Responding to user queries by context-based intelligent agents

A method comprising:

• receiving a user selection of one of a plurality of domain corpora for answering questions pertaining to a domain
• receiving a user query pertaining to the domain;
• extracting parts of speech (POS) data associated with tokens generated from the user query wherein the POS data includes a subject of the user query;
• using a term vector model scoring for determining a context from a set of contexts and relevant to the user query;
• providing the user query and the context determined as relevant to the user query to a machine comprehension (MC) model trained to identify an answer span responsive to the user query, wherein the MC model is trained on a plurality of documents specific to the domain and the plurality of documents include a section and a sub-section;
• receiving the answer span identified by the MC model from the relevant context as responsive to the user query;
• building a tree-structure using the subject from the user query and the answer span;
• generating a response to the user query by concatenating the subject and the answer span;
• and
• providing the response to the user as a reply to the user query via an output user interface.

A non-transitory processor-readable storage medium comprising machine-readable instructions that cause a processor to:

• receive a plurality of documents, wherein the plurality of documents include one or more sections and the one or more sections include one or more sub-sections and wherein the plurality of documents pertain to a specific domain;
• generate one or more candidate answers using key entities extracted from each of the sub-sections;
• automatically generate questions based on each of the candidate answers;
• train a machine comprehension (MC) model on at least the candidate answers and the questions, wherein the MC model is trained to identify answer spans to user queries from the plurality of documents and the MC model includes a 3-layer Question Answer (QA) architecture with a predetermined number of hidden neurons;
• receive a query from a user, the query pertaining to the specific domain;
• score a set of contexts using a term vector model scoring, wherein the scores indicate relevance of each of the set of contexts and the questions to the query from the user;
• identify a top-scoring context from the set of contexts;
• determine an answer span from the top-scoring context, the answer span being identified by the MC model;
• provide the answer span to the user as a reply to the query via an output user interface if the answer span forms a complete sentence,
• else provide a response generated from the answer span as the reply to the query.