Beijing Didi Infinity Technology and Development Co., Ltd.

Facebook, Inc.

Meta Platforms, Inc.

Kun Han

Fuchun Peng

Benoit F Dumoulin

Wenhai Yang

Ying Lyu

Emmanouil Koukoumidis

Zheng Zhou

Haiyang Xu

Bo Zeng

Vivek Narayanan

Cheng Zhang

Anuj Kumar

Michael Robert Hanson

Fei Sha

Embodiments of the disclosure provide a multi-class classification system. An exemplary system includes at least one processor and at least one non-transitory memory storing instructions that, when executed by the at least one processor, cause the system to perform operations. The operation includes applying a multi-class classifier to classify a set of objects into multiple classes and applying a plurality of binary classifiers to the set of objects, wherein the plurality of binary classifiers are decomposed from the multi-class classifier, each binary classifier classifying the set of the objects into a first group consisting of one or more classes selected from the multiple classes and a second group consisting of one or more remaining classes of the multiple classes. The operation also includes jointly classifying the set of objects using the multi-class classifier and the plurality of binary classifiers.

Adversarial multi-binary neural network for multi-class classification

In one embodiment, a method includes, by one or more computing systems, receiving, from a client system via an assistant xbot, a first audio input from a user, parsing the first audio input to identify a request for a content digest from an online social network, retrieving a plurality of content objects from the online social network, each of the content objects being accessible by the user, generating a customized newsfeed summary based on a determined semantical meaning of each of the plurality of content objects, and sending, to the client system via the assistant xbot, instructions for presenting an audio clip to the user responsive to the request for the content digest, wherein the audio clip comprises the customized newsfeed summary.

Content suggestions for content digests for assistant systems

The present disclosure is related to systems and methods for processing speech dialogue. The method includes obtaining target speech dialogue data. The method includes obtaining a text vector representation sequence, a phonetic symbol vector representation sequence, and a role vector representation sequence by performing a vector transformation on the target speech dialogue data based on a text embedding model, a phonetic symbol embedding model, and a role embedding model, respectively. The method includes determining a representation vector corresponding to the target speech dialogue data by inputting the text vector representation sequence, the phonetic symbol vector representation sequence, and the role vector representation sequence into a trained speech dialogue coding model. The method includes determining a summary of the target speech dialogue data by inputting the representation vector into a classification model.

Systems and methods for processing speech dialogues

In one embodiment, a method includes receiving an indication of a trigger action by a first user at a client system, wherein the trigger action is associated with a priming content object, identifying related content objects associated with the priming content object, selecting recommended content objects based on the priming content object, the related content objects, and profile information of the first user, wherein each of the selected recommended content objects comprises entity information of entities associated with the priming content object, and presenting content suggestions at the client system, wherein each content suggestion comprises one of the selected recommended content objects.

Engaging users by personalized composing-content recommendation

In some cases, lower quality, large scale training data can be automatically generated by automatic labeling. The generated training data can be used to pre-train a machine learning model. For instance, the model can be a model for detection of verbal harassment. Parameters of the pre-trained model can be refined or updated using another one or more higher-quality sets of training data, with which the model can be subsequently trained.

Training mechanism of verbal harassment detection systems

In some cases, a verbal harassment detection system may use machine learning models to detect verbal harassment in real-time or near real-time. The system may receive an audio segment comprising a portion of audio captured by a microphone located within a vehicle. Further, the system may convert the audio segment to a text segment. The system may provide at least the text segment to a prediction model associated with verbal harassment detection to obtain a harassment prediction. Further, the system may provide the audio segment to an emotion detector to obtain a detected emotion of a speaking user that made an utterance included in the audio segment. Based at least in part on the harassment prediction and the detected emotion, the system may automatically, and without user intervention, determine whether a user is being harassed.

Real-time verbal harassment detection system

In some cases, one or more heuristics can be automatically generated using a small dataset of segments previously labeled by one or more domain experts. The generated one or more heuristics along with one or more patterns can be used to assign training labels to a large unlabeled dataset of segments. A subset of segments representing an occurrence of verbal harassment can be selected using the assigned training labels. Randomly selected segments can be used as being indicative of a non-occurrence of verbal harassment. The selected subset of segments and randomly selected segments can be used to train one or more machine learning models for verbal harassment detection.

Generation of training data for verbal harassment detection

A computing device receives a message including a request for a recommendation. A representation of a hypothetical ideal recommendation to provide in response to the message is determined based on the message content. Data regarding entities that are potential recommendations are retrieved from a data store, the data regarding each entity including a representation of the entity (e.g., a vector) derived from factual information about the entity and opinions of other users of the entity. Ranking scores are determined for at least a subset of the entities based on the difference between the entity representations and the representation of the hypothetical ideal recommendation. An entity to recommend is selected based on the ranking scores and a reply to the message is sent that identifies the selected entity.

Combining machine-learning and social data to generate personalized recommendations

In one embodiment, a method includes, by one or more computing systems, receiving a request from a user for a content digest from an online social network, retrieving one or more content objects associated with the online social network that are accessible by the user, determining a semantical-embedding for each retrieved content object based on a query model, determining one or more categories for each retrieved content object, generating a set of content suggestions for each retrieved content object based on the one or more categories associated with the content object and the semantical-embedding of the content object, ranking for each retrieved content object, the one or more content suggestions in the respective set based on a comparison of a semantical-embedding associated with each content suggestion to the semantical-embedding of the content object, and sending instructions for presenting the content digest to the user.

In one embodiment, a method includes receiving an indication of a trigger action by a first user, wherein the trigger action is associated with a priming content object, identifying one or more related content objects associated with the priming content object, generating a first feature vector representing the priming content object, the one or more related content objects, and profile information of the first user, accessing a plurality of second feature vectors representing a plurality of recommended content objects, respectively, selecting one or more of the recommended content objects based on comparisons between the first feature vector and the respective second feature vectors representing the recommended content objects, and sending, to a client system in response to the trigger action, instructions for presenting one or more content suggestions to the first user, wherein each content suggestion comprises one of the selected recommended content objects.

In one embodiment, a method includes receiving a user request for a summarization of a particular type of content objects from a client system associated with a first user, determining one or more modalities associated with the user request, selecting a plurality of content objects of the particular type based on a user profile of the first user, wherein the user profile comprises one or more confidence scores associated with one or more subjects associated with the first user, respectively, and wherein the plurality of content objects are selected based on the one or more confidence scores, generating a summary of each content object based on the user profile and the determined modalities, and sending, to the client system in response to the user request, instructions for presenting the summaries of the plurality of content objects, wherein the summaries are presented via one or more of the determined modalities.

Generating personalized content summaries for users

An online system receives a request to generate presentation content for presentation to a user. The online system receives a set of content items and identifies a surface for presenting the presentation information to the user. For example, the surface may be a voice only surface, a voice and graphical display, a graphical display only. Based on the identified surface, the online system ranks the set of content items. The online system then determines presentation information for a subset of the content items and transmits instructions to present the presentation information at the surface.

Growing community of inventors

Sunnyvale, CA, United States of America

Kun Han