection tube to a test tube cabin corresponding to equipment; S2, reading detection item information of a patient; S3, distributing a premade barcode test tube to a printing tube-pasting region by the test tube cabin of the blood collection tube distribution machine, and performing premade barcode scanning to obtain premade barcode information; S4, binding the detection item information of the patient with the premade barcode information; and S5, detecting whether the barcode information is consistent with the information of the patient or not. According to the matching method provided by the invention, the information of the patient can be well bound with the premade barcode on the blood collection tube, so that the personal information and detection item information of the patient can be obtained only by reading the premade barcode information subsequently.

The invention provides a question generation method and system, and the method comprises the steps: recognizing a to-be-read understanding text based on a named entity recognition tool, and obtainingan answer part; substituting the to-be-read understanding text and the corresponding answer part into a pre-trained question generation model to generate a plurality of questions for answers; correcting the plurality of questions to obtain questions corresponding to the to-be-read understanding text; wherein the questions generate a model; based on an existing dialogue system and a reading understanding text, introducing a copying mechanism and a placeholder mechanism into an algorithm model of a multi-layer and multi-scale transformer network to replace a named entity in the reading understanding text. According to the method and system, the execution speed and accuracy of generating the questions are improved, the expansibility is improved, manual annotation is greatly reduced, and meanwhile, the readability and diversity of question generation are improved by utilizing an existing dialogue system.

Embodiments of the invention provide a document emotion analysis method and apparatus, an electronic device and a readable storage medium. The document emotion analysis method and apparatus can assistin improving an analysis effect, and enables emotion analysis to be closer to daily life. The method comprises the steps of obtaining a document, and preprocessing the document to obtain clauses andwords of the document; establishing index relationships between the clauses and the document, and between the words and the document; modeling the clauses and the words by utilizing a subject emotionmodel, generating emotions of the clauses and themes of the words in the document, and establishing corresponding relationships between the clauses and the words; according to the emotions of the clauses, the themes of the words, the corresponding relationships between the clauses and the words, and the index relationships between the clauses and the document, and between the words and the document, calculating probability distribution of "document-emotion-clause" and probability distribution of "document-theme-word"; and according to the probability distribution of "document-emotion-clause" and the probability distribution of "document-theme-word", calculating an emotional tendency of the document.

The embodiment of the invention discloses a system exception prediction model training method, a system state prediction method and device, equipment and a storage medium. The method comprises the steps of acquiring sample features; initializing the system anomaly prediction model according to set weight parameters; processing the sample features through the system anomaly prediction model to obtain prediction energy; constructing a target function based on the predicted energy; and in back propagation, updating weight parameters of the system anomaly prediction model through the objective function. Therefore, the abnormal condition of the system operation state can be predicted.

The invention belongs to the technical field of machine translation, and discloses a machine translation style migration performance improvement method based on iterative knowledge migration. The machine translation style migration performance improvement method based on iterative knowledge migration comprises the following steps: pre-training a translation model and a text style migration model, guiding the translation model by the text style migration model, constructing a pseudo-parallel sentence pair and performing data tuning, guiding the text style migration model by the translation model, and iteratively improving the translation style migration performance. According to the method, the problem of less training data in machine translation style migration is relieved. According to the data tuning model, grammar error correction is carried out by fully utilizing the original text and the text after style migration, so that the pseudo-parallel data can be smoother, and the quality of the pseudo-parallel data is effectively improved. The performance of the translation model and the text style migration model is improved.

The invention provides an abnormal behavior detection method and device, electronic equipment and a storage medium, and the method comprises the steps: inputting an abnormal behavior video image into a trained abnormal behavior detection model, and obtaining an abnormal behavior detection result; wherein the trained abnormal behavior detection model is obtained by inputting an abnormal behavior video sample image into a preset fusion network for training; wherein the abnormal behavior video sample image carries an abnormal behavior region information weak label; wherein the preset fusion network is obtained by fusing a classification model and a target detection model. According to the trained abnormal behavior detection model, the classification model and the detection model are fused to form a deep learning framework, and the detection model is used as a weak supervision model to assist in completing the task of image classification, so that excessive manual annotation is omitted, the classification model does not lose key information of a shallow network, and the detection efficiency is improved. Therefore, the classification accuracy of the model is improved.

The embodiment of the invention discloses a picture labeling method and device, computer equipment and a storage medium, and the method comprises the steps: obtaining a training sample picture, and adjusting the size of the training sample picture to a preset size threshold value; for every two training sample pictures, applying a preset feature point extraction algorithm to extract feature pointsof the training sample pictures and carrying out feature point pairing to determine feature point pairs; filtering the feature point pair by applying a preset noise filtering algorithm to obtain a first target feature point pair; if the number of the first target feature point pairs is greater than a preset feature point pair number threshold, selecting feature point pairs with the preset featurepoint pair number threshold as second target feature point pairs; calculating the distance between each feature point pair in the second target feature point pair and the sum of the distances; and ifthe sum of the distances is smaller than a preset distance threshold, marking the two training sample pictures to be similar. Labor cost of marking the training sample pictures in the picture recognition model training process based on artificial intelligence is reduced.

The embodiment of the invention provides a method and system for expanding a MOOC course concept, and the method comprises the steps of taking an online interactive game as a training environment, and carrying out the training to obtain a reinforcement learning model; performing concept extension based on an in-class concept set and the reinforcement learning model, obtaining user feedback in the extension process, and the in-class concept set being composed of course knowledge point content needing to be supplemented and explained in the MOOC course; and returning the feedback of the user to the interactive game to be expanded again until a preset target is reached, so as to obtain an expansion result. According to the embodiment of the invention, by using the training method of reinforcement learning, the invention can be applied to newly set courses on a large scale after training on some specific courses, and compared with a traditional method, a large amount of manual annotation is saved, so that the method has relatively high ductility. Meanwhile, due to the multi-level training mode, when courses related to the multidisciplinary field are processed, a high-quality extension result can be kept.

The invention provides an entity linking method based on context semantic relation and document consistency constraint in order to solve the problems of cost and the like caused by the fact that a large amount of manual annotation is needed in a traditional entity linking method. The method comprises three steps of data preprocessing, candidate entity generation and candidate entity disambiguation. In the data preprocessing stage, the noise problem existing in the data is solved; in the candidate entity generation stage, a candidate entity set with a high recall rate is obtained by using a filtering technology of a Wikipedia connection graph; in the candidate entity disambiguation stage, a candidate entity set is used as a weak supervision constraint, the relation between entities and local contexts and coherence information between entities in a document are considered, candidate entity disambiguation is carried out through a neural network, a final entity link result is obtained, and candidate entities correspond to a knowledge graph.

The invention provides a road structure prediction and target detection method based on a multi-task neural network, and relates to the fields of automatic driving, deep learning, computer vision and the like. The method comprises the steps: constructing a multi-task neural network of a context parameter sharing mechanism, wherein the multi-task neural network has the functions of predicting a road structure and detecting a target at the same time; building a loss function mathematical model through the loss between the road structure predicted value and the vehicle layout predicted value and the real value; making a data set through an image and a map, and carrying out closed-loop training on a prediction part of the network; and finally, deploying the method on an automobile, and applying the method to road structure prediction and target detection. According to the method, the multi-task neural network can complete road structure prediction and target detection functions only through image information, and road structure and target prediction can be carried out on invisible and shielded areas in the image.

The invention discloses a work order data classification method. The method comprises the steps of obtaining work order data; analyzing a logic relationship between at least two keywords in the work order data, and determining a specific rule matched with the logic relationship from a plurality of preset rules to obtain a matching result; determining a first target label of the work order data based on the matching result; wherein the first target label is used for representing the type of the work order data. The embodiment of the invention further discloses a terminal and a storage medium.

The invention relates to the technical field of deep learning, and provides a human body skeleton labeling method and device and electronic equipment, and the method comprises the steps: obtaining a to-be-labeled current frame of human body image and corresponding current frame skeleton key point information, image information of the current frame of human body image and skeleton key point information of the current frame are stored as a current frame annotation file, and the skeleton key point information comprises type information and position information of skeleton key points; labeling the current frame of human body image by using the current frame of labeling file to obtain an initial labeling image, and judging whether missing detection or false detection exists in the initial labeling image to obtain a judgment result; and correcting the skeleton key points with leak detection or false detection in the initial annotation image corresponding to the current frame of human body image according to the judgment result to obtain the current frame of annotation image. According to the embodiment of the invention, the human skeleton annotation image can be quickly and accurately obtained.

A method and apparatus for training a supervised machine learning model is disclosed. The method includes: generating a plurality of artificial images, each artificial image containing the motion state of the same target object at different time points in one or more time periods; recording the movement state of the same target object at one or more time points in the process of generating the plurality of artificial images; or motion-related annotation data over multiple time periods; generate a multimedia stream including motion based on multiple artificial images; use data from multiple frames of the multimedia stream as multiple input data to the model to perform operations in the model to Obtaining derived data related to the motion; and comparing the derived data with the labeled data to determine whether to adjust parameters of the model. Through this method, a large number of manual annotations required in the training process of the model can be omitted.

The invention discloses a power grid regulation and control unstructured table data extraction processing method and device based on semantic analysis and a storage medium. The method comprises the steps that a power grid unstructured original table text is acquired and preprocessed; automatically labeling the preprocessed original table text through a pre-constructed semantic recognition model; analyzing data in the preprocessed original table text based on a labeling result; sorting the analysis result into a preset structured table to generate a structured table text; according to the method, the problem of power grid regulation and control unstructured table data extraction processing can be solved, so that knowledge support is provided for intelligent control.

The invention discloses a junk corpus screening method, system and device based on an LGBM model and a BTM model, and the method comprises the steps of carrying out the comment extraction of differenttypes of commodities, and obtaining comment data; performing topic mining on the comment data by using a BTM model, and summarizing spam comment high-frequency words according to a mining result; training an LGBM model based on the comment data and the spam comment high-frequency words; and screening out spam comment corpora by using the trained LGBM model. According to the invention, junk comments irrelevant to commented commodities can be screened out under the conditions of ensuring the inference speed and reducing manual annotation.

The invention discloses a menu field seed word automatic extraction realization method and a storage medium. The method comprises the following steps: obtaining menu data; establishing a semantic model based on the word vectors and the document vectors; pre-defining menu field seed words, and inputting the field seed words into the semantic model to obtain the same group of synonyms of the field seed words. According to the invention, the semantic model is established; The method comprises the steps of defining menu field seed words in advance, inputting the field seed words into a semantic model to obtain the same group of synonyms of the field seed words, accelerating rapid extraction of the field seed words through automatic processing of a program algorithm, saving a large amount of manual annotations, and successfully applying the field seed words to the menu field; While a large amount of labor cost and time are saved, faster and more data sources are provided for downstream semantic understanding services.

The invention provides an article reading comprehension answer retrieval system and device based on machine learning. The method comprises the steps: extracting keywords of different sentences and question sentences in an article according to a semantic rule, and obtaining core words corresponding to the different sentences and question core words; vectorizing the core words of the sentences and the question core words according to a pre-trained statement model and obtaining core word vectors of the sentences and a question core word vector; calculating the similarity between the question coreword vector and the core word vectors of different sentences according to the cosine distance, and obtaining the similarities of different sentences; judging the similarities of different sentences;and taking sentences with high similarities as training corpora and inputting the training corpora into a neural network combined by a recurrent neural network and a multi-layer perceptron for training to obtain an answer retrieval neural network model. The technical problem of manually annotating corpora in the prior art is solved, machine annotation is generated by adopting a fixed rule, and thetechnical effects of moderate accuracy, no need of manual annotation and cost saving are achieved.