14th july 2010 uppsala, sweden
DESCRIPTION
Ryu Iida Shumpei Kobayashi Takenobu Tokunaga Tokyo Institute of Technology { ryu-i,skobayashi,take }@ cl.cs.titech.ac.jp. Incorporating Extra-linguistic Information into Reference Resolution in Collaborative Task Dialogue. 14th July 2010 Uppsala, Sweden. ACL 2010. - PowerPoint PPT PresentationTRANSCRIPT
Incorporating Extra-linguistic Information into Reference Resolution in Collaborative Task Dialogue
Ryu Iida Shumpei Kobayashi
Takenobu Tokunaga
Tokyo Institute of Technology{ryu-i,skobayashi,take}@cl.cs.titech.ac.jp
ACL 2010
1
14th July 2010 Uppsala, Sweden
Research background The task of identifying reference relations including
anaphora and coreference within texts has received a great deal of attention in NLP
Research trends for reference resolution have drastically shifted from hand-crafted rule-based approaches to corpus-based approaches Many researchers have examined ways for introducing
various linguistic clues(Ge et al. 1998, Soon et al. 2001, Ng and Cardie 2002, Yang et al. 2003, 2005, Poon and Domingos, 2008, etc.)
2
Typical problem setting of reference resolution
Annotated data sets provided by Message Understanding Conference (MUC) and Automatic Content Extraction (ACE) Limited version of coreference;
relations where expressions refer to named entities More information extraction-oriented
Coreference task as defined by MUC and ACE is geared toward only identifying coreference relations anchored to an entity within the text
3
Treatment of referential behavior in language generation community
Investigations of referential behaviour in real world situations (Di Eugenio et al. 2000, Byron 2005, van Deemter 2007, Foster 2008, Spanger et al. 2009) applications: e.g. human-robot interaction
Spanger et al. (2009): dialogues of two participants collaboratively solving Tangram puzzle Corpus includes extra-linguistic information synchronised
with utterances (e.g. operations on the puzzle pieces) They revealed that multi-modal perspective of reference is
needed for more practical reference understanding
4
Challenging issue Create a model bridging a referring expression in text
and its object in real world
Focus on incorporating extra-linguistic information into existing corpus-based approach Target corpus: Spanger et al. (2009)’s REX-J corpus
5
Table of contents Research background Collaborative work dialogue corpus: REX-J corpus Reference resolution model and use of extra-linguistic
information Empirical evaluation Summary and future work
6
REX-J corpus (Spanger et al. 2009)
Collaborative work dialogues in Japanese for solving Tangram puzzle Operations to solve the puzzle and situations updated by a
series of operations are recorded by a puzzle simulator on computer
Relationship between referring expressions and their referents on a computer display is manually annotated
7
8
Screenshot of Tangram simulatorGoal shape
area
Working area
3 operations on puzzle pieces:move, rotate,flip
Positions of every piece and every action are recorded at intervals of 10 msec
9
Experimental environment Share only working area and linguistic information in
dialogue Two different roles: “solver” and “operator”
operatorsolver
can see a certain goal shape
cannot manipulate pieces
cannot see the goal shape
can manipulate pieces
REX-J Corpus: statistics Recruited 12 Japanese graduate students
6 pairs * 4 different goal shapes 24 dialogues
10
Table of contents Research background REX-J corpus Reference resolution model and use of extra-linguistic
information Empirical evaluation Summary and future work
11
123
456 7
Task definition12
…A : move it more to the right.B : which triangle? Is this?
no antecedent in preceding utterances
Time piece operation...12:01:03 1 rotate12:01:05 3 move12:01:10 6 move12:01:12 6 rotate
referent of ‘it’: piece 6
Operation history
utterances
Task: select a piece out of a fixed set of pieces given a referring expression by referring to both preceding utterances and series of the recent operations
Ranking model to identify referents
Machine learning-based approaches (Soon et al. 2001, Ng and Cardie 2002, etc.) Take into account linguistic factors: relative salience Ranking candidate antecedents in preceding discourse
(Iida et al. 2003, Yang et al. 2003, Denis and Baldridge 2008) Denis and Baldridge (2008) reported appropriately constructing a
model for ranking all candidates achieved better performance than pairwise ranking.
Adopt a ranking-based model in which all candidates compete with one another Use ranking SVM instead of Maximum Entropy
13
Extra-linguistic information (1/2):history of mouse movement
Current position of mouse cursor and history of mouse movements
Represent the temporal salience of participant’s focus of attention and its transition
mouse cursor
12 3
45
67
14
Extra-linguistic information (1/2):Action history feature
mouse cursor was over a piece (i.e. a candidate referent) at the beginning of uttering a RE
a piece is the last piece that mouse cursor was over time distance after mouse cursor was over a piece:
x <10 sec / 10 sec ≤ x < 20 sec / 20 sec ≤ x mouse cursor is never over a piece in the preceding
utterances
15
Extra-linguistic information (2/2):history of series of operations
Recently manipulated pieces tend to be paid more attention than the other pieces
12 3
45
67
Time piece operation...12:01:03 1 rotate12:01:05 3 move12:01:10 2 move12:01:12 2 rotate
Operation history
16
Extra-linguistic information (2/2):Current operation feature
a piece is being manipulated at the beginning of utteringa RE
a piece is the most recently manipulated piece time distance after a piece was most recently
manipulated: x <10 sec / 10 sec ≤ x < 20 sec / 20 sec ≤ x
a piece has never been manipulated
17
Table of contents Research background REX-J corpus Reference resolution model and use of extra-linguistic
information Empirical evaluation Summary and future work
18
19
Empirical evaluation Investigate the impact of the extra-linguistic information
Data set: referring expressions in REX-J corpus (2,048 referring expressions in 40 dialogues) 13 expressions are excluded
Expressions referring to more than one object Vague expressions
E.g. “biggest triangle” in the situation where there are two biggest triangles on the display
2,035 expressions are used on 10-fold cross-validation
20
Two models Pronouns are likely to be more directly associated with actions
pointing to a piece Denis and Baldridge (2008)
the size of training instances is relatively small, the models induced by learning algorithms should be separately
created with regards to distinct features Separated model
Create two rankers; learn pronouns and non-pronouns independently Pronoun model: use the training instances whose REs are pronouns Non-pronoun model: use all other training instances
Combined model Create one ranker; induced from all training instances
Features21
3 types of features Action history features Current operation features Discourse history features
Acquired from the expressions of a given referring expression and its candidate antecedent in the preceding utterances
e.g. a piece is referred to by the most recent RE case makers (o (accusative) or ni (dative)) follow RE
Baseline model: use only discourse history features
Resultsmodel discourse
history(baseline)
+action history
+current operation
+action history, +current operation
separated model
0.664(1352/2035)
0.790(1608/2035)
0.685(1394/2035)
0.780(1587/2035)
a) pronoun model
0.648(660/1018)
0.886(902/1018)
0.692(704/1018)
0.875(891/1018)
b) non-pronoun model
0.680(692/1017)
0.694(706/1017)
0.678(690/1017)
0.684(696/1017)
combined model
0.664(1352/2035)
0.749(1524/2035)
0.650(1322/2035)
0.743(1513/2035)
22
Resultsmodel discourse
history(baseline)
+action history
+current operation
+action history, +current operation
separated model
0.664(1352/2035)
0.790(1608/2035)
0.685(1394/2035)
0.780(1587/2035)
a) pronoun model
0.648(660/1018)
0.886(902/1018)
0.692(704/1018)
0.875(891/1018)
b) non-pronoun model
0.680(692/1017)
0.694(706/1017)
0.678(690/1017)
0.684(696/1017)
combined model
0.664(1352/2035)
0.749(1524/2035)
0.650(1322/2035)
0.743(1513/2035)
24
0.227
0.004
Pronouns are more sensitive to the usage of the action history features
Resultsmodel discourse
history(baseline)
+action history
+current operation
+action history, +current operation
separated model
0.664(1352/2035)
0.790(1608/2035)
0.685(1394/2035)
0.780(1587/2035)
a) pronoun model
0.648(660/1018)
0.886(902/1018)
0.692(704/1018)
0.875(891/1018)
b) non-pronoun model
0.680(692/1017)
0.694(706/1017)
0.678(690/1017)
0.684(696/1017)
combined model
0.664(1352/2035)
0.749(1524/2035)
0.650(1322/2035)
0.743(1513/2035)
feature name
feature type Description
AH1 action history mouse cursor was over a piece at the beginning of uttering a RE
CO1 current operation
a piece is being manipulated at the beginning of uttering a RE
25
Partially overlapped
Other current operation features may have bad effects for ranking referents due to their ill-formed definitions
27
Summary and future directions
[Summary] We demonstrated our first result of incorporating extra-linguistic
clues into a corpus-based approach to reference resolution The performance increased by at most 12 points in comparison
to the baseline model. extra-linguistic information in this domain are useful
[Future work] Explore the effect of other extra-linguistic information
e.g. eye-gaze information Investigate general aspect between REs and their objects;
Further evaluation based on the different multimodal tasks