English. Lexical sets contain the words filling the argument positions of a verb in one of its senses. They can be extracted from corpora automatically. The purpose of this paper is demonstrating that their vector representation based on word embedding provides insights onto many linguistic phenomena, such as causativeinchoative verbs. A first experiment aims at investigating the internal structure of the sets, which are known to be radial and continuous categories cognitively. A second experiment shows that the distance between the intransitive subject set and transitive object set is correlated with the spontaneity of the event expressed by the verb, defined according to morphological coding and frequency.
Italiano. I set lessicali contengono le parole che occupano le posizioni argomentali di un verbo in una delle sue accezioni, e possono essere estratti in modo automatico dai corpora. L’obiettivo di questo articolo è dimostrare che la loro rappresentazione vettoriale illumina alcuni fenomeni linguistici, come i verbi ad alternanza causativo-incoativa. Un esperimento investiga la struttura interna degli insiemi, che a livello cognitivo sono ritenuti categorie radiali e continue. Inoltre, un secondo esperimento mostra che la distanza fra l’insieme dei soggetti intransitivi e l’insieme degli oggetti transitivi è correlata alla spontaneità dell’evento espresso dal verbo, definita secondo la marca morfologica e la frequenza.
Lexicographic attempts to cope with verb sense
disambiguation often rely on “lexical sets”
The full exploitation of the semantic information
inherent to argument fillers for verbs can take
advantage from some recent developments in
distributional semantics. Recently, efficient algorithms
have been devised mapping each word of a
vocabulary into a corresponding vector of n real
numbers, which can be thought as a sequence of
coordinates in a n-dimensional space
A possible testbed for the usefulness of representing the argument fillers as vectors are the verbs showing the so called causative-inchoative alternation. These verbs appear either as transitive or intransitive. In the first case, an agent brings about a change of state; in the second, the change of a patient is presented as spontaneous (e.g. to break, as in “Mary broke the key” vs. “the key broke”).
The two alternative forms of these verbs can
be morphologically asymmetrical: in this case,
one has a derivative affix and the other does not.
The first is labelled here as “marked”, the
second as “basic”. Italian verbs with an asymmetrical
alternation derive from the phenomenon of
anticausativization. The intransitive form is marked
since it is sometimes preceded by the clitic si
In this way, a correlation between the form and the meaning of these verbs was demonstrated. Moreover, Samardzic and Merlo (2012) and Haspelmath et al. (2014) argue that verbs that appear more frequently (intra- and crosslinguistically) in the inchoative form tend to morphologically derive the causative form, too. This time, the correlation holds between form and frequency. Vice versa, situations entailing agentive participation prefer to mark the inchoative form and occur more frequently in the causative form. 2
In the literature, many methods are available for
the automatic detection of verb classes, such as
causative-inchoative verbs. They exploit features
based on argument alternations, such as
subcategorization frames
These features were usually induced from
automatic parses of heterogeneous and wide corpora
The data are sourced from a sample of ItWac, a
wide Italian corpus gathered through web crawling
close > open > improve > break > fill > gather > connect > split > stop > go out > rise > rock > burn > freeze > turn > dry > wake > melt > boil > sink
The extraction step consisted in identifying
their argument fillers inside the sentences in the
sample. In particular, the arguments considered
were the subjects of intransitives (S) and objects
1LAS scores for the relevant dependency relations: 0.751
with dobj (direct object), 0.719 with nsubj (subject), 0.691
with nsubjpass (subject of a passive verb).
(O)
These operations resulted in a database where
each verb lemma had a single entry and was
associated with a list of fillers, divided by argument
type. With this procedure, lexical sets were
extracted automatically, although they were not
divided by verb sense. Afterwards, each of the
argument fillers was mapped to a vector relying on a
space model pre-trained through Word2Vec
In order to bring to light the linguistic
information concealed in the automatically extracted
lexical sets, we devised two experiments. One
investigates the internal structure of lexical sets. In fact,
previous works based on set theory treated them as
categoric sets, of which a filler is either a member
or not. Research in psychology, however, has long
since demonstrated that the members of a
linguistic set are found in a radial continuum where the
most central one is the prototype for its category,
and those at the periphery are less representative
2Subjects of forms with si were treated as intransitive subjects. Subjects of passive verbs were treated as objects.
3It was generated by a CBOW algorithm with negative sampling, 300 dimensions, a context window of 10 tokens, pruning of infrequent words and sub-sampling.
4For previous work on lexical sets considering prototypicality in the context of the notion of shimmering, see Jezek and Hanks (2010).
Once the fillers have been mapped to their respective vectors, a lexical set appears as a group of points in a multi-dimensional model. The centre of this group is the Euclidean mean among the vectors, which is a vector itself and is called centroid. In the first experiment, we calculated the coordinates of the centroid of the lexical sets S and O for any selected verb5. Then we evaluated the cosine similarity of every vector member of the sets from its centroid. The value of this metric goes from 0 (overlap) to 1 (maximum distance) and is useful to evaluate how far a filler is from its prototype. We obtained two sets of cosine similarity values for each verb: these can be plotted as boxes and whiskers, like in Figure 1. The example represents those of dividere ‘to split’. The rectangles stand for the values in the second and third quartiles, whereas the horizontal line for the median6. From all these distance values, we picked the median value for each lexical set. The plot of these medians for the S set and the O set of each verb ordered according to Haspelmath’s ranking is shown in Figure 2.
Two main results can be observed from these plots: the S lexical set lies in a more compact range of distances, whereas O is more scattered. On the other hand, the vectors of S tend farther from the centroid. This is demonstrated by the ranges where their distance values fall. Moreover, the averages of medians for the ten verbs on the left part of the scale (frequently transitive) and for the ten verbs on the right (frequently intransitive) were compared. The average median in S was 0.696567 for the former and 0.585263 for the latter. The average median in O was 0.556878 for the former and 0.522418 for the latter. This shows that the variation in O appears to be random. On the other hand, the median of the distances in S is normally lower for verbs that lie in the bottom half of the Haspelmath’s scale.
The second experiment consisted in estimating the cosine distance between the centroid of S and the centroid of O for each verb. This operation was aimed at finding to which extent the lexical sets of S and O overlap. In fact, Montemagni et al. (1995) and McCarthy (2000) assessed in a corpus some asymmetries between these lexical sets, which in principle should share all their members.
5Every filler was weighted proportionally to its absolute frequency.
6The median is the value separating the higher half of the ordered values from the lower half.
Inspecting our results, the distance between S
and O seems to behave as a measure of
spontaneity, intended as cross-linguistic frequency and
morphological markedness of a verb: the more the
centroids tend to be set apart, the more the verb
tends to have a morphologically unmarked and
more frequent intransitive form. In fact, we
compared the ranking of 20 alternating verbs
according to the ratio of their cross-linguistic frequency
of transitive and intransitive forms
Both scales display a common tendency. In particular a Spearman’s ranking test was performed over them, yielding a mild positive correlation of = 0:56391 with a quite strong confidence, i.e. with p < 0:01.7 5
The representation of lexical sets of Italian
causative-inchoative verbs as vectors was
demonstrated to provide insights into their internal
structure and their relation with spontaneity defined
according to morphological coding and frequency.
The distances of the objects appeared to be
distributed more uniformly, whereas those of the
intransitive subjects more densely and remotely
from the centroid. This difference cannot stem
from the frequency of anaphoric fillers (contrary
to transitive subjects), since both these argument
positions share the discursive function of
introducing new referents, and are hence occupied by fully
referential fillers
Moreover, the medians of the distances of the
subject fillers from their centroid were shown to
vary. An interpretation is that they are
sensible to the frequency scale: this implies that
frequently transitive (hence, non-spontaneous) verbs
have semantically less homogeneous sets of
referents, since they are farther from the prototype.
Possibly this discovery can be related with the
fact that non-spontaneous verbs impose less
selectional restrictions on subjects
The lack of a perfect correlation between these
vector distance and frequency measures is maybe
due to errors in the automatic extraction and data
sparseness for the former, or an insufficient sample
7An alternative measure was considered for the ranking:
the cardinality of the S-O intersection weighted by the set
union. In this case, Spearman correlation was = 0:42255,
but it was not significant because p 0:06.
of languages in the typological survey of
Haspelmath et al. (2014) for the latter. A possible
interpretation of the correlation is that the entities
capable of bringing about a change of state and
those that undergo it are indiscernible only for
non-spontaneous verbs. Studies on causer entities
related them not only with the feature of
agentivity, but also in general with the so-called
‘teleological capability’
Our work provided evidence that lexical sets of Italian causative-inchoative verbs are continuous and radial categories, whose distribution around the prototype vary to a great extent. It is sensitive to the grammatical role and sometimes to the position of the verb in the so-called spontaneity scale. Moreover, a correlation was discovered between the distance between transitive object and intransitive subject lexical sets of a given verb and its cross-linguistic tendency to appear more frequently as intransitive or as transitive. Figure 4 is a synopsis of this result in the context of the correlations established in previous works.
Distant S and O centres
In Figure 4, solid lines stand for correlations proven based on cross-linguistic evidence (frequency-form) and evidence from the Italian language (frequency-lexical sets). The dotted line, on the other hand, suggests the existence of and underlying motivation for the correlations, which nonetheless remains unproven and undetermined in its nature. Its possible validation is left to future research.
Future works should also choose different
pretrained vector models, in order to try and replicate
these results. In particular, the new vector models
could be optimized for similarity through semantic
lexica