The term Distributional semantic models (DSMs)
refers to a family of unsupervised corpus-based
approaches to semantic similarity computation.
These models rely on the distributional
hypothesis
In recent years, these models have become very
popular in a wide range of NLP tasks
This paper illustrates such a qualitative linguistic approach by investigating the presence of partof relations among distributionally similar French words. We compare distributional data and a set of part-of relations provided by humans in a lexical network. In order to assess the nature of the partof word pairs which can – or cannot – be found in DSMs, these words were sense-tagged using WordNet supersenses. Our results show considerable discrepancies between the representation of part-of sense pairs in distributional data. 2
As its name suggests, part-of relation – or meronymy1 – holds between a part – the meronym – and its whole – the holonym –, like in bed/pillow, armor/steel or ostrich/feather. It is one of the central relations used in knowledge representation.
Automatic extraction of part-of relations has
been addressed using many approaches, most of
which are pattern-based
Studies were conducted to assess the nature
of the semantic relations extracted by
distributional models – using human judges
It appears, though, that a certain proportion of them tend to do so. For example, in Baroni and Lenci (2010)’s DSM, player, pianist and musician are among the ten most distributionally similar words of orchestra. In the following of this study, we compare the semantic properties of the meronyms which can be extracted using a distributional approach and the properties of the meronyms which cannot.
1Some authors make a distinction between part-of relation
and meronymy
The first step consists in gathering a set of
meronyms. Although efforts are made to provide
expert-built lexical semantic resources for French
Although collaboratively-built lexical semantic
resources have shown to be valuable
One interesting characteristic of JDM part-of
relations is that a considerable number of them
do not fit into traditional typologies of meronymy
relations. For example, topological inclusions
(cell/prisoner), attachment relations (ear/earring)
or ownership (millionaire/money) are very
common among JDM part-of pairs although they are
considered to be non-meronymic relations
After filtering the pairs whose members do not appear in our DSM and removing most of the erroneous relations, there were 24 089 part-of pairs left in our dataset. 3.2
In a previous study
The above-mentioned lack of freely-available thesauri for French led us to use WordNet to perform this task. Words of our dataset were 1) translated to English, 2) mapped to WordNet synsets and 3) linked to their translation’s supersense(s). Supersenses – or lexicographer classes – are a set of 44 coarse semantic categories used to classify WordNet’s noun, verb and adjective entries3. Examples of the 25 noun supersenses are GROUP, LOCATION or FOOD. Supersenses were then manually disambiguated (drawer can both belong to the PERSON and ARTIFACT supersenses, but only the latter fits in the paircabinet/drawer). 3.3
We use a DSM4 generated from the frWaC corpus
Words in the DSM appear at least 20 times in the corpus and in at least 5 different contexts.
Syntactic dependencies were used as contexts
using the Talismane parser
The weighting of the contexts was made using the pointwise mutual information and the cosine measure was used to compute the similarity between the context vectors. The minimum similarity threshold has been set to 0.02. The total number of word pairs in the DSM is 3 674 254. 4
We then measure the proportion of semanticallyannotated part-of pairs – sense pairs – in our set which are present in the DSM. Sense pairs which occur less than 100 times in the dataset are discarded. Table 1 provides the list of the 22 re3http://wordnet.princeton.edu/man/ lexnames.5WN.html
4Provided by Franck Sajous from the CLLE-ERSS laboratory. maining sense pairs and, for each one, the ratio of part-of pairs present in the DSM. In this section, we describe the homogeneous sense pairs – whose semantic classes are identical – and the heterogeneous ones, then we provide a detailed analysis of some of the PERSON/BODY meronyms which have been extracted by the DSM. 4.1
As expected, part-of relations composed of two words of the same class are the most represented in the DSM. 84 % of the TIME/TIME part-of pairs were extracted by the DSM. This can be explained by the fact that the members of pairs like mois/jour ‘month/day’ both appear in contexts involving temporal prepositions like venir IL Y A ‘to come SINCE’, se de´rouler DURANT ‘to take place DURING’ or scrutin AVANT ‘election BEFORE’.
Likewise, the spatial dimension plays a crucial role in the extraction of meronyms (78.3 % of LOCATION/LOCATION pairs are extracted). This is due to the fact that, as for time, spatial information can be conveyed by specific prepositions. Thus, LOCATION/LOCATION meronyms’ shared contexts massively involve the DANS ‘IN’ relation.
SUBSTANCE pairs are the third best-extracted kind of pairs. The reason why 37.6 % of them has not been extracted can be illustrated by the comparison of acier ‘steel’ and two of its meronyms, namely fer ‘iron’ – which was extracted in the DSM – and carbone ‘carbon’ – which was not extracted: 1. acier and fer both appear in contexts like grille EN ‘grille COMP’, forge´ MOD ‘forged MOD’ or lame DE ‘blade COMP’. Thus, they appear as materials and, moreover, as materials which are used to build the same kind of things; 2. although being a material as well, carbone does not appear as such in the corpus. Rather, its contexts are chemical compounds like monoxyde DE ‘monoxide COMP’. It is also modified by adjectives likeinorganique MOD ‘inorganic MOD’, which describe chemical properties of carbone. These two kinds of contexts are not found among acier’s.
So, we can see that there is a discrepancy between the contexts in which acier appears in the corpus and the ones in which carbone appears: whereas holonym/meronym
TIME/TIME
LOC./LOC.
SUBST./SUBST.
OBJECT/OBJECT COMM./COMM.
GROUP/PERSON LOC./ARTIFACT
BODY/BODY ANIMAL/ANIMAL ARTIFACT/COMM.
ACT/ARTIFACT % 84 78.3 62.4 61 53.8 52.8 46.8 40.5 41 39.9 35.8 holonym/meronym ARTIFACT/PERSON ARTIFACT/ARTIFACT
ARTIFACT/LOC.
ARTIFACT/PLANT ARTIFACT/SUBST.
OBJECT/ANIMAL PLANT/PLANT GROUP/ANIMAL PERSON/ARTIFACT
ANIMAL/BODY PERSON/BODY acier – as well as fer – is used as a material, the representation of carbone that emerges from the corpus is that of a chemical element. At the other end of the scale, part-of relations composed of two words of different classes are – also logically – the less represented in the DSM.
Part-of pairs composed of words that refer to human beings or to animals and their body parts are barely present in the DSM (although being the most frequent sense pairs in our dataset). In frWaC, PERSON words appear as subjects of action (prendre ‘to take’, dire ‘to say’) or cognitive verbs (vouloir ‘to want’, savoir ‘to know’). They are frequently modified by nationality adjectives. Body parts do not appear in such contexts. The class of body parts was actually found to be quite heterogeneous, in that body parts’ distributions in the corpus differ from persons’, but not in the same way: • organ nouns mostly appear in noun compounds to indicate the location of medical interventions (radiographie DE ‘x-ray MOD’) or affections (cancer de ‘cancer COMP’ or le´sion de ‘injury COMP’); • limb nouns are modified by adjectives related to location and are objects of verbs like lever ‘to raise’ or e´tendre ‘to stretch’.
All these contexts are obviously incompatible with PERSON words.
A similar distributional discrepancy can be observed with the ANIMAL/BODY sense pair, except that animal nouns tend to appear in contexts like e´levage DE ‘farming COMP’ or espe`ce DE ‘species COMP’. They are also modified by size adjectives. It is interesting to note that many animal body parts like teˆte DE ‘head COMP’, peau DE ‘skin COMP’ or queue DE ‘tail COMP’ do appear among the closest contexts of animal nouns. This means that the meronymic relation between nouns referring to animals and their body parts is not a paradigmatic one. Thus, it is reasonable to say that, in order to extract this particular relation, the use of syntagmatic patterns would be a better strategy than the use of a paradigmatic DSM.
The sense pair GROUP/PERSON also presents an interesting situation. Of all the heterogeneous sense pairs, meronymic relations belonging to this one are the most likely to be extracted by the distributional method. This can be explained by a tendency to use the GROUP entities in a metonymic way: although an army is not the same kind of thing as a soldier, both words share contexts like tirer SUJ ‘to shoot SUBJ’ or tue´ PAR ‘killed BY’. Another reason is the transitivity of properties like nationality: arme´e ‘army’ and soldat ‘soldier’ are both modified by nationality adjectives because usually, members of the armed forces of a nation have to be citizens of this nation.
In the section 2, we mentioned the fact that three meronyms of orchestra were present among its ten most distributionally similar words in Baroni and Lenci (2010)’s DSM. In our data, the meronyms orchestre/musicien have also been extracted: as for army and soldier, these words share semantic features. They are related to the kind of music a musician and an orchestra can play (classique MOD ‘classical MOD’, traditionnel MOD ‘traditional MOD’ or jazz DE ‘jazz MOD’), the kind of actions they perform (interpre´te´ PAR ‘performed BY’, accompagne´ PAR ‘accompanied BY’) or their nationality. 4.3
In the previous subsection, we saw that meronyms belonging to the PERSON/BODY are the least likely to be extracted with the distributional approach. In this subsection, we provide further insight into this result by examining the nature of the few PERSON/BODY meronymic pairs that were successfully extracted.
The examination of the 5.5 % of PERSON/BODY meronymic pairs that were successfully extracted is disappointing: the vast majority of the contexts shared by the meronym and the holonym are quite random. For example, the meronyms homme/main ‘man/hand’ share contexts like nu MOD ‘bare MOD’ or dos DE ‘back COMP’, which are not very informative about their relation. On the other hand (!) some shared contexts like doigt DE ‘finger COMP’ and saisir SUJ ‘to grab SUBJ’ are more informative. The fact that these specific features are shared by the meronyms indicates some kind of similarity between them: when a man grabs a rock, it is actually his hand that completes the action of grabbing, as well as a man’s fingers are also his hand’s fingers.
The meronyms enfant/oeil ‘child/eye’ also share some interesting contexts: both the meronym and the holonym are subjects of verbs of visual perception like regarder ‘to look’, percevoir ‘to perceive’ or observer ‘to observe’. The metonymic interpretation is quite straightforward: although the eye is the child’s part that allows him to look/perceive/observe, this ability is extended to the whole child.
This phenomenon partially explains why such meronyms share semantic – thus distributional – features and are more likely to be extracted with a DSM. 5
The main goal of this study is to shed light on the linguistic phenomena at work in DSMs. By comparing a set of sense-tagged part-of relations and a distributional model, we show that the semantic class of the meronyms has a dramatic influence on their probability to be extracted by a DSM. We also highlight the – positive – influence of metonymy in the extraction of heterogeneous meronyms.
These results show that the part-of relation is not a monolithic entity but a collection of different kinds of relations between different kinds of words which may or may not be distributionally similar.
This work was partially supported by the ANSM (French National Agency for Medicines and Health Products Safety) through the Vigi4MED project under grant #2013S060.