Learning(to(InterpretNatural( Language(Commands(through ... fileCommanding(Robots(•...
Transcript of Learning(to(InterpretNatural( Language(Commands(through ... fileCommanding(Robots(•...
Learning to Interpret Natural Language Commands through
Human-‐Robot Dialog
Jesse Thomason Shiqi Zhang, Raymond Mooney, Peter Stone
The University of Texas at AusHn
1
Commanding Robots
• Autonomous robots in human environments
• Simplest to interact with via natural language
2
Our Task
• Command a robot operaHng in an office environment
• Robot autonomously wanders by default
• Robot can navigate to rooms and deliver items
3
System Goals
• Require liRle iniHal data – More domain independent
• Reason using composiHon – “Alice’s office”
• Robust to lexical variaHon – “bring”, “deliver”, “take”
• Execute the right acHon – Perform clarificaHons with user
4
Closest Previous Work
• Service robot that accepts commands (Kollar, 2013)
• SemanHcs match spans of words to known acHons/people/locaHons
• Can learn new referring expressions through dialog
• This system would explicitly match “Alice’s office” to room 3
5
Human Go to Alice’s office Robot Where is “Alice’s office”? Human Room 3
Closest Previous Work
• When system sees “Bob’s office”, will have to ask where that is
• Want to take advantage of composiHonality instead – Reason about possessive marker “’s” and what enHHes “office” picks out
• Need a more powerful formalism for represenHng sentence semanHcs – Want to keep iniHal training data light
6
Helpful Previous Work
• Augment a semanHc parser through conversaHon logs (Artzi, 2011)
• Key idea for us: use known system semanHc meanings to guess human uRerance word meanings
7
Human I would like to fly out of boston arriving to new york and back from new york to boston
System Leaving boston (CONFIRM:from(fl1,BOS)) on what date? (ASK:λx.departdate(fl1,x))
Tag Token Sequence
8
(S/PP)/NP : λx.λP.(ac7on(bring)∧pa7ent(bring,x)∧P(bring))
NP : coffee
PP/NP : λx.λy.(recipient(y,x))
bring
coffee
to
Bob NP : bob
syntax : seman7cs token(s)
University of Washington SemanHc Parsing Framework (SPF); (Artzi, 2011) Known possibiliHes for each token stored in a lexicon Use Combinatory Categorial Grammar (CCG)-‐driven parsing
Construct Meaning Hierarchically
9
(S/PP)/NP : λx.λP.(ac7on(bring)∧pa7ent(bring,x)∧P(bring)) NP : coffee PP/NP : λx.λy.(recipient(y,x))
bring coffee to Bob
NP : bob
S/PP : λP.(ac7on(bring)∧pa7ent(bring,coffee)∧P(bring)) PP : λy.(recipient(y,bob))
S : ac7on(bring)∧pa7ent(bring,coffee)∧recipient(bring,bob)
Tag Token Sequence – Missing Entry
10
(S/PP)/NP : λx.λP.(ac7on(bring)∧pa7ent(bring,x)∧P(bring))
?
PP/NP : λx.λy.(recipient(y,x))
bring
java
to
Bob NP : bob
Given semanHc form, can guess about missing token syntax/semanHcs
S : ac7on(bring)∧pa7ent(bring,coffee)∧recipient(bring,bob)
Human bring java to bob Robot what should I bring to bob? Human coffee
Tag Token Sequence – Missing Entry
11
(S/PP)/NP : λx.λP.(ac7on(bring)∧pa7ent(bring,x)∧P(bring))
?
PP/NP : λx.λy.(recipient(y,x))
bring
java
to
Bob NP : bob
:-‐ (S/PP)/NP : λx.λP.(ac7on(bring)∧pa7ent(bring,x)∧P(bring)) :-‐ (S/NP)/NP : λx. λy.(ac7on(bring)∧recipient(bring,x) ∧pa7ent(bring,y)) :-‐ NP : coffee :-‐ NP : bob
ac7on(bring)∧pa7ent(bring,coffee)∧ recipient(bring,bob)
bring :-‐ (S/PP)/NP : λx.λP.(ac7on(bring)∧pa7ent(bring,x)∧P(bring)) bring :-‐ (S/NP)/NP : λx.λy.(ac7on(bring)∧recipient(bring,x) ∧pa7ent(bring,y)) coffee :-‐ NP : coffee Bob :-‐ NP : bob
Given form:
Lexicon entries that produce parts of this form:
Candidates for ‘java’ lexical entry:
Tag Token Sequence – Missing Entry
(S/PP)/NP : λx.λP.(ac7on(bring)∧pa7ent(bring,x)∧P(bring))
NP : coffee
PP/NP : λx.λy.(recipient(y,x))
bring
java
to
Bob NP : bob
With new lexicon entry, we can construct the correct semanHc form
S : ac7on(bring)∧pa7ent(bring,coffee)∧recipient(bring,bob)
MeeHng System Goals
• Require liRle iniHal data – Bootstrap parser with 5 expressions, 105 words
• Handle composiHon used by speakers – Use CCG-‐driven semanHc parsing (Artzi, 2011)
• Robust to lexical variaHon – Incrementally train parser to obtain new words
• Execute the right acHon – Use dialog to clarify meanings with user (Kollar, 2013)
13
Mechanical Turk Experiment
• Users given one navigaHon and one delivery goal – Train/test goals chosen at random from possibiliHes
• Chat with robot’s dialog agent unHl goal is understood
14
Mechanical Turk Interface
15
Large-‐Scale Experiment
• Tested in 4 phases • ~50 users received test goals, ~50 train goals – Unique users in each phase
• System incrementally trained via train goal conversaHons only
16
Mechanical Turk Dialog Turns
17
2.7 2.5 3.4 3.5
18
5.6 5.2 3.8
0 2 4 6 8 10 12 14 16 18 20
Batch 0 Batch 1 Batch 2 Batch 3
Average Tu
rker Turns fo
r Success
NavigaHon task turns Delivery task turns
*
Mechanical Turk Survey Responses
18
2.2
2.6 2.7 2.8
2.2 2 1.9
1.5
0
0.5
1
1.5
2
2.5
3
Batch 0 Batch 1 Batch 2 Batch 3
Average
The robot understood me The robot frustrated me
Strongly Agree
Somewhat Agree
Somewhat Disagree
Strongly Disagree
* *
*
Robot Experiment
• Same setup, but real robot and fewer users – Users type to robot to mimic Mechanical Turk setup
• 10 users in iniHal test batch • System interacted freely with people on the floor for four days as training (34 conversaHons in total)
• 10 users in the second test batch, arer retraining
19
Office Robot Dialog CompleHon
20
90 90
20
60
0 10 20 30 40 50 60 70 80 90
100
Batch 0 Batch 1
Percen
t of U
sers Com
pleF
ng Task
NavigaHon task compleHon Delivery task compleHon
+
Office Robot Survey Responses
21
1.6
2.9
2.5
1.5
0
0.5
1
1.5
2
2.5
3
3.5
Batch 0 Batch 1
Average
The robot understood me The robot frustrated me
Strongly Agree
Somewhat Agree
Somewhat Disagree
Strongly Disagree
*
+
Conclusions
• Lexical acquisiHon reduces dialog lengths for mulH-‐argument predicates like delivery
• Causes users to perceive the system as more understanding
• Leads to less user frustraHon • Allows improving language understanding without large, annotated corpora
22
Learning to Interpret Natural Language Commands through
Human-‐Robot Dialog
Jesse Thomason Shiqi Zhang, Raymond Mooney, Peter Stone
The University of Texas at AusHn
23
Related Work
• Command processing has taken many forms • Specify tasks step-‐by-‐step (Meriçli, 2014) – Assumes parHcular words in parHcular order
• Specify low-‐level acHon sequences (Misra, 2014; Tellex, 2011) – Uses a parser trained on a huge corpus
• Map language to acHon specificaHons (Matuszek, 2013) – Cannot learn new words/expressions
24
Future Work
• Perceptual grounding (`blue’,`ler of’) • Predicate invenHon (`ruddy’) • Learning a mulH-‐objecHve dialog policy that trades off learning and user saHsfacHon
25