Sap Se

Sohyeong Kim

Christian Reisswig

Darko Velkoski

Hung Tu Dinh

Eduardo Vellasques

Ying Jiang

Cordula Guder

Faisal El Hussein

Xiang Yu

Urko Sanchez Sanz

Embodiments are described for a system comprising a memory and at least one processor coupled to the memory. The at least one processor is configured to receive optical character recognition (OCR) information of a document and determine a beginning, inside, and outside (BIO) tags and labels of the one or more word boxes based on the OCR information. The at least one processor is further configured to group a first word box and a second word box based on BIO tags of the first and the second word boxes and merge the first and the second word boxes into a combined word box based on a label of the first word box matching a label of the second word box. Finally, the at least one processor is configured to output the combined word box and the label of the first word box.

Efficient document information extraction system using optical character recognition (OCR) information

A method for training machine learning model is provided. The method comprises extracting texts and locations of the texts from a document, generating embeddings for the document, a first set of the embeddings characterizing a first subset of the texts and locations of the first subset of the texts and a second set of the embeddings characterizing a second subset of the texts that are masked and additional locations of the second subset of the texts that are masked, generating additional embeddings characterizing contents of the second subset of the texts, generating relevance values based on a comparison, identifying, for each of the additional locations of the second subset of the texts that are masked, a respective content of the second subset of the texts having a reference value that is higher than a remaining relevance values, and outputting each of the respective content.

Training of machine learning models using content masking techniques

Methods, systems, and articles of manufacture, including computer program products, are provided for synthesizing images for machine learning. The method may include selecting one or more image preprocessing transformations to apply on the foreground object image; applying the selected one or more image preprocessing transformations to the foreground object image; selecting a background image from a set of background images depicting a variety of different backgrounds which may be associated with the foreground object image; merging the selected background image with the foreground object image to form a synthesized image; selecting one or more image transformations to apply on the synthesized image; applying the selected one or more image transformations to the synthesized image; and storing the synthesized image in a collection of synthesized images to train a machine learning model.

Query image synthesis for image retrieval systems based on data augmentation techniques

Various examples are directed to systems and methods for classifying text. A computing device may access, from a database, an input sample comprising a first set of ordered words. The computing device may generate a first language model feature vector for the input sample using a word level language model and a second language model feature vector for the input sample using a partial word level language model. The computing device may generate a descriptor of the input sample using a target model, the input sample, the first language model feature vector, and the second language model feature vector and write the descriptor of the input sample to the database.

Contextualized text description

Various examples described herein are directed to systems and methods for analyzing text. A computing device may train an autoencoder language model using a plurality of language model training samples. The autoencoder language mode may comprise a first convolutional layer. Also, a first language model training sample of the plurality of language model training samples may comprise a first set of ordered strings comprising a masked string, a first string preceding the masked string in the first set of ordered strings, and a second string after the masked string in the first set of ordered strings. The computing device may generate a first feature vector using an input sample and the autoencoder language model. The computing device may also generate a descriptor of the input sample using a target model, the input sample, and the first feature vector.

Bi-directional contextualized text description

Disclosed herein are system, method, and computer program product embodiments for providing an image search training system. An embodiment operates by determining a query image on which to train an image search system, and a positive image visually similar to the query image. A set of negative images from a plurality of negative images visually dissimilar to the query image are selected, where the selected set of negative images includes both a first negative image and a second negative image. A first similarity measure between the first negative image and the positive image, and a second similarity measure between the second negative image and the positive image are calculated. The first negative image is selected based on the first similarity measure being less than the second similarity measure. The query image, the positive image, and the first negative image are provided to the image search system for training.

Image search and training system

Disclosed herein are system, method, and computer program product embodiments for robust key value extraction. In an embodiment, one or more hierarchical concepts units (HCUs) may be configured to extract key value and hierarchical information from text inputs. The HCUs may use a convolutional neural network, a recurrent neural network, and feature selectors to analyze the text inputs using machine learning techniques to extract the key value and hierarchical information. Multiple HCUs may be used together and configured to identify different categories of hierarchical information. While multiple HCUs may be used, each may use a skip connection to transmit extracted information to a feature concatenation layer. This allows an HCU to directly send a concept that has been identified as important to the feature concatenation layer and bypass other HCUs.

Growing community of inventors

Berlin, Germany

Sohyeong Kim