Yahoo! Inc.

The State University of New York

Microsoft Technology Licensing, LLC

Ruofei Zhang

Zhongfei Zhang

Ramesh R Sarukkai

Subodh Shakya

Ming Zhou

Jianchang Mao

John Thrall

Nan Duan

Jian Jiao

Yeyun Gong

Sai Surya Kiran Evani

Systems and methods are provided for generating a keyword sequence from an input query. A first text sequence corresponding to an input query may be received and encoded into a source sequence representation using an encoder of a machine learning model. A keyword sentence may then be generated from the source sequence representation using a decoder of the machine learning model. The decoder may generate a modified generation score for a plurality of prediction tokens, wherein the modified generation score is based on the respective prediction token generation score and a maximum generation score for a suffix of each prediction token. The decoder may then select the prediction token of the plurality of prediction tokens based on the modified generation score, and add the selected prediction token to the previously decoded partial hypothesis provided by the decoder.

Look ahead strategy for trie-based beam search in generative retrieval

Systems and Methods for multi-modal or multimedia image retrieval are provided. Automatic image annotation is achieved based on a probabilistic semantic model in which visual features and textual words are connected via a hidden layer comprising the semantic concepts to be discovered, to explicitly exploit the synergy between the two modalities. The association of visual features and textual words is determined in a Bayesian framework to provide confidence of the association. A hidden concept layer which connects the visual feature(s) and the words is discovered by fitting a generative model to the training image and annotation words. An Expectation-Maximization (EM) based iterative learning procedure determines the conditional probabilities of the visual features and the textual words given a hidden concept class. Based on the discovered hidden concept layer and the corresponding conditional probabilities, the image annotation and the text-to-image retrieval are performed using the Bayesian framework.

System and method for multimedia ranking and multi-modal image retrieval using probabilistic semantic models and expectation-maximization (EM) learning

Disclosed are apparatus and methods for detecting whether a video is adult or non-adult. In certain embodiments, a learning system is operable to generate one or more models for adult video detection. The model is generated based on a large set of known videos that have been defined as adult or non-adult. Adult detection is then based on this adult detection model. This adult detection model may be applied to selected key frames of an unknown video. In certain implementations, these key frames can be selected from the frames of the unknown video. Each key frame may generally correspond to a frame that contains key portions that are likely relevant for detecting pornographic or adult aspects of the unknown video. By way of examples, key frames may include moving objects, skin, people, etc. In alternative embodiments, a video is not divided into key frames and all frames are analyzed by a learning system to generate a model, as well as by an adult detection system based on such model.

Apparatus and methods for detecting adult videos

System and method for image annotation and multi-modal image retrieval using probabilistic semantic models comprising at least one joint probability distribution

Disclosed are apparatus and methods for facilitating the ranking of web objects. The method includes automatically adjusting a plurality of weight values for a plurality of parameters for inputting into a ranking engine that is adapted to rank a plurality of web objects based on such weight values and their corresponding parameters. The adjusted weight values are provided to the ranking engine so as to generate a ranked set of web objects based on such adjusted weight values and their corresponding parameters, as well as a particular query. A relevance metric (e.g., that quantifies or qualifies how relevant the generated ranked set of web objects are for the particular query) is determined. The method includes automatically repeating the operations of adjusting the weight values, providing the adjusted weight values to the ranking engine, and determining a relevance metric until the relevance metric reaches an optimized level, which corresponds to an optimized set of weight values. The repeated operations utilize one or more sets of weight values including at least one set that results in a worst relevance metric value, as compared to a previous set of weight values, according to a certain probability in order to escape local optimal solution to reach the global optimal solution.

Optimization framework for tuning ranking engine

Techniques are described herein for automatically evaluating the quality of digital images based on one or more color characteristics of the images. In some embodiments, a quality metric that indicates the likelihood that the digital images convey semantics is generated based on color characteristics of the digital images. The quality metric may be used, for example, to determine which keyframe to use to make a thumbnail to represent video data. In some embodiments, feature values are generated for an image based on color characteristics of the image, and the feature values are assigned to bins. In such embodiments, the quality metric may be generated to indicate how uniform the distribution of feature values is among the bins.

Automatically generating a content-based quality metric for digital images

The present invention provides for improving the search relevance of a search results page by including a perceived relevance factor. The system, device and method monitors user selection of elements in the search results page, where these selections indicate relevance of the element compared with the original search request. A perceived relevance factor for the element is then determined based on probabilistic-based computations accounting for the element, which may include a description, a thumbnail and/or meta data, and the position of the element on the search results page. Thereby, for future searches and search results page generation, relevance factors may be calculated based on various factors, including the element attribute based relevant scores and the perceived relevance factor.

System and method for modeling user selection feedback in a search result page

System and method for image annotation and multi-modal image retrieval using probabilistic semantic models

The present invention is directed towards systems and methods for generating one or more channels for the organization of content items. A method according to one embodiment comprises selecting a content item and one or more items of metadata for the selected content item. A determination is made to determine if the selected content item should be associated with a given channel on the basis of the metadata. A channel is generated on the basis of the determination, with the selected content item organized in association with the channel.

Growing community of inventors

Sunnyvale, CA, United States of America

Ruofei Zhang