A Machine Reading Comprehension (MRC) system that finds answers for open-domain natual language questions.
Additional Angles to Viewing Questions
Tabris has been created to support Magi's revolutionary feature of directly providing answers to users' natural language queries. Unlike optimizing for a typical SQuAD task, the model accentuates two types of results - passages that are relevant to and complete in addressing the question, and phrases that are suitable in logic and succinct in wording. Therefore, Tabris supports both subjective and objective questions better.
Millisecond-Level End-to-End Reasoning
The network structure of Tabris makes full use of the hardware's parallel computing ability. With vectorized in-graph tokenization and text span extraction, Tabris is capable of providing millisecond-level end-to-end reasoning on common hardwares, accomodating the performance requirement of real-time online queries.