公开(公告)号：US11625644B1

公开(公告)日：2023-04-11

申请号：US16793233

申请日：2020-02-18

Applicant: Amazon Technologies, Inc.

Inventor： Elad Haramaty ,  Liane Lewin-Eytan ,  David Carmel ,  Arnon Lazerson

IPC: G06N20/00 ,  G06F16/2457 ,  G06N7/01

Abstract: Devices and techniques are generally described for ranking of search results based on multiple objectives. In various examples, a first set of search results may be determined. A first objective and a second objective for ranking the first set of search results may be determined. A first label associated with the first objective may be selected for a first training data instance. A second label associated with the second objective may be selected for a second training data instance. A first machine learning model may be generated using the first training data instance and the second training data instance. In some examples, the first machine learning model may be effective to rank the first set of search results based at least in part on the first objective and the second objective.

公开(公告)号：US11262978B1

公开(公告)日：2022-03-01

申请号：US16446080

申请日：2019-06-19

Applicant: Amazon Technologies, Inc.

Inventor： Nachshon Cohen ,  David Carmel ,  Simone Filice ,  Elad Kravi ,  Liane Lewin-Eytan ,  Vitaly Mirkis ,  Sofia Tolmach

IPC: G06F3/0482 ,  G06F3/16 ,  G06F16/907 ,  G06F16/9038 ,  G06N3/08 ,  G06F16/9032 ,  G06F16/903 ,  G06N20/10 ,  G06F16/9532

Abstract: Techniques are disclosed. The techniques include receiving, by a computing device, an answer to a query, the answer comprising a content element and a metadata element. Based on the content element and the metadata element, the techniques generate a presentable sentence of a plurality of presentable sentences. The techniques then utilize a machine learning model to determine a relevance score for the presentable sentence based on the query and the presentable sentence, the relevance score being of a set of relevance scores and corresponding to a measure of relevance with respect to the presentable sentence answering the query. The techniques then select a portion of the plurality of presentable sentences based on a ranking of the set of relevance scores and transmit the portion to a user device for presentation on a voice-based interface of the user device.