1109 results about "Cosine similarity" patented technology

An organization's data records are often noisy: because of transcription errors, incomplete information, and lack of standard formats for textual data. A fundamental task during data cleansing and integration is matching strings—perhaps across multiple relations—that refer to the same entity (e.g., organization name or address). Furthermore, it is desirable to perform this matching within an RDBMS, which is where the data is likely to reside. In this paper, We adapt the widely used and established cosine similarity metric from the information retrieval field to the relational database context in order to identify potential string matches across relations. We then use this similarity metric to characterize this key aspect of data cleansing and integration as a join between relations on textual attributes, where the similarity of matches exceeds a specified threshold. Computing an exact answer to the text join can be expensive. For query processing efficiency, we propose an approximate, sampling-based approach to the join problem that can be easily and efficiently executed in a standard, unmodified RDBMS. Therefore the present invention includes a system for string matching across multiple relations in a relational database management system comprising generating a set of strings from a set of characters, decomposing each string into a subset of tokens, establishing at least two relations within the strings, establishing a similarity threshold for the relations, sampling the at least two relations, correlating the relations for the similarity threshold and returning all of the tokens which meet the criteria of the similarity threshold.

Methods and apparatus are provided for classifying a spoken utterance into at least one of a plurality of categories. A spoken utterance is translated into text and a confidence score is provided for one or more terms in the translation. The spoken utterance is classified into at least one category, based upon (i) a closeness measure between terms in the translation of the spoken utterance and terms in the at least one category and (ii) the confidence score. The closeness measure may be, for example, a measure of a cosine similarity between a query vector representation of said spoken utterance and each of said plurality of categories. A score is optionally generated for each of the plurality of categories and the score is used to classify the spoken utterance into at least one category. The confidence score for a multi-word term can be computed, for example, as a geometric mean of the confidence score for each individual word in the multi-word term.

The invention discloses a cross-age face verify method based on characteristic learning; the method comprises the following steps: 1, obtaining to-be compared two face images; 2, using a face characteristic point positioning method to carry out align operation for the two face images; 3, respectively carrying out feature extraction for each image, wherein the extraction method includes the following steps: a, automatically extracting high-level meaning characteristics through a depth convolution nerve network; b, calculating LBP histogram characteristics of the image; c, fusing the characteristics obtained in the a and b steps, and expressing characteristic vectors; 4, using a cosine similarity method to calculate a distance between the characteristic vectors obtained by the step3, and determining whether the two images are from a same person or not. The method firstly uses the depth network to cross-age face verify, and creatively combines the handwork design LBP histogram characteristics with depth network autonomous learning characteristics, thus realizing complementation between high-rise meaning characteristic and lower characteristics, and providing better accuracy.

The invention provides a text information similarity matching method and device on the basis of TF-IDF. The method includes the steps of acquiring text information; performing word segmentation on thetext information to obtain word segments w1, w2, ..., wn-1 and wn; using a CBOW model for calculating word vectors V(w1), V(w2), ..., V(wn-1) and V(wn) of the word segments respectively; using the TF-IDF algorithm for calculating TF-IDF values k1, k2, ..., kn-1 and kn of the word segments respectively; according to products of the word vectors of the word segments and the corresponding TF-IDF values, obtaining a sentence vector V; calculating the cosine similarity between the sentence vector V and sentence vectors of pre-stored statements, and determining the pre-stored statement with the maximum cosine similarity. Through the process above, the pre-stored statement which is the most similar to the text information can be found, the accuracy of problem recognition can be improved in the aspects of robot dialogue, information classification and the like, and therefore the dialogue efficiency or the classification efficiency is improved. A computer device and a storage medium are also provided.

The invention discloses a face recognition convolutional neural network training method based on a novel loss function. Put SoftMax loss, cosine similarity loss, Center Loss are combined used as the objective function in the training of a convolutional neural network for face recognition. The loss of cosine similarity is added to overcome the difference caused by different measurement methods of face feature alignment in training and testing while the class spacing is enlarged and the intra-class distance is reduced. The method includes such steps as 1, dividing face recognition data into training sample set and a test sample set; 2, carrying out data preprocessing on the face images in the training sample set; 3, constructing a new convolutional neural network structure for face recognition based on a new loss function; 4, inputting that training sample set into a face recognition convolution neural network for training; 5, saving that parameters of the face recognition model; 6, testing the face recognition model by using the test sample set after data preprocessing.

The invention relates to the field of automatic abstracting, and discloses a graph model-based automatic abstracting method. According to the technical scheme, an LDA probability topic model is applied to measurement of semantic correlation between sentences and improvement of the measurement effect of sentence correlation; and an idea of topic correlation and position sensitivity of the sentences is provided, so that abstract generation is relatively reasonable and effective. The method comprises the following steps: firstly, obtaining topic probability distribution of a document and word probability distribution of the topic through training the LDA topic model, determining the topic probability distribution of the sentences and effectively converting a semantic similarity measurement between the sentences into a similarity measurement problem of the topic probability distribution of the sentences; with the sentences as nodes, building edges by referring tothe cosine similarity and according to the semantic similarity between the sentences and generating a text graph representing the document; calculating the topic correlation between the sentences according to the topic probability distribution of the sentences and the topic probability distribution of the document; and calculating the position sensitivity and the like of the sentences according to the positions of the sentences in the document.

A method and system for determining view invariant spatial-temporal descriptors encoding details of both motion dynamics and posture interactions that are highly representative and discriminative. The method and system describe determining posture interactions descriptors and motion dynamics descriptors by utilizing cosine similarity approach thereby rendering the descriptors to be view invariant.

The invention discloses a method for extracting novel field words by the aid of combination of word2vec and Bootstrapping iteration. The method includes preprocessing field word corpus; segmenting preprocessed field texts by the aid of n-gram; counting word frequencies, left and right adjacent word numbers, left and right word entropy and mutual information six-dimensional statistics of segmented character strings; setting a group of parameters by the aid of kmeans; carrying out preliminary evaluation; carrying out filtering to obtain first round of results; respectively computing sums of cosine similarity of each candidate word and seed sets by the aid of word vector spaces and a group of field seed data; setting sum threshold values and carrying out secondary evaluation so as to extract the novel words of fields. The word vector spaces are obtained by means of word2vec training. The method has the advantages that the method is applicable to extracting the novel words from the large-scale field corpus and excellent in portability; the problem of difficulty in filtering non-field words with verb-object constructions, reduplication and the like can be fundamentally solved by the aid of the method.

The invention provides a countercheck method for automatically identifying a speaker aiming to voice deception, which is a voice anti-spoofing technology based on a method combining various features and a plurality of sub-systems. According to the invention, the serial features of the posterior probability of a phoneme in the phonological level and the MFCC features of voice level or MFDCC features of phase level are combined, thus the performance of the system is significantly enhanced. By combining the provided i-vector sub-system and OpenSMILE (open Speech and Music Interpretation by Large Space Extraction criterion containing voice and rhythmic information, the final presentation of the system is further enhanced. To a back-end model, the development datum are used; and under the situation of knowing deceptive attacks, a two-level support vector machine has better performance compared with one-level cosine similarity or PLDA evaluations, while the one-level evaluation approach has better robustness under the situation without seeing the test datum and knowing the deceptive conditions.

The invention discloses a face image feature extraction and comparison method based on deep learning. The method comprises the steps: performing local modularization on the images in a face library; constructing a convolution neural network, and inputting image modules into the convolution neural network to train; obtaining the image modules after performing local modularization on the images in a face library, using the trained convolution neural network model to extract the features of each image module, and obtaining the feature vector corresponding to each image; after performing local modularization on the face images to be compared, utilizing the trained convolution neural network model to extract the features of each image module, and obtaining the feature vectors corresponding to the face images to be compared; and using the cosine similarity to compare the feature vectors corresponding to the face images to be compared with the feature vector corresponding to each image in order. For the face image feature extraction and comparison method based on deep learning, face comparison in the OFD eastern face library ranks top 5, and the accuracy is 95%.

The invention relates to a vehicle type recognition method based on a CNN and domain adaptive learning. The method comprises steps of: establishing a CNN-based initial model by adding a rotation-invariant layer in an Alexnet network, distinguishing a discriminant layer and designing a new objective function; separately extracting the feature maps of different-domain sample convolution layers by using the established initial model, calculating the cosine similarity between the sample feature maps, determining the shared convolution kernel or the non-shared convolution kernel of the CNN, retaining the weight and the offset of the shared convolution kernel, and updating the weight and the offset of the non-shared convolution kernel; based on a target-domain training sample, calculating the cosine similarity between respective feature map layers and the average similarity of the entire target domain, and clustering each type of similar feature maps according to the average similarity; expanding source-domain samples with similar distribution characteristics in the target domain to new samples in the target domain, adjusting the entire CNN model by using the new samples in the target domain, and then using a softmax classifier to classify the vehicle types of the test samples in the target domain.

The invention provides a short text similarity calculation method. The method comprises the steps as follows: S1, performing word segmentation on training corpus, obtaining the word vector of each word with a word2vec algorithm, and combining the word vectors into a word vector set; S2, performing word segmentation on to-be-calculated short texts respectively, finding the word vector of each wordof the to-be-calculated short texts from the word vector set, and combining the word vectors into a short text vector set; S3, calculating the cosine similarity between each word vector in the word vector set and each word vector in the short text vector set, obtaining the maximum similarity value of each word vector, and combining the maximum similarity values into short text sentence vectors; S4, calculating the similarity between two short text sentence vectors to calculate the similarity between two short texts. The invention also provides a short text similarity calculation system. According to the similarity algorithm, short text sentences are represented by sentence vectors, the semantic similarity between short text sentences is effectively depicted, and the accuracy is high.

The invention discloses a face image comparison method, a device, a terminal device and a storage medium. The method comprises the following steps: obtaining a face image to be compared and extractinga feature vector to obtain a feature vector to be compared; the feature vector to be compared is normalized to obtain a comparison normalized vector; obtaining a base library matrix, calculating thecosine similarity between the feature vectors to be compared and the N standard feature vectors of the base library by a parallel calculation method based on the comparison normalized vector and the base library matrix, and obtaining N cosine similarity values; the maximum cosine similarity value is retrieved from the N cosine similarity values, and the comparison result is obtained according to the maximum cosine similarity value and the preset similarity threshold. The face image comparison method calculates the similarity of the face images by using a matrix operation library and a parallelcalculation mode, thereby speeding up the running speed of the computer and improving the efficiency of the face image comparison.

The invention discloses a healthy diet knowledge network construction method based on a neural network and a graph structure. The method comprises the steps that word vector modeling is performed on a text corpus, so that each non-stop word in the text corpus corresponds to one word vector with a fixed length; a cosine similarity between two word vectors is used to measure the relational degree between entities corresponding to the two word vectors; food material entity nodes and symptom entity nodes are extracted, the two types of entity nodes are regarded as entity nodes in a topological structure, edge relations between the entity nodes are constructed to form the graph structure, and all the edge relations between the entity nodes are described by one group of representative words; vector expressions corresponding to each representative word are arranged to obtain a representative matrix of the edge relations between the entity nodes; and a classification framework based on a deep neural network is designed, the representative matrix is input, and polarities of the edge relations between the entity nodes are classified. Through the method, the problems that a traditional healthy diet knowledge base is not high in automation degree and obvious in domain limitation are effectively solved.

Media content is recommended based on suitability for a designated activity. A vector engine is trained using a plurality of lists, each of the lists containing metadata associated with a plurality of media objects. The vector engine includes a neural network trained with corpus data including (i) the plurality of lists (ii) a plurality of titles, each one of the titles associated with one of the lists, and (iii) the metadata associated with the plurality of media objects. Training the vector engine involves initializing, using the vector engine, a plurality of feature vectors representing each of the lists, each of the media objects, and each of a plurality of words in the titles of the lists. The training then further involves nudging, using the vector engine, the feature vectors based on a plurality of co-occurrences of the lists, the media objects, the words in the titles of the lists, or a combination thereof. A feature vector corresponding to an activity is identified among the feature vectors. At least one of the media objects, (ii) at least one of lists or (iii) a combination thereof suitable for the activity is selected based on cosine similarities between the feature vector corresponding to the activity and others of the feature vectors.

The invention discloses an LDA (latent dirichlet allocation) and VSM (vector space model) based similar Chinese herb literature recommendation method. The method includes: adopting an IKAnalyzer to perform word segmentation on topics and summary information of literature on the basis of a terminological dictionary for Chinese herbs, constructing a vector space, performing dimensionality reduction on the vector space, constructing a semantic dictionary, numbering all lexical items in the dictionary in sequence, performing vectorization through each document on the basis of the semantic dictionary, constructing term vectors of each document, utilizing LDA and a Gibbs sampling algorithm to perform training to obtain probability distribution of each document on themes, then computing a value of similarity between every two documents by the aid of KL divergence, computing cosine similarity of the term vectors of each document on the basis of term frequency, performing joint weighting on the two kinds of similarities prior to performing similarity sorting, and then making recommendation. By the method, the literature, similar both in content and theme, in the Chinese herb literature can be recommended to users, and recommendation results are closer to user requirements.

A zero-sample classifying method based on class transfer comprises the steps of acquiring a vision characteristic of C kinds of training samples, a class semantic characteristic of the training sampleand a true label matrix; calculating a semantic similarity matrix by means of cosine similarity or Gaussian similarity through the class semantic characteristic; calculating a diagonal matrix of a class semantic similarity matrix; calling a Sylvester equation in an MATLAB toolset for obtaining a mapping matrix; inputting the vision characteristic of the training sample, the corresponding class semantic characteristic and the true label matrix into a target function, continuously adjusting the value of a model regularization parameter, calculating the least value of the target function, and finishing model training; and in a testing period, inputting the vision characteristic of the testing sample and the corresponding semantic characteristic, calculating scores of the classes, and determining the class with highest score as the predicated class of the testing sample. The zero-sample classifying method based on class transfer has advantages of sufficiently digging the semantic relationbetween different classes, realizing knowledge transfer between a known class classifier and an unknown class classifier, and realizing high convenience in application in image classification.

The invention discloses an indoor fusion positioning method based on multi-features. The whole process of the method is divided into three parts such as Wi-Fi fingerprint probability calculation, inertial navigation calculation and particle filtering fusion, wherein the Wi-Fi fingerprint probability calculation uses a cosine similarity matching algorithm, the inertial navigation calculation is divided into three parts such as step number calculation, step size calculation and walking direction estimation, the step number calculation uses methods such as walking detection, center moving equating, window wave peak detection, dynamic time regularization, derivative analysis filtering and wave peak resurrection, the step size calculation uses a step size and step frequency relationship model, the walking direction estimation uses the time and frequency zone analysis of a horizontal direction acceleration and strategy adjusting based on a gravity direction, and finally in the particle filtering fusion part, the hidden markov model constructed based on the inertial navigation and the Wi-Fi fingerprint probability uses the particle filtering method and combines with the map information to calculate the position probability so as to finally obtain the positioning result.

The present invention discloses a cross-scene pedestrian searching method based on depth learning. The method comprises a step of carrying out preprocessing on each image in a sample library, a step of constructing and training a convolutional neural network, a step of extracting an upper half body local feature vector set and a lower half body local feature vector set from two groups of preprocessed image sets, and then the two local feature vector sets are fused to obtain a global feature vector, a step of carrying out preprocessing on an image to be searched, extracting an upper half body local feature vector and a lower half body local feature vector and fusing the two vectors to obtain a global feature vector, a step of orderly comparing the global feature vector corresponding to the image to be searched and the global feature vectors corresponding to the sample library images through a cosine similarity, outputting a group of similarity values, and sorting the similarity values according to a sorting algorithm. The method has the advantages that with the pedestrian images obtained in a monitoring video as the sample library, the design of features is not needed, the robustness is high, and the accuracy rate of actual searching is high.

The invention discloses an item-based explicit and implicit feedback mixing collaborative filtering recommendation algorithm. The method comprises the following steps of obtaining the information of interest of users on every item and establishing the score matrix of every user on all the items; calculating the average score of every user, the quantity of the scoring users of every item and the average score of every item; calculating a common comment user quantity matrix; calculating the Pearson similarity and the modified cosine similarity of between any two items; calculating the similarity based on explicit feedback; calculating the cosine similarity based on implicit feedback; calculating a final similarity; obtaining the nearest neighbor set I of a current item; when providing a recommendation list to a target user u, according to the score matrix, obtaining the scored items and the unscored items of the target user u; calculating the prediction scores of the unscored items of the target user u and selecting N items with the highest scores inside the unscored items of the target user u to the user. The item-based explicit and implicit feedback mixing collaborative filtering recommendation algorithm can effectively improve the accuracy of prediction recommendation.

The invention relates to the field of artificial intelligence, discloses a method, device and equipment for obtaining a supervision recognition result in multiple modes and a storage medium, and solves the problem of semantic similarity matching of current business supervision terms and business products. The method comprises the following steps: creating a knowledge graph; processing the knowledge graph according to a first preset rule, a second preset rule and the entity relationship file to obtain an entity and an entity relationship; updating the knowledge graph according to the entities and the entity relationship to obtain a target knowledge graph; analyzing the target knowledge graph and the training text through an encoder to obtain fused to-be-processed information; performing random masking processing on the fused to-be-processed information according to a preset strategy to obtain training data; performing word embedding vector processing and self-made force mechanism processing on the training data to obtain a target sentence vector and a target word vector; and calculating a weighted average value of the semantic cosine similarity and the character string similarity ofthe target sentence vector and the target word vector according to a preset weight ratio to obtain a supervision and identification result.