=Paper=
{{Paper
|id=None
|storemode=property
|title=Automated Speech Act Classification For Online Chat
|pdfUrl=https://ceur-ws.org/Vol-710/paper22.pdf
|volume=Vol-710
|dblpUrl=https://dblp.org/rec/conf/maics/MoldovanRG11
}}
==Automated Speech Act Classification For Online Chat==
https://ceur-ws.org/Vol-710/paper22.pdf
Automated Speech Act Classification For Online Chat
Cristian Moldovan and Vasile Rus Arthur C. Graesser
Department of Computer Science Department of Psychology
Institute for Intelligent Systems Institute for Intelligent Systems
The University of Memphis Institute for Intelligent Systems
Memphis, TN 38152, USA The University of Memphis
cmldovan|vrus@memphis.edu art.graesser@gmail.com
Abstract which we map the previously described model into a set of
features and then use machine learning algorithms to learn
In this paper, we present our investigation on using super-
vised machine learning methods to automatically classify on- the parameters of the model from annotated training data.
line chat posts into speech act categories, which are seman- We test in this paper this hypothesis and report how well the
tic categories indicating speakers’ intentions. Supervised ma- first 2 to 6 words of an utterance can diagnose its speech
chine learning methods presuppose the existence of annotated act. The tuned models are then evaluated on separate test
training data based on which machine learning algorithms can data sets. In particular, we work with online chat conver-
be used to learn the parameters of some model that was pro- sations in which participants in online chatrooms converse
posed to solve the task at hand. In our case, we used the an- with each other via computer networks. Each online chat-
notated Linguistic Data Consortium chat corpus to tune our room participant can see everyone else’s dialogue turns, or
model which is based on the assumption that the first few to- chat posts, and respond.
kens/words in each chat post are very predictive of the post’s
speech act category. We present results for predicting the
The rest of the paper is organized as in the followings.
speech act category of chat posts that were obtained using The next section presents theoretical background on speech
two machine learning algorithms, Naı̈ve Bayes and Decision acts as well as an overview of various speech act tax-
Trees, in conjunction with several variants of the basic model onomies. The Approach section offers the details of our
that include the first 2 to 6 words and their part-of-speech tags approach. The following section describes related work ad-
as features. The results support the validity of our initial as- dressing the task of speech act classification in similar con-
sumption that the first words in an utterance can be used to texts, e.g. online chats. The Experiments and Results sec-
predict its speech act category with very good accuracy. tion provides a summary of the experiments and results. The
Conclusions section ends the paper.
Introduction
The task of speech act classification involves classifying a Language As Action - Speech Acts
discourse contribution, e.g. an utterance, into a speech act Speech act theory has been developed based on the lan-
category selected from a set of predefined categories that guage as action assumption which states that when people
fulfill particular social discourse functions. Examples of say something they do something. Speech act is a term in
speech act categories are Questions, Statements, or Greet- linguistics and the philosophy of language referring to the
ings. For instance, the hearer infers from the following ut- way natural language performs actions in human-to-human
terance How did you do that? that the speaker is asking a language interactions, such as dialogues. Its contemporary
Question, which informs the hearer to prepare an answer. use goes back to John L. Austin’s theory of locutionary, illo-
Sometimes the speaker just states something as in the fol- cutionary and perlocutionary acts (Austin 1962). According
lowing Statement, The situation is getting worse every day. to Searle (Searle 1969), there are three levels of action car-
or greets someone as in Hello! . ried by language in parallel: first, there is the locutionary act
In this paper, we propose an automated method to clas- which consists of the actual utterance and its exterior mean-
sify online chat posts into speech act categories. The pro- ing; then, there is the illocutionary act, which is the real in-
posed automated method relies on a model that emphasizes tended meaning of the utterance, its semantic force; finally,
the use of the first tokens or words in an utterance to de- there is the perlocutionary act which is the actual effect of
cide their speech act category. For instance, a Question can the utterance, such as scaring, persuading, encouraging, etc.
be distinguished from a Statement based on the first words It is interesting to notice that the locutionary act is a fea-
because usually a Question starts with question words such ture of any kind of language, not only natural ones, and
as How which is followed by an auxiliary verb such as did. that it does not depend on the existence of any actor. In
Our model is based on the assumption that humans do in- contrast, an illocutionary act needs the existence of an envi-
fer speakers’ intentions early on when they hear the first few ronment outside language and an actor that possesses inten-
words of an utterance. To automate the process, we framed tions, in other words an entity that uses language for act-
our problem as a supervised machine learning problem in ing in the outside environment. Finally, a perlocutionary
act needs the belief of the first agent in the existence of a a diet and daily exercise.” the direct speech act is the actual
second entity and the possibility of a successful communi- statement of what happened ”They did this by doing that.”,
cation attempt: the effect of language on the second entity, while the indirect speech act could be the encouraging ”If
whether the intended one or not, is taking place in the en- you do the same, you could lose a lot of weight too.”
vironment outside language, for which language exists as In our work presented here, we assume there is one speech
a communication medium. As opposed to the locutionary act per utterance and the set of speech acts used are all at the
act, the illocutionary and perlocutionary acts do not exist in same level of depthness forming a flat hierarchy. These sim-
purely descriptive languages (like chemical formulas), nor plification assumptions are appropriate for a first attempt at
in languages built mainly for functional purposes (like pro- automating the speech act classification process and testing
gramming languages). They are an indispensable feature of our leading tokens model. Furthermore, the LDC data set
natural language but they are also present in languages built imposed further constraints on our experiments as the LDC
for communication purposes, like the languages of signs or corpus does assume only one speech act per chat posts and
the conventions of warning signals. also uses a flat set of speech act categories.
In a few words, the locutionary act is the act of say-
ing something, the illocutionary act is an act performed Speech Act Taxonomies
in saying something, and the perlocutionary act is an act The task of speech act classification, the focus of our paper,
performed by saying something. For example, the phrase requires the existence of a predefined set of speech act cate-
”Don’t go into the water” might be interpreted at the three gories or speech act taxonomy. Researchers have proposed
act levels in the following way: the locutionary level is the various speech act taxonomies over the years. We present
utterance itself, the morphologically and syntactically cor- next a summary of the most important important ones as
rect usage of a sequence of words; the illocutionary level is judged from historical and relevance perspectives.
the act of warning about the possible dangers of going into The classic categorization of (Austin 1962) postulates five
the water; finally, the perlocutionary level is the actual per- major speech act classes based on five categories of perfor-
suasion, if any, performed on the hearers of the message, to mative verbs: Expositives - verbs asserting or expounding
not go into the water. In a similar way, the utterance ”By views, clasifying usages and references; Exercitives - verbs
the way, I have a peanut butter sandwich with me; would issuing a decision that something is to be so, as distinct from
you like to have a bite?” can be decomposed into the three a judgement that it is so; Verdictives - verbs delivering a
act levels. The locutionary act is the actual expressing of finding, official or unofficial, upon evidence or reason as to
the utterance, the illocutionary act is the offer implied by the value or fact; Commissives - verbs commiting the speaker to
phrase, while the perlocutionary act, namely the intended ef- some course of action; and Behabitives - verbs involving the
fect on the interlocutor, might be impressing with own self- attitudinal reaction of the speaker to someone’s conduct or
lessness, creating a gesture of friendliness, or encouraging fortunes (D’Andrade and Wish 1985).
an activity, in this case eating. The taxonomy proposed by (Searle 1969) consists of six
The notion of speech act is closely linked to the illocu- major classes: Representatives - committing the speaker to
tionary level of language. The idea of an illocutionary act something’s being the case; Directives - attempt by speaker
can be best captured by emphasizing that ”by saying some- to get the hearer to do something; Commissives - commit-
thing, we do something” (Austin 1962). Usual illocutionary ting the speaker to some course of action; Expressives -
acts are: greeting (”Hello, John!”), describing (”It’s snow- expressing the psychological state specified; Declarations -
ing.”), asking questions (”Is it snowing?”), making requests bringing into existence the state described in the proposition
(”Could you pass the salt?”), giving an order (”Drop your and Representative Declarations - giving an authoritative de-
weapon!”), making a warning (”The floor is wet!”), or mak- cision about some fact.
ing a promise (”I’ll return it on time.”). The illocutionary The category scheme proposed by (D’Andrade and Wish
force is not always obvious and could also be composed of 1985) treats most utterances as conveying more than a
different components. As an example, the phrase ”It’s cold speech act and does not attempt to establish a hierarchical
in this room!” might be interpreted as having the intention of order among multiple speech acts. The primary motiva-
simply describing the room, or criticizing someone for not tion for the speech act coding system was a desire to inves-
keeping the room warm, or requesting someone to close the tigate correspondences between speech acts and adjectival
window, or a combination of the above. A speech act could ”dimensions” descriptive of interpersonal behavior. In order
be described as the sum of the illocutionary forces carried for a classifying system to be useful for measuring interper-
by an utterance. It is worth mentioning that within one ut- sonal communication, the distinctions reflected by the cod-
terance, speech acts can be hierarchical, hence the existence ing scheme should be relevant to native speakers’ percep-
of a division between direct and indirect speech acts, the lat- tions and evaluations of interaction. Their classes are: As-
ter being those by which one says more than what is literally sertions (Expositives), Questions (Interrogatives), Requests
said, in other words, the deeper level of intentional mean- and Directives (Exercitives), Reactions, Expressive Evalua-
ing. In the phrase ”Would you mind passing me the salt?”, tions (Behabitives), Commitments (Commissives) and Dec-
the direct speech act is the request best described by ”Are larations (Verdictives, Operatives).
you willing to do that for me?” while the indirect speech act While there seems to be some consensus on the existence
is the request ”I need you to give me the salt.” In a similar of some speech acts, like greetings, questions, answers, etc.,
way, in the phrase ”Bill and Wendy lost a lot of weight with the efficiency of a particular taxonomy for solving a particu-
Table 1: Literature Speech Act Taxonomies
Name Main Classes
Austin Expositives, Exercitives, Verdictives, Commissives, Behabitives
Searle Representatives, Directives, Commissives, Expressives, Declarations, Representative Declarations
D’Andrade and Wish Expositives, Interrogatives, Exercitives, Reactions, Verdictives, Commissives, Behabitives
VerbMobil Request, Suggest, Convention, Inform, Feedback
lar problem ultimately rests on the task at hand. For instance in the case of n-grams, to data sparseness problems because
(Olney, et al. 2003) uses a taxonomy that divides questions of the exponential increase in the number of features. Be-
into 16 subcategories and has only 3 classes for the rest of sides the computational challenges posed by such feature-
the utterances, which is suitable for an Intelligent Tutoring rich methods, it is not clear whether there is need for so
environment. The 16 subclasses of Questions are: Verifi- many features to solve the problem of speech act classifi-
cation, Disjunctive, Concept Completion, Feature Specifica- cation.
tion, Quantification, Definition, Example, Comparison, In- We believe that humans infer speakers’ intention after
terpretation, Causal Antecedent, Causal Consequence, Goal hearing only few of the leading words of an utterance. One
Orientation, Instrumental/Procedural, Enablement, Expecta- argument in favor of this assumption is the evidence that
tional and Judgmental. hearers start responding immediately (within milliseconds)
In the case of Verbmobil, a longterm interdisciplinary or sometimes before speakers finish their utterances ((Juraf-
Language Technology research project with the aim to de- sky and Martin 2009) - pp.814). This paper is a first step
velop a system that can recognize, translate and produce nat- towards exploring the validity of such a hypothesis within
ural utterances, the taxonomy used takes into consideration the context of automated speech act classification of online
in which of the five dialogue phases the actual speech acts chat posts.
occur. The main classes of their taxonomy tree are: Request, Intuitively, the first few words of a dialog utterance are
Suggest, Convention, Inform and Feedback which all ramify very informative of that utterances speech act. We could
into subclasses. For instance, the Convention class is com- even show that some categories follow certain patterns. For
posed of the following subclasses: Thank, Deliberate, Intro- instance, Questions usually begin with a wh- word while
duce, Politeness Formula and Greeting. (Alexandersson, et speech acts such as Answers, Accepting, or Rejecting, con-
al. 1997) tain a semantic equivalent of yes or no among the first words,
A summary of the theoretical speech act taxonomies and and Greetings use a relatively small bag of words and ex-
the Verbmobil taxonomy mentioned above are presented in pressions. In the case of other classes, distinguishing the
Table 1. In our work, we will use the LDC set of speech act speech act after just the first few words is not trivial, but
categories, which are described later. possible. It should be noted that in typed dialogue, which is
a variation of spoken dialogue, some information is lost. For
The Approach instance, humans use spoken indicators such as the intona-
As we already mentioned, we adopted a supervised machine tion to identify the speech act of a spoken utterance.
learning method to automate the process of speech act clas- We must also recognize that the indicators allowing hu-
sification. Machine learning methods imply the design of mans to classify speech acts also include the expectations
a feature set which can then be used together with various created by previous speech acts, which are discourse pat-
machine learning algorithms. We used two such algorithms, terns learned naturally. For instance, after a first greeting an-
Naı̈ve Bayes and Decision Trees, to learn the parameters of other greeting, that replies to the first one, is more likely. We
the basic model and induce classifiers that can categorize ignored such intonational and contextual clues so far in our
new utterances into speech act categories. Naı̈ve Bayes are work in order to explor the potential of classifying speech
statistical classifiers that make the naı̈ve assumption of fea- acts based on words alone. We do plan to incorporate con-
ture independence. While this assumption means models textual clues in future experiments.
that are too simplistic at times, it helps with better estimat- A key decision when developing methods to classify
ing the parameters of the model which in turn leads to good speech acts is choosing the speech act taxonomy. In our
classifiers in general. Decision Trees are based on the idea of work presented in this paper, we adopted the taxonomy pro-
organizing the features in a hierarchical decision tree based posed by the developers of the LDC chat corpus (Forsyth
on information gain. More informative features are always and Martell 2007). The taxonomy is presented in Table 2.
higher in the tree. We will use chat posts and their speech acts from the LDC
In the automated speech act classification literature, re- corpus to illustrate the basic idea of our leading tokens ap-
searchers have considered rich feature sets that include the proach. We picked examples of posts labeled as Yes/No
actual words (possibly lemmatized or stemmed) and n- Questions from the corpus. Selecting the first few words as
grams (sequences of consecutive words). In almost every features seems to be a good approach after seeing the follow-
such case, researchers apply feature selection methods be- ing 12 randomly selected instances of the Yes/No Questions
cause considering all the words might lead to overfitting and, class: ”is 10-19-20sUser68 back yet”, ”Any women from
Table 2: Speech act taxonomy and frequencies in the LDC online chat corpus
Classification Percent Example
Statement 34.50% 10-19-40sUser11...some people have a lot of blank pages
System 17.02% JOIN
Greet 13.40% Hey You
Emotion 11.52% lmao
Wh-Question 5.33% where from@11-09-adultsUser12
Yes/No Question 5.22% wisconsin?
Continuer 3.48% but i didnt chance it
Accept 2.45% ok
Reject 2.14% I can’t do newspaper.. I can’t throw that far and stairs give me problems
Bye 1.57% goodnite
Yes Answer 1.17% yeah
No Answer 0.94% nope 11-09-adultsUser27
Emphasis 0.48% Ok I’m gonna put it up ONE MORE TIME 10-19-30sUser37
Other 0.43% 0
Clarify 0.34% i mean the pepper steak lol
Nashville in here?”, ”are you a male?”, ”hey any guys with or two words long and labeling the missing positions with
cams wanna play?”, ”any guyz wanna chat”, ”any single a none tag means encouraging a classifier to find common
white females?”, ”r u serious”, ”can’t sleep huh?”, ”re- features between all short utterances, regardless of their dif-
ally?”, ”any girls wanna chat with 24/m”, ”22/m/wa any ferent words. We must introduce artificial values such as
ladies want to chat”, ”can i talk to him!!”. The word ”any” none for missing positions in short posts in order to generate
seems to appear often and so are the forms of the auxiliary values for all the six features, for instance, in models where
verb ”to be” and modal verbs. It would also seem very use- we use the first 6 words in chat posts to predict the speech
ful to use a lemmatizer or stemmer, that map morphological acts. We only used the first 6 leading words as the average
variations of the same word to a canonical form, adapted to length in our LDC corpus was 4.67 words meaning models
the specific environment of online chat. For instance, we with 6 words should use up all the words in the posts, on
would like to automatically decide that ”guyz” is the same average, to make predictions.
as ”guys” and that the words ”r” and ”u” may in fact be an
abbreviation for ”are you”. Without this additional knowl- Related Work
edge many resemblences would be lost. Also, the post ”re-
Forsyth and Martell (Forsyth and Martell 2007) developed
ally?” has no common feature with the others, except for the
a speech act classifier on the LDC corpus, using the taxon-
question mark.
omy of (Wu, Khan, Fisher, Shuler and Pottenger 2005). The
Some other speech act classes are even more suitable to corpus consisted of online chat sessions in English between
this approach. As an example, we will provide 12 randomly speakers of different ages. Their prediction model relied on
selected instances labeled as Yes Answer in the same cor- a set of 22 features that include: the number of chat posts
pus: ”yes 10-19-20sUser30”, ”sure 10-19-20sUser126”, ago the user last posted something, number of chat posts in
”yes 10-19-20sUser115!!!!”, ”yes”, ”yep”, ”yes....”, ”yes the future that contain a yes/no pattern, total number of users
i sleep”, ”yeah...”, ”U are Yes”, ”Yes i would 10-19- currently logged on, the number of posts ago that a post was
30sUser12”, ”yep....cool...kool...”, ”yep...”. The word a JOIN (System message), total number of tokens in post,
”yes”, usually on the first position in the post, is a powerful first token in post contains ”hello” or variants, first token in
common feature, as well as the relatively short length of the post contains conjunctions such as ”and”, ”but”, ”or”, etc.,
posts. A common feature is also the usage of pronouns, es- number of tokens in the post containing one or more ”?”
pecially ”I”. However, without knowing that ”yes”, ”yep” and number of tokens in the post in all caps. The values for
and ”yeah” are variants of the same word, any automated all the features were normalized. The first 9 features were
classification method would lose a significant amount of ac- based on the distance of the post to specific posts around it,
curacy. while the rest of the features were based on the density of
A previous attempt by (Marineau, et al. 2000) explored some key words in the post or in the first token of the post
classification using the first three words of each utterance. belonging to a specific speech act category. The machine
We extended the range and used from the first two words learning algorithms they used were Backpropagation Neu-
up to the first six words of each post. Using more words ral Network and Naı̈ve Bayes, with the former performing
does provide more information and thus an easier way to better. Neither method seemed to make a reasonable clas-
differentiate between classes. However, due to the nature sification unless the frequency of the class was higher than
of the corpus we used, sometimes considering too many 3%.
words is a disadvantage, because many posts are only one Obviously, in the classification system of Forsyth and
Table 3: 10-fold cross-validation on LDC online chat corpus
Naı̈ve Bayes Decision Trees
n Accuracy Kappa Precision Recall F-Measure Accuracy Kappa Precision Recall F-Measure
2 74.14 .676 .719 .741 .714 78.33 .727 .772 .783 .772
3 73.05 .662 .698 .731 .697 78.35 .727 .772 .784 .772
4 72.57 .656 .690 .726 .690 77.27 .711 .755 .773 .746
5 72.17 .651 .671 .722 .683 77.27 .711 .755 .773 .746
6 71.70 .645 .662 .717 .677 77.32 .711 .755 .773 .746
Table 4: 10-fold cross-validation on LDC online chat corpus without ”System” posts
Naı̈ve Bayes Decision Trees
n Accuracy Kappa Precision Recall F-Measure Accuracy Kappa Precision Recall F-Measure
2 66.64 .558 .646 .666 .634 71.80 .622 .702 .718 .702
3 65.28 .543 .627 .653 .615 71.87 .623 .704 .719 .703
4 64.46 .533 .618 .645 .604 71.82 .622 .702 .718 .703
5 64.03 .527 .605 .640 .598 71.77 .621 .701 .718 .702
6 63.51 .520 .585 .635 .591 71.82 .622 .702 .718 .703
Martell, the order of posts in the chat and automatic system larger research community. All the user screen names were
messages (like JOIN or PART) played a major role. As far replaced with a mask, for example killerBlonde51 was re-
as syntactical information is concerned, they started from placed by 101930sUser112.
the assumption that the first word of a post is very impor- The original motivation for the development of the cor-
tant for determining the speech act of the post, especially in pus was an attempt to automatically determine the age and
the case of Wh-Question, Yes/No Question, Continuer, Yes gender of the poster based on their chat style, using features
Answer and No Answer. Also, the question mark and the like average number of words per post, vocabulary breadth,
exclamation mark were considered indicative. use of emoticons and punctuation. Subsequently, the cor-
In order to automatically classify speech acts, pus was manually annotated with part of speech labels for
(Samuel, Carberry and Vijay-Shanker 1998) applied a each word and a speech act category per post. An automatic
Transformation-Based Learning machine learning al- speech act classifier was used for this purpose and then each
gorithm on Reithinger and Klessen’s training set (143 post was manually verified (Forsyth and Martell 2007). We
dialogues, 2701 utterances) and on a disjoint testing set take advantage in our experiments of the part of speech in-
(20 dialogues, 328 utterances) (Reithinger and Klesen formation available in the LDC corpus by incorporating this
1997). The features investigated were punctuation marks, information in our basic model. We report results with and
speaker direction (provided by the corpus), number of without part of speech information, which was included in
words in utterances, speech acts of previous and following the basic model in the form of part of speech tags for each
utterances, and a feature called dialogue act cues. The word considered in a particular instance of the model.
latter is finding the n-grams for n = 1,2,3 that minimize
The part of speech (POS) tagging used the Penn Treebank
the entropy of the distribution of speech acts in a training
tagset with some additions specific to the problems related
corpus. Other processing steps they used included filtering
to a chat corpus. Abbreviations such as ”lol” and emoti-
out irrelevant dialogue act cues and clustering semantically-
cons such as ”:)” are frequently encountered and since they
related words. The results showed a comparison between
all convey emotion they were treated as individual tokens
features: manually selected cue phrases, word n-grams,
and tagged as interjections (”UH”). Also, some words that
and entropy-minimization cues, all combined with the
would normally be considered misspelled and were practi-
additional processing steps. The best results were obtained
cally standard online were treated as correctly spelled words
using entropy minimization with filtering and clustering.
and tagged according to the closest corresponding word
class. For example, the word ”wont” if treated as a mis-
Experiments and Results spelling would normally be tagged as ”ˆMDˆRB”, the char-
The LDC online chat corpus is a product of the Naval Post- acter ˆ referring to a misspelling. The same word would be
graduate School (Lin 2007). It contains 10,567 posts from tagged as ”MD” and ”RB” when referring to ”modal” and
different online chat rooms in English. All the posts had ”adverb”, respectively. However, since it was highly fre-
to go through a sanitizing process in order to protect user quent in the chat domain, ”wont” was tagged as ”MD”. In
privacy, so that the corpus could be made available to the contrast, words that were just plain misspelled and did not
Table 5: 10-fold cross-validation on LDC online chat corpus without ”System” posts and without POS tags
Naı̈ve Bayes Decision Trees
n Accuracy Kappa Precision Recall F-Measure Accuracy Kappa Precision Recall F-Measure
2 69.40 .574 .641 .694 .641 71.79 .622 .701 .718 .703
3 66.88 .546 .632 .669 .613 71.82 .622 .703 .718 .703
4 65.74 .532 .618 .657 .598 71.85 .623 .703 .719 .703
5 64.57 .517 .594 .646 .584 71.80 .623 .702 .718 .703
6 63.89 .507 .587 .639 .576 71.78 .622 .701 .718 .702
appear frequently were tagged with the misspelled version ing finding is the fact that best results using Naı̈ve Bayes are
of the tag, for example the word ”intersting” was tagged as obtained when using only the first two leading words in a
”ˆJJ” (Forsyth and Martell 2007). chat post instead of more. When using Decision Trees, re-
In the LDC corpus, each post was assigned a single speech sults obtained with the first two leading words are as good as
act category from the 15 categories of the chat taxonomy when using even 6 words. Thus, we can conclude that our
proposed by (Wu, Khan, Fisher, Shuler and Pottenger 2005). hypothesis that the first few leading words of an utterance
Those categories along with examples and their frequencies are very diagnostic of that utterances’ speech act is true, at
in the corpus are represented in Table 2. least for online chat posts, the focus of our experiments.
In order to implement the machine learning approach, we
extracted for each of the 10,567 posts the first n tokens
(n = 2..6) and their part of speech (POS) tags. Further- Conclusions
more, we recorded the annotated speech act category as the
correct class of the post, which is needed during training. We Our results acknowledge the fact that the first few to-
then use Naı̈ve Bayes and Decision Trees (J48) from WEKA kens of a chat post are indicative of the speech act of
(Witten and Frank 2005) to induce classifiers based on the the post. It is worth mentioning the chat language could
leading n tokens and their POS tags. We experimented with be considered more challenging than natural conversation
several variants of the basic model by generating an instance language. Indeed, online chat being an environment that
for each n = 2..6. The accuracy of the induced classifiers apparently encourages extreme creativity and exhibits a
was measured using 10-fold cross-validation. In the cases in very high tolerance to misspellings and breaking language
which the post had less than n tokens, we replaced the empty rules. For instance, ”hey, heya, heyheyhey, heys, heyy,
feature slots with a dummy token and a dummy POS tag. A heyyy, heyyyy, heyyyyy, heyyyyyy, heyyyyyyy, heyyyyyyyy,
summary of results is shown in Table 3. heyyyyyyyyy, heyyyyyyyyyy, heyyyyyyyyyyyyyy” are in the
The System class clearly increases the performance of the LDC corpus chat-specific variants of the same greeting
classifier due to the large number of instances and their sim- word. To this point, we did not use a lemmatizer-like tool
plicity. Practically, the System posts are ”PART”, ”JOIN”, that could reduce the high number of variants for the same
and just a few other variants. In a second round of experi- token, especially in the case of speech acts with a high emo-
ments, we wanted to investigate the validity of our approach tional content, such as rejections, accepting, emotion and
on real posts only, i.e. we did not take into account the Sys- emphasis, and also in the case of greetings and yes/no ques-
tem posts. As already mentioned, the actual System mes- tions/answers, which in a literary corpus would usually be
sages are too specific to a particular chat system and they are represented by a relatively small number of expressions, but
not natural language. As before, we extracted for each post which in the chat environment are especially targeted by lan-
the first n tokens and the speech act category. On the remain- guage creativity.
ing 7,935 posts, we applied the same Naı̈ve Bayes and De- One future extension we plan to do is utilizing word n-
cision Trees (J48) classifiers with 10-fold cross-validation. grams for detecting speech acts. N-grams could better cap-
The results are presented in Table 4. A significant drop in ac- ture word order and thus better differentiate between pat-
curacy can be noticed. Still, the performance of the proposed terns such as ”do you” (a bigram), which most likely indi-
approach is very good on the natural, non-System posts. We cates a question, and ”you do”, which indicates a Command.
also observe that there is no major difference among the var- Furthermore, we plan on using the dialog act cues proposed
ious models when the number of leading words is vare. This by (Samuel, Carberry and Vijay-Shanker 1998) for detect-
may be explain in the case of chat posts by the relative short ing n-grams that minimize the entropy of the distribution of
nature of these posts. speech acts in our training corpus and apply them as features
In Table 5, we provide the results obtained by the induced for speech act classification. We also plan to test our leading
classifiers without System posts and without parts of speech words hypothesis on dialogue data that is naturally longer
tags for the words. These results reveal the power of our than chat posts. We hope to understand better from such
basic model alone (without part of speech information) on data whether only two or three leading words are enough as
natural posts (System posts are not included). An interest- opposed to six.
References Olney, A.; Louwerse, M.; Mathews, E.; Marineau, J.; Hite-
Mitchell, H.; and Graesser, A. 2003. Utterance Classifica-
Alexandersson, J.; Buschbeck-Wolf, B.; Fujinami, T.; tion in AutoTutor. Building Educational Applications using
Maier, E.; Reithinger, N.; Schmitz, B.; Siegel, M. 1997. Natural Language Processing: Proceedings of the Human
Dialogue Acts in VerbMobil-2. volume 226, Verbmobil Language Technology, Philadelphia, PA.
Report, German Research Center for Artificial Intelligence
(DFKI), Saarbrcken, 1998 Reithinger, N.; and Klesen, M. 1997. Dialogue act
classification using language models. In Proceedings of
Austin, J.L. 1962. How to do Things with Words. Oxford EuroSpeech-97, pages 2235-2238.
University Press, 1962.
Samuel, K.; Carberry, S.; and Vijay-Shanker, K. 1998.
D’Andrade, R.G.; and Wish, M. 1985. Speech Act Theory Dialogue Act Tagging with Transformation-Based Learn-
in Quantitative Research on Interpersonal Behavior. Dis- ing. In Proceedings of the 36th Annual Meeting of the As-
course Processes 8:2:229-258, 1985. sociation for Computational Linguistics, pages 1150-1156,
Forsyth, E.N.; and Martell, C.H. 2007. Lexical and Dis- Montreal, 1998.
course Analysis of Online Chat Dialog. International Con- Searle, J.R. 1969. Speech Acts. Cambridge University
ference on Semantic Computing (ICSC 2007), pp. 19-26. Press, GB, 1969.
Jurafsky, Dan.; and Martin, J.H. 2009. Speech and Lan- Witten, I. H.; and Frank, E. 2005. Data Mining: Prac-
guage Processing. Prentice Hall, 2009. tical machine learning tools and techniques, 2nd Edition.
Lin, J. 2007. Automatic Author profiling of Online Chat Morgan Kaufmann, San Francisco.
Logs. M.S.Thesis, Naval Postgraduate School, Monterey. Wu, T.; Khan, F.M.; Fisher, T.A.; Shuler, L.A.; and
Pottenger, W.M. 2002. Posting Act Tagging us-
Marineau, J.; Wiemer-Hastings, P.; Harter, D.; Olde, B.;
ing Transformation-Based Learning. In Proceedings of
Chipman, P.; Karnavat, A.; Pomeroy, V.; and Graesser, A.
the Workshop on Foundations of Data Mining and Dis-
2000. Classification of speech acts in tutorial dialog. In
covery, IEEE International Conference on Data Mining
Proceedings of the workshop on modeling human teach-
(ICDM’02), December 2002.
ing tactics and strategies at the Intelligent Tutoring Systems
conference.