Transcript From-Loc - Computer Science

Tense and Implicit Role Reference
Joel Tetreault
University of Rochester
Department of Computer Science
Implicit Role Reference


Verb phrases have certain required roles –
NP’s that are expected
For example: “take”:






Something to take (theme)
A place to take it from (from-loc)
A place to take it to (to-loc)
Something to do the taking (agent)
Possibly a tool to do the taking (instrument)
Very little work has been has been done
(Poesio, 1994; Asher and Lascarides, 1998)
Goal





Resolving IRR’s important to NLP
To investigate how implicit roles work
Develop an algorithm for resolving them
Evaluation of algorithm for empirical
results
Use temporal information and discourse
relations to improve results
Outline





Implicit Roles
Annotation
Algorithm
Results
Discussion
Example
(1) Take engine E1 from Avon to Dansville
(2a) Pick up the boxcar and take it to Broxburn
[from ?]
(2b) And then take the boxcar from Corning
[to ?]
(3a) Leave E1 there but move the boxcar down
the road to Evansville. [from ?]
(3b) Leave the boxcar there.
Statistics: Role Distributions
Role
Total
Explicit
Implicit
Instrument
25
6
19
Theme
40
30
10
From-Loc
39
13
26
To-Loc
27
20
7
Total
131
69
62
Corpus



Annotated a subset of the TRAINS-93
Corpus (Heeman and Allen, 1994)
86 utterance task-oriented dialog
between two humans
Task: move commodities around in a
virtual world
Annotation




Used sgml style annotation scheme
NP’s annotated with ID and its class
(engine, tanker, location, food)
VP’s annotated with ID, event time, and
roles
Roles for each verb are taken from
TRIPS lexicon (Allen et al., 2000)
Temporal Annotation


An event time was assigned each
utterance, such as: t0, t1, u1, etc.
And constraints upon the event time are
imposed:



t9>t1 (t9 comes after t1)
t9<t1 (t9 precedes t1)
t9>t1 & t9<t10 (t9 comes after t1 and
before t10)
Sample Annotation




U1: Take Engine E1 from Avon to Dansville.
U2: Pick up the boxcar
<ve id=ve122 time=t0 theme=ne12 fromloc=ne5 to-loc=ne6>Take <ne
id=ne12>engine E1</ne> from <ne
id=ne5>Avon</ne> to <ne
id=ne6>Dansville</ne></ve>.
<ve id=ve123 time=t1>t0 theme=ne13
from-loc=ne6i>Pick up<ne id=ne13>the
boxcar</ne></ve>.
Statistics





Most implicit roles have antecedents found
locally (0-2 sentences back over 90% of the
time)
Instrument: 79% Instr, 10% theme, 10% ID
Theme: 88% Theme, 12% %ID
From-Loc: 62% From-Loc, 38% To-Loc
To-Loc: 57% To-Loc, 29% From-Loc, 14% Theme
Algorithm

For each utterance u, process u left to
right:


If NP is encountered, push it on
appropriate focus stack
If VP is encountered:


place all explicit roles on top of appropriate
focus stack
If role is implicit….
Algorithm Example
U1: Take engine E1 from Avon to Dansville
Engine E1
Theme
Avon
Dansville
From-Loc To-Loc
[empty]
Instrument
U2: Also take the boxcar
boxcar
Avon
Dansville
Theme From-Loc To-Loc
[empty]
Instrument
Implicit Role Algorithm

Type determines method. If role is:



Instrument: search through current utterance first
for an entity that meets verb’s constraints, else go
back through past utterance’s instrument and
theme focus lists
Theme: same as above except search theme
before instrument for each past utterance
From/To-Loc: use temporal reasoning to
determine what order to search past To-Loc and
From-Loc lists for each utterance:
Temporal Algorithm



For two utterances uk and uj,with k > j,
determine rel(uk, uj):
If time(uk) > time(uj) then rel(uk, uj) =
narrative
Else rel(uk, uj) = parallel
Experiment



Developed LISP system that automates the
algorithm
For each marked implicit role, system tries to
find an antecedent
Notations:



R-L – each focus list searched right to left (order
of recency)
L-R – search is left-to-right (sentence order)
Time – algorithm augmented with temporal
algorithm
Results
Algorithm
Instrument Theme From-Loc
To-Loc
Overall
Baseline
42.1%
11.1%
88.5%
44.5%
57.1%
R-L
78.9%
55.6%
65.4%
22.2%
61.9%
L-R
78.9%
44.4%
88.5%
44.5%
73.0%
Time, L-R
78.9%
55.6%
61.5%
55.6%
65.1%
Time, R-L
78.9%
44.5%
69.3%
55.6%
66.7%
Total
19
9
26
9
63
Discussion




From-Loc’s – naive version is better
To-Loc – any strategy better than naïve
Top pronoun resolution algorithms perform
around 70-80% accuracy
Problems:




Corpus size – hard to make concrete conclusions
or find trends
Annotation scheme is basic
Need to handle ‘return verbs’ properly
Augment system to identify whether implicit roles
should be resolved or not (ignore general cases)
Current Work


Building a larger corpus that can be
annotated automatically using the
TRIPS parser
Domain is much more varied and has
different types of verbs