Samira Shaikh
, Tomek Strzalkowski
and Nick Webb
State University of New York, University at Albany, New York, U.S.A.
Polish Academy of Sciences, Warsaw,Poland
Keywords: Dialogue-Act Tagging, Urdu, Classification, Multi-Party Discourse.
Abstract: Classification of dialogue acts constitutes an integral part of various natural language processing
applications. In this paper, we present an application of this task to Urdu language online multi-party
discourse. With language specific modifications to established techniques such as permutation of word
order in detected n-grams and variation of n-gram location, we developed an approach that is novel to this
language. Preliminary performance results when compared to baseline are very encouraging for this
Urdu, belonging to the family of Indo-European
languages, has an estimated 487 million speakers
worldwide, next in line after English. In our context,
Urdu is the modern variant, more generally referred
to as Hindi/Urdu. The need arises for the
development of robust natural language tools aimed
at understanding and investigation of the language.
Social interaction in an increasingly online world
also provides a rich resource for research. The
dynamics of small group interaction have been well
studied for spoken and face-to-face conversation.
However, for a reduced-cue environment such as
online chat in a virtual chat room, these dynamics
are obtained distinctly, and require explicit linguistic
devices to convey social and cultural nuances.
Indeed, how are social behaviours exhibited and
conveyed when the only medium of communication
is language?
Our overall objective is to develop computational
models of how certain social phenomena are
manifested in language through the choice of
linguistic, semantic and conversational forms by
discourse participants. The social behaviors we are
currently studying include, among others, Topic
Control, Task Control, Disagreement and
Involvement. These are, in turn, utilized to predict
higher-level social phenomena such as leadership
and group cohesion. Dialogue act tagging forms an
essential component of our prototype system. Using
dialogue acts to model the functional aspect of an
utterance in discourse, we can arrive at
determinations of socio-linguistic behaviors by the
participants. For example, we posit that an equal
amount of agreement and disagreement between all
participants of a discourse, points towards a more
cohesive group; as opposed to a discourse
characterized by an inordinate amount of
disagreement or agreement. It is essential that the
data corpus used contain the discourse features we
are interested in modeling, which led us to collect
our own data. Another requirement was that the
discourse participants be native speakers of the
target language, so that natural and spontaneous
discourse may be obtained. We also developed a
hierarchy of dialogue acts that are tuned
significantly towards dialogue pragmatics and
eschew syntactic variations.
This paper pertains to Urdu online chat
conversations; we selected Roman Urdu, as this is
the preferred form of writing used in most Urdu chat
rooms. We use a cue-phrase based method, using n-
grams as features and enhance it by adding a word
order alteration feature specifically targeting the
Urdu grammar structure. Classification of dialogue
acts in the Urdu language is a novel task that has not
been hitherto addressed. While our approach is
preliminary, we are quite encouraged by the
Shaikh S., Strzalkowski T. and Webb N..
DOI: 10.5220/0003637303980404
In Proceedings of the International Conference on Knowledge Discovery and Information Retrieval (KDIR-2011), pages 398-404
ISBN: 978-989-8425-79-9
2011 SCITEPRESS (Science and Technology Publications, Lda.)
We model our cue-based approach in line with the
work by researchers in the field. Fraser (1990)
showed that discourse markers are “part of the
grammar of a language”. Grosz and Sidner (1986)
proposed ways in which discourse may be
segmented into constituent structures. Several
researchers (Heeman et al., 1998; Hirschberg and
Litman, 1993; Marcu, 1997; Reichman, 1985;
Schirin, 1987; Warner, 1985; Zukerman and Pearl,
1986) identified and selected cue phrases in dialogue
that are generally useful for dialogue processing.
Samuel, Carberry and Vijay-Shanker (1999) used n-
grams in utterances to automatically detect cue
phrases. Webb, Hepple and Wilks (2005) had a
similar approach of identifying cue phrases using
intra-utterance features and determining the n most
likely dialogue acts for an utterance. All of these
works point to learning features that are specific to
the language under consideration. A
morphologically rich language such as Urdu requires
techniques that both exploit and overcome its
Work in dialogue act classification in languages
such as Urdu is still nascent. In fact, we have been
able to discover no prior work towards dialogue act
classification in Urdu discourse. Somewhat related is
the use of n-grams in Urdu for authorship attribution
in Urdu poetry (Raza et al., 2009a). Word
segmentation in Urdu is an issue that affects
machine-learning algorithms (Durrani and Hussain,
2010). However, the use of Roman Urdu in our
corpus mitigates this issue.
Our initial focus has been on on-line chat dialogues.
Chat data, although plentiful on-line, its adaptation
for research purposes present a number of
challenges. On the one hand there are users’ privacy
issues, and their complete anonymity on the other.
Furthermore, most data that may be obtained from
public chat-rooms is of limited value for the type of
modeling tasks we are interested in due to its high-
level of noise, lack of focus, and rapidly shifting,
chaotic nature, which makes any longitudinal studies
virtually impossible. To derive complex models of
conversational behavior, we needed the interaction
to be reasonably focused on a task and/or social
objectives within a group.
Few data collections exist covering multiparty
dialogue, and even fewer with on-line chat.
Moreover, the few collections that exist were built
primarily for the purpose of training dialogue act
tagging and similar linguistic phenomena; few if any
of these corpora are suitable for deriving pragmatic
models of conversation, including socio-linguistic
phenomena. Existing resources include a multi-
person meeting corpus ICSI-MRDA (Janin et al.,
2003) and the AMI Meeting Corpus (Carletta, 2007),
which contains 100 hours of meetings captured
using synchronized recording devices. Still, all of
these resources look at spoken language rather than
on-line chat. Some corpora exist such as the NPS
Internet chat corpus (Forsyth and Martell, 2007),
which has been hand-anonymized and labeled with
part-of-speech tags and dialogue act labels. The
StrikeCom corpus (Twitchell et al., 2007) consists of
32 multi-person chat dialogues between players of a
strategic game, where in 50% of the dialogues one
participant has been asked to behave ‘deceptively’.
These are resources in the English language; some
of the corpora that exist in Urdu are aimed towards
tasks such as part of speech tagging and lexicon
building (Hussain, 2008; Raza et al., 2009b; Ijaz and
Hussain, 2007). Few, if any of these corpora are
suitable for deriving pragmatic models of
conversation, including socio-linguistic phenomena.
It is thus more typical that those interested in the
study of Internet chat compile their own corpus on
an as needed basis, e.g., Khan et al. (2002), Kim et
al. (2007).
We designed a series of experiments in which
recruited subjects were invited to participate in a
series of on-line chat sessions in a specially designed
secure chat-room. The experiments were carefully
designed around topics, tasks, and games for the
participants to engage in so that appropriate types of
behavior, e.g., disagreement, power play,
persuasion, etc. may emerge spontaneously.
Discussions were centered on a range of topics that
included issues relevant to native speakers of Urdu,
such as the “Value of the Burka in Modern-Day
Women” and “Politics of Pakistan under Prime
Minister Zardari” as well as task-oriented topics
such as choosing the best candidate for a given job
from an array of resumes. These experiments and the
resulting corpus have been described in a separate
publication. We assembled a corpus of 20 hours of
Urdu chat, consisting of 40,000+ words, 6000+ turns
and 9 different participants, over the course of
fourteen 90-minute chat sessions.
Figure 1 shows a fragment of one Urdu dialogue,
where 6 participants in the chat session discuss the
selection of a candidate for given job description.
Note the use of short sentences, lack of punctuation
and capitalization and typically improper grammar.
In addition, emoticons (e.g. , ), misspellings, and
abbreviations are also common.
Figure 1: A fragment of Urdu conversation.
We have annotated 5 Urdu dialogues (2000+
turns) in total which were used for training our Urdu
modules. There is need of more training data, and
we intend to collect and annotate more Urdu
dialogues. All annotation was done using a specially
designed annotation tool for the purpose, by two
trained annotators who are native speakers of the
language. The inter-annotator agreement for
dialogue acts is 0.82 alpha (Krippendorf, 2005),
which is adequate for the training we need to
The functional or dialogic aspect of an utterance has
to do with its role or purpose in conversation.
Statements, questions, answers, offers, acceptances
and rejections, as well as expressions of thanks are
all examples of such functions in a dialogue. Our
objective is to capture how an utterance functions in
dialogue, which may or may not be directly related
to its form. For example, the utterance “Can you
close the window?” may function as a question or as
a directive, depending upon the context in which it is
We developed a hierarchy of 15 dialogue acts
adapted to better capture significant social nuances
within conversation. Syntactic distinctions between
categories, e.g., wh-questions vs. yes-no questions,
etc are avoided. The tagset we adopted is based on
DAMSL (Allen and Core, 1997) and SWBD
(Jurafsky, Shriberg and Biasca, 1997), but
compressed to 15 tags tuned significantly towards
dialogue pragmatics and away from more surface
characteristics of utterances.
Our classification of dialogue acts is bi-fold
(Figure 2). At the Information-Level (Figure 3), we
seek to identify the purpose of an utterance in
relation to the task given to the participants.
Figure 2: Dialogue act levels.
Figure 3: Information-Level dialogue act categories.
At the Functional-Level, we classify Dialogue Acts
into three hierarchical categories (see Table 1
below): (a) Statements-and-Responses, (b)
Questions-and-Directives, and (c) Conversational-
Norms. Each of these categories consists of several
top-level tags and may also contain specialized tags
under these. This makes a total of 15 dialogue acts at
the Functional-Level. In addition, there are 3
dialogue acts at the Information-Level.
The Assertion-Opinion category contains four
specialized tags under it, A.1.1 Response-Answer,
A.1.2 Response-Non-Answer, A.1.3 Agree-Accept
and A.1.4 Disagree-Reject. For an utterance, a
specialized tag is preferably applied wherever
pertinent. For example, the utterance “mein aapse
sehmat hu us baat par” (I agree with you on that)
functions as an assertion, as well as an agreement;
and is assgined the tag Agree-Accept rather than
Assertion-Opinion. A full description of these
dialogue acts is beyond the scope of this paper, and
will be the subject of a future publication. It is
important to note that the annotation and categories
have been developed to support the objectives of our
project and do not necessarily conform to other
similar classification systems used in the past.
Each utterance in dialogue is assigned two
dialogue acts, one at the Information-Level and one
at the Functional-Level. Figure 4 shows the
annotation applied to the conversation fragment in
Figure 1.
5. MM: mujhe carla ka resume pasand hai
(MM: I like resume of carla )
6. MM: experience ke hisab se
(MM: based on experience)
7. RI: mujhe bhi
(RI: me too)
8. SA: ji carla ka tek hai
(SA: yes, (resume) of carla is fine
KDIR 2011 - International Conference on Knowledge Discovery and Information Retrieval
Table 1: Functional-Level dialogue act categories.
Category Top Level
A. Statements and
A.1 Assertion-Opinion
-A.1.1 Response-Answer
-A.1.2 Response-Non-Answer
-A.1.3 Agree-Accept
-A.1.4 Disagree-Reject
A.2 Offer-Commit
A.3 Acknowledge
B. Questions and
B.1 Information-Request
B.2 Confirmation-Request
B.3 Action-Directive
C. Conversational
C.1 Conventional-Opening
C.2 Conventional-Closing
C.3 Other-Conventional-Phrase
C.4 Correct-Misspelling
Figure 4: Annotation of conversation fragment showing
Information-Level and Functional-Level tags.
We use annotated data to learn cue phrases in each
dialogue act category. We generate n-grams of
varying length from the utterances, discarding stop
words, emoticons (e.g. , etc) and some
prepositions. This process generated ~11500 n-
grams. The n-grams of length no greater than 3 are
saved and ranked in order of their frequency and
length. The threshold has been determined
experimentally and varies with the dialogue act
under consideration. Some dialogue acts appear
more frequently in data and generate higher number
of n-grams, than those that are infrequent. Frequency
and n-gram length are generally inversely
proportional to each other. However, n-grams of
length greater or equal to 2 are preferable, due to
their high accuracy and predictive nature and thus
we use a lower threshold for longer length n-grams.
We use the most frequent n-grams that appear in
utterances tagged with a specific dialogue act and
the most predictive ones. Frequency values fluctuate
significantly. For example, for the Conventional-
Opening category of dialogue act in Conversational-
Norms, the n-grams are stable and highly predictive.
To give an idea about the spread of frequency, the
most frequent Information-Level dialogue act tag
assigned in our annotated corpus is Task. The
frequency of the most frequent n-gram learned for
this tag is 392. The least applied dialogue act tag in
our corpus is Signal-Non-Understanding (at the
Functional-Level). The frequency of the most
frequent n-gram for this tag is 2. Note the frequency
distribution of learned n-grams shown in Figure 5,
which follows Zipf’s law (1949) with a long tail of
the curve. To get the best performance, we select the
most frequent n-grams from the head of the curve,
and the highly predictive (i.e. greater length) yet less
frequent n-grams from the tail of the curve. Very
low frequency unigrams are not selected.
Figure 5: Frequency of learned n-grams.
We are currently using absolute frequency counts
of n-grams in our determinations, we may replace
them with normalized counts or percentages in the
next prototype.
Some dialogue act classification systems (Stolcke
et al., 2000; Samuel, Carberry and Vijay-Shanker,
1999), place <start> and <end> tags, to determine
the position in utterance where the n-gram should
occur. While this provides a salient handle over the
utterance in the English language, Urdu grammar is
not restrictive on word order, and using such a
mechanism presents a challenge.
To illustrate, consider the sentences below:
1. mein aapse sehmat hu us baat par
(I agree with you on that)
2. us baat par mein aapse sehmat hu
n-grams -->
5. MM: mujhe carla ka resume pasand hai
(MM: I like resume of carla
DA Info-Level: Task; Func-Level: Assertion-Opinion
6. MM: experience ke hisab se
(MM: based on experience)
DA Info-Level: Task; Func-Level: Assertion-Opinion
7. RI: mujhe bhi
(RI: me too)
DA Info-Level: Task; Func-Level: Agree-Accept
8. SA: ji carla ka tek hai
(SA: yes, (resume) of carla is fine
DA Info-Level: Task; Func-Level: Agree-Accept
(On that I agree with you)
Both are valid utterances in Urdu grammar,
wherein the tri-gram “mein aapse sehmat” (I agree
with you) occurs in different positions. This non-
restrictive word order voids the use of markers.
Accordingly, we do not utilize the <start> and
<end> markers for the learned n-grams. They may
occur at any position in an utterance.
Another modification, made specifically for Urdu
is adding new n-grams by changing the word order
in the learned n-grams. This is a modification to
overcome the lack of training data. The post-
positions applied as suffixes to Urdu words, are a
parallel to English prepositions.
To that end, the two phrases below:
1. chalo karte hain
(come let’s do)
2. karte hain chalo
(let’s do come)
are both likely to occur in an utterance.
We learn additional n-grams by deriving
permutations of existing n-grams. These then add to
the frequency count of the original n-gram, although
there are ways in which this frequency assigment
can be refined. Adding to the frequency of the
original n-gram, instead of treating them as separate
instances is practical, since there may be
permutations that occur very rarely in text. “hain
karte chalo” (one other permuation of the above
sentences) does not usually occur in the type of
colloquial or informal text we are looking at; it may
occur in highly stylized forms of text such as Urdu
poetry and as such does not warrant treatment as
separate a n-gram.
Notably, both modifications described above may
notionally be applied for a similar task to any
language with a similar grammar and post-positional
suffixes. (Turkish and Japanese are examples).
The n-grams that have a frequency above a certain
threshold act as cue phrases for that dialogue act.
Using the above mechanism to extract cue phrases,
we use a method where these cue phrases act as
features for machine-learning algorithms. Other
researchers (Samuel et al., 1999) have also used this
method of passing their cue phrases as a feature to a
machine learning method. If the extracted cues are
reliable in identifying dialogue acts, then a classifier
that uses these cues directly should perform
reasonably well.
Table 2 shows a few n-grams that have been
learned for the dialogue act category Action-
Directive with their English translations. Note that a
phrase in Urdu may have different meanings in
English, depending on the context. A total of ~180
n-grams were selected as cue phrases for the various
dialogue acts in our corpus.
Table 2: Examples of n-grams learned for the Action-
Directive dialogue act, their English translations and
Urdu n-gram English Frequency
kar sakte ho
will you do/
you will do/
you may do
karo will do 31
aap log
you people/
you guys/
chalo karte hain
let’s do/
come let’s do/l
let’s do come
will you do/
you will do
padh lo
you read/
read it
Using the most frequently applied dialogue act tags
in the corpus, we can derive a baseline result. We
simply assign the Task tag for the Information-Level
and the Assertion-Opinion tag for the Functional-
Level on the test data set and compute performance
accuracy. This serves as the baseline to compare
results against. The results are shown in the first
column in Table 3.
We use Weka (
machine-learning software to run our classification
algorithms and use the NaiveBayesMultinomial
classifier given in the software kit. Results are
shown in the third column in Table 3, using 10-fold
cross validation on the data corpus. To assess the
improvement in performance by the addition of cue
phrases, we first ran the classifier using simply the
entire utterance as a feature (results shown in the
second column of Table 3 below). Performance at
the Information-Level is much better, as there are 3
classes (categories) of dialogue acts, as opposed to
the Information-Level where there are 15 categories.
And we note that using the entire utterance as a
feature to predict dialogue acts performs poorly.
This is to be expected, due to the noise presented by
the extraneous words in the utterance. However, it
KDIR 2011 - International Conference on Knowledge Discovery and Information Retrieval
does perform better than the baseline.
Table 3: Accuracy of dialogue act classifier using our
approach vs. a simple baseline.
utterance as
cue phrases as
57.2% 60.3% 88.0%
29.6% 37.4% 75.1%
In the above Table 3, the cue phrases that act as
features do not utilize the modification of frequency
counts by using permuted n-grams technique we
described earlier. Even so, there is a significant gain
over the baseline, which indicates that the selected
cue phrases are highly predictive indicators for
dialogue acts in our corpus. Table 4 shows the
performance after the addition of the n-gram
permutation module. There is a solid albeit small
increase in performance accuracy. This is likely due
to the small size of training data corpus. A big
percentage of the cue-phrases we have used are
unigram words, whose frequencies are unaffected by
this modification. However, the 3% of the total
trigrams from our selected n-grams show an increase
in frequency counts. This may push some of the n-
grams above the threshold and result in their
selection as cue phrases.
As an example, the trigrams:
1. kaise hai aap?
(how are you?)
2. aap kaise hai?
(you how are?)
both valid sentences and both occur as indicators for
the Conventional-Opening tag. Since they are
essentially permutations of each other, we can add
their frequency counts. This results in the selection
of this very accurate trigram as a cue-phrase for the
Conventional-Opening tag, where it was not
previously chosen.
Table 4: Accuracy of dialogue act classifier using
permutation of n-grams modification.
of n-grams
With permutation
of n-grams
Info-Level 88.0% 89.4%
Func-Level 75.1% 77.3%
Our goal was to establish the strength of
cue-phrases as features for this task. Using n-grams
is a natural choice since they provide an
understanding of the characteristics of the
underlying utterance. Our method overcomes the
challenges presented by the highly irregular
structure of language used in virtual chat rooms by
filtering out noise present in superfluous words,
emoticons and stop words and using only the crucial
words that are also highly predictive to act as cue-
phrases. While the cue-phrases are significant intra-
utterance features, we plan to expand feature set by
adding inter-utterance features as well. This serves
to add the context of conversation to the classifier.
We have also annotated the corpus for
communication links which indicate who is speaking
to whom and whether the utterance is addressed to a
subset of speakers, a response to a prior utterance or
a continuation of the speaker’s own prior utterance.
This can provide additional evidence for the
We presented an approach to the dialogue act
classification task in Urdu language. This is an
application novel for this language. We also
described certain modifications designed to address
Urdu grammar. Lack of sufficient data for training
and testing is an issue. However, current
performance results are encouraging and provide
insight towards future modifications. One
enhancement would be to improve the selection of
cue phrases, using additional indicators that
complement the frequency counts we currently use.
We hope to test our algorithms on a significantly
larger data set to further validate the hypotheses and
mechanisms. Our contention is that this is a very
promising first attempt at the dialogue act
classification task in a language and grammar
previously uncharted for this task.
Allen, J. M. Core. 1997. Draft of DAMSL: Dialog Act
Markup in Several Layers.
Carletta, J. 2007. Unleashing the killer corpus: experiences
in creating the multi-everything AMI Meeting Corpus.
Language Resources and Evaluation Journal 41(2):
Durrani, N., Hussain, S. 2010. Urdu Word Segmentation.
In the 11th Annual Conference of the North American
Chapter of the Association for Computational
Linguistics (NAACL HLT 2010), Los Angeles, US,
Eric N. Forsyth and Craig H. Martell. 2007. Lexical and
Discourse Analysis of Online Chat Dialog. First IEEE
International Conference on Semantic Computing
(ICSC 2007), pp. 19-26.
Fraser, B. 1990. An Approach to Discourse Markers.
Journal of Pragmatics. 14:383–395
Grosz, B. and C. Sidner., 1986. Attention, Intentions, and
the Structure of Discourse. Computational Linguistics.
12 (3):175-204.
Heeman, P., D. Byron, and J. Allen. 1998. Identifying
Discourse Markers in Spoken Dialog. In Applying
Machine Learning to Discourse Processing: Papers
from the 1998 American Association for Articial
Intelligence Spring Symposium. 44–51. Stanford,
Hirschberg, J. and D. Litman. 1993. Empirical Studies on
the Disambiguation of Cue Phrases. Computational
Linguistics. 19(3):501–530.
Hussain, S. 2008. Resources for Urdu Language
Processing. In Proceedings of the 6th Workshop on
Asian Language Resources. IJCNLP’08, IIIT
Hyderabad, India.
Ijaz, M and Hussain, S. 2007. Corpus Based Urdu Lexicon
Development, In Proceedings of Conference on
Language Technology (CLT07), University of
Peshawar, Pakistan.
Janin, A., Baron, D., Edwards D., Gelbart D., Morgan N.,
Peskin B., Pfau T., Shriberg E., Stolcke A., Wooters
C. 2003. The ICSI Meeting Corpus. In Proc. ICASSP.
Hong Kong.
Jurafsky, Dan, Elizabeth Shriberg, and Debra Biasca.
1997. Switchboard SWBD-DAMSL Shallow-
Discourse-Function Annotation Coders Manual.
Khan F. M., T. A. Fisher, L. Shuler, T. Wu, and W. M.
Pottenger. 2002. Mining chatroom conversations for
Social and Semantic Interactions. Technical Report
LU-CSE-02-011, Lehigh University.
Kim, Jihie., Shaw, Erin., Chern, Grace. and Donghui
Feng. 2007. An Intelligent Discussion-Bot for Guiding
Student Interactions in Threaded Discussions. AAAI
Spring Symposium on Interaction Challenges for
Intelligent Assistants.
Krippendorff, K. 2005. Computing Krippendorff’s alpha-
reliability. Technical Report. University of
Pennsylvania. PA.
Marcu, D. 1997. The Rhetorical Parsing, Summarization,
and Generation of Natural Language Texts. Ph.D.
thesis. University of Toronto, Toronto, Canada. Tech
Report #CSRG-371.
Raza, A. A., Athar, A., Nadeem, S. 2009a. N-GRAM
Based Authorship Attribution in Urdu Poetry. In the
Proceedings of the Conference on Language and
Technology 2009 (CLT09), FAST NU, Lahore,
Pakistan, 22-24 Jan 2009
Raza, A.A., Hussain, S., Sarfraz, H., Ullah, I., Sarfraz, Z.
2009b. Design and development of phonetically rich
Urdu Speech Corpus. In Proceedings of O-
COCOSDA'09. School of Information Science and
Engineering of Xinjiang University, Urunqi, China
Reichman, R. 1985. Getting Computers to Talk Like You
and Me: Discourse Context, Focus, and Semantics.
MIT Press, Cambridge, Massachusetts.
Samuel, K.; Carberry, S.; and Vijay-Shanker, K. 1999.
Automatically selecting useful phrases for dialogue act
tagging. In Proceedings of the Fourth Conference of
the Pacific Association for Computational Linguistics,
Waterloo, Ontario, Canada.
Schiffrin, D. 1987. Discourse Markers. Cambridge
University Press, London, England.
Stolcke A., K. Ries, N. Coccaro, E. Shriberg, R. Bates, D.
Jurafsky, P. Taylor, R. Martin, C. Van Ess-Dykema, &
M. Meteer. 2000. Dialogue Act Modeling for
Automatic Tagging and Recognition of Conversational
Speech. Computational Linguistics. 26(3), 339-373.
Twitchell, Douglas P., Jay F. Nunamaker Jr., and Judee K.
Burgoon. 2004. Using Speech Act Profiling for
Deception Detection. Intelligence and Security
Informatics. LNCS, Vol. 3073
Warner, R. 1985. Discourse Connectives in English.
Garland Publications, New York, New York.
Webb, N., M. Hepple and Y. Wilks. 2005. Dialogue Act
Classification using Intra-Utterance Features. In
Proceedings of the AAAI Workshop on Spoken
Language Understanding, Pittsburgh, USA.
Zipf, G. 1949. Human Behavior and the Principle of Least
Effort. Addison-Wesley.
Zukerman, I. and J. Pearl. 1986. Comprehension-Driven
Generation of Meta-Technical Utterances in Math
Tutoring. In Proceedings of the Sixth National
Conference of the American Association for Articial
Intelligence. Philadelphia, Pennsylvania.
KDIR 2011 - International Conference on Knowledge Discovery and Information Retrieval