Proceedings of the Workshop on Multiword Expressions: Identifying and Exploiting Underlying Properties, pages 45–53,
Sydney, July 2006. c©2006 Association for Computational Linguistics
Classifying Particle Semantics in English Verb-Particle Constructions
Paul Cook
Department of Computer Science
University of Toronto
Toronto, ON M5S 3G4
Canada
pcook@cs.toronto.edu
Suzanne Stevenson
Department of Computer Science
University of Toronto
Toronto, ON M5S 3G4
Canada
suzanne@cs.toronto.edu
Abstract
Previous computational work on learning
the semantic properties of verb-particle
constructions (VPCs) has focused on their
compositionality, and has left unaddressed
the issue of which meaning of the compo-
nent words is being used in a given VPC.
We develop a feature space for use in clas-
sification of the sense contributed by the
particle in a VPC, and test this on VPCs
using the particle up. The features that
capture linguistic properties of VPCs that
are relevant to the semantics of the par-
ticle outperform linguistically uninformed
word co-occurrence features in our exper-
iments on unseen test VPCs.
1 Introduction
A challenge in learning the semantics of mul-
tiword expressions (MWEs) is their varying de-
grees of compositionality—the contribution of
each component word to the overall semantics
of the expression. MWEs fall on a range from
fully compositional (i.e., each component con-
tributes its meaning, as in frying pan) to non-
compositional or idiomatic (as in hit the roof ). Be-
cause of this variation, researchers have explored
automatic methods for learning whether, or the de-
gree to which, an MWE is compositional (e.g.,
Lin, 1999; Bannard et al., 2003; McCarthy et al.,
2003; Fazly et al., 2005).
However, such work leaves unaddressed the ba-
sic issue of which of the possible meanings of a
component word is contributed when the MWE is
(at least partly) compositional. Words are notori-
ously ambiguous, so that even if it can be deter-
mined that an MWE is compositional, its meaning
is still unknown, since the actual semantic contri-
bution of the components is yet to be determined.
We address this problem in the domain of verb-
particle constructions (VPCs) in English, a rich
source of MWEs.
VPCs combine a verb with any of a finite set
of particles, as in jump up, figure out, or give in.
Particles such as up, out, or in, with their literal
meaning based in physical spatial relations, show
a variety of metaphorical and aspectual meaning
extensions, as exemplified here for the particle up:
(1a) The sun just came up. [vertical spatial movement]
(1b) She walked up to him. [movement toward a goal]
(1c) Drink up your juice! [completion]
(1d) He curled up into a ball. [reflexive movement]
Cognitive linguistic analysis, as in Lindner (1981),
can provide the basis for elaborating this type of
semantic variation.
Given such a sense inventory for a particle,
our goal is to automatically determine its mean-
ing when used with a given verb in a VPC. We
classify VPCs according to their particle sense,
using statistical features that capture the seman-
tic and syntactic properties of verbs and particles.
We contrast these with simple word co-occurrence
features, which are often used to indicate the se-
mantics of a target word. In our experiments, we
focus on VPCs using the particle up because it is
highly frequent and has a wide range of meanings.
However, it is worth emphasizing that our feature
space draws on general properties of VPCs, and is
not specific to this particle.
A VPC may be ambiguous, with its particle oc-
curring in more than one sense; in contrast to (1a),
come up may use up in a goal-oriented sense as in
45
The deadline is coming up. While our long-term
goal is token classification (disambiguation) of a
VPC in context, following other work on VPCs
(e.g., Bannard et al., 2003; McCarthy et al., 2003),
we begin here with the task of type classification.
Given our use of features which capture the statis-
tical behaviour relevant to a VPC across a corpus,
we assume that the outcome of type classification
yields the predominant sense of the particle in the
VPC. Predominant sense identification is a useful
component of sense disambiguation of word to-
kens (McCarthy et al., 2004), and we presume our
VPC type classification work will form the basis
for later token disambiguation.
Section 2 continues the paper with a discussion
of the features we developed for particle sense
classification. Section 3 first presents some brief
cognitive linguistic background, followed by the
sense classes of up used in our experiments. Sec-
tions 4 and 5 discuss our experimental set-up and
results, Section 6 related work, and Section 7 our
conclusions.
2 Features Used in Classification
The following subsections describe the two sets of
features we investigated. The linguistic features
are motivated by specific semantic and syntactic
properties of verbs and VPCs, while the word co-
occurrence features are more general.
2.1 Linguistically Motivated Features
2.1.1 Slot Features
We hypothesize that the semantic contribution
of a particle when combined with a given verb is
related to the semantics of that verb. That is, the
particle contributes the same meaning when com-
bining with any of a semantic class of verbs.1 For
example, the VPCs drink up, eat up and gobble up
all draw on the completion sense of up; the VPCs
puff out, spread out and stretch out all draw on the
extension sense of out. The prevalence of these
patterns suggests that features which have been
shown to be effective for the semantic classifica-
tion of verbs may be useful for our task.
We adopt simple syntactic “slot” features which
have been successfully used in automatic seman-
tic classification of verbs (Joanis and Stevenson,
1Villavicencio (2005) observes that verbs from a seman-
tic class will form VPCs with similar sets of particles. Here
we are hypothesizing further that VPCs formed from verbs
of a semantic class draw on the same meaning of the given
particle.
2003). The features are motivated by the fact
that semantic properties of a verb are reflected
in the syntactic expression of the participants in
the event the verb describes. The slot features
encode the relative frequencies of the syntactic
slots—subject, direct and indirect object, object of
a preposition—that the arguments and adjuncts of
a verb appear in. We calculate the slot features
over three contexts: all uses of a verb; all uses of
the verb in a VPC with the target particle (up in our
experiments); all uses of the verb in a VPC with
any of a set of high frequency particles (to capture
its semantics when used in VPCs in general).
2.1.2 Particle Features
Two types of features are motivated by proper-
ties specific to the semantics and syntax of par-
ticles and VPCs. First, Wurmbrand (2000) notes
that compositional particle verbs in German (a
somewhat related phenomenon to English VPCs)
allow the replacement of their particle with seman-
tically similar particles. We extend this idea, hy-
pothesizing that when a verb combines with a par-
ticle such as up in a particular sense, the pattern
of usage of that verb in VPCs using all other par-
ticles may be indicative of the sense of the target
particle (in this case up) when combined with that
verb. To reflect this observation, we count the rel-
ative frequency of any occurrence of the verb used
in a VPC with each of a set of high frequency par-
ticles.
Second, one of the striking syntactic properties
of VPCs is that they can often occur in either the
joined configuration (2a) or the split configuration
(2b):
(2a) Drink up your milk! He walked out quickly.
(2b) Drink your milk up! He walked quickly out.
Bolinger (1971) notes that the joined construction
may be more favoured when the sense of the par-
ticle is not literal. To encode this, we calculate the
relative frequency of the verb co-occurring with
the particle up with each of a0 –a1 words between
the verb and up, reflecting varying degrees of verb-
particle separation.
2.2 Word Co-occurrence Features
We also explore the use of general context fea-
tures, in the form of word co-occurrence frequency
vectors, which have been used in numerous ap-
proaches to determining the semantics of a target
46
word. Note, however, that unlike the task of word
sense disambiguation, which examines the context
of a target word token to be disambiguated, here
we are looking at aggregate contexts across all in-
stances of a target VPC, in order to perform type
classification.
We adopt very simple word co-occurrence fea-
tures (WCFs), calculated as the frequency of any
(non-stoplist) word within a certain window left
and right of the target. We noted above that the
target particle semantics is related both to the se-
mantics of the verb it co-occurs with, and to the
occurrence of the verb across VPCs with different
particles. Thus we not only calculate the WCFs of
the target VPC (a given verb used with the parti-
cle up), but also the WCFs of the verb itself, and
the verb used in a VPC with any of the high fre-
quency particles. These WCFs give us a very gen-
eral means for determining semantics, whose per-
formance we can contrast with our linguistic fea-
tures.
3 Particle Semantics and Sense Classes
We give some brief background on cognitive
grammar and its relation to particle semantics, and
then turn to the semantic analysis of up that we
draw on as the basis for the sense classes in our
experiments.
3.1 Cognitive Grammar and Schemas
Some linguistic studies consider many VPCs to be
idiomatic, but do not give a detailed account of
the semantic similarities between them (Bolinger,
1971; Fraser, 1976; Jackendoff, 2002). In con-
trast, work in cognitive linguistics has claimed that
many so-called idiomatic expressions draw on the
compositional contribution of (at least some of)
their components (Lindner, 1981; Morgan, 1997;
Hampe, 2000). In cognitive grammar (Langacker,
1987), non-spatial concepts are represented as spa-
tial relations. Key terms from this framework are:
Trajector (TR) The object which is conceptually
foregrounded.
Landmark (LM) The object against which the
TR is foregrounded.
Schema An abstract conceptualization of an ex-
perience. Here we focus on schemas depict-
ing a TR, LM and their relationship in both
the initial configuration and the final config-
uration communicated by some expression.
TR
TR
LM LM
Initial Final
Figure 1: Schema for Vertical up.
The semantic contribution of a particle in a VPC
corresponds to a schema. For example, in sen-
tence (3), the TR is the balloon and the LM is the
ground the balloon is moving away from.
(3) The balloon floated up.
The schema describing the semantic contribution
of the particle in the above sentence is shown
in Figure 1, which illustrates the relationship be-
tween the TR and LM in the initial and final con-
figurations.
3.2 The Senses of up
Lindner (1981) identifies a set of schemas for each
of the particles up and out, and groups VPCs ac-
cording to which schema is contributed by their
particle. Here we describe the four senses of up
identified by Lindner.
3.2.1 Vertical up (Vert-up)
In this schema (shown above in Figure 1), the
TR moves away from the LM in the direction of
increase along a vertically oriented axis. This in-
cludes prototypical spatial upward movement such
as that in sentence (3), as well as upward move-
ment along an abstract vertical axis as in sen-
tence (4).
(4) The price of gas jumped up.
In Lindner’s analysis, this sense also includes ex-
tensions of upward movement where a vertical
path or posture is still salient. Note that in some of
these senses, the notion of verticality is metaphor-
ical; the contribution of such senses to a VPC may
not be considered compositional in a traditional
analysis. Some of the most common sense exten-
sions are given below, with a brief justification as
to why verticality is still salient.
47
Initial
TR
LM = goal LM = goal
TR
Final
Figure 2: Schema for Goal-Oriented up.
Up as a path into perceptual field. Spatially
high objects are generally easier to perceive.
Examples: show up, spring up, whip up.
Up as a path into mental field. Here up encodes
a path for mental as opposed to physical objects.
Examples: dream up, dredge up, think up.
Up as a path into a state of activity. Activity is
prototypically associated with an erect posture.
Examples: get up, set up, start up.
3.2.2 Goal-Oriented up (Goal-up)
Here the TR approaches a goal LM; movement
is not necessarily vertical (see Figure 2). Proto-
typical examples are walk up and march up. This
category also includes extensions into the social
domain (kiss up and suck up), as well as exten-
sions into the domain of time (come up and move
up), as in:
(5a) The intern kissed up to his boss.
(5b) The deadline is coming up quickly.
3.2.3 Completive up (Cmpl-up)
Cmpl-up is a sub-sense of Goal-up in which the
goal represents an action being done to comple-
tion. This sense shares its schema with Goal-up
(Figure 2), but it is considered as a separate sense
since it corresponds to uses of up as an aspectual
marker. Examples of Cmpl-up are: clean up, drink
up, eat up, finish up and study up.
3.2.4 Reflexive up (Refl-up)
Reflexive up is a sub-sense of Goal-up in which
the sub-parts of the TR are approaching each other.
The schema for Refl-up is shown in Figure 3; it is
unique in that the TR and LM are the same object.
Examples of Refl-up are: bottle up, connect up,
couple up, curl up and roll up.
LM = TR LM = TR
Initial Final
Figure 3: Schema for Reflexive up.
Vertical up Goal-Oriented up
Completive up
Reflexive up
Figure 4: Simplified schematic network for up.
3.3 The Sense Classes for Our Study
Adopting a cognitive linguistic perspective, we as-
sume that all uses of a particle make some compo-
sitional contribution of meaning to a VPC. In this
work, we classify target VPCs according to which
of the above senses of up is contributed to the ex-
pression. For example, the expressions jump up
and pick up are designated as being in the class
Vert-up since up in these VPCs has the vertical
sense, while clean up and drink up are designated
as being in the class Cmpl-up since up here has
the completive sense. The relations among the
senses of up can be shown in a “schematic net-
work” (Langacker, 1987). Figure 4 shows a sim-
plification of such a network in which we connect
more similar senses with shorter edges. This type
of analysis allows us to alter the granularity of our
classification in a linguistically motivated fashion
by combining closely related senses. Thus we can
explore the effect of different sense granularities
on classification.
4 Materials and Methods
4.1 Experimental Expressions
We created a list of English VPCs using up, based
on a list of VPCs made available by McIntyre
(2001) and a list of VPCs compiled by two human
judges. The judges then filtered this list to include
only VPCs which they both agreed were valid, re-
sulting in a final list of 389 VPCs. From this list,
training, verification and test sets of sixty VPCs
each are randomly selected. Note that the expense
of manually annotating the data (as described be-
low) prevents us from using larger datasets in this
initial investigation. The experimental sets are
48
chosen such that each includes the same propor-
tion of verbs across three frequency bands, so that
the sets do not differ in frequency distribution of
the verbs. (We use frequency of the verbs, rather
than the VPCs, since many of our features are
based on the verb of the expression, and moreover,
VPC frequency is approximate.) The verification
data is used in exploration of the feature space and
selection of final features to use in testing; the test
set is held out for final testing of the classifiers.
Each VPC in each dataset is annotated by the
two human judges according to which of the four
senses of up identified in Section 3.2 is contributed
to the VPC. As noted in Section 1, VPCs may
be ambiguous with respect to their particle sense.
Since our task here is type classification, the
judges identify the particle sense of a VPC in its
predominant usage, in their assessment. The ob-
served inter-annotator agreement is a0a1a0a3a2 a0 for each
dataset. The unweighted observed kappa scores
are a0a1a0a5a4a7a6 , a0a1a0a3a8a10a9 and a0a1a0 a1 a1 , for the training, verifica-
tion and test sets respectively.
4.2 Calculation of the Features
We extract our features from the 100M word
British National Corpus (BNC, Burnard, 2000).
VPCs are identified using a simple heuristic based
on part-of-speech tags, similar to one technique
used by Baldwin (2005). A use of a verb is con-
sidered a VPC if it occurs with a particle (tagged
AVP) within a six word window to the right. Over
a random sample of 113 VPCs thus extracted, we
found 88% to be true VPCs, somewhat below the
performance of Baldwin’s (2005) best extraction
method, indicating potential room for improve-
ment.
The slot and particle features are calculated us-
ing a modified version of the ExtractVerb software
provided by Joanis and Stevenson (2003), which
runs over the BNC pre-processed using Abney’s
(1991) Cass chunker.
To compute the word co-occurrence features
(WCFs), we first determine the relative frequency
of all words which occur within a five word win-
dow left and right of any of the target expressions
in the training data. From this list we eliminate
the most frequent 1% of words as a stoplist and
then use the next a11 most frequent words as “fea-
ture words”. For each “feature word”, we then cal-
culate its relative frequency of occurrence within
the same five word window of the target expres-
#VPCs in Sense Class
Sense Class Train Verification Test
Vert-up 24 33 27
Goal-up 1 1 3
Cmpl-up 20 23 22
Refl-up 15 3 8
Table 1: Frequency of items in each sense class.
#VPCs in Sense Class
Sense Class Train Verification Test
Vert-up 24 33 27
Goal-up a12 21 24 25
Cmpl-up
Refl-up 15 3 8
Table 2: Frequency of items in each class for the
3-way task.
sions in all datasets. We use a11a14a13a16a15 a0 a0 and a11a14a13 a1 a0 a0
to create feature sets WCFa17a19a18a19a18 and WCFa20a19a18a19a18 respec-
tively.
4.3 Experimental Classes
Table 1 shows the distribution of senses in each
dataset. Each of the training and verification sets
has only one VPC corresponding to Goal-up. Re-
call that Goal-up shares a schema with Cmpl-up,
and is therefore very close to it in meaning, as in-
dicated spatially in Figure 4. We therefore merge
Goal-up and Cmpl-up into a single sense, to pro-
vide more balanced classes.
Since we want to see how our features per-
form on differing granularities of sense classes, we
run each experiment as both a 3-way and 2-way
classification task. In the 3-way task, the sense
classes correspond to the meanings Vert-up, Goal-
up merged with Cmpl-up (as noted above), and
Refl-up, as shown in Table 2. In the 2-way task, we
further merge the classes corresponding to Goal-
#VPCs in Sense Class
Sense Class Train Verification Test
Vert-up 24 33 27
Goal-up a12 36 27 33
Cmpl-up a12
Refl-up
Table 3: Frequency of items in each class for the
2-way task.
49
up/Cmpl-up with that of Refl-up, as shown in Ta-
ble 3. We choose to merge these classes because
(as illustrated in Figure 4) Refl-up is a sub-sense of
Goal-up, and moreover, all three of these senses
contrast with Vert-up, in which increase along a
vertical axis is the salient property. It is worth em-
phasizing that the 2-way task is not simply a clas-
sification between literal and non-literal up—Vert-
up includes extensions of up in which the increase
along a vertical axis is metaphorical.
4.4 Evaluation Metrics and Classifier
Software
The variation in the frequency of the sense classes
of up across the datasets makes the true distri-
bution of the classes difficult to estimate. Fur-
thermore, there is no obvious informed baseline
for this task. Therefore, we make the assumption
that the true distribution of the classes is uniform,
and use the chance accuracy a0a2a1a4a3 as the baseline
(where a3 is the number of classes—in our exper-
iments, either a15 or a6 ). Accordingly, our measure
of classification accuracy should weight each class
evenly. Therefore, we report the average per class
accuracy, which gives equal weight to each class.
For classification we use LIBSVM (Chang and
Lin, 2001), an implementation of a support-vector
machine. We set the input parameters, cost
and gamma, using 10-fold cross-validation on the
training data. In addition, we assign a weight of
a5 a6a8a7a10a9a12a11a14a13a16a15a18a17a20a19a22a21 a7a10a15a18a15a23a5
a5a24a19a22a21 a7a10a15a18a15a26a25a23a5 to each class
a27 to eliminate the ef-
fects of the variation in class size on the classifier.
Note that our choice of accuracy measure and
weighting of classes in the classifier is necessary
given our assumption of a uniform random base-
line. Since the accuracy values we report incorpo-
rate this weighting, these results cannot be com-
pared to a baseline of always choosing the most
frequent class.
5 Experimental Results
We present experimental results for both
Ver(ification) and unseen Test data, on each
set of features, individually and in combination.
All experiments are run on both the 2-way and
3-way sense classification, which have a chance
baseline of 50% and 33%, respectively.
3-way Task 2-way Task
Features Ver Test Ver Test
Slots 41 51 53 67
Particles 37 33 65 47
Slots a12 Particles 54 54 59 63
Table 4: Accuracy (%) using linguistic features.
5.1 Experiments Using the Linguistic
Features
The results for experiments using the features that
capture semantic and syntactic properties of verbs
and VPCs are summarized in Table 4, and dis-
cussed in turn below.
5.1.1 Slot Features
Experiments using the slot features alone test
whether features that tap into semantic informa-
tion about a verb are sufficient to determine the
appropriate sense class of a particle when that verb
combines with it in a VPC. Although accuracy on
the test data is well above the baseline in both the
2-way and 3-way tasks, for verification data the
increase over the baseline is minimal. The class
corresponding to sense Refl-up in the 3-way task
is relatively small, which means that a small vari-
ation in classification on these verbs may lead to
a large variation in accuracy. However, we find
that the difference in accuracy across the datasets
is not due to performance on VPCs in this sense
class. Although these features show promise for
our task, the variation across the datasets indicates
the limitations of our small sample sizes.
5.1.2 Particle Features
We also examine the performance of the parti-
cle features on their own, since to the best of our
knowledge, no such features have been used be-
fore in investigating VPCs. The results are dis-
appointing, with only the verification data on the
2-way task showing substantially higher accuracy
than the baseline. An analysis of errors reveals no
consistent explanation, suggesting again that the
variation may be due to small sample sizes.
5.1.3 Slot + Particle Features
We hypothesize that the combination of the slot
features with the particle features will give an in-
crease in performance over either set of linguis-
tic features used individually, given that they tap
into differing properties of verbs and VPCs. We
find that the combination does indeed give more
50
3-way Task 2-way Task
Features Ver Test Ver Test
WCFa17a19a18a19a18 45 42 59 51
WCFa20a19a18a19a18 38 34 55 48
Table 5: Accuracy (%) using WCFs.
consistent performance across verification and test
data than either feature set used individually. We
analyze the errors made using slot and particle fea-
tures separately, and find that they tend to classify
different sets of verbs incorrectly. Therefore, we
conclude that these feature sets are at least some-
what complementary. By combining these com-
plementary feature sets, the classifier is better able
to generalise across different datasets.
5.2 Experiments Using WCFs
Our goal was to compare the more knowledge-rich
slot and particle features to an alternative feature
set, the WCFs, which does not rely on linguistic
analysis of the semantics and syntax of verbs and
VPCs. Recall that we experiment with both 200
feature words, WCFa17a19a18a19a18 , and 500 feature words,
WCFa20a19a18a19a18 , as shown in Table 5. Most of the exper-
iments using WCFs perform worse than the cor-
responding experiment using all the linguistic fea-
tures. It appears that the linguistically motivated
features are better suited to our task than simple
word context features.
5.3 Linguistic Features and WCFs Combined
Although the WCFs on their own perform worse
than the linguistic features, we find that the lin-
guistic features and WCFs are at least somewhat
complementary since they tend to classify differ-
ent verbs incorrectly. We hypothesize that, as with
the slot and particle features, the different types
of information provided by the linguistic features
and WCFs may improve performance in combina-
tion. We therefore combine the linguistic features
with each of the WCFa17a19a18a19a18 and WCFa20a19a18a19a18 features;
see Table 6. However, contrary to our hypothesis,
for the most part, the experiments using the full
combination of features give accuracies the same
or below that of the corresponding experiment us-
ing just the linguistic features. We surmise that
these very different types of features—the linguis-
tic features and WCFs—must be providing con-
flicting rather than complementary information to
the classifier, so that no improvement is attained.
3-way Task 2-way Task
Features Ver Test Ver Test
Combineda17a19a18a19a18 53 45 63 53
Combineda20a19a18a19a18 54 46 65 49
Table 6: Accuracy (%) combining linguistic fea-
tures with WCFs.
5.4 Discussion of Results
The best performance across the datasets is at-
tained using all the linguistic features. The lin-
guistically uninformed WCFs perform worse on
their own, and do not consistently help (and in
some cases hurt) the performance of the linguis-
tic features when combined with them. We con-
clude then that linguistically based features are
motivated for this task. Note that the features are
still quite simple, and straightforward to extract
from a corpus—i.e., linguistically informed does
not mean expensive (although the slot features do
require access to chunked text).
Interestingly, in determining the semantic near-
est neighbor of German particle verbs, Schulte im
Walde (2005) found that WCFs that are restricted
to the arguments of the verb outperform simple
window-based co-occurrence features. Although
her task is quite different from ours, similarly re-
stricting our WCFs may enable them to encode
more linguistically-relevant information.
The accuracies we achieve with the linguistic
features correspond to a 30–31% reduction in er-
ror rate over the chance baseline for the 3-way
task, and an 18–26% reduction in error rate for
the 2-way task. Although we expected that the
2-way task may be easier, since it requires less
fine-grained distinctions, it is clear that combining
senses that have some motivation for being treated
separately comes at a price.
The reductions in error rate that we achieve with
our best features are quite respectable for a first
attempt at addressing this problem, but more work
clearly remains. There is a relatively high variabil-
ity in performance across the verification and test
sets, indicating that we need a larger number of
experimental expressions to be able to draw firmer
conclusions. Even if our current results extend to
larger datasets, we intend to explore other feature
approaches, such as word co-occurrence features
for specific syntactic slots as suggested above, in
order to improve the performance.
51
6 Related Work
The semantic compositionality of VPC types has
recently received increasing attention. McCarthy
et al. (2003) use several measures to automati-
cally rate the overall compositionality of a VPC.
Bannard (2005), extending work by Bannard et al.
(2003), instead considers the extent to which the
verb and particle each contribute semantically to
the VPC. In contrast, our work assumes that the
particle of every VPC contributes composition-
ally to its meaning. We draw on cognitive lin-
guistic analysis that posits a rich set of literal and
metaphorical meaning possibilities of a particle,
which has been previously overlooked in compu-
tational work on VPCs.
In this first investigation of particle meaning in
VPCs, we choose to focus on type-based clas-
sification, partly due to the significant extra ex-
pense of manually annotating sufficient numbers
of tokens in text. As noted earlier, though, VPCs
can take on different meanings, indicating a short-
coming of type-based work. Patrick and Fletcher
(2005) classify VPC tokens, considering each as
compositional, non-compositional or not a VPC.
Again, however, it is important to recognize which
of the possible meaning components is being con-
tributed. In this vein, Uchiyama et al. (2005)
tackle token classification of Japanese compound
verbs (similar to VPCs) as aspectual, spatial, or
adverbial. In the future, we aim to extend the
scope of our work, to determine the meaning of
a particle in a VPC token, along the lines of our
sense classes here. This will almost certainly re-
quire semantic classification of the verb token (La-
pata and Brew, 2004), similar to our approach here
of using the semantic class of a verb type as indica-
tive of the meaning of a particle type.
Particle semantics has clear relations to prepo-
sition semantics. Some research has focused on
the sense disambiguation of specific prepositions
(e.g., Alam, 2004), while other work has classi-
fied preposition tokens according to their seman-
tic role (O’Hara and Wiebe, 2003). Moreover,
two large lexical resources of preposition senses
are currently under construction, The Preposi-
tion Project (Litkowski, 2005) and PrepNet (Saint-
Dizier, 2005). These resources were not suitable
as the basis for our sense classes because they do
not address the range of metaphorical extensions
that a preposition/particle can take on, but future
work may enable larger scale studies of the type
needed to adequately address VPC semantics.
7 Conclusions
While progress has recently been made in tech-
niques for assessing the compositionality of VPCs,
work thus far has left unaddressed the problem of
determining the particular meaning of the compo-
nents. We focus here on the semantic contribution
of the particle—a part-of-speech whose seman-
tic complexity and range of metaphorical mean-
ing extensions has been largely overlooked in prior
computational work. Drawing on work within
cognitive linguistics, we annotate a set of 180
VPCs according to the sense class of the particle
up, our experimental focus in this initial investiga-
tion. We develop features that capture linguistic
properties of VPCs that are relevant to the seman-
tics of particles, and show that they outperform
linguistically uninformed word co-occurrence fea-
tures, achieving around 20–30% reduction in er-
ror rate over a chance baseline. Areas of on-going
work include development of a broader range of
features, consideration of methods for token-based
semantic determination, and creation of larger ex-
perimental datasets.

References
S. Abney. 1991. Parsing by chunks. In R. Berwick,
S. Abney, and C. Tenny, editors, Principle-
Based Parsing: Computation and Psycholin-
guistics, p. 257–278. Kluwer Academic Pub-
lishers.
Y. S. Alam. 2004. Decision trees for sense dis-
ambiguation of prepositions: Case of over. In
HLT-NAACL 2004: Workshop on Computa-
tional Lexical Semantics, p. 52–59.
T. Baldwin. 2005. The deep lexical acquisition of
English verb-particle constructions. Computer
Speech and Language, Special Issue on Multi-
word Expressions, 19(4):398–414.
C. Bannard. 2005. Learning about the meaning of
verb-particle constructions from corpora. Com-
puter Speech and Language, Special Issue on
Multiword Expressions, 19(4):467–478.
C. Bannard, T. Baldwin, and A. Lascarides. 2003.
A statistical approach to the semantics of verb-
particles. In Proceedings of the ACL-2003
Workshop on Multiword Expressions: Analysis,
Acquisition and Treatment, p. 65–72.
D. Bolinger. 1971. The Phrasal Verb in English.
Harvard University Press.
L. Burnard. 2000. The British National Cor-
pus Users Reference Guide. Oxford University
Computing Services.
C.-C. Chang and C.-J. Lin. 2001. LIBSVM: a
library for support vector machines. Soft-
ware available at http://www.csie.ntu.
edu.tw/˜cjlin/libsvm.
A. Fazly, R. North, and S. Stevenson. 2005. Au-
tomatically distinguishing literal and figurative
usages of highly polysemous verbs. In Proceed-
ings of the ACL-2005 Workshop on Deep Lexi-
cal Acquisition.
B. Fraser. 1976. The Verb-Particle Combination in
English. Academic Press.
B. Hampe. 2000. Facing up to the meaning of
‘face up to’: A cognitive semantico-pragmatic
analysis of an English verb-particle construc-
tion. In A. Foolen and F. van der Leek, edi-
tors, Constructions in Cognitive Linguistics. Se-
lected Papers from the fifth International Cog-
nitive Linguistics Conference, p. 81–101. John
Benjamins Publishing Company.
R. Jackendoff. 2002. English particle construc-
tions, the lexicon, and the autonomy of syntax.
In N. Dehe, R. Jackendoff, A. McIntyre, and
S. Urban, editors, Verb-Particle Explorations.
Mouton de Gruyter.
E. Joanis and S. Stevenson. 2003. A general fea-
ture space for automatic verb classification. In
Proceedings of the Conference of the European
Chapter of the Association for Computational
Linguistics (EACL-2003), p. 163–170.
R. W. Langacker. 1987. Foundations of Cognitive
Grammar: Theoretical Prerequisites, volume 1.
Stanford University Press, Stanford.
M. Lapata and C. Brew. 2004. Verb class disam-
biguation using informative priors. Computa-
tional Linguistics, 30(1):45–73.
D. Lin. 1999. Automatic identification of non-
compositional phrases. In Proceedings of the
37th Annual Meeting of the Association for
Computational Linguistics, p. 317–324.
S. Lindner. 1981. A lexico-semantic analysis of
English verb particle constructions with out and
up. Ph.D. thesis, University of California, San
Diego.
K. C. Litkowski. 2005. The Preposition Project. In
Proceedings of the Second ACL-SIGSEM Work-
shop on the Linguistic Dimensions of Preposi-
tions and their Use in Computational Linguis-
tics Formalisms and Applications.
D. McCarthy, B. Keller, and J. Carroll. 2003.
Detecting a continuum of compositionality in
phrasal verbs. In Proceedings of the ACL-
SIGLEX Workshop on Multiword Expressions:
Analysis, Acquisition and Treatment.
D. McCarthy, R. Koeling, J. Weeds, and J. Carroll.
2004. Finding predominant word senses in un-
tagged text. In Proceedings of the 42nd Annual
Meeting of the Association for Computational
Linguistics, p. 280–287.
A. McIntyre. 2001. The particle verb list.
http://www.uni-leipzig.de/
˜angling/mcintyre/pv.list.pdf.
P. S. Morgan. 1997. Figuring out figure out:
Metaphor and the semantics of the English
verb-particle construction. Cognitive Linguis-
tics, 8(4):327–357.
T. O’Hara and J. Wiebe. 2003. Preposition se-
mantic classification via Penn Treebank and
FrameNet. In Proceedings of CoNLL-2003, p.
79–86.
J. Patrick and J. Fletcher. 2005. Classifying verb-
particle constructions by verb arguments. In
Proceedings of the Second ACL-SIGSEM Work-
shop on the Linguistic Dimensions of Preposi-
tions and their use in Computational Linguistics
Formalisms and Applications, p. 200–209.
P. Saint-Dizier. 2005. PrepNet: a framework for
describing prepositions: Preliminary investiga-
tion results. In Proceedings of the Sixth Interna-
tional Workshop on Computational Semantics
(IWCS’05), p. 145–157.
S. Schulte im Walde. 2005. Exploring features to
identify semantic nearest neighbours: A case
study on German particle verbs. In Proceed-
ings of the International Conference on Recent
Advances in Natural Language Processing.
K. Uchiyama, T. Baldwin, and S. Ishizaki.
2005. Disambiguating Japanese compound
verbs. Computer Speech and Language, Special
Issue on Multiword Expressions, 19(4):497–
512.
A. Villavicencio. 2005. The availability of verb-
particle constructions in lexical resources: How
much is enough? Computer Speech and Lan-
guage, Special Issue on Multiword Expressions,
19(4):415–432.
S. Wurmbrand. 2000. The structure(s) of particle
verbs. Master’s thesis, McGill University.
