Abstract: A method for detecting small obstacles includes: acquiring a first 3D point cloud corresponding to image data acquired by a cleaning robot; extracting, from the first 3D point cloud, a second 3D point cloud of a ground region; extracting, from the second 3D point cloud, a third 3D point cloud having a height value in a set height range; calculating a ground projection point cloud of the third 3D point cloud; and, determining morphologically-connected regions in the ground projection point cloud, and using a morphologically-connected region having an area less than a preset value as a region where a small obstacle is located. By effectively recognizing the ground, acquiring the ground projection point cloud and processing the point cloud by image processing, accuracy and timeliness of the algorithm can be improved compared with directly processing discrete 3D point clouds.

Abstract: An annotation method and device and a storage medium are provided. The annotation method includes operations as follows. A first probability value that a first sample image is annotated with an Nth tag when the first sample image is annotated with an Mth tag is determined based on first tag information of a first image set. M and N are unequal and are positive integers. The first probability value is added to second tag information of a second sample image annotated with the Mth tag in a second image set.

Abstract: A method for acquiring push information can be applied to a terminal device and include: transmitting a device model of the terminal device to a server; receiving a push software development kit corresponding to the device model, and a configuration file corresponding to the push software development kit transmitted by the server; initializing the push software development kit based on the configuration file, and completing a registration with the server; and receiving, through the push software development kit, push information transmitted by the server. The terminal device can therefore acquire its corresponding push software development kit according to the device model of the terminal device through a plug-in method, thereby increase push arrival rate and improve quality of operational data.

Abstract: A method for training a neural network, includes: training a super network to obtain a network parameter of the super network, wherein each network layer of the super network includes multiple candidate network sub-structures in parallel; for each network layer of the super network, selecting, from the multiple candidate network sub-structures, a candidate network sub-structure to be a target network sub-structure; constructing a sub-network based on target network sub-structures each selected in a respective network layer of the super network; and training the sub-network, by taking the network parameter inherited from the super network as an initial parameter of the sub-network, to obtain a network parameter of the sub-network.

Abstract: A keyword extraction method includes: extracting candidate words from an original document to form a first word set; acquiring a first association degree between each first word thereof and the original document, and determining a second word set according to the first association degree; for each second word in the second word set, inquiring, in a word association topology, at least one node word satisfying a condition of association with the second word and forming a third word set, the word association topology indicating an association relation among multiple node words in a predetermined field; and determining a union set of the second and third word sets, acquiring a second association degree between each candidate keyword in the union set and the original document, and selecting, according to the second association degree, at least one candidate keyword from the union set, to form a keyword set of the original document.

Abstract: The present disclosure relates to a method and an apparatus based on a neural network model. The method includes acquiring a plurality of training samples, each of the training samples comprising an identifier of an input word, an identifier of an output word, and position information, wherein in one of the training samples, the output word is a context of the input word, and the position information indicates a relative position between the output word and the input word; calling a position relation-based Continuous Skip-gram Model (Skip-Gram), with each of the training samples as an input, to obtain an output result, the output result comprising a word matrix embedding of the input word, a word matrix embedding of the output word and a position matrix embedding of the position information; and updating the position relation-based Skip-Gram based on the output result to train the position relation-based Skip-Gram.

Abstract: A search method for a neural network model structure, includes: generating an initial generation population of network model structure based on multi-objective optimization hyper parameters, as a current generation population of network model structure; performing selection and crossover on the current generation population of network model structure; generating a part of network model structure based on reinforcement learning mutation, and generating a remaining part of network model structure based on random mutation on the selected and crossed network model structure; generating a new population of network model structure based on the part of network model structure generated by reinforcement learning mutation and the remaining part of network model structure generated by random mutation; and searching a next generation population of network model structure based on the current generation population of network model structure and the new population of network model structure.

Abstract: A method for compressing a neural network model, includes: obtaining a set of training samples including a plurality of pairs of training samples, each pair of the training samples including source data and target data corresponding to the source data; training an original teacher model by using the source data as an input and using the target data as verification data; training intermediate teacher models based on the set of training samples and the original teacher model, one or more intermediate teacher models forming a set of teacher models; training multiple candidate student models based on the set of training samples, the original teacher model, and the set of teacher models, the multiple candidate student models forming a set of student models; and selecting a candidate student model of the multiple candidate student models as a target student model according to training results of the multiple candidate student models.

Abstract: A method for compressing a neural network model includes: obtaining a first trained teacher model and a second trained teacher model based on N training samples, N being a positive integer greater than 1; for each of the N training samples, determining a first guide component of the first teacher model and a second guide component of the second teacher model respectively, determining a sub optimization target corresponding to the training sample and configured to optimize a student model according to the first guide component and the second guide component, and determining a joint optimization target based on each of the N training samples and a sub optimization target corresponding to the training sample; and training the student model based on the joint optimization target.

Abstract: A text processing method includes: word segmentation is performed on a first text to be processed to determine multiple words of the first text; one word is removed from the first text to obtain one second text; each word is removed from the first text one by one to obtain a second text set; the first text and each second text in the second text set are analyzed respectively to obtain a first tree structure of the first text and a second tree structure of each second text; and a summary text of the first text is determined according to the first tree structure and each second tree structure.

Abstract: A training method for a machine translation model, includes: obtaining a multi-domain mixed training data set; performing data domain classification on a plurality of training data pairs in the training data set to obtain at least two domain data subsets; based on each domain data subset, determining at least two candidate optimization targets for the domain data subset, and training at least two candidate single domain models corresponding to each domain data subset based on the at least two candidate optimization targets, respectively; testing the at least two candidate single domain models corresponding to each domain data subset separately, and selecting a candidate optimization target with a highest test accuracy as a designated optimization target for the domain data subset; and training a hybrid domain model based on each domain data subset in the training data set and the designated optimization target corresponding to each domain data subset.

Abstract: A method for optimizing a training set for text classification includes: the training set for text classification is acquired; part of samples are selected from the training set as a first initial training subset, and an incorrectly tagged sample in the first initial training subset is corrected to obtain a second initial training subset; a text classification model is trained according to the second initial training subset; the samples in the training set are predicted by the trained text classification model to obtain a prediction result; an incorrectly tagged sample set is generated according to the prediction result; a key incorrectly tagged sample is selected from the incorrectly tagged sample set, and a tag of the key incorrectly tagged sample is corrected to generate a correctly tagged sample corresponding to the key incorrectly tagged sample; and the training set is updated by using the correctly tagged sample.

Abstract: A clustering method includes: multiple data to be clustered are acquired, the data to be clustered including exact matching information; multiple data sets consisting of data to be clustered with same exact matching information are determined; a clustering center is determined according to the amount of data to be clustered in each of the multiple data sets; and the multiple data to be clustered are clustered according to the clustering center to obtain target clusters. Accordingly, multiple data sets each consisting of data to be clustered with same exact matching information are determined, and a clustering center is determined according to the amount of the data to be clustered in each of the multiple data sets, so that unreasonable determination of the clustering centers can be improved, and accuracy of a clustering algorithm can further be improved.

Abstract: A method for natural language, includes: determining a slot tagging result output by a Bi-directional Long Short-Term Memory-Conditional Random Field algorithm (BiLSTM-CRF) model after slot tagging on conversation data input by a user; determining reward information based on the slot tagging result and a reward of the user for the slot tagging result; and performing reinforcement learning on the BiLSTM-CRF model according to the reward information.

Abstract: A method for information processing, includes: obtaining a bilingual vocabulary containing N original bilingual word pairs, N being a positive integer; obtaining an original bilingual training set containing multiple original bilingual training sentence pairs; selecting at least one original bilingual training sentence pair matching any original bilingual word from the original bilingual training set as a bilingual sentence pair candidate; constructing a generalized bilingual sentence pattern based on at least one bilingual sentence pair candidate; and obtaining an augmented bilingual training set containing multiple augmented bilingual training sentence pairs, based on the bilingual vocabulary and the generalized bilingual sentence pattern.

Abstract: A hypernetwork training method includes: acquiring a multipath neural subnetwork based on a preconstructed initial hypernetwork; training the multipath neural subnetwork to update a weight parameter of each substructure in the multipath neural subnetwork; synchronizing the weight parameter of each substructure in the multipath neural subnetwork to the preconstructed initial hypernetwork; and determining whether the preconstructed initial hypernetwork converges, and if it is determined that the preconstructed initial hypernetwork does not converge, re-executing the acquiring, the training, the synchronizing, and the determining, to obtain a target hypernetwork.

Abstract: A message processing method includes: acquiring an unread message in the communication software; identifying contents of the unread message to obtain summary information of the unread message; and outputting the summary information.

Abstract: A method and apparatus for controlling a device are disclosed. The method includes: performing voice recognition on a received sound signal to obtain a voice recognition result; determining keywords using the voice recognition result; determining a target intelligent device having attribute information matched with the keywords from intelligent devices, where relationships between the intelligent devices and attribute information of the intelligent devices are constructed in advance, and the attribute information characterizes a device operation provided by the intelligent device corresponding to the attribute information; and controlling the target intelligent device to perform an operation indicated by the voice recognition result.

Abstract: The present disclosure relates to a method and a terminal for performing word segmentation on text information, and a storage medium. The method includes: acquiring the text information and configuration information, in which the configuration information includes at least two first word segmentation rules; converting the first word segmentation rules into second word segmentation rules according to a predetermined rule; in response to determining that an intersection exists between character strings of the text information matched by two of the second word segmentation rules, determining that two first word segmentation rules corresponding to the two of the second word segmentation rules associated with the intersection conflict; and processing the text information according to the configuration information, and outputting a result of the word segmentation on the text information.

Abstract: A method for processing untagged data includes: similarity comparison is performed on a semantic vector of untagged data and a semantic vector of each piece of tagged data to obtain similarities corresponding to respective pieces of tagged data; a preset number of similarities are selected according to a preset selection rule; the untagged data is predicted with a tagging model obtained by training through the tagged data, to obtain a prediction result of the untagged data; and the untagged data is divided into untagged data that can be tagged by a device or untagged data that cannot be tagged by the device according to the preset number of similarities and the prediction result.