Adobe, Inc.

Kushal Chawla

Niyati Himanshu Chhaya

Balaji Vasan Srinivasan

Sopan Khosla

Tanya Goyal

Hrituraj Singh

Projjal Chanda

Jaya Singh

Mithlesh Kumar

Arijit Pramanik

Atanu Ranjan Sinha

Ritwik Sinha

Gaurush Hiranandani

Deepak Pai

Moumita Sinha

Cedric Huesler

Dhruv Singal

Somak Aditya

Soumya Vadlamannati

Vaishnav Pawan Madandas

Aman Deep Singh

Yash Shrivastava

Aditya Jain

Sharmila Nangi Reddy

Pranil Joshi

Aarushi Agrawal

Bhavy Khatri

Abhinav Mishra

Certain embodiments involve using a gated convolutional encoder-decoder framework for applying affective characteristic labels to input text. For example, a method for identifying an affect label of text with a gated convolutional encoder-decoder model includes receiving, at a supervised classification engine, extracted linguistic features of an input text and a latent representation of an input text. The method also includes predicting, by the supervised classification engine, an affect characterization of the input text using the extracted linguistic features and the latent representation. Predicting the affect characterization includes normalizing and concatenating a linguistic feature representation generated from the extracted linguistic features with the latent representation to generate an appended latent representation. The method also includes identifying, by a gated convolutional encoder-decoder model, an affect label of the input text using the predicted affect characterization.

Detecting affective characteristics of text with gated convolutional encoder-decoder framework

Systems and methods for generating a tuned summary using a word generation model. An example method includes receiving, at a decoder of the word generation model, a training data learned subspace representation of training data. The method also includes identifying tunable linguistic characteristics of the word generation model and training the decoder to output a training tuned summary of the training data learned subspace representation based on at least one of the tunable linguistic characteristics. The method further includes receiving an input text and a target characteristic token, and generating, by the trained decoder of the word generation model, each word of a tuned summary of the input text from a learned subspace representation and from feedback about preceding words of the tuned summary, wherein the tuned summary is tuned to target characteristics represented by the target characteristic token.

Generating summary content tuned to a target characteristic using a word generation model

Systems and methods for natural language processing (NLP) are described. The systems may be trained by identifying training data including clean data and noisy data; predicting annotation information using an artificial neural network (ANN); computing a loss value for the annotation information using a weighted loss function that applies a first weight to the clean data and at least one second weight to the noisy data; and updating the ANN based on the loss value. The noisy data may be obtained by identifying a set of unannotated sentences in a target domain, delexicalizing the set of unannotated sentences, finding similar sentences in a source domain, filling at least one arbitrary value in the similar delexicalized sentences, generating annotation information for the similar delexicalized sentences using an annotation model for the source domain, and applying a heuristic mapping to produce annotation information for the sentences in the target domain.

Predicting joint intent-slot structure

Certain embodiments involve using a gated convolutional encoder-decoder framework for applying affective characteristic labels to input text. For example, a method for identifying an affect label of text with a gated convolutional encoder-decoder model includes receiving, at an encoder, input text. The method also includes encoding the input text to generate a latent representation of the input text. Additionally, the method includes receiving, at a supervised classification engine, extracted linguistic features of the input text and the latent representation of the input text. Further, the method includes predicting an affect characterization of the input text using the extracted linguistic features and the latent representation. Furthermore, the method includes identifying an affect label of the input text using the predicted affect characterization. The gated convolutional encoder-decoder model is jointly trained using a weighted auto-encoder loss associated with a reconstruction engine and a weighted classification loss associated with the supervised classification engine.

A web experience augmentation system predicts, during a web browsing session of a user, augmentation data that the user is likely to want to view during the web browsing session. This prediction is based on both local content preferences for the user and global content preferences. The local content preferences for the user refer to an indication of the webpages accessed during the current web browsing session of the user. The global content preferences refer to analytics for webpages on a website obtained over an extended period of time that extends prior to the web browsing session of the user. The web experience augmentation system also modifies a webpage to which the user navigates to include the predicted augmentation data.

Web experience augmentation based on local and global content preferences

Forecasting a potential audience size and an unduplicated audience size for a digital campaign includes receiving an audience segment input and a time period input. The audience segment input is converted into multiple atomic target specifications. For each of the multiple atomic target specifications, a potential audience size is determined during the time period input by selecting a time series model based on a frequency of attribute values from the atomic target specification and combining the selected time series model with a frequent item set model. The potential audience size for each of atomic target specifications is aggregated over the time period input into a total potential audience size. The total potential audience size is output. The time series model and the frequent item set model are obtained using data from a historic bid request database.

Forecasting potential audience size and unduplicated audience size

Certain embodiments involve tuning summaries of input text to a target characteristic using a word generation model. For example, a method for generating a tuned summary using a word generation model includes generating a learned subspace representation of input text and a target characteristic token associated with the input text by applying an encoder to the input text and the target characteristic token. The method also includes generating, by a decoder, each word of a tuned summary of the input text from the learned subspace representation and from a feedback about preceding words of the tuned summary. The tuned summary is tuned to target characteristics represented by the target characteristic token.

Certain embodiments involve facilitating natural language processing through enriched distributional word representations. For instance, a computing system receives an initial word distribution having initial word vectors that represent, within a multidimensional vector space, words in a vocabulary. The computing system also receives a human-reaction lexicon indicating human-reaction values respectively associated with words in the vocabulary. The computing system creates an enriched word distribution by modifying one or more of the initial word vectors such that the distance between the pair of initial word vectors representing a pair of words is decreased based on a human-reaction similarity between the pair of words.

Affect-enriched vector representation of words for use in machine-learning models

Techniques are disclosed to assist an author in creating content variations of a given input text to better suit the mood or the affect preferences of the target audience. Affect distribution in the content is utilized to capture these psycholinguistic preferences. According to one embodiment, in a first phase the optimal/idea psycholinguistic preference for text content aimed at a particular audience segment is determined. In a second phase, a given text content is modified to align to a target language distribution, which was determined in the first phase. In one example case, word level replacement, insertions and deletions are executed to generate a modified and coherent version of the input text. The output text thus reflects the psycholinguistic requirements of the audience.

Content optimization for audiences

An affective summarization system provides affective text summaries directed towards affective preferences of a user, such as psychological or linguistic preferences. The affective summarization system includes a summarization neural network and an affect predictor neural network. The affect predictor neural network is trained to provide a target affect level based on a word sequence, such as a word sequence for an article or other text document. The summarization neural network is trained to provide a summary sequence based on the target affect level and on the word sequence for the text document.

Machine learning techniques for generating document summaries targeted to affective tone

Techniques are disclosed for predicting a tone of a text communication using psycholinguistic features of the text communication. In some examples, a method may include generating a feature vector for a text communication using features, including psycholinguistic features, extracted from the text communication, and predicting a tone of the text communication based on the feature vector. The tone is predicted by a trained prediction module that is trained using psycholinguistic features of text communications in a training set used to train the trained prediction module. The predicted tone is at least one of a predicted measure of frustration, a predicted measure of formality, and a predicted measure of politeness.

Prediction of tone of interpersonal text communications

Systems and methods are disclosed herein for determining user behavior in an augmented reality environment. An augmented reality application executing on a computing system receives a video depicting a face of a person. The video includes a video frame. The augmented reality application augments the video frame with an image of an item selected via input from a user device associated with a user. The augmented reality application determines, from the video frame, a score representing an action unit. The action unit represents a muscle on the face of the person depicted by the video frame and the score represents an intensity of the action unit. The augmented reality application calculates, from a predictive model and based on the score, an indicator of intent of the person depicted by the video.

Growing community of inventors

Bengaluru, India

Kushal Chawla