Abstract: Techniques are disclosed relating to text sanitization. Given textual data, a computer system identifies tokens predicted to constitute sensitive information. Multi-field data structures (e.g., triplets) are generated for the identified tokens that include questions, answers, and corresponding context. These data structures are supplied to a pre-trained multiple-choice question (MCQ) reading comprehension model. The model outputs, for each data structure, a probability that the question and answer for a given data structure, provided the context, is accurate. A post-processing module can then rank probabilities in this set of probabilities and select the multi-field data structure with the highest probability (in some cases, a programmable threshold must also be met). The selected multi-field data structure is then used to select category information to be used in sanitizing the textual data.

Abstract: A method of categorizing text entries on a document can include determining, for each of a plurality of text bounding boxes in the document, respective text, respective coordinates, and respective input embeddings. The method may further include defining a graph of the plurality of bounding boxes, the graph comprising a plurality of connections among the plurality of bounding boxes, each connection comprising a first and second bounding box and zero or more respective intermediate bounding boxes. The method may further include determining a respective attention value for each connection according to a quantity of intermediate bounding boxes in the connection and, based on a the respective attention values and a transformer-based machine learning model applied to the respective input embeddings and respective coordinates, determining output embeddings for each bounding box and, based on the respective output embeddings, generating a bounding box label for each bounding box.

Abstract: Embodiments may provide improved techniques for mental health screening and its provision. For example, a method may comprise receiving input data relating to communications among persons, the input data comprising a plurality of modalities, extracting features relating to the plurality of modalities from the received input data, performing multimodal fusion on the extracted features, wherein the multimodal fusion is performed on at least some of the features relating to individual modalities and on at least some combinations of features relating to a plurality of modalities, classifying the fused features using a trained model for detection of at least one mental disorder, and generating a representation of a disorder state based on the classified fused features. For the multimodal fusion, a late fusion scheme instead of early fusion may be used to make the model more interpretable and explainable without compromising the performance.