58results about How to "Avoid manual labeling" patented technology

The invention discloses a semi-supervised automatic aspect extraction method based on domain information. The semi-supervised automatic aspect extraction method comprises the steps of network information crawling, information pre-processing, keyword extraction, comment document recombination and fine-grit mark LDA learning. The invention further discloses a semi-supervised automatic aspect extraction system based on the domain information. The semi-supervised automatic aspect extraction system based on the domain information comprises a network information crawling module, an information pre-processing module, a keyword extraction module, a comment document recombination module and a fine-grit mark LDA learning module. By the adoption of the semi-supervised automatic aspect extraction method and system based on the domain information, all extracted aspects of a commodity are more clear and more definite, and the differences between the aspects are more clear; a generated aspect structure (order and content) generated through the semi-supervised automatic aspect extraction method and system can be kept consistent with a commodity aspect structure which is predefined in a seed word set, so that the semi-supervised automatic aspect extraction method and system have the advantages that semantic clustering can be conducted on different expressions used by a consumer for description of the same commodity aspect, and human interference can be reduced in the process of opinion mining of the commodity.

The invention discloses a method for identifying a text font in a natural scene picture. With a picture synthesis method, lots of natural scene pictures that are integrated with different font texts and are similar to real effects specifically are obtained; picture training is carried out to obtain a font identifying device and a text location device; more text pictures are located in an internetpicture by using the text location device; and on the basis of a migration learning method, learning is carried out in the pictures by using the font identifying device, so that the identification precision is improved. According to the invention, the high-quality training picture is synthesized automatically, so that manual marking needing lots of time and efforts is avoided and the cost is lowered substantially; with the font classification device, the high identification precision is realized. Besides, because of introduction of the migration learning method, the accuracy of the font classification device is enhanced by using lots of font-tag-free pictures.

The invention provides a generation method and device of a word segmentation training set. The generation method of the word segmentation training set comprises the following steps: obtaining training corpus, adopting different word segmentation devices to independently carry out word segmentation on different training corpuses to obtain word segmentation results corresponding to different word segmentation devices; dividing the word segmentation results into accurate matching word segmentation results and non accurate matching word segmentation results; and according to the word segmentation results, carrying out noise reduction processing on the non accurate matching word segmentation results to obtain the word segmentation training set. The method can lower time and cost for the generation of the word segmentation training set, is low in implementation cost and improves effect.

The invention provides a text processing method and device, a computer storage medium and electronic equipment, and relates to the field of artificial intelligence. The method comprises the followingsteps: obtaining text to be processed, inputting the text to be processed into a multi-label classification model, wherein the multi-label classification model is obtained by training based on an unbalanced text sample set and an unbalanced attenuation loss function, the unbalanced text sample set is a text sample set in which the number of label positive samples and the number of label negative samples are unbalanced, and the unbalanced attenuation loss function comprises a first loss part, a second loss part and recall loss; performing attribute extraction on the to-be-processed text throughthe multi-label classification model to obtain a label corresponding to the to-be-processed text; and obtaining a corresponding entity from the to-be-processed text according to the label, and constructing a triad according to the label and the entity so as to update the knowledge graph according to the triad. According to the invention, the on-call rate of the text label can be improved and thecost is reduced.

The invention discloses a user query intention understanding method, system and computer terminal. The method comprises the steps that a seed concept of a predetermined intention domain is taken as astarting point, concepts and classification labels related to the domain in a knowledge base are crawled; according to the crawled concepts and classification labels, a concept link graph of the predetermined intention domain is established; the probability of nodes in the concept link graph belonging to the predetermined intention domain is calculated; a user query statement is acquired, the concept link graph is retrieved, if the user query statement covers concept nodes, the concepts are returned; if no, a deep learning model is used to calculate the similarity between the user query statement and each concept node, and the first K concepts most similar to the user query statement are returned; according to a returned result and the probability of the concept nodes belonging to the predetermined intention domain, the probability of the user query statement belonging to the predetermined intention domain is calculated, and compared with a predetermined threshold, it is judged whetheror not the user query statement belongs to the predetermined intention domain. The user query intention understanding method, system and computer terminal solve the problem of the concept feature coverage and semantic expression of the query statement.

The invention discloses a short text opinion excavation method based on complementation corpus, and belongs to attribute opinion excavation; the method comprises the following steps: firstly selecting training corpus from certain segment of weibo corpus, segmenting words, tagging part-of-speech, and selecting; tagging attribute words for training corpus according to opinion words; using part of speech tags as a characteristic training maximum entropy model; then, building a cross-corpus topic model according to weibo corpus and news corpus of certain event, combining the maximum entropy model to parse the topic to which the event belongs, and extracting corresponding attribute word distribution and opinion word distribution; finally, using an emotion classifier to carry out polarity analysis according to all opinion words of certain specific sharing topic or all opinion words of certain specific exclusive topic. The short text opinion excavation method can carry out attribute analysis and opinion excavation on certain public opinion event, has high effectiveness, robustness and usability, and can provide important application values on opinion excavation and public opinion monitoring fields.

The invention relates to an unsupervised medical image segmentation method based on an adversarial network, belongs to medical care informatics, and particularly relates to the technical field of medical image segmentation. According to the technical scheme, the method comprises the following steps: firstly, randomly generating or utilizing a third-party data set to obtain a group of auxiliary masks according to shape prior information, and sending the auxiliary masks and an unlabeled training image into a cyclic consistency adversarial network to generate binary masks; and a discriminator based on variational self-encoding and a generator correction module based on discriminator feedback are utilized to improve the quality of the binary mask. And after the binary mask of the training image is obtained, iterative training is performed by using a noise weighted Dice loss function, so that a final high-precision segmentation model can be obtained. According to the method, the problem that the convolutional neural network needs a large number of manual annotations in the training process of medical image segmentation can be solved, the problems of low performance, poor robustness and the like of an unsupervised segmentation method are solved, and the performance of an unsupervised medical image segmentation algorithm is effectively improved.

The invention discloses a color image searching method based on a quaternion exponential moment. The method comprises the following steps: using the quaternion exponential moment for decomposing a color image during a color image searching process; taking a moment value acquired by utilizing the quaternion exponential moment to decompose the color image as an image feature; utilizing a calculation method taking Euclidean distance as the similarity between images to calculate the similarity between different images. The quaternion exponential moment is simultaneously combined with a color feature and a texture feature, so that the color and shape features of the image can be effectively represented. The method has the characteristics of simple design, easiness in implementation and no need of manual marking during an image feature extracting process, so that the complex and easily mistaken manual marking process is avoided and the searching precision and efficiency are greatly increased.

The invention provides an ophthalmic surgery navigation system. The ophthalmic surgery navigation system provides eye position and rotation information for an intraoperative doctor through integrationwith a surgical microscope, and assists a doctor in performing surgery. According to the system, surgery video is obtained in real time based on an artificial intelligent image processing algorithm,the current frame of image is tracked based on extraction of a plurality of target areas or features, accumulated errors are corrected according to reference images, and finally, information such as eye scaling, displacement, rotation and non-rigid deformation is determined. Eye position information is projected into a surgical microscope, real-time, continuous, automatic and accurate navigation of surgery is achieved, manual marking is avoided, personalized ophthalmic surgery is achieved, surgical quality is improved, and postoperative visual recovery of patients is improved.

The invention discloses a method and device for generating dress collocation data, and relates to the technical field of computers. One specific embodiment of the method comprises the following steps:carrying out dress segmentation on each input image to obtain one or more single dress images of each dress collocation image, the input images comprising dress matching images; evaluating the similarity between each single dress image and a dress commodity image in a commodity library so as to select a similar dress commodity image of each single dress image; and matching the similar dress commodity images of the single dress image in each dress collocation image according to a dress collocation principle of the dress collocation image to obtain one or more groups of dress collocation data corresponding to each dress collocation image. The implementation mode can automatically, quickly and accurately accumulate a large amount of dress collocation data with various styles, avoids time andlabor consumption, overcomes personal aesthetic tendency limitation, fully utilizes social media and network resources, and enables the dress collocation data to accord with fashion trends.

The invention discloses an unsupervised multi-document abstract generation method for public opinion analysis. The method comprises the steps: 1, collecting online public opinion news in real time, and automatically dividing news sets according to network hotspots; 2, extracting a single document abstract of each piece of public opinion news in the set in an unsupervised manner; and 3, analyzing all the extracted single-document abstracts in the set to obtain an unsupervised multi-document abstract. According to the method, the problems that an existing multi-document abstract method is relatively low in effect, relatively poor in generative abstract practicability and lack of Chinese public opinion abstract training corpora are solved, so that public opinion news is monitored.

The invention discloses a label gumming and pasting device for manufacturing and producing roller type sound boxes. The device comprises a base, motor seats are installed on two sides of the lower endof the base, a driving motor is connected to the middle of the lower ends of the motor seats, a driving belt pulley is installed at the middle of the front end of the driving motor, a conveying frameis installed at the middle of the upper end of the base, a driven belt pulley is installed at the top end of the conveying frame, a transmission wheel is installed on the rear end of a coaxial position of the driven belt pulley, supporting frames are installed on the two sides of the upper end of the base, a fixing frame is installed at the top end of each supporting frame, a rotating shaft is installed on the right side of the lower end of the fixing frame, a roller is installed at the middle of the rotating shaft, a rolling groove is formed in the outer side of the roller, a rolling ball isarranged in the rolling groove, a lifting rod is connected to the left end of the rolling ball, a lifting shaft is connected to the left end of the lifting rod, and a lifting plate is connected to the lower end of the lifting shaft. By adoption of the label gumming and pasting device disclosed by the invention, automatic gumming and label pasting of sound boxes are achieved, and manual label pasting is avoided, thereby not only improving the label pasting efficiency of the sound boxes, but also reducing the labor intensity of workers, and meeting the manufacture and production demands of large-scale sound box label pasting.

The invention relates to a bidding data named entity recognition method based on a pre-training model. The method specifically comprises the following steps: S1, obtaining an open source pre-training model; S2, acquiring an unlabeled corpus, and performing data preprocessing; S3, training the pre-training model in the S1; S4, performing supervised training by using the labeled data to obtain a reference model M; S5, enabling the reference model M to predict the unlabeled data to obtain pseudo-label data; S6, adding the pseudo label data into the training set, adding the real label data into the training set, and jointly training to obtain a model M'; S7, constructing a fragment decoding network; S8, encoding the text input model M'; S9, inputting the text code into a fragment decoding network; and S10, extracting entity fragments and categories thereof. According to the invention, after the model is pre-trained, model decoding is carried out in a fragment identification mode to predict the beginning and ending positions of the entity, the decoding speed can be increased, and an entity result with good precision can be obtained.

The invention belongs to the technical field of target detection, and discloses a target detection data set generation method and device, equipment and a storage medium. The method comprises the following steps: traversing a preset background image folder path to obtain a traversed current background image; generating a blank annotation file according to a file name corresponding to the current background image; and storing the current background image and the blank annotation file into the initial data set to obtain a target data set. Through the above mode, the data set is generated according to the background image and the blank annotation file, in the case of target detection and identification errors, the background image with the identification errors is stored to the preset background image folder path, and the target data set can be generated through the data set generation process, so that manual annotation is avoided, the data set generation efficiency is high, and thus quickly responding to model optimization work.

The embodiment of the invention discloses a text data relationship extraction method, device and equipment and a readable storage medium. The method comprises the steps of obtaining text data with a keyword tag; constructing sentences according to the keyword tags and a predetermined entity relationship, and taking the sentences and text data as input data of a reading understanding model; and judging whether an answer output by the reading understanding model is a correct entity in the text data or not, and if so, marking the two entities and an entity relationship in the text data. Accordingto the technical scheme, sentences formed by keyword tags and predetermined entity relations and text data serve as input data of a reading understanding model, whether answers of the reading understanding model to the sentences are correct entities in the text data or not is judged, and if yes, the entities in the text data and the entity relations are marked; the problem of lack of training corpora in the application field of relation extraction is solved, and the cost is reduced.

The invention discloses an automatic test tube label pasting device and belongs to the field of medical devices. The device comprises a bottom plate, a storage unit, an isolation unit, a clamping wheel, a label separating unit, a signal processing unit and a plurality of induction units. The isolation unit is used for separating single test tubes. The clamping wheel is used for moving the test tubes so as to perform label pasting. According to the automatic test tube label pasting device, the test tubes are separated from one another after being stored through the isolation unit, the labels are separated from label paper by means of the label separating unit, then the clamping wheel is used for driving the test tubes to run and to be pasted with the labels, the labels are pressed tightly by means of elastic assemblies, the test tube labels are further pasted automatically, in addition, the structure is compact, the device is fast and efficient to use, simplex manual label pasting is avoided, and the label pasting speed and quality are guaranteed. Furthermore, the signal processing unit and a plurality of sensors are arranged, and the applicability of the automatic test tube label pasting device is guaranteed to a greater extent.

The invention discloses a cloud platform abnormal log analysis method. The method comprises the following steps of S1, collecting historical operation logs of a cloud platform and performing preprocessing; s2, converting the obtained historical log content into digital vector representation; s3, clustering division is carried out on the logs represented by the digital vectors; s4, constructing a deep neural network to learn the encoded log; and S5, analyzing the newly generated log by using the trained deep neural network to obtain alarm information, tracing the alarm information, and judging a fault point. According to the invention, the historical logs are learned by constructing the deep neural network, and then the trained deep neural network is applied to the analysis of the existing logs, so that the possible faults in the cloud platform operation process can be pre-warned by analyzing the generated log content. The abnormity in cloud parallel operation can be positioned and solved in advance, and the operation stability of the cloud platform is improved.

The invention belongs to the technical field of image recognition in particular to a paper job page number recognition method based on a Faster-RCNN, and aims to solve the problems that in the prior art, a training set is not rich, and a page number training set of a graph and/or image combination style does not exist, so that the page number recognition accuracy is low, and some page numbers cannot be recognized. The method comprises the following steps: calculating a page number center coordinate in a page picture through a paper job page number positioning method, and obtaining the page number picture by using a rectangular frame; and obtaining a corresponding page number category through the page number identification model. The page number identification model is constructed based ona Faster-RCNN network, and the training sample set, the sample label and the to-be-identified page number picture are selected from the same book. The page numbers of the same book are used as the data source, sample expansion is carried out, the sample sets with different effects are generated for the page numbers of different styles, the labels corresponding to the samples are automatically generated, the page number recognition accuracy is high, the robustness is high, and the efficiency is high.

The invention discloses an unsupervised image classification method based on an automatic encoder. The method comprises the following steps: S1, designing an automatic encoder model based on a convolutional neural network architecture; S2, enabling an auto-encoder model to extract feature information in the image through multi-task loss; S3, preparing to-be-classified image data, and carrying out the self-encoder model training; S4, after the auto-encoder model is trained, enabling the auto-encoder model to complete the encoding of feature information in the image; S5, filtering the noise and background information in the original image through coding of the auto-encoder model, and completing the image classification. According to the method, learning can be carried out without any label or model, so that the image classification task is completed, and the labor cost is reduced.

The invention relates to a hole marking method and device, computer equipment and a readable storage medium. The method comprises the steps of obtaining a first parameter of a wall body and a second parameter of a hole in a two-dimensional wall body drawing through a Revit platform, obtaining a three-dimensional contour model of the wall body according to the first parameter, editing the three-dimensional contour model to obtain a three-dimensional wall body model, and based on the second parameter, carrying out information labeling on the hole in the three-dimensional wall body model. By adopting the method, the three-dimensional wall body model can be constructed through the first parameter of the wall body in the two-dimensional wall body drawing, the wall body model does not need to be manually drawn, the drawing efficiency of the three-dimensional wall body drawing is improved, information labeling is automatically performed on the hole in the three-dimensional wall body model through the second parameter of the hole, manual labeling can be avoided, the labor cost is saved, and the hole marking efficiency is improved.

An automatic detection method for a minimum repetition unit of a printed fabric image includes scanning printing fabric into an image; manually selecting a rectangular area containing a plurality of pixels in the image, and selecting pixels of at least two vertices as key points; detecting a rectangular sub-image similar to the rectangular area image in the image, marking key points of the rectangular sub-image, adjusting pixel distances to minimize the number of key points, namely similarity points; identifying the minimum repeat units based on the number of similar points. The method can realize the detection of the minimum repetition unit for processing the color printed fabric and the complex printed fabric with various patterns, and greatly reduces the difficulty and workload of colorseparation and plate making.

The invention relates to a visual navigation guide point marking method and device, computer equipment and a storage medium. The method comprises the following steps of: acquiring sequential image information acquired from the environment when an unmanned vehicle runs, and determining a first target point in a first image according to a preset rule, performing calculation through image feature matching and affine coordinate transformation to obtain a second target point, projected to a second image, of the target point of the first image, using the second target point as a guide point of the second image, and acquiring a control strategy when no person performs autonomous navigation in the scene according to the guide point. By adopting the method, the guide points corresponding to a series of images can be automatically generated, manual marking is avoided, and the method has the advantages of high efficiency and good consistency.