Supervised Classification of Aspectual Verb Classes in German
Subcategorization-Frame-Based vs Window-Based Approach: A Comparison
J
¨
urgen Hermes
1
, Michael Richter
2
and Claes Neuefeind
1
1
Institute for Digital Humanities, Cologne University, Albertus-Magnus-Platz, Cologne, Germany
2
Department of Automatic Language Processing, Leipzig University, Augustusplatz 10, Leipzig, Germany
Keywords:
Machine Learning, Classification, Aspectual Verb Classes, Subcategorization Frame based Approach,
Window based Approach.
Abstract:
The present study examines the results of two experiments: the aspectual classification of German verbs
within a window-based distributional framework and the classification within a subcategorization-frame-based
framework. The predictive power of pure, unstructured co-occurrences of verbs is compared against that of
linguistically motivated, well defined co-occurrences which we denote as informed distributional framework.
Using a support vector machine classifier and a classification into an extended Vendler classification (Vendler,
1967) as the gold standard, we observe excellent results in both frameworks which perform almost on a par.
1 INTRODUCTION
This paper presents experiments on the automatic
assignment of German verbs to aspectual classes
(Vendler, 1967) using corpus data in a distributional
framework (Harris 1954, see also Rubenstein and
Goodenough 1965, Sch
¨
utze and Pedersen 1993, Lan-
dauer and Dumais 1997, Pantel 2005, Turney and
Pantel 2010). The primary motivation for undertaking
this research is the (almost complete) lack of studies
on the automatic aspectual classification in typolog-
ical research on German. In our study, we compare
two approaches: i. an aspectual verb classification in
a framework which utilizes subcategorization frames
of verbs (henceforth informed distributional frame-
work), extracting classified nouns in the argument po-
sitions of verbs (Richter and Hermes 2015, Hermes
et al. 2015) and ii. an aspectual verb classification
in a purely distributional framework, thus considering
co-occurrences of all types.
We pose the question of whether classification in
a purely distributional framework would yield better
classification results than classification in a linguis-
tically well grounded informed distributional frame-
work. The former approach employs verb vectors
of very high dimensionality consisting of a consid-
erably higher amount of linguistic material than the
vectors used in an informed distributional framework.
This could be a point in favor of the purely distribu-
tional framework. On the other hand, the studies of
Richter and Hermes (2015) and Hermes et al. (2015)
with sets of 35 and 95 German verbs, respectively,
achieved promising classification results within an in-
formed distributional framework. A classification in-
spired by Vendler was used as gold standard. This
classification is the extension of the Vendlerian ty-
pology through the addition of one class (henceforth
Vendler + 1): The additional aspectual class accom-
plishments with an affected subject. The studies men-
tioned above (see also Richter and van Hout 2016)
provide evidence for this class. The accomplishments
with an affected subject class differs from the classi-
cal accomplishments in the semantic role of the sub-
ject. Instead of exclusively assigning the agent role to
the subject, the subject in the accomplishments with
an affected subject class is assigned both a patient
role and an agent role. Consider verbs such as drink,
where an agent subject also has the semantic prop-
erties of a patient since the drinker-agent is affected
and undergoes a change of state (temporarily puts on
weight, gets drunk etc.). Naess (2007) refers to these
semantic roles as volitional undergoers.
In the present study, we exclusively focus on lex-
ical aspect that is, aspectual properties of bare verbs
(or the fundamental aspectual category in the termi-
nology of Siegel and McKeown 2000). Thus, aspec-
tual properties of sentences and VPs as results of as-
pectual coercion or aspectual shift, respectively, are
not subject of this study, rather it is our aim to pre-
dict the aspectual classes of verbs from their con-
Hermes, J., Richter, M. and Neuefeind, C.
Supervised Classification of Aspectual Verb Classes in German - Subcategorization-Frame-Based vs Window-Based Approach: A Comparison.
DOI: 10.5220/0006728106530662
In Proceedings of the 10th International Conference on Agents and Artificial Intelligence (ICAART 2018) - Volume 2, pages 653-662
ISBN: 978-989-758-275-2
Copyright © 2018 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
653

texts.
1
The aspectual classification of verb classes
from contexts might give indications how language
learners manage to build up aspectual verb classes in
their mental lexicon. From that perspective, research
on verb classes is vital due to their relevance for the
processing of natural language by human beings and,
in addition, for the theory of natural language acqui-
sition (see Tomasello 2000, Goldberg 1995, Naigles
et al. 1992, Naigles et al. 1995, Naigles et al. 1993,
Wittek 2002, Richter and van Hout 2013). Research
on aspectual classes is of particular relevance because
it models the temporal and causal structures of events
(see Vendler 1967, Dowty 1979, Dowty 1991, Roth-
stein 2004, Fernando 2004, Gruender 2008). Theory-
driven work by Klein (2009) and experimental stud-
ies by Siegel (1997) and Siegel and McKeown (2000)
highlight the potential of aspect for classifying lin-
guistic units such as verbs and documents.
2 RELATED WORK
There are hardly any studies which address automatic
classifications of the complete Vendlerian typology,
let alone studies which compare subcategorization-
frame-based against window-based approaches. By
focusing on tense forms of verbs, Klavans and
Chodorow (1992) determined gradual state-properties
of verbs. Siegel (1997) and Siegel and McKeown
(2000) classified verbs into states and events using
temporal and modal indicators from contexts such as
temporal adverbs, tense forms and manner and evalu-
ation adverbs. Zarcone and Lenci (2008) presented an
automatic classification of the four Vendlerian aspect
classes in Italian utilizing, amongst others, syntactic
and semantic features of the arguments of the target
verbs and verb tense. The authors, however, aimed
at modelling aspectual shift and consequently focus
on aspectual properties of sentences, decomposing
the components of sentential event types. Friedrich
1
As an example of aspectual coercion, consider an atelic
verb such as walk, which can be combined with a PP denot-
ing a destination as in he walks to the store and expressing
a telic event. The sentence walks to the store is no longer an
activity, instead, it is an accomplishment. Aspectual coer-
cion can also be triggered by quantification Krifka (1989).
A prototypical accomplishment verb such as kill can occur
in a sentence expressing an activity, as in he is killing carpet
moths(note the present progressive form of the verb) which
stands classical tests of activities, e.g. he is killing carpet
moths for an hour, permanently/forever. The direct object is
a bare plural, expressing cumulative objects (Quine, 1960)
which combine well with atelic verbs. With a quantized di-
rect object (Krifka, 1989) however, the sentence is clearly
telic: he kills two carpet moths in one hour/*for hours.
and Palmer (2014) presented an automatic classifica-
tion of aspectual verb classes in English using con-
textual features including tense forms, albeit only dis-
tinguishing between stative, dynamic and mixed type
verbs.
Studies on the automatic assignment of non-
aspectual verb classes within a distributional frame-
work from Dorr and Jones (1996), Merlo and Steven-
son (2001), Joanis et al. (2008), Vlachos et al. (2009),
Schulte im Walde and Brew (2002), Schulte im Walde
(2003) and Schulte im Walde (2006) for German
verbs provide corpus based evidence that argument
frames, syntactic subcategorization information and,
in addition, aspect (Joanis et al., 2008) are reliable
predictors.
3 VENDLER’S TYPOLOGY
Vendler (1967) defines four aspectual classes
2
:
States, activities, achievements and accomplishments,
based on the time schemata of verbs and verb phrases.
He gives the following illustrative examples (Vendler,
1967, 149): Activities such as A was running at time
t are true if the instantiation of t is on a time stretch
throughout A was running. An accomplishment such
as A was drawing the circle at t is true if t is on the
time stretch in which A drew that circle. An achieve-
ment such as A won the race between t
1
and t
2
means
that the time at which A won that race is between t
1
and t
2
; and a statement such as A loves somebody from
t
1
and t
2
means that at any instant between t
1
and t
2
A
loved that person.
In (1) below we give Pustejovsky (1991)’s de-
scription of the four Vendler classes and in addition
a description of the additional class accomplishments
with an affected subject which extends the description
of the accomplishment class with a subject variable.
In line with Vendler (1967), Bach (1986), and Dowty
(1979), Pustejovsky (1991) distinguished the event
types process, state and transition (see also Jackend-
off 1972, Lakoff 1970 and Wright 1963). The latter
is a function from any event type to its opposite. For
instance, x closes the door expresses a transition from
an event type e
1
, the open door, to an event type e
2
, the
closed door, by acting of agent x and e
2
is the opposite
of e
1
(¬e
1
). The combinatorial variations within the
three event types process, state and transition allows
for the formal description of the complete Vendlerian
typology.
2
The Vendlerian quadripartition has been modified and
extended: Dowty (1979) added degree achievements, Smith
(1991) added semelfactices, Verkuyl (2005) in contrast de-
fined a tripartition consisting of states, processes and events.
NLPinAI 2018 - Special Session on Natural Language Processing in Artificial Intelligence
654

The abbreviation ’LCS’ in (1) means the lexical
conceptual structure which gives a decomposition of
predicates (Dowty 1991, Jackendoff 1983, Levin and
Rapoport 1992, Pustejovsky 1991). Hence, LCS is
the minimal decomposed event structure of verbs.
(1) Accomplishments
ES : process
transition
−−−−−→ state
LCS
0
: [act(x, y)&¬Q(y)] [Q(y)]
LCS : cause([act(x, y)], become(Q(y)))
Achievements
ES : process
transition
−−−−−→ state
LCS
0
: [¬P(y)] [P(y)]
LCS : become(P(y))
Activities
ES : process
LCS
0
: [act(x)]
LCS : act(x)
States
ES : state
LCS
0
: [Q(x)]
LCS : Q(x)
Accomplishments
with an affected subject
ES : process
transition
−−−−−→ state
LCS
0
: [act(x, y)&¬P(x)&¬Q(y)] [P(x)&Q(y)]
LCS : cause([act(x, y)], become(P(x), become(Q(y)))
The classes Accomplishments and accomplish-
ments with an affected subject can be interpreted as a
combination of activities and achievements. The lat-
ter express a result, the former a result preceded by an
activity. Pulman (1997) gives a slightly different de-
scription of the achievements class: he formulates a
transition from the event type point instead of process
to state where point is an atomic event whose internal
temporal structure (if it may have any) is ignored.
3
4 METHODOLOGY
In order to answer the question of whether informed
distributional methods perform better with regards to
3
Consider the achievement verb find. According to Pul-
man (1997) the event of finding has a resultant state, the
finding itself however is an atomic event, ignoring a possi-
ble complex event structure consisting for instance of dis-
covering something on the ground, taking a decision to pick
it up, bending down etc.
the prediction of aspectual verb classes than distri-
butional methods on shallow input, we tested differ-
ent workflows to classify a selection of 95 common
German verbs taken from Schumacher (1986). Schu-
macher defines seven lexical semantic macrofields:
Verben der allgemeinen Existenz (verbs of general ex-
istence), Verben der speziellen Existenz (verbs of spe-
cial existence), Verben des sprachlichen Ausdrucks
(verbs of linguistic expression), Verben der Differenz
(verbs of difference), Verben der Relation und des
geistigen Handelns (verbs of relation and mental pro-
cessing), Verben des Handlungsspielraums (verbs of
freedom of action) and Verben der vitalen Bedrfnisse
(verbs of vital needs). The macrofields are split into
30 subfields. We chose the verbs randomly from the
thirty subfields, the only criterion being the inclusion
of every subfield in order to cover the complete se-
mantic range of Schumachers typology.
Figure 1 shows the workflow of our analyses,
starting with raw sentence data taken from the
SdeWaC corpus (Faaß and Eckert, 2013) at the top,
to sets of classified verbs at the bottom. The different
methods are applied in four experiments, each
employing an individual process chain combining a
different set of components. The four experiments
are depicted as varying paths in figure 1, all starting
at the top (sDeWaC) but ending with four different
sets of verb classes at the bottom.
An overview showing which method combination
is unique in each experiment is given in table 1.
Table 1: Combination of workflow elements.
Combination - N Cluster + N Cluster
- Parsed Input 1 2
+ Parsed Input 3 4
We implemented each process chain on the basis
of combined and configurable components within the
workflow management tool Tesla
4
so that every ex-
periment performed can be reproduced by other re-
searchers.
For the classification of the 95 verbs, we used a
Support Vector Machine classifier (Joachims, 1998)
with a non-linear kernel. For 35 verbs, we adapted
the aspectual classification as the gold standard which
was validated in Richter and van Hout (2016), i.e.
Vendler + 1, and we assigned 60 verbs to aspectual
4
Tesla (Text Engineering Software LAboratory), see
http://tesla.spinfo.uni-koeln.de is an open source virtual re-
search environment, integrating both a visual editor for con-
ducting text-engineering experiments and a Java IDE for
developing software components (Hermes and Schwiebert,
2009).
Supervised Classification of Aspectual Verb Classes in German - Subcategorization-Frame-Based vs Window-Based Approach: A
Comparison
655

Figure 1: Overview of the complete workflow of all performed experiments. The experiments were realized via 4 process
chains which can be identified by the numbers at the bottom of the figure. Each process chain uses a different combination of
components (see table 1 for an overview and text for detailed explanations.
classes applying the criteria in Richter and van Hout
(2016). We trained the SVM using this aspectual clas-
sification as training data and tested it with a 10-fold
cross-validation: The data were randomly split into a
training and a test set (proportion 90 percent (train-
ing set), 10 percent (test set). The classifier was both
trained and tested on each of the 10 combinations of
training and test set. Thus, in total we obtained 10 sets
of class predictions and took the mean accuracy as fi-
nal result. We used a SVM classifier with a polyno-
mial kernel (which turned out to outperform other ker-
nel types) and a multiclass classifier (instead of train-
NLPinAI 2018 - Special Session on Natural Language Processing in Artificial Intelligence
656

ing a single classifier per class, we employed just one
classifier for the complete set of our aspectual classes.
We give some examples of the verb classes of the
aspectual gold standard classification below; the com-
plete list of all 95 pre-classified verbs used in this
study can be found in the appendix.
1. Accomplishments: aufbauen auf (to build on / to
be based on), herstellen (to produce), schneiden
(to cut), zers
¨
agen (to saw into pieces), verl
¨
angern
(to extend), mitteilen (to tell / to inform), bermit-
teln (to communicate / to forward), verhindern
(to prevent), abgrenzen (mark off / to define),
ver
¨
andern (to change)
2. Accomplishments with affected subject: un-
tersuchen (to examine), bedenken (to consider),
er
¨
ortern (to debate), nachpr
¨
ufen (to ascertain /
to check), aufessen (to eat up), essen (to eat),
beachten (to note), kaufen (to buy)
3. Activities: laufen (to walk / to run), eingehen auf
(to respond to so. / sth.), hmmern (to hammer),
sich orientieren an (to be geared to) ansteigen (to
increase), fallen (to fall), richten auf (to direct to-
wards / to focus), denken (to think), stattfinden (to
take place), wachsen (to grow)
4. Achievements: einschlafen (to fall asleep), verge-
hen (to go (by) / to pass / to diasappear),
¨
ubersehen (to overlook), verlieren (to lose), an-
fangen (to begin), abweichen (to deviate)
5. States: existieren (to exist), fehlen (to lack),
m
¨
ussen (to must), halten fr (to take so. / sth. for
so./ sth.), folgen aus (to follow from), angeh
¨
oren
(to belong to),
¨
ubereinstimmen (to agree), betre-
ffen (to concern), sein (to be), vorherrschen (to
predominate)
4.1 Classification of Verbs using
Co-occurence Vectors
The first experiment (marked with 1, left side of fig-
ure 1) is taken as a benchmark for purely distribu-
tional methods: We extracted 2000 sentences for each
verb from the SdeWaC corpus and collected the most
frequent co-occurrences using the frequency-based
heuristics described in Levy and Bullinaria (2001),
simply taking the k most frequent types of our cor-
pus as vector features.
5
The vectors were com-
5
We decided for the heuristics because of economy con-
siderations (Ockhams razor), giving preference to the sim-
pler method that performs on a par with more complex
ones: As Levy and Bullinaria (2001) show in their paper,
performance in tasks like synonym detection is compara-
ble to more sophisticated methods of feature selection, such
puted in three different configurations. As a base-
line, we first took the 200 most frequently occurring
elements (mostly closed class function words such
as und and, zu to, weil because / since, etc.), and a
context window of size 1 (henceforth 1-200), accept-
ing only the direct neighbors as co-occurrences. In
the second configuration, co-occurrences were com-
puted against the 2000 most frequently occurring el-
ements within a fixed context window of 5 items to
both sides (henceforth 5-2k). In addition, we em-
ployed a positional weighting scheme using the HAL
model (Hyperspace Analogue to Language, Lund and
Burgess 1996). In a third configuration, we took the
10.000 most frequently occurring words and a win-
dow size of 10 words (henceforth 10-10k), again us-
ing the HAL-weighting scheme. While the restriction
to function words within a narrow window mainly
reflects grammar-related distributional properties, the
consideration of content words in combination with a
broader window and position weighting emphasizes
the more semantically oriented aspects of their distri-
bution. The resulting verb vectors were normalized
and weighted with the TF-IDF measure before they
were passed to the final classification step.
4.2 Classification of Verbs using
Co-occurence Vectors with Reduced
Dimensionality
As a first step towards more informed methods,
we restricted the vector features to nominal co-
occuurrences (tagged as NN and NE in the SdeWaC
corpus). In order to reduce the feature space and
to increase the allocation density of the vectors, we
clustered all of these nominal co-occurrences. At this
stage we set foot on path 2 in figure 1. Here, we com-
puted co-occurrence vectors based on the same sub-
set of the SdeWaC corpus that we used to determine
the verb features, again using frequency-based feature
selection. The resulting vectors (10-10k) were again
weighted by the TF-IDF measure and passed to the
cluster analysis.
For cluster analysis we used three different imple-
mentations from the ELKI Data Mining API
6
, namely
KmeansLloyd with cluster sizes of k = 10.
as taking the most variant elements (see Lund and Burgess
1996), the most ’reliable’ (see Lowe and McDonald 2000,
or to perform a dimensionality reduction (e.g. by singular
value decomposition as done in the LSA model, see Lan-
dauer and Dumais 1997).
6
The open source framework ELKI (Environment
for DeveLoping KDD-Applications Supported by Index-
Structures) was developed at the LMU Munich, see
http://elki.dbs.ifi.lmu.de
Supervised Classification of Aspectual Verb Classes in German - Subcategorization-Frame-Based vs Window-Based Approach: A
Comparison
657

4.3 Classification of Verb Vectors with
Nominal Fillers and Aspectual
Features
For the remaining experiments (paths 3 and 4 in figure
1) we preprocessed the sentences from the SdeWaC
corpus with the Mate Dependency Parser
7
(Bohnet,
2010) to determine subjects and objects (accusative,
dative, and prepositional) for each verb and, in ad-
dition, to determine their aspectual features, which
Vendler (1967) suggested as a means of distinguish-
ing aspectual verb classes. To that end we collected
structures in the sentences such as adverbial fillers in
dependent or governing positions, which distinguish
the aspectual behaviors of the verbs being investi-
gated. The aspectual features are:
1. verb in imperative form
2. verb complex with aufh
¨
oren / stoppen (to stop / to
finish) as governing verbs,
3. verb complex with
¨
uberzeugen (to convince) as
governing verb,
4. matrix verb with time adverbials for durations,
like minutenlang (for minutes), in einer Minute
(in a minute),
5. matrix verb with time units, like Minute (minute),
Jahrhundert (century),
6. matrix verbs with seit (since), combined with unit
of time,
7. matrix verb with adverbials sorgf
¨
altig / mit
Sorgfalt (careful / with care),
8. matrix verb with adverbials absichtlich / mit Ab-
sicht (on purpose),
9. matrix verb with adverbials fast / beinahe (al-
most).
We generated vectors for each feature combina-
tion (subjects, direct objects, dative objects, preposi-
tional objects, and adverbials) in order to determine
which combination of fillers has the best predictive
power concerning the aspectual verb classes defined
by Vendler.
4.4 Classification of Verb Vectors with
Nominal Fillers and Aspectual
Features with Reduced
Dimensionality
Finally (path 4 in figure 1), we combined the informed
distributional method based on parsed input with the
7
See: https://code.google.com/p/mate-tools/
dimensionality reduction based on cluster analysis.
We constructed the verb vectors as described in chap-
ter 4.1 and clustered the nouns as described in chapter
4.2. Then we reassembled the verb vectors using the
generated noun classes from the noun clustering sub-
workflow. The verb vectors could then be reduced to
39 or fewer dimensions (up to nine aspectual features
complemented by ten clusters for each argument po-
sition).
5 RESULTS
The results are given in reverse order (from path 4 on
the right to path 1 on the left of figure 1, because the
method described in 4.4 was the starting point of our
study.
5.1 Results of the Classification of Verbs
using Dimension Reduced Nominal
Fillers and Aspectual Features
The workflow is a slightly modified version of the
workflow in Hermes et al. (2015). Thus, the results
were practically the same: Feature combinations ex-
clusively comprising aspect features yielded high ac-
curacy values (see figure 2).
Input to the classification were vectors with 39 di-
mensions (each 10 for subjects, direct objects, and
prepositional objects and 9 for the adverbial fea-
tures reflecting the aspectual behaviour). Taking the
classification with five aspectual verb classes as the
gold standard, ten noun classes per argument posi-
tion clearly outperform the approaches with fewer
features. Additionally, counting every noun token
leads to better results than counting only the noun
types. Medium length vectors (2000 dimensions),
constructed on the basis of a medium context width
(window size of five elements) achieve the best out-
comes; the verb vectors of the KMeansLloyd noun
clustering in particular show the best performance.
Figure 2 depicts the accuracy of feature combinations
and is subject to the following result description.
The combinations aspect/subject/direct object
and aspect/subject/ direct object/prepositional object
outperform the remaining feature combinations with
.95 accuracy, and .94 accuracy respectively. To
determine the significance values for the accuracy
level for the classification in five classes, we cal-
culated Cohens kappa. Kappa values above .81
are characterized as almost perfect agreement and
therefore highly significant. With the two feature
combinations described above we reached kappa
NLPinAI 2018 - Special Session on Natural Language Processing in Artificial Intelligence
658

values higher than .90. The feature combinations
aspect/direct object/prepositional object with .88
accuracy, κ = .84, and aspect/prepositional object
with .86 accuracy, κ = .81, also achieve almost
perfect agreements. Substantial agreements, with
above .61, can be observed with the combinations
aspect/subject/prepositional object, .84 accuracy, κ
= .78, aspect/direct object, .92 accuracy, κ = .75,
and aspect/subject, .82 accuracy, κ = .75. None of
the feature combinations without aspectual features
achieves a satisfactory result. The single features
achieve only fair agreements: Aspect achieves .57
accuracy, κ = .34, subject achieves .52, κ = .27, and
direct object and prepositional object achieve .51
accuracy and κ = .24 each.
Figure 2: Results of process chain 4.
5.2 Results of the Classification of Verbs
using Nominal Fillers and Aspectual
Features (no Dimension Reduction)
To reduce the complexity of our workflow, we classi-
fied the verbs using feature vectors where all nominal
fillers were taken into account instead of clustering
them into groups (processing chain 3 in figure 1).
Instead of the 39-dimensional vectors in process-
ing chain 4, we obtained vectors with almost 40000
dimensions (11257 subjects, 12450 direct objects,
16196 prepositional objects, 9 for adverbial fillers).
For the most part, the results were comparable to
the results of processing chain 4 (see figure 3): Em-
ploying all nominal and adverbial fillers (marked as
asdp in figure 3), we achieved .92 accuracy. In com-
parison with the +cluster-workflow we achieved the
same accuracy value when we left out the adverbial
fillers (sdp), but an even worse value when we left out
the objects (as). This is not an overly surprising re-
sult because the adverbial fillers were limited to only
nine dimensions. Within very high-dimensional vec-
tors they should lose ground. We see here a first in-
dication that the paradigm the more the merrier fits
well with our results. We consistently achieved bet-
ter results when we normalized the vectors by cosine.
Without length normalization accuracy apparently de-
creases (at least .07, see figure 3, last bar).
Figure 3: Results of process chain 3.
5.3 Results of the Classification of Verbs
using Dimension Reduced
Co-occurrences
In the third experiment (path 2 in figure 1) we left out
the parsing step. Within process chain 2 we built verb
vectors from co-occurrences of nouns that were clus-
tered as illustrated above. Firstly, we generated 39
clusters to provide the same dimensionality as in pro-
cess chain 4, which gave us a very poor result (see fig-
ure 4, first bar). By increasing the number of clusters,
the results became better, cumulating in .90 accuracy
for 500 clusters. Here again, the maxim seems to be
the more the merrier. Consequently, we should try to
expand the number of clusters to the number of fea-
tures. That is actually what we did in process chain 1,
which will be described in the next section.
5.4 Results for the Classification of
Verbs using Pure Co-occurrences
In the last experiment, we classified the verbs without
parsing and without clustering the noun fillers.
Instead, we built the vectors by simply collecting the
most frequent co-occurrences of a verb, irrespective
of whether they are nominal elements or not. Al-
though this was meant to be the baseline analysis,
Supervised Classification of Aspectual Verb Classes in German - Subcategorization-Frame-Based vs Window-Based Approach: A
Comparison
659

Figure 4: Results of process chain 2.
we obtained excellent results up to .98 accuracy,
especially with higher dimensionality of the vectors
(see figure 5).
Figure 5: Results of process chain 1.
6 CONCLUSION
The present study provides evidence that Vendlerian
aspectual verb classes plus a class of accomplish-
ments with an affected subject, i.e. Vendler + 1 classi-
fication, can be inferred from the contexts of the target
verbs. We observed that two out of our four models
outperformed the remaining ones and achieved excel-
lent agreements between the automatically inferred
aspectual classes and the Vendler + 1 gold standard
classification: (i) a model within an informed distri-
butional framework, considering structured, language
theoretically well grounded (and thus restricted) con-
text material that is, clustered nouns in the argument
positions of verbs in combination with aspectual fea-
tures and (ii) a model within a non informed frame-
work, considering large amounts of data, i.e. com-
pletely unstructured co-occurrence material of verbs.
This outcome can be interpreted as a manifestation of
the principle ’the more, the merrier’. On the one hand,
preprocessed linguistic information, i.e. small units
of linguistic information, were used for the construc-
tion of verb vectors (informed distributional). The re-
sulting verb vectors of the latter model have a small
dimensionality, however a lot of preprocessing steps
were necessary. On the other hand, as much pure
data as possible was included in the induction, lead-
ing to high dimensional verb vectors (non informed
distributional). This model needs relatively few pre-
processing steps, which means that the experiment’s
workflow is easy to design. When combining the two
approaches, that is using noun clusters and aspectual
features in a (restricted) distributional framework or
left out the clustering of noun arguments in a (re-
stricted) informed distributional framework, the qual-
ity of classification decreased.
Which model is preferable depends on the aims
and preconditions of the analysis. When a very large
set of verbs has to be classified, one might run into
trouble with the very high dimensionalities of the
verb vectors, making analysis slow, if not impossi-
ble. In that case the informed distributional method
which requires preprocessing tools like parsers would
be preferable. The choice of the method thus depends,
in general, on the amount of data that have to be an-
alyzed and on the availability of preprocessing tools
for the language which is being studied.
The results of the present study fit well into the-
ories of language acquisition. Naigles et al. (1992),
Naigles et al. (1995) and Naigles et al. (1993) de-
scribe frame compliance as an essential strategy used
by young children to interpret sentences. In order to
classify verbs into classes, a task which is mastered
quite late in the acquisition process (Wittek 2002,
Richter and van Hout 2013), children need to build up
knowledge about sentence types, for instance about
transitivity properties of sentences, and this knowl-
edge is utilized to learn verb classes (Brooks and
Tomasello 1999, Brooks et al. 1999). At the begin-
ning of the acquisition process knowledge of sen-
tence types is sometimes presumed to be item based
(Tomasello, 2000) meaning that it depends on specific
verbs and their contexts, while other studies provide
evidence of a linguistic maturation process: A con-
stant development of linguistic knowledge over time.
Our study provides evidence for Tomasello’s item
based approach. Starting with unstructured contexts
of verbs, learners identify very early item based proto-
typical verb/arguments structures and subsequently in
the course of the acquisition process they use context
materials to come to analogies. When learners recog-
nize a context, they classify a new verb according to
that context which is the environment of a previous
learned and eventually previously classified verb. In
order to manage the task of classifying verbs, linguis-
tic experience is required: The more language learn-
ers build up linguistic knowledge and manage to iden-
NLPinAI 2018 - Special Session on Natural Language Processing in Artificial Intelligence
660

tify arguments of verbs and other verb dependent el-
ements in the sentence, the more they build up the
structural knowledge of sentences types which in turn
is a prerequisite in order to induce verbs classes from
their contexts.
In summary, the results of this study show that,
for German, the Vendler + 1 typlogy can be inferred
from context materials of verbs. We found that both
classification frameworks that is, the informed dis-
tributional subcategorization-frame-based framework
and the non informed window-based framework per-
formed almost on a par. As future work, our experi-
ments should be complemented by clustering experi-
ments which means switching to unsupervised learn-
ing by induction. This would mean a step towards
simulation of induction based language acquisition
processes (see Bruner et al. 1956) on rule based learn-
ing of concepts).
REFERENCES
Bach, E. (1986). The algebra of events. Linguistics and
Philosophy, 9:5–16.
Bohnet, B. (2010). Very high accuracy and fast dependency
parsing is not a contradiction. In Proceedings of the 23rd
International Conference on Computational Linguistics,
COLING ’10, pages 89–97, Stroudsburg, PA, USA. As-
sociation for Computational Linguistics.
Brooks, P. and Tomasello, M. (1999). How children con-
strain their argument structure constructions. Language,
75(4):720–738.
Brooks, P., Tomasello, M., Dodson, K., and Lewis, L.
(1999). Young childrens overgeneralizations with fixed
transitivity verbs. Child Development, 7(6):1325–1337.
Bruner, J., J., G. J., and A., A. G. (1956). A Study of Think-
ing. Wiley.
Dorr, B. J. and Jones, D. A. (1996). Role of word sense
disambiguation in lexical acquisition: Predicting seman-
tics from syntactic cues. In 16th International Confer-
ence on Computational Linguistics, Proceedings of the
Conference, COLING 1996, Center for Sprogteknologi,
Copenhagen, Denmark, August 5-9, 1996, pages 322–
327.
Dowty, D. (1979). Word Meaning and Montague Grammar.
D. Reidel, Dordrecht.
Dowty, D. (1991). Thematic proto-roles and argument se-
lection. Language, 67(3):547–619.
Faaß, G. and Eckert, K. (2013). Sdewac - a corpus of
parsable sentences from the web. In Proceedings of
the International Conference of the German Society for
Computational Linguistics and Language Technology,
Darmstadt,Germany.
Fernando, T. (2004). A finite-state approach to events in
natural language semantics. Journal of Logic and Com-
putarion, 14(1):79–92.
Friedrich, A. and Palmer, A. (2014). Automatic prediction
of aspectual class of verbs in context. In Proceedings
of the 52nd Annual Meeting of the Association for Com-
putational Linguistics (Volume 2: Short Papers), pages
517–523, Baltimore, Maryland. Association for Compu-
tational Linguistics.
Goldberg, A. E. (1995). Constructions: a construction
grammar approach to argument structure. University of
Chicago Press, Chicago.
Gruender, S. (2008). An algorithm from adverbial aspect
shift. In Proceedings of the 22nd International Confer-
ence on Computer Linguistics (Coling08), pages 289–
296.
Harris, Z. (1954). Distributional structure. Word,
10(23):146–162.
Hermes, J., Richter, M., and Neuefeind, C. (2015). Au-
tomatic induction of german aspectual verb classes in a
distributional framework. In International Conference of
the German Society for Computational Linguistics and
Language Technology. Proceedings of the Conference
(GSCL 2015), pages 122–29.
Hermes, J. and Schwiebert, S. (2009). Classification of text
processing components: The tesla role system. In Ad-
vances in Data Analysis, Data Handling and Business
Intelligence, pages 285–294. Springer.
Jackendoff, R. (1972). Semantic Interpretation in Genera-
tive Grammar. MIT press, Cambridge MA.
Jackendoff, R. (1983). Semantics and Cognition. MIT
press, Cambridge MA.
Joachims, T. (1998). Text categorization with support vec-
tor machines: Learning with many relevant features.
Joanis, E., Stevenson, S., and James, D. (2008). A general
feature space for automatic verb classification. Natural
Language Engineering, 14(3):337–367.
Klavans, J. L. and Chodorow, M. (1992). Degrees of sta-
tivity: The lexical representation of verb aspect. In Pro-
ceedings of the 14th Conference on Computational Lin-
guistics - Volume 4, COLING ’92, pages 1126–1131,
Stroudsburg, PA, USA. Association for Computational
Linguistics.
Klein, W. (2009). How time is encoded. In The expression
of time, pages 39–82. Mouton de Gruyter, Berlin, New
York.
Krifka, M. (1989). Nominal reference, temporal constitu-
tion, and quantification in event semantics. In Bartsch,
R., van Benthem, J., and van Emde Boas, P., editors,
Semantics and Contextual Expressions, pages 75–115.
Foris.
Lakoff, G. (1970). Irregularity in Syntax. Rinehart and
Winstons, New York.
Landauer, T. K. and Dumais, S. T. (1997). A solution to
Plato’s problem: The latent semantic analysis theory of
acquisition, induction, and representation of knowledge.
Psychological review, 104(2):211–240.
Levin, B. and Rapoport, T. R. (1992). Lexical subordina-
tion. In Papers from the 24th Regional Meeting of the
Chicago Linguistic Society, pages 275–289.
Supervised Classification of Aspectual Verb Classes in German - Subcategorization-Frame-Based vs Window-Based Approach: A
Comparison
661

Levy, J. P. and Bullinaria, J. A. (2001). Learning lexi-
cal properties from word usage patterns: Which con-
text words should be used? In Connectionist Models
of Learning Development and Evolution: Proceedings of
the 6 th Neural Computation and Psychology Works hop,
pages 273–282. Springer.
Lowe, W. and McDonald, S. (2000). The direct route: me-
diated priming in semantic space. In Proceedings of
the annual Conference of the Cognitive Science Society
(CogSci-2000).
Lund, K. and Burgess, C. (1996). Hyperspace analogue to
language (HAL): A general model semantic representa-
tion. In Brain and Cognition, volume 30, page 265.
Merlo, P. and Stevenson, S. (2001). Automatic verb clas-
sification based on statistical distributions of argument
structure. Comput. Linguist., 27(3):373–408.
Naess, A., editor (2007). Prototypical Tranisitivity. Typo-
logical Studies in Language 72. Benjamins, Amsterdam.
Naigles, L., Fowler, A., and Helm, A. (1995). Syntactic
bootstrapping from start to finish with special reference
to Down syndrome. Beyond names for things: Young
childrens acquisition of verbs, pages 299–330.
Naigles, L., Gleitman, L., and Gleitman, H. (1993). Chil-
dren acquire word meaning components from syntactic
evidence. Language and cognition: A developmental
perspective. Norwood, NJ: Ablex, 5:104–140.
Naigles, L. G., Fowler, A., and Helm, A. (1992). Develop-
mental shifts in the construction of verb meanings. Cog-
nitive Development, 7(4):403–427.
Pantel, P. (2005). Inducing ontological co-occurrence vec-
tors. In Proceedings of the 43rd Annual Meeting on As-
sociation for Computational Linguistics, ACL ’05, pages
125–132, Stroudsburg, PA, USA. Association for Com-
putational Linguistics.
Pulman, S. G. (1997). Aspectual shift as type coercion.
In Transactions of the Philological Society 95(2), pages
279–317.
Pustejovsky, J. (1991). The synatax of event structure. Cog-
nition, 41:47–81.
Quine, W. v. O. (1960). Word and Object. MIT press, Cam-
bridge MA.
Richter, M. and Hermes, J. (2015). Classification of ger-
man verbs using nouns in argument positions and aspec-
tual features. NetWordS 2015 Word Knowledge and Word
Usage, pages 177–181.
Richter, M. and van Hout, R. (2013). Interpreting resultative
sentences in German: Stages in l1 acquisition. Linguis-
tics, 51(1):117–144.
Richter, M. and van Hout, R. (2016). A classification of ger-
man verbs using empirical language data and concepts of
vendler and dowty. Sprache und Datenverarbeitung - In-
ternational Journal for language data processing, 38(1-
2):81–117.
Rothstein, S. (2004). Structuring Events: A Study in the
Semantics of Aspect. Explorations in Semantics. Wiley.
Rubenstein, H. and Goodenough, J. B. (1965). Contextual
correlates of synonymy. Commun. ACM, 8(10):627–633.
Schulte im Walde, S. (2003). Experiments on the choice
of features for learning verb classes. In Proceedings of
the Tenth Conference on European Chapter of the Asso-
ciation for Computational Linguistics - Volume 1, EACL
’03, pages 315–322, Stroudsburg, PA, USA. Association
for Computational Linguistics.
Schulte im Walde, S. (2006). Experiments on the automatic
induction of german semantic verb classes. Computa-
tional Linguistics, 32(2):159–194.
Schulte im Walde, S. and Brew, C. (2002). Inducing german
semantic verb classes from purely syntactic subcategori-
sation information. In Proceedings of the 40th Annual
Meeting on Association for Computational Linguistics,
ACL ’02, pages 223–230, Stroudsburg, PA, USA. Asso-
ciation for Computational Linguistics.
Schumacher, H. (1986). Verben in Feldern:
Valenzw
¨
orterbuch zur Syntax und Semantik deutscher
Verben, volume 1. Walter de Gruyter.
Sch
¨
utze, H. and Pedersen, J. (1993). A vector model for
syntagmatic and paradigmatic relatedness. pages 104–
113.
Siegel, E. V. (1997). Learning methods for combining lin-
guistic indicators to classify verb. In Proceedings of
the 2nd Conference on Empirical Methods in Natural
Language Processing, EMNLP, Brown University, Prov-
idence, RI. cmp-lg/9707015.
Siegel, E. V. and McKeown, K. R. (2000). Learning meth-
ods to combine linguistic indicators: Improving aspec-
tual classification and revealing linguistic insights. Com-
put. Linguist., 26(4):595–628.
Smith, C. (1991). The Parameter of Aspect. Kluwer, Dor-
drecht.
Tomasello, M. (2000). Do young children have adult syn-
tactic competence? Cognition, 74(3):209 – 253.
Turney, P. D. and Pantel, P. (2010). From frequency to
meaning: Vector space models of semantics. J. Artif.
Int. Res., 37(1):141–188.
Vendler, Z. (1967). Linguistics in philosophy. G - Refer-
ence, Information and Interdisciplinary Subjects Series.
Cornell University Press.
Verkuyl, H. J. (2005). Aspectual composition: Surveying
the ingredients. pages 19–39.
Vlachos, A., Korhonen, A., and Ghahramani, Z. (2009).
Unsupervised and constrained dirichlet process mixture
models for verb clustering. In Proceedings of the Work-
shop on Geometrical Models of Natural Language Se-
mantics, GEMS ’09, pages 74–82, Stroudsburg, PA,
USA. Association for Computational Linguistics.
Wittek, A. (2002). Learning the meaning of change-of-
state verbs: A case study of German child language, vol-
ume 17. Walter de Gruyter.
Wright, G. H. v. (1963). Norm and Action. Routledge and
Kegan Paul, London.
Zarcone, A. and Lenci, A. (2008). Computational models
of event type classification in context. In Proceedings of
the 6th International Conference on Language Resources
and Evaluation, pages 1232–1238, Marrakech, Morocco.
NLPinAI 2018 - Special Session on Natural Language Processing in Artificial Intelligence
662