Xerox Corporation

Naver Corporation

Stéphane Clinchant

Gabriela Csurka

Boris Chidlovskii

Jean-Michel Renders

Vassilina Nikoulina

Florent C Perronnin

Marc Dymetman

Guillaume M Bouchard

Jutta Willamowski

Nikolaos Lagos

Hady Elsahar

Julien Ah-Pine

German Kruszewski

Sofia Michel

Badr Youbi Idrissi

Bryan Eikema

Christophe Legras

Thibault Formal

Methods and systems for training a neural language model. Clean sequence pairs are received including clean source and target sequences. For each clean sequence pair, a noisy version is sampled with an adversarial generator to generate a noisy sequence pair. Parameters of the neural language model are optimized on the clean and noisy sequence pairs. Parameters of the adversarial generator are optimized to minimize a modeling loss of the adversarial generator and maximize a neural language loss of the neural language model using backpropagation.

Adversarial generation method for training a neural model

A sampling system includes: an energy-based model (EBM) configured to generate non-negative scores of an input having discrete classifications, respectively; and a sampling module configured to: generate a sample from a probability distribution of the EBM using a proposal distribution; set a probability of acceptance of the sample based on a minimum of (a) 1 and (b) an acceptance value determined based on the sample, a score of the sample from the EBM, the proposal distribution, and an upper boundary value; determine a distribution value between 0 and 1 using a uniform distribution; and discard the sample when the distribution value is greater than the probability of acceptance of the sample.

Sampling from discrete energy-based models with quality/efficiency trade-off

A system and method perform cross-modal information retrieval, by generating a graph representing the set of media objects. Each node of the graph corresponds to a media object and is labeled with a set of features corresponding to a text part of the respective media object. Each edge between two nodes represents a similarity between a media part of the two nodes. A first relevance score is computed for each media object of the set of media objects that corresponds to a text-based score. A second relevance score is computed for each media object by inputting the graph into a graph neural network. The first relevance score and the second relevance score are combined to obtain a final ranking score for each media object.

System and method for performing cross-modal information retrieval using a neural network using learned rank images

Generating a pedestrian tour includes: receiving a query from a user device via a network, the query indicative of a request for a pedestrian tour and including a location and preferences including a target length for the pedestrian tour; obtaining a graph of a geographical area around the location of the user device, the graph including nodes indicative of path crossings and arcs indicative of paths connecting pairs of the nodes; determining scores for the arcs based on characteristics of the arcs, respectively, and the preferences; selecting connecting ones of the arcs based on the scores of the connecting ones of the arcs, lengths of the selected connecting ones of the arcs, and the target length; adding the selected connecting ones of the arcs to the pedestrian tour to initialize the pedestrian tour; and transmitting the pedestrian tour to the user device via the network for display.

System and method for generating pedestrian tours

A classification method includes receiving a collection of samples, each sample comprising a multidimensional feature representation. A class label prediction for each sample in the collection is generated with one or more pretrained classifiers. For at least one iteration, each multidimensional feature representation is augmented with a respective class label prediction to form an augmented representation, a set of corrupted samples is generated from the augmented representations, and a transformation that minimizes a reconstruction error for the set of corrupted samples is learned. An adapted class label prediction for at least one of the samples in the collection is generated using the learned transformation and information is output, based on the adapted class label prediction. The method is useful in predicting labels for target samples where there is no access to source domain samples that are used to train the classifier and no access to target domain training data.

Transductive adaptation of classifiers without source data

A domain-adapted classification system and method are disclosed. The method includes mapping an input set of representations to generate an output set of representations, using a learned transformation. The input set of representations includes a set of target samples from a target domain. The input set also includes, for each of a plurality of source domains, a class representation for each of a plurality of classes. The class representations are representative of a respective set of source samples from the respective source domain labeled with a respective class. The output set of representations includes an adapted representation of each of the target samples and an adapted class representation for each of the classes for each of the source domains. A class label is predicted for at least one of the target samples based on the output set of representations and information based on the predicted class label is output.

Adapted domain specific class means classifier

Documents of a set of documents are represented by bag-of-words (BOW) vectors. L labeled topics are provided, each labeled with a word list comprising words of a vocabulary that are representative of the labeled topic and possibly a list of relevant documents. Probabilistic classification of the documents generates for each labeled topic a document vector whose elements store scores of the documents for the labeled topic and a word vector whose elements store scores of the words of the vocabulary for the labeled topic. Non-negative matrix factorization (NMF) is performed to generate a document-topic model that clusters the documents into k topics where k&#x3e;L. NMF factors representing L topics of the k topics are initialized to the document and word vectors for the L labeled topics. In some embodiments the NMF factors representing the L topics initialized to the document and word vectors are frozen, that is, are not updated by the NMF after the initialization.

Joint approach to feature and document labeling

A machine learning method operates on training instances from a plurality of domains including one or more source domains and a target domain. Each training instance is represented by values for a set of features. Domain adaptation is performed using stacked marginalized denoising autoencoding (mSDA) operating on the training instances to generate a stack of domain adaptation transform layers. Each iteration of the domain adaptation includes corrupting the training instances in accord with feature corruption probabilities that are non-uniform over at least one of the set of features and the domains. A classifier is learned on the training instances transformed using the stack of domain adaptation transform layers. Thereafter, a label prediction is generated for an input instance from the target domain represented by values for the set of features by applying the classifier to the input instance transformed using the stack of domain adaptation transform domains.

Domain adaptation by multi-noising stacked marginalized denoising encoders

Processing methods and systems are provided for representing documents relative to importance of words in the document. A processor comprising a weighting model of word importance in a document in a collection relative to an importance of the word in other documents in the collection computes a deviation of distribution of the word from a probability distribution of the word in other documents in the collection, where the deviation distribution is weighted in accordance with a concavity control function. A concavity control parameter is adjustable relative to word frequency.

Measuring informative content of words in documents in a document collection relative to a probability function including a concavity control parameter

A set of word embedding transforms are applied to transform text words of a set of documents into K-dimensional word vectors in order to generate sets or sequences of word vectors representing the documents of the set of documents. A probabilistic topic model is learned using the sets or sequences of word vectors representing the documents of the set of documents. The set of word embedding transforms are applied to transform text words of an input document into K-dimensional word vectors in order to generate a set or sequence of word vectors representing the input document. The learned probabilistic topic model is applied to assign probabilities for topics of the probabilistic topic model to the set or sequence of word vectors representing the input document. A document processing operation such as annotation, classification, or similar document retrieval may be performed using the assigned topic probabilities.

Document processing employing probabilistic topic modeling of documents represented as text words transformed to a continuous space

A multimedia information retrieval system includes a storage and an electronic processing device. The latter is configured to perform a process including: computing values of a pairwise similarity measure quantifying pairwise similarity of documents of a multimedia reference repository; storing the computed values in the storage; performing an initial information retrieval process respective to the multimedia reference repository to return a set of initial repository documents; and identifying a set of top ranked documents of the multimedia reference repository based at least on the stored computed values pertaining to the set of initial repository documents.

Cross-media similarity measures through trans-media pseudo-relevance feedback and document reranking

A translation method is adapted to a domain of interest. The method includes receiving a source text string comprising a sequence of source words in a source language and generating a set of candidate translations of the source text string, each candidate translation comprising a sequence of target words in a target language. An optimal translation is identified from the set of candidate translations as a function of at least one domain-adapted feature computed based on bilingual probabilities and monolingual probabilities. Each bilingual probability is for a source text fragment and a target text fragment of the source text string and candidate translation respectively. The bilingual probabilities are estimated on an out-of-domain parallel corpus that includes source and target strings. The monolingual probabilities for text fragments of one of the source text string and candidate translation are estimated on an in-domain monolingual corpus.

Lexical and phrasal feature domain adaptation in statistical machine translation

Cross-lingual information retrieval is disclosed, comprising: translating a received query from a source natural language into a target natural language; performing a first information retrieval operation on a corpus of documents in the target natural language using the translated query to retrieve a set of pseudo-feedback documents in the target natural language; re-translating the received query from the source natural language into the target natural language using a translation model derived from the set of pseudo-feedback documents in the target natural language; and performing a second information retrieval operation on the corpus of documents in the target natural language using the re-translated query to retrieve an updated set of documents in the target natural language.

Query translation through dictionary adaptation

A translation system and method suited to use in Cross Language Information Retrieval employ a retrieval-based scoring function for reranking candidate translations. The method includes translating an input source language query to generate a set of the candidate translations in a target language. The candidate translations are scored with the scoring function, which allows them to be reranked, and an optimal one or more selected for use in querying a domain-specific collection of documents in the target language. The scoring function applies weights to features extracted from the candidate translations. The weights have been learned on features extracted from translated queries, each generated by translation of an original query into the target language, and a measure of information retrieval performance of each the translated queries. One or more of the features used is a domain-specific feature which relies on a corpus of documents in the domain of interest.

Domain adaptation for query translation

A system and method for information retrieval are disclosed. The method includes querying a multimedia collection with a first component of a multimedia query (e.g., a text-based part of the query) to generate first comparison measures between the first component of the query and respective objects in the collection for a first media type (e.g., text). The multimedia collection is queried with a second component of the multimedia query (e.g., an image-based part of the query) to generate second comparison measures between the second component of the query and respective objects in the collection for a second media type (e.g., visual). An aggregated score for each of a set of objects in the collection is computed, based on the first comparison measure and the second comparison measure for the object. This includes applying an aggregating function to the first and second comparison measures in which a first confidence weighting is applied to the first comparison measure and a second confidence weighting is applied to the second comparison measure. The first confidence weighting is independent of the second comparison measure. The second confidence weighting is dependent on the first comparison measure. Information based on the aggregated scores is output.

Growing community of inventors

Grenoble, France

Stéphane Clinchant