Proceedings of the ACL Student Research Workshop, pages 31–36,
Ann Arbor, Michigan, June 2005. c©2005 Association for Computational Linguistics
Automatic Discovery of Intentions in Text and its Application to Question
Answering
Marta Tatu
Human Language Technology Research Institute
Department of Computer Science
University of Texas at Dallas
Richardson, TX 75080, USA
marta@hlt.utdallas.edu
Abstract
Semantic relations between text concepts
denote the core elements of lexical se-
mantics. This paper presents a model for
the automatic detection of INTENTION se-
mantic relation. Our approach first identi-
fies the syntactic patterns that encode in-
tentions, then we select syntactic and se-
mantic features for a SVM learning classi-
fier. In conclusion, we discuss the appli-
cation of INTENTION relations to Q&A.
1 Introduction
1.1 Problem description
Intentions comprise of semantic relationships that
express a human’s goal-oriented private states of
mind, including intents, objectives, aims, and pur-
poses. As a relation, it encodes information that
might not be explicitly stated in text and its detec-
tion might require inferences and human judgment.
The answer to the question What was Putin trying
to achieve by increasing military cooperation with
North Korea? is found in the sentence Putin is at-
tempting to restore Russia’s influence in the East
Asian region. Extracting the exact answer to restore
Russia’s influence in the East Asian region becomes
easier if this is recognized as Putin’s intention which
matches the question’s expected answer.
In this paper, we describe a method that identi-
fies intentions in domain independent texts. We em-
ployed two machine learning algorithms to create
models that locate intentions in a given paragraph
using a set of six syntactic and semantic features.
1.2 Motivation
The current state-of-the-art NLP systems cannot ex-
tract intentions from open text and, as we saw in the
example, their detection benefits Question Answer-
ing. An intention is the answer to general questions
like What is the goal of X?, What does X plan to do?,
or What does X aim for? The INTENTION seman-
tic relation is one of the most challenging relations
because text fragments may convey unstated inten-
tions. These are most pervasive in dialogues, com-
munication specific to humans. For example, in the
following conversation, the vendor infers the client’s
unstated intention of buying the cups.
Customer: Where do you have the $1 cups?
Salesman: How many do you want?
Intentions are closely related to other semantic re-
lations such as beliefs, motives, desires, or plans.
In the above example, the context tells us that this
takes place in a superstore, well-known as a place
where people buy things from. The clerk’s an-
swer emerges from our common beliefs and back-
ground knowledge as well as from his desire to
help a customer. Intentions are the framework for
plans. Many philosophers and artificial intelligence
researchers studied the intentions as parts of coor-
dinating plans (Bratman, 1987; Pollack, 1990) be-
cause people establish plans for future times.
In this paper, we regard intentions as expres-
sions of a particular action that shall take place in
the future, in which the speaker is some sort of
agent (Anscombe, 1957). For example, the sentence
Mary is going to buy a TV set shows Mary’s in-
tention. Anscombe (1957) considers intentions as
a subclass of predictions, besides commands and
31
prophecies. John is going to be sick is usually a
prophecy, John, go for a walk! is an order, and John
plans to take a walk expresses an intention.
1.3 Previous work
Various methodologies have been proposed and used
over the years for the task of extracting semantic
relations from text. Purely probabilistic models,
empirical methods, or hand-coded constraints were
some of the approaches that do not use machine
learning algorithms. Later on, methods that use de-
cision tree, neural networks, memory-based learn-
ing, or support vector machines were introduced.
Currently, there is also a increased interest in shal-
low semantic parsing of open texts and automatic la-
beling of semantic roles. Wiebe et al. (2004) focused
on the detection of subjective language such as opin-
ions, evaluations, or emotions in text. Using clues
of subjectivity (low-frequency words, collocations),
they identify opinion piece texts such as editorials,
letters to the editor, or arts and leisure reviews.
There exists an immense literature in philoso-
phy about the different types of intentions and their
characteristics. Bratman (1987) tries to find the re-
lationship between the two distinct phenomena of
doing something intentionally and intending to do
something. Numerous philosophical studies dis-
cuss how intentions relate to other psychological
concepts, such as, beliefs, desires, hopes, or ex-
pectations (Audi, 1973; Bratman, 1981; Bratman,
1987). Intentions are consistent with the person’s
beliefs, and, unlike ordinary desires, require con-
sistency (Bratman, 1987). They can generate rea-
sons for or against future intentions (Bratman, 1981;
Bratman, 1987). As plan elements, intentions re-
quire a certain stability. Their side effects need not
be intended, even if they were taken into considera-
tion in the first place1 (Bratman, 1990).
2 Syntax and Semantics of Intention
2.1 Syntactic patterns
Because, in all the cases that we encountered, inten-
tions were conveyed by phrases, we took a closer
look at how intentions can be expressed in the writ-
ten text. For our investigations, we chose the Sem-
1Due to space limitations, we couldn’t include detailed ex-
amples. Please see the cited articles for examples.
Cor text collection (Miller et al., 1993), a subset of
the Brown corpus manually tagged with WordNet
senses (37,176 sentences in 352 newspaper articles).
After manually classifying the first 2,700 sentences
from SemCor into sentences that contain or not in-
tentions, only 46 examples were identified. The
syntactic patterns listed in Table 1 cover 95.65% of
them. Because the first pattern comprises more than
half of the studied examples, our algorithm focuses
on detecting intentions encoded by a0a2a1a4a3 to a0a5a1a7a6 .
We note that this pattern is ambiguous and may con-
vey other semantics. For instance, Mary began to
play with the dog, He told her to meet you are en-
coded by our pattern, but do not express intentions.
Pattern Example Frequency
a8a10a9a10a11 to a8a12a9a14a13 plan to go for a walk 27 (58.69)
NN to VB strivings to give up drink 6 (13.04)
VB PP VP He resigned so that he can work 5 (10.87)
for the school campaign
goal/purpose is to VB his goal is to leave the country 4 (8.69)
ADJ to VB eager to end a pitching slump 2 (4.34)
Table 1: INTENTION syntactic patterns
2.2 Semantics of intentions
From the semantic point of view, an intention may
be very specific, it may contain a future time or a
location (John intends to meet Mary today), but ev-
ery intention must specify a future action. Hence,
we propose the following representation for the IN-
TENTION semantic relation: INT(a15 a3a17a16a19a18a20a3a21a16 a15 a6 ) where
a15
a3 is the event denoting the intention, a18 a3 denotes the
person that has the intention and a15 a6 is the intended
action or event. If the intention is more specific
then we will identify instances of other semantic re-
lations2. a22a24a23a26a25a14a27a29a28 a18a20a3a31a30a33a32 INTa28a34a15 a3a35a16a19a18a20a3a21a16 a15 a6a17a30a33a32a37a36 a15a17a15a21a38a39a28a34a15 a6a35a30
a32a41a40a43a42a45a44a47a46
a28
a18 a6 a30a48a32
a38a49a23a47a50
a42a51a46
a28
a18a53a52a17a30a48a32 THEME
a28a34a15
a6 a16a19a18 a6 a30a54a32
TIMEa28a34a15 a6 a16a19a18a53a52a17a30 represents a more specific intention.
The semantics of the INTENTION relation allows
the derivation of inference rules which show that IN-
TENTION dominates other semantic relations such as
PURPOSE, ENTAIL, or ISA. For example, if a person
a18a24a3 intends to perform action
a15
a6 and this action has
a purpose a15 a52 , then we can say that a18 a3 intends to do
a15
a52 3. Formally, we can express the above relations
2The list of semantic relations that can specialize an INT
includes THEME, LOCATION, TEMPORAL, MANNER, INSTRU-
MENT, SOURCE, MEANS, and FREQUENCY. Their arguments
are a55a57a56 , the intention verb, and a corresponding a58a60a59 .
3Similar statements can be made for the ENTAIL and ISA
32
with the following set of implications4:
INTa0 a55a2a1a4a3a34a58a5a1a6a3a34a55 a56a8a7a10a9 PURPOSEa0 a55a57a56a11a3 a55a13a12a14a7a16a15 INTa0 a55a13a17a18a3a34a58a5a1a6a3a34a55a11a12a6a7
INTa0 a55 a1 a3a34a58 a1 a3a34a55 a56 a7a10a9 ENTAILa0 a55 a56 a3 a55 a12 a7a19a15 INTa0 a55 a17 a3a34a58 a1 a3 a55 a12 a7
INTa0 a55 a1 a3a34a58 a1 a3a34a55 a56 a7a10a9 IS-Aa0 a55 a56 a3 a55 a12 a7a19a15 INTa0 a55 a17 a3 a58 a1 a3 a55 a12 a7
INTa0 a55a2a1a4a3a34a58a5a1a6a3a34a55 a56a8a7a10a9 PURPOSEa0 a55a13a12a20a3 a55a57a56a6a7a22a21a15 INTa0 a55a13a17a18a3a34a58a5a1a6a3a34a55a11a12a6a7
INTa0 a55 a1 a3a34a58 a1 a3a34a55 a56 a7a10a9 CAUSEa0 a55 a56 a3 a55 a12 a7a22a21a15 INTa0 a55 a17 a3 a58 a1 a3 a55 a12 a7
The first three implications formalize the above
inference rules. If John intends to start his car to
go to the park, then John intends to go to the park.
Similarly, if John intends to buy a car, then we can
say that he intends to pay for it. The sentences John
intends to go to the park. He’s starting his car right
now express John’s intention to go to the park (a15 a6 ).
The purpose of starting the car (a15 a52 ) is to go to the
park. We cannot say that John intends to start his
car. This is just an intentional action done to achieve
his objective. The fifth rule tries to eliminate the ef-
fects (a15 a52 ) of an intention (a15 a6 ) from being considered
as intentions or objectives. If John intends to swim
in the pool (a15 a6 ) even if he knows that he is going to
catch a cold (a15 a52 ) because the water is too cold, we
cannot say that John intends to catch a cold.5 The
traditional relational properties (reflexivity, symme-
try, or transitivity) do not hold for the INTENTION
semantic relation.
3 Learning Model
3.1 Experimental data
We applied the most frequent syntactic pattern that
expresses intentions in text (a0a5a1 a3 to a0a5a1 a6 ) on the
first 10,000 sentences of the SemCor2.0 collection
and we extracted 1,873 sentences. These sentences
contain 115 intentions (manually identified by a
graduate student, not the author). The data consist-
ing of these positives and 258 arbitrarily selected
negative examples, was randomly divided into a
training set that contains 80% of the examples and
a test set with the remaining 20% instances. The
statistics are shown in Table 2.
Intentions Non-Intentions Total
Training 92 208 300
Testing 23 50 73
Table 2: Experiments Data Division
semantic relations.
4
a55a23a1 and a55a13a17 represent different intentions of the same person.
5A more detailed example can be found in (Bratman, 1990).
3.2 Features for intention
After analyzing our training data, we pinpointed a
set of features to help us identify the intentions en-
coded by the pattern a0 a1 a3 to a0a5a1 a6 . The WordNet
senses needed to extract the semantic features were
taken from SemCor. We will use Mary intends to
revise the paper to show each feature’s value.
The semantic class of the the a0a5a1 a3 verb’s agent
or specializations of it. Intentions and objectives
are specific to humans. Thus, the semantic class of
the a0a5a1 a3 agent bears a high importance. We used
an in-house semantic parser to retrieve the AGENT
of the a0 a1 a3 verb. The feature’s value is its WordNet
semantic class. Mary names a person. Thus, the
semantic class that we are seeking is entity#1.
We chose this semantic generalization because
nouns and verbs belong to open part-of-speech
classes. There can be an enormous number of pos-
sibilities and any models built using them as fea-
ture values will not be able to generalize beyond the
training examples. Therefore, we introduce a bias
in our learning framework based on the assumption:
noun and verb concepts will semantically behave
as the concepts that subsume them in the WordNet
structures. But, by generalizing concepts, we lose
some of their semantic properties. Hence, we spe-
cialize the semantic class a24 of a concept a25 by re-
placing it with its immediate hyponym (a25 ) that sub-
sumes a25 . We can further increase the semantic level
by specializing a25 . We note that the number of values
is still finite even though we specialized the general
concepts. As the specialization level increases, there
will be words a25 that cannot be further specialized
(entity#1 cannot be specialized even once). In such
cases, we add a25 to the set of feature values.
The semantic class of the a0a5a1 a3 verb or its spe-
cializations. The intention phrase is subordinated
to a verb (a0a5a1 a3 ). The semantic class of this verb is
the system’s second feature. In our example, a0 a1 a3
(intend#1) semantic class is wish#3.
The semantic class of the a0a2a1a2a6 verb’s agent, if
this agent differs from the a0a5a1 a3 verb’s agent; other-
wise, a common value (equal) is given. We identify
the AGENT of the a0a5a1 a6 verb. The specializations of
its semantic class will be used if the top noun proves
to be too general. In the sample sentence, the agent
of revise is Mary. We can have a different agent for
33
Semantic Semantic class of the a8a53a9 a11 verb (%)
class of no specialization a0a2a1 a3 level of specialization a4a2a5
a6
level of specialization
the a8a10a9 a11 ’s Semantic class of the a8a53a9 a13 verb Semantic class of the a8a12a9 a13 verb Semantic class of the a8a12a9 a13 verb
agent no spec. a0a7a1 a3 level a4a7a5
a6
level no spec. a0a7a1 a3 level a4a7a5
a6
level no spec. a0a7a1 a3 level a4a2a5
a6
level
no spec. 87.67 80.82 87.67 90.41 87.67 87.67 86.30 83.56 84.93
a0a7a1
a3 level 89.04 82.19 87.67 87.67 89.04 87.67 87.67 86.30 84.93
a4a7a5
a6
level 87.67 83.56 87.67 90.41 90.41 89.04 89.04 87.67 86.30
Table 3: Accuracy of models using the a8a10a9a12a11 specialization level for the a0a2a1 a6 agent semantic class
the a0a2a1a7a6 verb (Mary intends John to revise the pa-
per). Let’s assume that Mary is John’s supervisor
and she can make him revise the document. The sen-
tence expresses Mary’s intention of persuading John
to revise the paper, but this objective is not encoded
by the pattern we considered.
The semantic class of the a0a5a1 a6 verb or its spe-
cializations. The a0a5a1 a6 verb expresses the future ac-
tion or behavior that the agent intends. We extract
this feature using WordNet hierarchies. Revise#1 be-
longs to the act#1 semantic class.
A flag indicating if the a0 a1 a3 verb has an affir-
mative or a negative form. We want to differen-
tiate between sentences like John wants to go for a
walk and John doesn’t want to go for a walk. The
first sentence expresses John’s intention, while, in
the second one, no intention can be identified.
The type of the analyzed sentence. This feature
is primarily concerned with questions. A question
like Where do you plan to go for a walk? indicates
the intention of going for a walk, unlike the question
Do you plan to go for a walk? which might express
an intention if the answer is “yes”. This feature’s
values are the wh-words that begin a question or n/a
for the other types of English sentences.
We did not analyze the affirmative versus the neg-
ative form of the a0a2a1 a6 verb because it does not affect
the objective attribute of the intention. The sentence
John intends not to go for a walk expresses a nega-
tive intention. This sentence is much stronger than
John doesn’t intend to go for a walk. In the former
context, John has set a goal for himself , while in the
second sentence, the objective does not exist.
4 Experimental Results
4.1 Impact of specialization
The first experiment was performed using the LIB-
SVM package6 and the WordNet semantic classes.
6http://www.csie.ntu.edu.tw/˜cjlin/libsvm/index.html
These features yield an accuracy of 87.67%. Try-
ing to improve the performance, we specialized the
semantic classes. When the a0 a1 a6 ’s agent semantic
class was specialized, the accuracy remained con-
stant. If we replace the a0a2a1 a6 ’s semantic class with
its direct hyponyms, the accuracy drops 5.48%. But,
the specialization of the a0a5a1 a3 agent’s semantic class
brings an improvement of 1.37% and the special-
ization of the a0a2a1 a3 ’s class produces an increase in
accuracy of 2.74%. Given this fluctuation in per-
formance, we performed 81 different experiments
which create SVM models using the same training
data annotated with more general or more specific
feature values. For each feature, we analyzed the
first two semantic specialization levels.
From our experiments, we noticed that the spe-
cialization of the a0 a1a33a6 ’s agent semantic class does
not influence the performance. Out of the 27 ex-
periment triplets in which this specialization level
changes, in only 4, it influences the result and, in
3 of them, the accuracy increases with the special-
ization level. Thus, our third feature is the second
specialization level of the a0a5a1 a6 ’s agent class. Ta-
ble 3 shows the results obtained when the values of
the radial kernel parameters were chosen to optimize
the 5-fold-cross-validation on the training data. The
best models are described in Table 4.
Model Level of specialization for the features
A semantic class of the a8a53a9 a11 agent, a0a13a1
a3 level of specialization for
the a8a10a9 a11 ’s semantic class, and semantic class of the a8a10a9 a13 verb
B a4a2a5
a6
semantic level for the a8a53a9a53a11 agent class, a0a7a1 a3 level of the
a8a53a9 a11 ’s semantic class, and the semantic class of the a8a10a9 a13 verb
C a4a2a5
a6
level of the a8a12a9 a11 agent’s semantic class and a0a7a1 a3
specialization levels for the a8a12a9a53a11 and a8a10a9a14a13 semantic classes
Table 4: The best three intention classifiers
4.2 Learning curves
We further analyzed our data and models and tried
to see how many training examples are needed to
reach 90.41% accuracy. We varied the training data
34
Semantic class of the Semantic class of the Semantic class of the Semantic class of the a8a53a9 a11 verb Sentence
a8a10a9 a11 ’s agent a8a10a9 a11 verb a8a10a9 a13 ’s agent a8a10a9 a13 verb form type
Model A 2.74 16.44 1.37 0 2.74 4.11
Model B 2.74 15.07 1.37 0 4.11 2.74
Model C 1.37 16.44 4.11 0 4.11 2.74
Table 5: The improvement (%) brought by each feature to the three best SVM models
size and validated the new models using our previ-
ous test set. Figure 1 shows the performance varia-
tion of three models that use feature sets identical in
terms of specialization levels to the ones of the A, B,
and C classifiers. All three models exhibit a similar
behavior with respect to the change in the training
set size. Therefore, our features create a stable al-
gorithm. The highest accuracy models use all 300
training examples. Thus, we did not reach the satu-
ration point, but, considering the performance curve,
this point is not very far.
 30
 40
 50
 60
 70
 80
 90
 100
 50  100  150  200  250  300
SVM model accuracy
Number of training examples
Model A
Model B
Model C
Figure 1: Testing set is constant
4.3 Feature impact on the SVM models
All our previous experiments used the entire set of
features. Now, we investigate the relative contribu-
tion of each feature. We performed experiments that
use only five out of the six features. In Table 5, we
list the accuracy increase that is gained by the inclu-
sion of each feature. The most influential attribute is
the a0a5a1 a3 verb’s semantic class or its specializations.
The intention’s description verb does not influence
the classification result. Because intentions consist
of a future action and verbs express actions, there
are very few verbs, such as dream or snore (invol-
untary actions) that cannot occupy the a0a2a1 a6 verb’s
position. The syntactic features bring an average in-
crease in accuracy of 3.50%.
4.4 Impact of word sense disambiguation
Perfect word sense disambiguation might be a too
strong assumption. In this section, we examine the
effects of weaker disambiguation. Table 6 shows the
accuracies of the best three models when each con-
cept is tagged with its first WordNet sense (No WSD)
and when the senses are given by an in-house WSD
system with an accuracy of 69% computed on the
SemCor data (Automatic WSD).
No WSD Automatic WSD Gold WSD
Model A 72.60 79.45 90.41
Model B 73.97 79.45 90.41
Model C 72.60 80.82 90.41
Table 6: Best models performance (%)
4.5 C5 results
After examining the SVM results, we applied the C5
machine learning algorithm (Quinlan, 2004) to the
same training data annotated with the same feature
set, in a similar manner. Again, we specialized the
four semantic classes, independently, and tested the
decision trees against the testing data. Table 7 shows
their accuracy. The highest values were obtained for
the first level of specialization of the a0a2a1 a3 verb se-
mantic class. The specialization levels of the other
semantic classes do not influence the accuracy of
the decision trees. The most tested attribute is the
a0a5a1 a3 verb. This further substantiates our observa-
tion, made during our SVM models analysis, that this
feature has the greatest importance in the intention
classification process. Our error analysis of the C5
results indicates that, because of the relatively small
numbers of training instances, C5 ignores some of
the features and makes wrong decisions.
5 Application to Question Answering
Questions involving intentions cannot be answered
only by keyword-based or simple surface-level
matching techniques. Table 8 lists two questions for
35
a0 a11 : What was Putin trying to achieve by increasing military cooperation with North Korea?
a0a2a1a4a3 a11 : Putin
a5a7a6
a11a9a8 & INT
a5a11a10a13a12a15a14a16a6
a11
a14a16a17
a8 & ANS
a5a7a17
a8 & MANNER
a5a7a17a18a14a19a10
a13a20a8 & increase
a5a11a10
a13
a14a16a6
a11
a14a21a6
a13a20a8 & military
a5a7a6
a13a19a8 & cooperation
a5a11a6
a13a20a8 & with
a5a7a6
a13
a14a16a6a23a22
a8
& North Koreaa5a11a6 a22 a8
a24 a11 : Putin is attempting [to restore Russia’s influence in the East Asian region][
INT]. The report said, the possibility remains that Russia could
increase military cooperation with North Korea based on their treaty.
a24 a1a4a3 a11 : Putin
a5a7a6
a11a9a8 & INT
a5a11a10a13a12a15a14a16a6
a11
a14a21a10
a11a25a8 & restore
a5a11a10
a11
a14a21a6
a11
a14a16a6
a13a20a8 & Russia
a5a7a6a26a22
a8 & ’s
a5a7a6a26a22a27a14a16a6
a13a25a8 & influence
a5a7a6
a13a25a8 & LOCATION
a5a7a6
a13
a14a21a6a29a28
a8 & East
a5a7a6a23a28
a8 &
Asiana5a7a6a23a28 a8 & regiona5a7a6a23a28 a8 & reporta5a7a6a26a30 a8 & saya5a11a10 a13 a14a16a6a26a30a31a14a32a10a13a22 a8 & possibilitya5a7a6a26a33 a8 & remainsa5a11a10a13a22a29a14a16a6a26a33a31a14a21a10a32a28 a8 & increasea5a11a10a13a28a34a14a16a6a26a22a31a14a21a6a26a35 a8 &
militarya5a7a6 a35 a8 & cooperationa5a7a6 a35 a8 & witha5a7a6 a35 a14a16a6a23a36 a8 & North Koreaa5a7a6a26a36 a8 & basea5a11a10 a28 a14a16a6a26a37 a8 & treatya5a11a6a23a37 a8
a0 a13 : From where does al Qaeda intend [to purchase weapons of mass destruction][INT]?
a0a2a1a4a3 a13 : alQaeda
a5a7a6
a11a32a8 & INT
a5a38a10a13a12a39a14a16a6
a11
a14a21a10
a11a9a8 & ANS
a5a11a17
a8 & LOCATION
a5a38a10
a11
a14a16a17
a8 & purchase
a5a11a10
a11
a14a16a6
a11
a14a16a6
a13a25a8 & weapons of mass destruction
a5a7a6
a13a19a8
a24 a13 : It is known that Osama bin Laden’s al Qaeda network has tried [to buy ingredients for weapons of mass destruction in Russia][
INT].
a24 a1a4a3 a13 : Osama bin Laden
a5a7a6
a11a13a8 & ’s
a5a7a6
a11
a14a21a6
a13a20a8 & al Qaeda
a5a7a6
a13a9a8 & network
a5a7a6a26a22
a8 & IS-A
a5a11a6
a13
a14a16a6a23a22
a8 & INT
a5a11a10a13a12a26a14a16a6
a13
a14a32a10
a11a13a8 & buy
a5a11a10
a11
a14a16a6a23a22a34a14a16a6a23a28
a8 &
ingredienta5a11a6 a28 a8 & PURPOSEa5a7a6 a28 a14a16a6 a30 a8 & weapons of mass destructiona5a11a6 a30 a8 & LOCATIONa5a11a10 a11 a14a21a6 a33 a8 & Russiaa5a7a6 a33 a8
Table 8: Question and answer pair examples
Semantic class of Semantic class of the a40a42a41 a1 verb
the a40a42a41 a1 ’s agent no spec. a43a31a44a21a45 level a46a29a47a23a48 level
no spec. 79.45 87.67 84.93
a43a27a44a32a45 level 68.49 87.67 84.93
a46 a47a23a48 level 79.45 87.67 84.93
Table 7: C5 models accuracy (%)
which finding the correct answer primarily depends
on the discovery of the INTENTION relation.
The answer type for the question a49 a3 is the IN-
TENTION argument itself. The question processing
module will detect that the answer being sought is
Putin’s intention. The semantic relations module
processes a50 a3 ’s text and discovers the INTENTION
relation. The question is searching for the intent of
Putin with regards to North Korea and the answer
text reveals Putin’s intention to restore Russia’s in-
fluence in the area. Question a49 a6 is searching for a
location as its answer type and the correct answer is
one which involves al Qaeda intending to purchase
weapons of mass destruction. The candidate answer
text (a50
a6 ) reveals the organization’s past intent to buy
(synonym with purchase) weapons in Russia. Be-
cause the two intentions have the same agent, future
action and theme, the two semantically enhanced
logic forms can now be unified and we can pin down
the location of the intent (Russia).
6 Conclusions
We proposed a method to detect the INTENT rela-
tion encoded by the sentence-level pattern a0a5a1 a3 to
a0a5a1 a6 with a 90.41% accuracy. We plan to investi-
gate the other INTENTION patterns as well as other
semantic relations such as MOTIVE, IMPLICATION,
or MEANING which, currently, cannot be identified
by the state-of-the-art NLP systems. These relation-
ships need to be analyzed to provide a complete cov-
erage of the underlying semantics of text documents.
We intend to incorporate our INTENTION detection
module into a Question Answering system and show
its impact.
References
Anscombe, G.E.M. 1957. Intention. Cornell University
Press, Ithaca, New York.
Audi, Robert. 1973. Intending. The Journal of Philoso-
phy, 70(13):387–403.
Bratman, Michael E. 1981. Intention and means-end
reasoning. The Philosophical Review, 90(2):252–265.
Bratman, Michael E. 1987. Intention, Plans, and Prac-
tical Reason. Harvard University Press, Cambridge,
Massachusetts.
Bratman, Michael E. 1990. What is intention? In Inten-
tions in Communication. MIT Press.
Miller, George A., Claudia Leacock, Randee Tengi, and
Ross T. Bunker. 1993. A semantic concordance. In
Proceedings of the ARPA Human Language Technol-
ogy Workshop
Miller, George A. 1995. Wordnet: A lexical database.
Communication of the ACM, 38(11):39–41.
Pollack, Martha E. 1990. Plans as complex mental atti-
tudes. In Intentions in Communication. MIT Press.
Quinlan, Ross. 2004. Data Mining Tools See5 and C5.0.
http://www.rulequest.com/see5-info.html
Wiebe, Janyce M., Theresa Wilson, Rebecca F. Bruce,
Matthew Bell, and Melanie Martin. 2004. Learn-
ing subjective language. Computational Linguistics,
30(3):277–308.
36
