International Business Machines Corporation

Priscilla Santos Moraes

Lakshminarayanan Krishnamurthy

Donna Karen Byron

Michael C Smith

Devendra Goyal

Ravi S Sinha

Amrish V Chaubal

Kadriye E Eyigoz

Shunguo Yan

David Konopnicki

Sridhar Sudarsan

Niyati Parameswaran

Dan G Tecuci

Devin A Conley

Kathryn V Banks

A computerized method comprising receiving, from a question answering system, a minimal answer value to a query submitted by a user. Also received are electronic documents based on the minimal answer value, and a document score value, associated with the query, for each of the electronic documents. The method comprises extracting entities and attributes from electronic documents, and for each computing one or more associated score value, and aggregating the document score value with the associated score values. The method comprises selecting some of entities and attributes based on the respective aggregated score value, thereby producing selected associated elements. The method comprises generating, using a computerized natural language (NL) generating system, a comprehensive NL answer, wherein the generating is based on the minimal answer value and the selected associated elements, and sending the comprehensive NL answer for presentation to the user.

Generating natural language answers automatically

Mechanisms are provided for implementing a candidate answer variant engine. The mechanisms receive an indication of a structured portion of content in a corpus, generate a plurality of groupings of elements of the structured portion of content, and generate, for each grouping of elements in the plurality of groupings of elements, and for each element in the grouping of elements, a corresponding grouping vector representation, corresponding to the element. The mechanisms, for each grouping vector representation of each grouping of elements in the plurality of groupings of elements perform a similarity measure calculation between the grouping vector representation and a vector representation of an input question, and select an element corresponding to the grouping vector representation for inclusion as a candidate answer variant based on results of the similarity measure calculation. The mechanisms perform question answering operations based on an analysis of one or more candidate answer variants.

Generating answer variants based on tables of a corpus

Electronic natural language processing in a natural language processing (NLP) system, such as a Question-Answering (QA) system. A receives electronic text input, in question form, and determines a readability level indicator in the question. The readability level indicator includes at least a grammatical error, a slang term, and a misspelling type. The computer determines a readability level for the electronic text input based on the readability level indicator, and retrieves candidate answers based on the readability level.

Readability awareness in natural language processing systems

An engagement classifier for a group chatbot is trained by leveraging the implicit dataset generated by humans engaging in both direct messages as well as group conversations. Human-to-human direct messages are used as an approximate representation of the domain knowledge and expertise of each user. The decision to engage in a group conversation is assumed to be based on that domain knowledge. The knowledge representations and instances of engagements in group conversations yields an effective set of features and labels which can be used to model the engagement decision. The same transfer learning technique is used to generate a knowledge representation for the group chatbot. Given this representation of the domain knowledge of the chatbot, the classifier can predict whether it should engage in any particular group conversation.

Unsupervised technique for training an engagement classifier in chat-based group conversation

To augment an image caption, a caption graph containing entity nodes corresponding to entities contained in the image and relationship edges between entity nodes corresponding to relationships between entities as illustrated in the image is generated. In addition, a contextual graph containing one or more of entity nodes corresponding to entities contained in the image and described in text associated with the image, textual entity nodes corresponding to textual entities described in text associated with the image and textual relationship edges between entity node pairs, textual entity node pairs and entity node and textual entity node pairs is generated. The textual relationship edges correspond to relationships described in the text associated with the image between entity pairs, textual entity pairs or entity and textual entity pairs. From the contextual graph, an augmented caption graph containing entity nodes, relationship edges, textual entities and textual relationship edges is generated.

Image captioning augmented with understanding of the surrounding text

Ground truth for a cognitive system is generated from a structured resource such as a table by identifying a subject of the structured resource and field headers. Linguistic analysis is performed on a given header to establish an interrogative context, and a question is generated relating to the subject based on the interrogative context, including an implementation of one or more mathematical operators. The question is generated using a question template, and has a question phrase based on the interrogative context, an operator phrase based on the selected operator, and a keyword phrase based on the subject. An answer to the question is determined by carrying out a computation that applies the selected operator(s) to one or more of the data values, to form a question-and-answer pair that is added to the ground truth. A filtering step is preferably used to ensure that the question-and-answer pair is valid.

Generating ground truth for questions based on data found in structured resources

Answer effectiveness evaluations include providing, by a computing device, an answer to a search query received from a user, and in response to receiving a subsequent search query from the user, determining by the computing device a level of effectiveness of the answer to the search query with respect to the user. The determination includes comparing aspects of the search query to aspects of the subsequent search query, calculating, based on the comparing, a relevance score that indicates a measure of similarity between the aspects of the search query and the aspects of the subsequent search query, and determining that the answer effectively answers the search query when the relevance score exceeds a threshold value.

Growing community of inventors

Pflugerville, TX, United States of America

Priscilla Santos Moraes