82results about How to "Alleviate the sparsity problem" patented technology

The invention discloses a text classification method based on a cyclic convolution network. The text classification method comprises the following steps: step 1, representing context vectors of all words by using a bidirectional cyclic network; step 2, combining the context vectors and a word vector of a current word into the representation of the current word; step 3, extracting the most important context information by using a maximum pool technology to obtain text representation; step 4, carrying out text classification by using the text representation. According to the method disclosed by the invention, more word order information in a text can be kept and a long-distance text dependence relation is captured; semantics of the words can be accurately described and the words and phrases, which have the greatest influences on the text classification, are found by the maximum pool technology, so that the text classification accuracy rate is effectively improved. A test shows that the efficiency of the method is averagely improved by 1% on the aspect of a plurality of text classification data sets.

The invention is proposed for solving the problem of low quality of translated texts and poor translation effect of a current Mongolian and Chinese machine. Because the Mongolian language belongs to low-resource language, a large quantity of Mongolian and Chinese parallel bilingual corpus can be collected difficultly. A transfer learning strategy is a method for solving the problems in the different but relevant fields by using existing knowledge. The method comprises the following steps: firstly, performing training by utilizing the large-scale English-Chinese parallel corpus based on a neural machine translation frame; secondly, transferring translation model parameters trained by the large-scale English-Chinese parallel corpus into the Mongolian and Chinese neural machine translation frame, and training a neural machine translation model by utilizing the existing English-Chinese parallel corpus; and finally, performing contrast and evaluation on texts translated by a neural machinebased on the transfer learning strategy and texts translated by a statistical machine in accordance with BLEU values and the fluency of the translated texts. Through operation of a control variable method, the transfer learning strategy is obtained, so that the translation performance of a Mongolian and Chinese machine is effectively improved.

The invention discloses an entity relationship prediction method and a prediction system based on knowledge representation learning. The method comprises four modules, namely, knowledge preparation, knowledge representation model construction, knowledge representation model training and entity relationship prediction. The knowledge preparation module completes the data preparation and builds the knowledge map. The knowledge representation model construction module completes the construction of the model, which eliminates the semantic differences among different types of entities through projection operation; the training module of knowledge representation model forms the final knowledge representation model based on the parameters of the iterative training knowledge representation model ofknowledge map. Entity relationship prediction module can predict the possible relationship between any given two entities. The method of the invention predicts entity relationship based on knowledgemap, projects different types of entities to the same semantic space through a spatial projection algorithm, and performs calculation operation, thereby achieving high reliability of prediction results.

The invention discloses a statistic machine translation method and a statistic machine translation system. The method comprises the following steps that: firstly, phrase division of a source language statement is performed and bilingual phrases are searched from a bilingual phrase table according to divided phrases; secondly, the matching degree of the divided phrases and the bilingual phrases is inspected; under the condition of complete matching, the bilingual phrases are added into a candidate phrase table and the fourth step is executed; under the condition of partial matching, the third step is executed; thirdly, a translation template is constructed according to the divided phrases and the bilingual phrases, and word translations of the divided phrases which are different from the bilingual phrases are filled into the translation template to generate novel bilingual phrases which are then added into the candidate phrase table; fourthly, the source language statement to be translated is translated according to the candidate phrase table. The statistic machine translation method and the statistic machine translation system can effectively improve the translation quality under the condition that bilingual corpuses are limited and solve the problem of data sparseness of the statistic machine translation system.

The invention relates to a service recommendation method based on trust extension and listwise rank learning, which comprises the steps of firstly representing each user to be the probability distribution of invoked service set arrangement by utilizing service ranking position information and referring to a Plackett-Luce model, and calculating the probabilistic user similarity based on the KL distance; taking a direct trust relationship and an indirect trust relationship between the users into consideration at the same time, calculating the direct trust degree by using a Beta trust model, calculating the indirect trust degree by using transfer characteristics of the trust relationships, obtaining the comprehensive trust degree, and constructing a trusted neighbor set of target users; integrating the trusted neighbor set into a matrix decomposition model, taking a cross entropy between a predicted ranking list and a correct ranking list as a loss function, designing a listwise rank learning algorithm to obtain an optimal ranking model, and outputting a recommendation list which best conform to user interests. The service recommendation method has high recommendation accuracy, and can satisfy potential functional requirements of the users to be greatest extent while ensuring the service recommendation quality.

The invention designs a sequence recommendation method fusing a dynamic knowledge graph. The item sequence of the user in the latest session is utilized to predict the next item which is possibly interacted by the user, so that the short-term preference of the user is better captured according to the item sequence in the session. The feature representation of the item sequence and the operation sequence is learned by adopting a gating graph neural network and a recurrent neural network respectively; and meanwhile, related project knowledge (such as commodity brands, rating, categories and the like) is utilized to construct a dynamic knowledge graph, a dynamic knowledge graph method is fused, a graph convolutional network (GCN) is adopted for modeling in the time dimension to learn embedding of nodes, and parameters (weight matrixes) in the graph convolutional network are evolved by using long short-term memory (LSTM) so as to capture the dynamic state of a graph sequence, therefore, the next interaction item of the user can be predicted more effectively.

The invention discloses a social recommending method based on classification. The social recommending method comprises the steps that an evaluation matrix of a user on classification is constructed; an initial user classification matrix is constructed according to evaluation data of the user on a project and classification information of the project; normalization processing is carried out on the initial user classification matrix, and the user classification matrix is reconstructed through a matrix decomposition method; the credibility between the friends of the user is constructed through the friend information of the user; the parameters of an obtained model are learnt and obtained through a stochastic gradient descent method according to the evaluation of the user on the classification of the project in the user classification matrix, the evaluation of the user on the project predicted and obtained through a socialization model and the credibility between the friends of the user, and therefore the final evaluation of the user on the project is predicted. According to the method, project classification information is guided into the social recommending method for the first time, the socialization information of the user and the classification information of the project are integrated on the basis of original collaborative filtering recommendation, recommending precision is improved, and the problems of data sparseness and cold starting in a recommending system are solved.

The invention discloses a word frequency based skip language model training method, relates to the technical field of machine translation and aims at solving the OOV problem of a statistical language model caused by linguistic data shortage in the prior art. The word frequency based skip language model training method comprises the steps that Chinese sentences are collected; the Chinese sentences are segmented; a learning set corpus is generated; statistics is conducted on vocabulary and word frequency in the learning set corpus to generate a Chinese vocabulary wt; statistics is conducted on phrases and the emerging times of phrases in the corpus to generate a 1-n Chinese phrase table pt0; a selective skip standard k is set, and k judgment is performed according to the statistical results of the word frequency in the Chinese vocabulary wt, and when the sum of the number of all the vocabulary with the emerging times k not greater than i accounts for above 60% of the number of all the vocabulary, k = i; linguistic model training is performed according to a Chinese sentence table pt2 to obtain a skip-ngram linguistic model. The word frequency based skip language model training method is used for obtaining a linguistic model probability table.

The invention discloses a social interest recommendation method and system based on graph convolution matrix decomposition. The method comprises the steps of obtaining a user social project scoring matrix, a user social adjacency matrix and a user comment text feature matrix of a to-be-recommended user; inputting the user social project scoring matrix, the user social adjacency matrix and the usercomment text feature matrix of the to-be-recommended user into a trained graph convolution matrix decomposition model; outputting a user potential feature matrix and a project potential feature matrix of the to-be-recommended user by the trained graph convolution matrix decomposition model ; and recommending potential projects to the to-be-recommended user according to the user potential featurematrix and the project potential feature matrix.

The invention discloses a user movement behavior prediction method, device and equipment based on a multi-granularity neural network and a computer readable storage medium, and the method comprises the steps: mapping a prediction space of a user movement position to a map, and obtaining a map region corresponding to the prediction space; respectively dividing the map area according to different granularities, and respectively converting the current movement track of the user into a plurality of grid ID sequences corresponding to the different granularities under the different granularities; using a multi-granularity neural network model; and respectively predicting the plurality of grid ID sequences by using different prediction models under the different granularities to obtain a plurality of prediction results of the user moving position under the different granularities, and fusing the plurality of prediction results to obtain a target prediction result of the user moving position.According to the method, the device, the equipment and the computer readable storage medium provided by the invention, the accuracy of predicting the movement behavior of the user is improved.

The invention discloses an interest point recommendation method and system based on a co-occurrence graph. The recommendation method includes the five processes of data collection and processing, co-occurrence graph establishment, similarity calculation, position factor modeling and user recommendation. The interest point recommendation method based on the co-occurrence graph is characterized in that the co-occurrence graph between interest points is established through attendance data of users, two types of similarity between nodes is calculated based on the co-occurrence graph, and then the users are subjected to interest point recommendation according to the similarity between the nodes in the graph. Recommendation is carried out based on the novel perspective, namely the similarity between the interest points, and by means of the recommendation method, information in existing data is fully mined, and the problems of sparsity in the recommendation process of the interest points and scalability of recommendation methods in a large-scale social network are solved.

The invention discloses an interpretable recommendation method integrating user implicit article preference and implicit trust, and aims to solve the problems of sparse scoring data and cold start ofa recommendation system, and by mining the implicit article preference and implicit trust of a user by utilizing user comments, the accuracy and interpretability of scoring prediction can be greatly improved. According to the method, user aspect preferences and article aspect quality are mined in comments by utilizing a topic model LDA and an aspect matrix decomposition model AMF, implicit articlepreferences of a user are captured by utilizing the user aspect preferences and the article aspect quality, and implicit trust of the user is captured through user scoring and comments. And finally,the predicted score is corrected according to the implicit article preference of the user and the implicit trust of the user, so that the recommendation accuracy and interpretability can be improved.

The invention provides a collaborative filtering recommendation algorithm fusing social trust influence, which comprises the following steps: step 1, preprocessing score matrix data of a project by users, filling and deleting unnecessary data, and calculating similarity among the users according to scores of the users; step 2, calculating a neighbor user set of the users; and 3, calculating the correct recommendation number of the user j to the user i, the local trust degree between the users and the global trust degree of each user according to the scoring matrix data of the user to the project. According to the invention, sparse user scoring information is fully mined; an implicit trust network is constructed, a local trust degree and a global trust degree are calculated through the scoring data; according to the method, the local trust degree and the global trust degree are combined to obtain the implicit trust degree, trust propagation of the users is considered for trust data between the users, the display trust relationship of the users is effectively expanded, and the problems of score matrix data sparseness and user cold start are relieved.

The invention belongs to the technical field of natural language processing methods, and in particular relates to a Bayesian word sense disambiguation method based on synonym expansion. The Bayesian word sense disambiguation method disclosed by the invention is used for mainly solving the problem that the current word sense disambiguation method has the problems of poor disambiguation effect, wasting time and energy to obtain disambiguation knowledge and the like. The Bayesian word sense disambiguation method based on synonym expansion disclosed by the invention comprises the following steps of: (1), expanding the context of a training corpus by adopting the Chinese thesaurus, and generating a lot of pseudo training corpuses; (2), removing noise in the pseudo training corpuses by utilizing a word collocation corpus, and generating a pseudo training corpus; (3), training a Bayesian disambiguation model by adopting the training corpus and the pseudo training corpus simultaneously; and (4), inputting a test corpus into the Bayesian disambiguation model, and co-determining word senses of ambiguous words by comprehensively utilizing the disambiguation knowledge in the two corpuses.

The invention provides a new method for representing crowd movement through a context-related graph embedding model. The method comprises the following steps: firstly, generating a fully-connected context correlation graph according to a historical track of a user; applying a word embedding technology to a graph model for the first time, learning sign-in point vector representation fused with context semantic information, then obtaining vector representation of each track through a recurrent neural network, and then introducing a reinforcement learning method to find out a current track generator. In the training process of the whole model, the track with the label and the track without the label are used for training synchronously, meanwhile, a strategy gradient method in adversarial learning is used for updating model parameters, and the model performance is further enhanced.

The invention discloses an information recommendation method and system based on a convolutional neural network and a noise reduction auto-encoder. Two deep learning models of a convolutional neural network and a noise reduction auto-encoder are used to learn user preferences from scores, trust, comments and other data. Meanwhile, a new correlation regularization method is provided to establish relationships of user preferences in different aspects, so that the performance is improved. Firstly, compared with a previous model, rich comment information is fused; then, preliminary processing is conducted on the comment text through a convolutional neural network model, the extracted effective features are put into a noise reduction auto-encoder model to extract hidden features of the commenttext, and more effective and compact representation of the comment text is obtained; finally, two noise reduction auto-encoders are added and used for processing scores and trust information respectively, corresponding prediction vectors are obtained through the three noise reduction auto-encoders respectively, weighted fusion is carried out, and therefore user preferences are modeled more accurately.

The invention relates to the field of position prediction and recommendation, in particular to a user space-time similarity measuring method in interest point prediction and recommendation. The user space-time similarity measuring method in interest point prediction and recommendation includes the steps of preprocessing data; measuring space-time cooccurrence mode similarity; measuring user space-time similarity. The invention provides the user space-time similarity measuring method, the space-time cooccurrence mode and similarity thereof are defined through check-in time close degree and interest point position close degree, then the user space-time cooccurrence mode and similarity thereof are defined, user space-time similarity among users is defined based on the weight of interest point check-in frequency through the most similar space-time cooccurrence mode and similarity thereof, and the user space-time similarity measuring method is provided for a user-based collaborative filtering method in interest point prediction and recommendation.

The invention discloses a scoring prediction method and device thereof. The method comprises the following steps of: a generating step: generating a representation of both a user and an item by constructing a neural network model for a user review of an item; a prediction step: optimizing a matrix decomposition model by utilizing the representation of a user and an item, and training the matrix decomposition model to predict target score, wherein the neural network model is a long and short time memory network model; extracting features of the user and the item by using an attention mechanismin the long and short time memory network model, and generating representation according to features; and adding the representation of the user and the item to the matrix decomposition model by nonlinear transformation. The scoring prediction method and device thereof provided by the invention is low in cost, and migration can be well performed among different data fields; the unknown score can bepredicted more comprehensively and correctly by combining the user scoring numerical information and the user comment text information; the prediction results can be obtained in a faster time with anoptimized matrix decomposition algorithm.