International Business Machines Corporation

Saneem Ahmed Chemmengath

Samarth Bharadwaj

Jaydeep Sen

Vishwajeet Kumar

Karthik Sankaranarayanan

Suranjana Samanta

Kuntal Dey

Pankaj Satyanarayan Dayama

Mustafa Canim

L Venkata Subramaniam

Amit Dhurandhar

Dinesh Garg

Pooja Aggarwal

Amar Prakash Azad

Ashish Mittal

Parag Jain

Sreyash Kenkre

Dhinakaran Vinayagamurthy

Shreya Khare

Prateeti Mohapatra

Anirban Laha

Saswati Dana

Ronny Luss

Saravanan Krishnan

Sandeep Nishad

Ioannis Katsis

Harsh Chaudhari

Retrieval aware natural language question generation for open domain document retrieval is provided. In one aspect, a system for retrieval aware question generation includes: a question decontextualizer configured to decontextualize a question generated from a context of a target document by adding terms from the context into the question itself to create a decontextualized question, where the decontextualized question alone enables open domain document retrieval without a need for also providing the context. The system can also include a detect document identifier configured to find the terms in the context; and a retriever configured to retrieve documents from the corpus of documents using the decontextualized question. A method for retrieval aware question generation using the present system is also provided.

Retrieval aware question generation

A method, computer program product and system are provided to generate perturbed text is provided. A processor receives a string of text from a user. A processor determines one or more classifications for at least one word in the string of text by a classification model. A processor determines a plurality of perturbations of the at least one word based on the one or more classifications, where the plurality of perturbations do not share the same one or more classifications as the at least one word in the string of text. A processor selects a perturbation of the string of text based on (i) an edit distance between the string of text and the plurality of perturbations, and (ii) a fluency metric for each of the plurality of perturbations. A processor provides the perturbation of the string of text to the user.

System and method for generating contrastive explanations for text guided by attributes

A method, system, and computer program product for automated artificial intelligence (AI) factsheet generation for modeling and model customization in speech to text (STT) services. The method receives audio data for a user. The audio data contains human speech. Text data is generated, using a first speech to text model, to represent the human speech of the audio data. A set of transcription errors of the first speech to text model are identified. A set of AI factsheets are generated to describe model metadata for the first speech to text model. Based on the set of transcription errors and the set of AI factsheets, the method generates a second speech to text model customized to the user.

Artificial intelligence factsheet generation for speech recognition

Techniques for enhanced table and text question answering based on multi-instance, multi-answer training are presented. An answer extractor component can determine answer scores associated with candidate answer data items based on analysis of a set of data, comprising row data items of a table and passage data items associated with the table, and a context of a query of the set of data. The answer extractor component can be trained based on application of denoised single-instance and multiple-instance answer matching data associated with contexts to an answer extractor model to generate a trained answer extractor model of the answer extractor component. A query response component can determine a correct answer data item responsive to the query from the candidate answer data items based on the answer scores associated with the candidate answer data items, wherein the candidate answer data items can be reranked based on reweighted answer scores.

Multi-instance, multi-answer training for table and text question answering

A question answering bot that digests non-relational data tables is provided. A processor receives a question regarding a non-relational data table. A processor extracts at least one feature of the question using a natural language processing (NLP) model. A processor extracts at least one similar feature of the non-relational data table to the extracted at least one feature of the question. A processor determines at least one relevant cell in the non-relational data table based on the at least one feature of the question and the at least one similar feature of the non-relational data table. A processor provides an answer to the question, where the answer is based on the at least one relevant cell.

Natural language question answering using non-relational tables

Methods, systems, and computer program products for providing a contextual dialogue framework over dynamic tables are provided herein. A computer-implemented method includes maintaining a context space for a natural language conversation of a user, wherein the context space comprises a dynamic set of one or more tables used for processing at least one query of the natural language conversation; obtaining an additional table associated with an additional query of the natural language conversation; discovering one or more implicit links between the additional table and the dynamic set of tables; updating the context space with the one or more implicit links; and answering the additional query based at least in part on the updated context space.

Contextual dialogue framework over dynamic tables

One embodiment provides a method, including: receiving a query from a user; providing the query to data owners, wherein each of the data owners has a local machine learning model and wherein the plurality of data owners train a meta-model; secret sharing model output from the data owners between the other data owners, wherein the model output comprises an output responsive to the query computed using the local machine learning model; receiving, from each of the plurality of data owners, a set of meta-features corresponding to the query; and generating a response to the query, wherein the generating comprises determining, by evaluating the meta-model using the set of meta-features received from each of the plurality of data owners, weights for outputs from the local machine learning models and aggregating the outputs in view of the weights.

Secure ensemble training and inference using heterogeneous private machine learning models

One embodiment provides a method, including: receiving an indication of an addition of a new participant in a textual communication between at least two existing participants; identifying at least one confidential topic contained within the textual communication by (i) parsing the textual communication and (ii) identifying at least one topic contained within the textual communication; the identifying comprising (i) accessing a confidentiality graph comprising (a) nodes representing participants and (b) edges representing confidential concepts that are acceptable discussion topics between participants connected by a corresponding edge and (ii) determining that an edge corresponding to the at least one confidential topic does not connect the new participant with both of the existing participants; and alerting one of the existing participants that the at least one confidential topic is included in the textual communication to be sent to the new participant.

Confidential information identification based upon communication recipient

Methods, systems, and computer program products for explaining cross domain model predictions are provided herein. A computer-implemented method includes providing a test data point to a domain adaptation model to obtain a prediction, wherein the domain adaptation model is trained on a set of labeled data points and a set of unlabeled data points. The method includes generating a task specific explanation for the prediction that includes one or more data points from among the sets that satisfy a threshold score for influencing the prediction. Additionally, the method includes generating a domain invariant explanation for the prediction. The domain variation explanation is generated by ranking pairs of data points based on a statistical similarity to the test data point, wherein each pair includes a data point from the set of labeled data points and a data point from the set of unlabeled data points.

Explaining cross domain model predictions

Methods, systems, and computer program products for generating concept images of human poses using machine learning models are provided herein. A computer-implemented method includes identifying events from input data by applying a machine learning recognition model to at least a portion of the input data, wherein the identifying comprises (i) detecting multiple entities from the input data and (ii) determining behavioral relationships among at least a portion of the multiple entities; generating, using a machine learning interpretability model and at least a portion of the identified events, images illustrating human poses related to at least a portion of the identified events; outputting at least a portion of the generated images to a user; and updating the machine learning recognition model based at least in part on (i) at least a portion of the generated images and (ii) input from the user.

Generating concept images of human poses using machine learning models

Methods, systems, and computer program products for generating concept images of human poses using machine learning models are provided herein. A computer-implemented method includes identifying one or more events from input data by applying a machine learning recognition model to the input data, wherein the identifying comprises (i) detecting multiple entities from the input data and (ii) determining one or more behavioral relationships among the multiple entities in the input data; generating, using a machine learning interpretability model and the identified events, one or more images illustrating one or more human poses related to the identified events; outputting the one or more generated images to at least one user; and updating the machine learning recognition model based at least in part on (i) the one or more generated images and (ii) input from the at least one user.

Techniques for script modification are provided including receiving a script and parsing the script to identify at least one attribute of the script. The identified at least one attribute is presented to a user in a graphical user interface via a display device and an adjustment of at least one element in the graphical user interface that corresponds to the at least one attribute is received from the user via an input device. Modification data corresponding to the at least one attribute are received from a data repository and at least one attribute of the script is modified based on the received adjustment and the obtained modification data corresponding to the at least one attribute. A modified script is generated based on the modified at least one attribute.

Growing community of inventors

Kozhikode, India

Saneem Ahmed Chemmengath