unified pragmatic models for generating and following ...dfried/talks/fak...real-world navigation...

52
Unified Pragmatic Models for Generating and Following Instructions Daniel Fried, Jacob Andreas, and Dan Klein UC Berkeley 1

Upload: others

Post on 17-Feb-2021

1 views

Category:

Documents


0 download

TRANSCRIPT

  • Unified Pragmatic Models for Generating and Following Instructions

    Daniel Fried, Jacob Andreas, and Dan KleinUC Berkeley

    1

  • Interpreting instructions

    Instruction walk along the blue carpet and you pass two objects

    2

  • Interpreting instructions

    Instruction walk along the blue carpet and you pass two objects

    3

  • Interpreting instructions

    Instruction walk along the blue carpet and you pass two objects

    4

  • Interpreting instructions

    Instruction walk along the blue carpet and you pass two objects

    5

  • Interpreting instructions

    Instruction walk along the blue carpet and you pass two objects

    6

  • Listener: reasoning about routes

    7

    walk along the blue carpet and you pass

    two objects

    ?walk along the blue carpet and you pass

    two objects

  • Listener: reasoning about routes

    8

    walk along the blue carpet and you pass

    two objects

    walk along the blue carpet and you pass

    two objects

  • Listener: reasoning about routes

    9

    walk along the blue carpet and you pass

    two objects

    walk along the blue carpet and you pass

    two objects

  • Listener: reasoning about routes

    walk along the blue carpet and you pass

    two objects 10

    walk along the blue carpet and you pass

    two objects

  • Generating instructions

    11

    GeneratedInstruction:

    go forward four segments to the intersection with the bare concrete hall

  • Speaker: reasoning about interpretation

    12

    ?

  • Speaker: reasoning about interpretation

    13

    go forward four segments to the

    intersection with the bare concrete hall

  • Explicit pragmatic reasoning

    14

    Interpretation

    Frank and Goodman, 2012; Goodman and Stühlmuller, 2013;

    Wang et al., 2016

    Generation

    Golland et al., 2010; Monroe and Potts, 2015; Andreas and Klein, 2016;

    Mao et al., 2016; …

    “blue”

    Monroe et al. 2017Our work

    Both

    “The sun is in the sky”

  • Explicit pragmatic reasoning

    14

    Interpretation

    Frank and Goodman, 2012; Goodman and Stühlmuller, 2013;

    Wang et al., 2016

    Generation

    Golland et al., 2010; Monroe and Potts, 2015; Andreas and Klein, 2016;

    Mao et al., 2016; …

    “blue”

    Monroe et al. 2017Our work

    BothLarge, structured

    domains

    “The sun is in the sky”

  • Learned models of listeners and speakers

    BaseListener

    walk along the blue carpet and

    you pass two objects

    Both models trained independently on action sequences annotated with human instructions

    15

    BaseSpeaker

    walk along the blue carpet and

    you pass two objects

  • 16

    walk along the blue carpet and you pass

    two objects

    Building a pragmatic listener

    BaseSpeaker

    Pragmatic Listener

  • 16

    walk along the blue carpet and you pass

    two objects

    Building a pragmatic listener

    BaseSpeaker

    Pragmatic Listener

    ?walk along the blue carpet and you pass

    two objects

  • 16

    walk along the blue carpet and you pass

    two objects

    Building a pragmatic listener

    BaseSpeaker

    Pragmatic Listener

    proposes

    BaseListener ?

    walk along the blue carpet and you pass

    two objects

  • 16

    walk along the blue carpet and you pass

    two objects

    Building a pragmatic listener

    BaseSpeaker

    proposes

    BaseListener ?

    walk along the blue carpet and you pass

    two objects

  • 16

    walk along the blue carpet and you pass

    two objects

    Building a pragmatic listener

    BaseSpeaker

    proposes

    BaseListener

    walk along the blue carpet and you pass

    two objects

  • 16

    walk along the blue carpet and you pass

    two objects

    Building a pragmatic listener

    BaseSpeaker

    proposes

    BaseListener

    rescores

    walk along the blue carpet and you pass

    two objects

  • 16

    walk along the blue carpet and you pass

    two objects

    Building a pragmatic listener

    BaseSpeaker

    proposes

    BaseListener

    rescores

    0.1 walk along the blue carpet and you pass

    two objects

  • 16

    walk along the blue carpet and you pass

    two objects

    Building a pragmatic listener

    BaseSpeaker

    proposes

    BaseListener

    rescores

    0.1

    0.1 walk along the blue carpet and you pass

    two objects

  • 17

    walk along the blue carpet and you pass

    two objects

    Building a pragmatic listener

    BaseSpeaker

    proposes

    BaseListener

    rescores

    0.1

    walk along the blue carpet and you pass

    two objects

  • 17

    walk along the blue carpet and you pass

    two objects

    Building a pragmatic listener

    BaseSpeaker

    proposes

    BaseListener

    rescores

    0.1

    walk along the blue carpet and you pass

    two objects

  • 17

    walk along the blue carpet and you pass

    two objects

    Building a pragmatic listener

    BaseSpeaker

    proposes

    BaseListener

    rescores

    0.1

    0.2 walk along the blue carpet and you pass

    two objects

  • 17

    walk along the blue carpet and you pass

    two objects

    Building a pragmatic listener

    BaseSpeaker

    proposes

    BaseListener

    rescores

    0.1

    0.2

    0.2

    walk along the blue carpet and you pass

    two objects

  • 18

    walk along the blue carpet and you pass

    two objects

    Building a pragmatic listener

    BaseSpeaker

    proposes

    BaseListener

    rescores

    0.1

    0.2

    walk along the blue carpet and you pass

    two objects

  • 18

    walk along the blue carpet and you pass

    two objects

    Building a pragmatic listener

    BaseSpeaker

    proposes

    BaseListener

    rescores

    0.1

    0.2

    walk along the blue carpet and you pass

    two objects

  • Unified pragmatic inference

    Pragmatic Listener

    19

    BaseSpeaker

    BaseListener ?

    Pragmatic Speaker

    BaseSpeaker

    BaseListener?

    actions

    instructions

    walk along the blue carpet …

    walk along the blue carpet …

  • proposes

    Building a pragmatic speaker

    BaseSpeaker

    20

  • proposes

    Building a pragmatic speaker

    BaseSpeaker

    walk forward past the stool

    go forward four segments to the intersection with the bare concrete

    hall

    20

  • proposes rescores

    Building a pragmatic speaker

    BaseSpeaker

    BaseListener

    walk forward past the stool

    go forward four segments to the intersection with the bare concrete

    hall

    20

  • proposes rescores

    Building a pragmatic speaker

    BaseSpeaker

    BaseListener

    walk forward past the stool

    go forward four segments to the intersection with the bare concrete

    hall

    20

    0.4

  • proposes rescores

    Building a pragmatic speaker

    BaseSpeaker

    BaseListener

    walk forward past the stool

    go forward four segments to the intersection with the bare concrete

    hall

    0.4

    20

    0.4

  • proposes rescores

    Building a pragmatic speaker

    BaseSpeaker

    BaseListener

    walk forward past the stool

    go forward four segments to the intersection with the bare concrete

    hall

    0.4

    21

    0.8

  • proposes rescores

    Building a pragmatic speaker

    BaseSpeaker

    BaseListener

    walk forward past the stool

    go forward four segments to the intersection with the bare concrete

    hall

    0.4

    21

    0.8

    0.8

  • proposes rescores

    Building a pragmatic speaker

    BaseSpeaker

    BaseListener

    walk forward past the stool

    go forward four segments to the intersection with the bare concrete

    hall

    0.4

    22

    0.8

  • Base model implementations

    ……

    +

    LSTM Encoder

    BaseListener

    [Mei et al., 2016]

    walk along the blue carpet …

    ……

    LSTM Decoder with Attention

    23

  • Base model implementationsBase

    Speaker

    24

    +

    LSTM Encoder……

    LSTM Decoder with Attention……

    walk

    walk along the blue

    along the blue

  • Listener tasks

    Listenerwalk along the blue carpet and

    you pass two objects

    SAIL navigation [MacMahon et al., 2006; Chen and Mooney, 2011]

    Sequential Context-dependent Execution (SCONE) [Long et al. 2016]

    1. a red guy appears on the far left2. then to orange’s other side

    Listener

    25

  • Listener results, SAIL

    59.6

    64.4 65.3

    50

    60

    70Model accuracy at following human instructions

    Artzi & Zettlemoyer

    26

    Base listener

    Pragmatic listener

  • Instruction walk along the blue carpet and you pass two objects

    BaseListener ✗

    Pragmatic Listener ✔

    Listener example, SAIL

    27

  • Listener results, SCONE

    69.7 70.9 69.672 72.7

    69.6

    50

    60

    70

    80

    Alchemy Scene Tangrams

    Model accuracy at following human instructionsBase listener Pragmatic listener

    28

  • Listener example, SCONE

    a red guy appears on the far leftthen to orange’s other side

    Instruction

    Pragmatic Listener

    ✔Base

    Listener

    29

  • Speaker task and evaluation

    Speakerwalk along the

    wood path to the chair

    walk along the wood path to the

    chair

    Human directionfollowers (MTurk)

    Humans try to interpret it

    Speaker produces an instruction

    30

  • Speaker results

    62.8

    29.3 31.3

    60.0

    75.2 75.369.3

    88.0

    0

    25

    50

    75

    100

    SAIL Alchemy Scene Tangrams

    Human accuracy at following instructions from:

    31

    83.378.0

    66.073.2

    Base speaker Pragmatic speakerOther humans

  • Speaker example, SCONE

    BaseSpeaker

    remove the last figureadd it back

    Pragmatic Speaker

    remove the last figureadd it back in the 3rd position

    Humantake away the last itemundo the last step

    ✗32

  • Real-world navigation

    human description:walk through the kitchen. go right into the living room and stop by the rug.

    base speaker:walk past the dining room table and chairs and wait there .

    pragmatic speaker:walk past the dining room table and chairs and take a right into the living room . stop once you are on the rug .

    Matterport3D vision-and-language navigation dataset [Anderson et al., 2018]

    33

  • Conclusions

    Unified inference for sequential interpretation and generation

    Reasoning counterfactually, and about likely interpretations

    Pragmatics helps for complex tasks in structured domains

    34

    Speaker

    Listener

  • Thanks!http://github.com/dpfried/pragmatic-instructions

    35