Microsoft Technology Licensing, LLC

Linkedin Corporation

Bing Zhao

Jeffrey William Pasternack

Arpit Dhariwal

Nimesh Madhavan Chakravarthi

Nandeesh Channabasappa Rajashekar

Ethan Zhang

Bo Long

Kin Fai Kan

Birjodh S Tiwana

Romer E Rosales-Delmoral

Pi-Chuan Chang

Adam Leon

Baolei Li

Ada Yu

Chung Yu Wang

Tianhua Duan

Daniel Bikel

Techniques for generating diverse smart replies using a synonym hierarchy are disclosed herein. A computer system may detect that a first set of one or more messages having first content has been transmitted from a first computing device of a first user to a second computing device of a second user, determine a plurality of candidate replies based on the first content of the first set of one or more messages, and then select a plurality of smart replies from the plurality of candidate replies using a hierarchical graph data structure and at least one diversity rule. The selecting of the plurality of smart replies comprises omitting at least one of the plurality of candidate replies from selection based on the at least one diversity rule, which limits a number of the plurality of smart replies that have a common parent node.

Generating diverse smart replies using synonym hierarchy

Disclosed are systems, methods, and non-transitory computer-readable media for generating personalized smart responses. A messaging system receives a message transmitted as part of a communication session between a first user and a second user and determines a set of candidate responses for replying to the message. The messaging system filters the set of candidate responses based on a personalized vocabulary for the first user, yielding a set of personalized candidate responses. The personalized vocabulary for the first user is determined based on member profile data associated with the first user. The messaging system determines, based on a set of personalized candidate selection rules, a subset of the candidate responses, yielding a set of recommended responses to the message, and causes the set of recommended responses to be presented on the second client device.

Generating personalized smart responses

This disclosure provides for systems and methods that generate personalized electronic messages for members of a networked communication service. The personalized electronic messages are generated according to commonalities between member profiles. In one approach, an electronic message template is referenced based on the commonalities. The electronic message template is populated with member profile attribute values selected from the member profiles. In another approach, a long-short term memory (LSTM) recurrent neural network (RNN) is used to generate the electronic messages. Under this approach, a sequence-to-sequence model is trained using previous electronic messages labeled with one or more member profile attributes and/or member profile attribute values. When provided with one or more member profile attribute values associated with matching member profiles, the LSTM RNN outputs a relevant and appropriate electronic message designed to create an interest in the recipient of the electronic message.

Writing personalized electronic messages using template-based and machine-learning approaches

Disclosed are systems, methods, and non-transitory computer-readable media for a sequence to sequence to classification model for generating recommended messages. A messaging system generates, using a sequence to sequence encoder, an embedding vector from a message being transmitted as part of a communication session, the sequence to sequence encoder having been trained based on historical message data that includes messages transmitted between users of the messaging system. The messaging system determines, based on the embedding vector, a set of candidate responses for replying to the first message, the set of candidate responses being a subset of a set of available responses. The messaging system selects, from the set of candidate responses, a set of recommended responses to the first message, and causes the set of recommended responses to be presented by a client device of a recipient user of the first message.

Sequence to sequence to classification model for generating recommended messages

Disclosed are systems, methods, and non-transitory computer-readable media for generating recommended responses based on historical data. A messaging system receives a message as part of a communication session between a first client device and a second client device. The message originated from the first client device. The messaging system determines, using the message as input in a statistical model, a set of candidate responses for replying to the message. The statistical model was generated based on historical message data transmitted as part of previous communication sessions between a plurality of client devices. The set of candidate responses is a subset of a set of available candidate responses. The messaging system determines, based on a set of candidate selection rules, a subset of the candidate responses yielding a set of recommended responses to the message, and causes the set of recommended responses to be presented on the second client device.

Generating recommended responses based on historical message data

Techniques for identifying multilingual cognates and using the multilingual cognates are provided. In one technique, multilingual cognates identified from multiple user profiles are used to train one or more translation models. In another technique, multilingual cognates identified from a single user's profile are used to translate text provided by that user. In another technique, multilingual cognates from a single user are used to align sentences in one language to sentences in another language and the aligned sentences are used to train a language model. In another technique, multilingual cognates identified from multiple user profiles are used to expand search queries. In another technique, multilingual cognates identified from multiple user profiles are used to translate other users' profiles into a target language so that users associated with a source language are viewing the other users' profiles.

Data mining multilingual and contextual cognates from user profiles

Techniques are provided to allow users of a social network to have multilingual profiles (or profiles in second languages that are different than the users' native, or original, profile languages). In one technique, a translation model is applied to one or more data items (in a first language) in a user's profile to generate translated data items in a second language. The translated data items are displayed to the user (or an online social 'friend' of the user) and the user is allowed to select one of the translated data items to include in the user's 'second language' profile. The selection may then be used to improve the translation model.

Generating multi-language social network user profiles by translation

Techniques for training a tokenizer (or word segmenter) are provided. In one technique, a tokenizer tokenizes a token string to identify individual tokens or words. A language model is generated based on the identified tokens or words. A vocabulary about an entity, such as a person or company, is identified. The vocabulary may be online data that refers to the entity, such as a news article or a profile page of a member of a social network. Some of the tokens in the vocabulary may be weighted higher than others. The language model accepts the weighted vocabulary as input and generates pseudo sentences. Alternatively, regular expressions are used to generate the pseudo sentences. The pseudo sentences are used to train the tokenizer.

Growing community of inventors

Sunnyvale, CA, United States of America

Bing Zhao