公开(公告)号：US20240428787A1

公开(公告)日：2024-12-26

申请号：US18340342

申请日：2023-06-23

Applicant: Amazon Technologies, Inc.

Inventor： Mahdi Namazifar ,  Di Jin ,  Yang Liu ,  Devamanyu Hazarika ,  Dilek Hakkani-Tur ,  Yubin Ge

IPC: G10L15/22 ,  G06F40/295 ,  G10L15/183

Abstract: Techniques for constraining the results of a generative language model to valid information using knowledge-grounded documentation. A generative language model may generate invalid results, including compound entities and incorrect entity relations. The techniques include, for a given user inquiry, determining a set of documented information, from a particular knowledge base, that corresponds to the user inquiry. The techniques further include determining a subgraph from a knowledge graph representing the knowledge base, as well as determining a trie data structure representation of the set of documented information. The user inquiry and subgraph are provided as input to a trained generative language model for generating a response to the user inquiry. The techniques include using the trie data structure to validate that the generated response corresponds to real information from the set of documented information.