What’s in a Definition? An Investigation of Semantic Features in Lexical
Dictionaries
Luigi Di Caro
a
Department of Computer Science, University of Turin, Italy
Keywords:
Lexical Semantics, Word Meaning, Word Definitions, Backward Dictionaries, Lexical Ambiguity.
Abstract:
Encoding and generating word meanings as short definitions for user- and machine-consumption dictionaries is
still a usually adopted strategy within interpretable lexical-semantic resources. Definitions have the property
of being natural to be created by humans, while several techniques have been proposed for the automatic
extraction from large corpora. However, the reversed process of going back to the words (i.e., onomasiological
search) is all but simple for both humans and machines. Indeed, definitions show context- and conceptual-
based properties which influence their quality. In this contribution, I want to draw the attention to this problem,
through a simple content-to-form experimentation with humans in the loop. The results give some first insight
on the relevance of the problem from a computational perspective. In addition, I analyzed (both quantitatively
and qualitatively) a set of 1,901 word definitions taken from different sources, towards the modeling of features
for their generation and automatic extraction.
1 INTRODUCTION
Nowadays, Natural Language Processing is becom-
ing more than one among the research areas within
the Artificial Intelligence field. With the advent of
large data repositories and new enabling technolo-
gies, the fascinating dream of machines interacting
through natural language is experiencing renewed at-
tention and horizons.
While research on Semantics has been always fo-
cusing on the philosophical aspects of meaning and
its relationship with language, Computational Seman-
tics works towards machine-based encoding of mean-
ing providing human-like capabilities such as simi-
larity and reasoning processes. Being short and ap-
proximate, Formal Semantics models inferences at
the symbolic level, while Distributional Semantics en-
ables the computation of semantic similarity between
symbols (i.e., lexical units) through corpora analysis.
In other words, on the one hand, one can define se-
mantics as the way of inferring knowledge given a
set of facts and rules expressed through symbolic el-
ements. On the other hand, semantics can be seen
as the embodied meaning within such symbols, irre-
spective of their power to enable logical inferences
about some global knowledge. Distantly from this du-
a
https://orcid.org/0000-0002-7570-637X
alism, Lexical Semantics represents a highly studied
area aiming at developing semantic resources such as
lexical inventories and ontologies to support a huge
variety of semantic analysis tasks. However, the en-
coded meaning in such resources are often still of lex-
ical type, such as definitions, glosses, and examples
of use. In this paper, I first study the fragility of such
representation of meaning through a simple content-
to-form experimentation with 20 participants. In par-
ticular, I asked some of them to provide individual
definitions on few concepts (of different types, as de-
tailed in the paper). Then, the remaining participants
had to guess by going back to the described words.
This task is often associated with the name of onoma-
siologic search, and it relates with the well-known tip-
of-the-tongue problem (Brown and McNeill, 1966).
What I found is that definitions resulted to be very
fragile encodings of lexical meaning, even with sim-
ple concepts, and in a controlled scenario. Then, we
looked deeper at the result of the experiment by an-
alyzing each definition in terms of different criteria,
trying to make some measurement of their quality,
and considering its effectiveness in correctly indicat-
ing the unveiled words.
Then, I carried out an experimentation with a
dataset of 1,901 word definitions about 300 random
concrete concepts, extracted with the help of Babel-
Net (Navigli and Ponzetto, 2010), belonging to re-
Di Caro, L.
What’s in a Definition? An Investigation of Semantic Features in Lexical Dictionaries.
DOI: 10.5220/0010116902250232
In Proceedings of the 16th International Conference on Web Information Systems and Technologies (WEBIST 2020), pages 225-232
ISBN: 978-989-758-478-7
Copyright
c
2020 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
225
sources such as Wordnet (Miller, 1995), Wikipedia
1
,
Wiktionary
2
, OmegaWiki
3
, and others. In particular,
I quantitatively and qualitatively analyzed the type of
information contained in the definitions, highlighting
possible features to be used for the creation of better
definitions rather than their automatic extraction from
large corpora.
2 MOTIVATIONS AND
RESEARCH QUESTIONS
Lexical Semantics is about encoding lexical meaning.
This is crucial in many multilingual NLP tasks and
applications. However, interpretable (as opposed to
statistical and vector-based) lexical resources mostly
rely on the paradox of providing word meanings
through, again, the use of words. This forces the
Word Sense Disambiguation (WSD) process to work
on non-machine-based representations of lexical con-
texts and definitions. However, some questions arise
and have already been approached computationally,
e.g., (Muresan and Klavans, 2002; Klavans et al.,
2003; Thorat and Choudhari, 2016). Some of them
may be the following:
1. How consistent and reliable are the definitions of
words?
2. How good are the definitions within lexical re-
sources? Or, more generally, is it possible to mea-
sure the quality of a definition?
3. Are there any features that may rule the effective-
ness of definitions?
4. Which semantic features and relations are typi-
cally covered by textual definitions?
In this paper, I carried out a simple experimentation
with humans in the loop to put some light on these
aspects.
In my perspective, we can roughly define the gen-
eral quality of a definition as its ability to let people
identify the described concept. If a definition carries
to wrong guesses, then we could say it does not make
its work properly (and it is difficult to think it could
do a better job in automatic machine-based methods).
While there exists significant work on the prin-
ciples behind definition writing, such as the genus-
differentia mechanism (e.g., (Strehlow, 1983)), to the
best of my knowledge, little effort has been put to-
wards computational-oriented modeling of definitions
1
https://www.wikipedia.org
2
https://en.wiktionary.org/
3
http://www.omegawiki.org/
quality. In this paper, I propose a first model, evalu-
ating it with a restricted set of concepts and through a
test with non-expert participants.
3 RELATED WORK
Lexical semantics resources usually fall into cate-
gories such as computational lexicons, corpus-based
models, semantic frames, and common-sense knowl-
edge bases. They all define the meaning of words
in terms of (sometimes categorized) textual descrip-
tions. In this section, I briefly overview them under a
content-to-form perspective.
3.1 Computational Lexicons
WordNet (Miller, 1995) may be considered as the
most referenced and used computational lexicon for
English. Counterparts in other languages (Bond and
Foster, 2013) are also available. WordNet is produced
by humans for humans, as concepts are described
through word definitions, and contextualized in terms
of paradigmatic relations such as hyperonymy and
meronymy. BabelNet (Navigli and Ponzetto, 2010) is
the result of a large-scale integration of WordNet with
Wikipedia and other sources of semantic informa-
tion. Most Word Sense Disambiguation (WSD) sys-
tems use these resources and their gloss-based model.
However, glosses are often short and disclose very
few semantic information from which is difficult to
go back to the words. In Section 6 I present an exper-
iment that puts some light on this issue.
3.2 Frames
(Fillmore, 1977) proposed semantic frames to encode
meanings through slot-filler structures, and FrameNet
(Baker et al., 1998) represents the largest frame-based
resource available. Slots and fillers represent at-
tributes and values respectively. While such represen-
tation could be used to better go back to the described
items, actually, frames do not encode the individual
meaning of concepts but their situational use. An in-
teresting and novel slot-filler approach was presented
by (Moerdijk et al., 2008) with the ANW dictionary
and the introduction of the concept of semagram. A
semagram is a conceptual structure that describes a
lexical entity on the basis of a wide range of charac-
teristics, defined with a rich slot-filler structure. The
semagrams provided in the ANW dictionary are, how-
ever, limited in coverage, often expressed with a frag-
mented set of semantic slots and written in Dutch. In
WEBIST 2020 - 16th International Conference on Web Information Systems and Technologies
226
(Leone et al., 2020), the authors revised the semagram
structure to overcome these limitations.
3.3 Corpus-based Distributional Models
Corpus-based semantic models are based on the Dis-
tributional Hypothesis (Harris, 1954), i.e., words co-
occurring in similar contexts tend to have similar
meanings. Latent Semantic Analysis (Dumais, 2004),
Latent Dirichlet Allocation (Blei et al., 2003), and,
more recently, embeddings of words (Mikolov et al.,
2013; Pennington et al., 2014; Bojanowski et al.,
2016) and word senses (Huang et al., 2012; Iacobacci
et al., 2015) represent vectorial / geometrical repre-
sentations of words. However, the relations holding
between vector representations are not typed, nor are
they organized systematically. While these represen-
tations work well for semantic similarity computa-
tion, vector dimensions are not interpretable concept
descriptions. Conceptual Spaces (G
¨
ardenfors, 2004)
provide a geometric approach to meaning where vec-
tor dimensions are instead qualitative features (e.g.,
colors may be represented through hue, saturation,
and brightness). However, the encoded knowledge
does not define concepts explicitly, and dimensions
usually represent perceptual mechanisms only.
3.4 Common-sense Knowledge
Common-sense knowledge resources may be de-
scribed as a set of shared and general facts or views
of a set of concepts. ConceptNet (Speer and Havasi,
2012; Speer et al., 2016) is one of the largest re-
sources of this kind. However, terms in ConceptNet
are not disambiguated, which leads to the confusion
of lexical-semantic relations involving concepts de-
noted by ambiguous words. NELL (Carlson et al.,
2010) matches entity pairs from seeds to extract re-
lational phrases from a Web corpus, but it is mostly
oriented to named entities rather than concept de-
scriptions. Property norms (McRae et al., 2005; De-
vereux et al., 2014) represent a similar kind of re-
source, which is more focused on the cognitive and
perception-based aspects of word meaning. Norms
are based on empirically constructed semantic fea-
tures via questionnaires asking people to produce fea-
tures they think as important for some target concept
(e.g., a crocodile is often associated with the norm
is-dangerous). The problem with norms is that they
do not represent complete descriptions (usually, only
immediate and common-sense facts are reported).
4 FEATURES OF DEFINITIONS
As earlier mentioned, the main aim is to understand
and evaluate the robustness of definitions as they
still lie at the core of most lexical resources, e.g.,
WordNet. Differently from existing features related
to writing quality such as (Witte and Faigley, 1981)
and more recently (McNamara et al., 2010), defini-
tions have not been yet analyzed and modeled from
a computational perspective and under a backward-
dictionary view. However, the actual connection be-
tween a word and its meaning is, at least by humans,
strictly dependent on the robustness of its definition.
Thus, this might also have some impact on compu-
tational approaches. In this contribution, I propose a
simple model made up of three features:
1. clarity;
2. richness; and
3. readability.
For clarity, I mean the (main) feature of a definition
of being non-ambiguous with respect to similar con-
cepts (e.g., hypernyms). This is actually what I empir-
ically measured with the experment that will follow. I
then define richness as the quantity of semantic infor-
mation
4
contained and readability as the lexical and
syntactic simplicity and shortness of the definition.
While the three features are slightly interconnected,
they may have diverging scores within a single defini-
tion.
For example, a definition may contain several se-
mantic relations without eliminating ambiguity. For
example, the following definition of terminal can be
erroneously associated with the concept computer:
In networking, a device consisting of a video
adapter, a monitor, and a keyboard. The
adapter and monitor and, sometimes, the key-
board are typically combined in a single unit.
Contrariwise, another (very short) definition of
terminal, while revealing little semantic content, is in-
stead able to clearly identify the concept:
A device communicating over a line.
In the next sections, I will present the results of
two experiments aiming to highlight features and dy-
namics of definitions from both a human and a com-
puter perspective.
4
For semantic information I mean the set of semantic
relations that can be grasped from the definition, such as
paradigmatic ones (e.g., hypernyms, meronyms, etc.), phys-
ical (e.g., size, shape, color, etc.), behavioral (e.g., purpose,
ways of use, etc.), and others.
What’s in a Definition? An Investigation of Semantic Features in Lexical Dictionaries
227
5 HUMAN-IN-THE-LOOP
EXPERIMENT
In this section, I describe how I conducted the experi-
ment with the 20 participants, and the used criteria for
the selection of the concepts. The aim of the exper-
iment was to analyze the descriptions of simple con-
cepts by (even non-expert) people under a computa-
tional perspective, and specifically towards an auto-
matic content-to-form approach.
5.1 Participants
Participants were 20-35 years old students with differ-
ent background (linguists, computer scientists, math-
ematicians, engineers). They were not aware of the
goals of the experiment, while they have been intro-
duced with some knowledge on Computational Lin-
guistics, and specifically, on Lexical Semantics.
5.2 Methodology
The idea of the experiment was to test the capabil-
ity of word definitions to uniquely identify the un-
derlying concepts. In order to be significantly sure
about the independence from single subjective views,
I asked 12 out of the 20 participants (def -participants,
from now on) to create definitions for all the concepts,
leaving the remaining 8 (test -participants) for the later
test phase. This way, by having 12 definitions for each
concept, we can be rather confident that the results
were not influenced by single and unfortunate defini-
tions (the entire set of definitions has been given to
the test-participants to make a single choice). I finally
asked the test-participants to mark the best definition
which has been mostly useful for giving the answer.
This last step allowed us to correlate some features
of the best-selected definitions with the accuracy ob-
tained by the participants during the experiment.
5.3 Concepts Selection
Due to the choice of employing most of the partici-
pants in the creation of definitions, I had to limit the
number of concepts. I have chosen 8 concepts and
identified two criteria for the (hard) task of selecting
them: generality (as opposed to specificity) and con-
creteness (as opposed to abstractness). Table 1 shows
the selected concepts along with their characteristics,
while Table 2 shows the obtained definitions for the
concept Screw.
Table 1: Concepts used for the experiment, along with their
characteristics.
General Specific
Abstract politics, justice greed, patience
Concrete food, vehicle screw, radiator
5.4 Results
In this section I report some insights gained from
the analysis of the experiment results. The collected
96 definitions
5
have an average length of 56 charac-
ters (16 and 225 for the shortest and longest defini-
tions respectively). No significant difference emerged
from the different concept types (abstract / concrete /
generic / specific).
In order to capture the effectiveness of the defi-
nitions, I carried out different measurement: 1) the
percentage of correct guesses given by the test-
participants on the 12 definitions
6
provided by the
def -participants, aggregated by concept type; 2) the
correlation between definitions features and the ob-
tained accuracy levels.
Wrong guesses were of different types: 1) hyper-
nyms or hyponyms (e.g., vehicle wheeled vehicle
motor vehicle); 2) sister terms (e.g., calm instead
of patience, and 3) less related concepts (e.g., stove
instead of radiator). The lexical overlap among def-
initions for the same concepts is very low (less than
20% on average, using stemming and stopwords re-
moval only), as it can be also deduced from the ex-
ample of Table 2. Although semantic-aware lexical
enrichment might increase such value, I left the test-
participants to directly guess the underlying concepts.
Table 3 shows the accuracy values aggregated by con-
cept type. The obtained overall accuracy is 58.75%,
meaning that even with more than ten times of lexical
context at disposal (12 definitions plus 1 from Word-
Net), the participants were often not able to make the
right guess.
Since the scale of the experiment can only have a
limited significance value, this is however indicative
of the fragility of definition-based lexical resources.
While abstract vs concrete concepts revealed small re-
ciprocal differences in the results, definitions of spe-
cific (rather than generic) concepts generally carried
to more correct guesses. Intuitively, physical objects
could be described in terms of specific words whereas
abstractness generally requires steps of generalization
involving a larger set of lexical items and syntactic
5
The dataset will be made available in case of accep-
tance.
6
Actually, the definitions given to the test-participants
were 13 since I also included the WordNet gloss.
WEBIST 2020 - 16th International Conference on Web Information Systems and Technologies
228
Figure 1: Methodology of the human-in-the-loop experiment.
Table 2: Collected definitions for the concept Screw. The most frequent terms (marked in bold) represent taxonomical
information (object, element, item, pin, fastener ), materials (metal), and usages as a tool (used to *). Other reported semantic
information are related to parts (slotted head), size (little), shape (helical, spiral).
Screw (WN 1:06:00)
[WordNet] A fastener with a tapered threaded shank and a slotted head.
Item used to connect artificial parts together.
Metal pin with raised helical thread running around it.
Little metal object which can be inserted in a support.
Threaded metal object used to produce other artifacts.
Metal object used to fix combinable elements.
Object that is used to look and join other components.
Object useful to fix other objects on some surfaces, for example a painting
on the wall.
Metal object with the shape of a spiral used to put things together.
A short, slender, sharp-pointed metal pin with a raised helical thread running
around it and a slotted head, used to join things together by being rotated
so that it pierces wood or other material and is held tightly in place.
Structural element needed to fix two parts.
Long and thin pointy item piercing two objects to hold them together.
structures.
Finally, by manually looking at the best defini-
tions selected by the test-participants, I discovered
the following insights. First, they converged, on av-
erage, on 4.08 definitions among the 13 at disposal
(variance = 1.24). This is indicative of the fact that
definitions do embody significant features which are
recognized to be effective for clearly identifying the
concepts. Second, best definitions contain a vari-
able number of semantic relations (no correlation with
richness). Third, they are on average 65% longer
than non-selected ones (rough estimation of correla-
tion with readability).
What’s in a Definition? An Investigation of Semantic Features in Lexical Dictionaries
229
Table 3: Obtained accuracy values (percentage of correct
answers given by the 10 test-participants), aggregated by
category.
Accuracy General Specific Total
Abstract 50% 75% 62.5%
Concrete 45% 65% 55%
Total 47.5% 70% 58.75%
6 COMPUTATIONAL
EXPERIMENT
In this section, I describe the method and the re-
sults of a semantic analysis of a 1,901-sized corpus of
word definitions covering 300 concrete (noun) con-
cepts. The aim is to find some insights on the type
and statistics related to the semantic information usu-
ally contained within definitions.
6.1 Methodology
As already mentioned in the Introduction, I made use
of BabelNet (Navigli and Ponzetto, 2010) for the ex-
traction of English definitions associated with an in-
put set of 300 concepts (see the next Section 6.2 for
details about the selection process). A total set of
1,901 word definitions have been retrieved, with an
average number of definitions per concept of 6, 34 and
an average number of tokens per definition of 14,55.
Examples of definitions for the concept salad are re-
ported below:
D.1 (WordNet): Food mixtures either ar-
ranged on a plate or tossed and served with
a moist dressing; usually consisting of or in-
cluding greens.
D.2 (Wikipedia): A salad is a dish consisting
of a mixture of small pieces of food, usually
vegetables.
D.3 (WikiData): Dish of raw vegetables.
D.4 (OmegaWiki): Any of a broad variety of
dishes, consisting of (usually raw) chopped
vegetables, most commonly including lettuce.
D.5 (Wiktionary): A food made primarily of
a mixture of raw or cold ingredients, typi-
cally vegetables, usually served with a dress-
ing such as vinegar or mayonnaise.
For each definition, I collected all its nouns and
searched for the following semantic information:
Presence of synonyms (e.g., plane airplane);
Presence of hypernyms (e.g., plane aircraft);
Presence of meronyms (e.g., plane wing);
Presence of purpose-related information (e.g.,
plane transportation).
For the discovery of the last phenomenon, I made use
of simple patterns (e.g., used for, used to, and used
as).
6.2 Concepts Selection
As expected, the accuracy values reached in the first
experiment were higher for concrete concepts (see
Section 5). For this reason, I decided to use con-
crete concepts only. In particular, I manually created
a set of 10 categories
7
covering different conceptual
aspects, then picking 30 concepts for each category
by making use of the WordNet hierarchy.
6.3 Results
An overview of the results of this experiment is shown
in Table 4. As expected, there is a significant use
of hypernyms (according to WordNet), which can be
easily seen as the genus part of the definition. How-
ever, only about 30% of the definitions contain direct
hypernyms, while most of the times more general hy-
pernyms are used. Despite the selection of concrete
objects, only the 10.57% makes use of meronyms.
Similarly, almost the 11% contains purpose-related
semantic information. Finally, an important issue
raised by the experiment regards the usage of syn-
onyms. This phenomenon often indicates the pres-
ence of circularity in the definitions, which can be
problematic (even in not all cases, as demonstrated
in (Burgess, 2007)).
7 CONCLUSIONS
In this contribution, I tried to open a discussion over
the soundness of lexical word definitions as they still
represent the main type of meaning encoding within
semantic resources. A simple content-to-form experi-
ment with different concepts is presented, testing their
general capability to actually uncover the underlying
concepts. Results show that definitions are very frag-
ile means for going back to the concepts. This calls
for further research on the quality and the features of
definitions, with a need for novel interpretable encod-
ing strategies. In addition, a further quantitative and
qualitative analysis on 1,901 word definitions coming
7
The categories have been taken from (Silberer et al.,
2013) and include: animals, musical instruments, tools, arti-
facts, vehicles, food, clothes, home utensils, appliance, con-
tainers.
WEBIST 2020 - 16th International Conference on Web Information Systems and Technologies
230
Table 4: Semantic analysis results of the 1,901 word definitions.
Phenomenon N. of definitions Perc. (%)
Presence of synonyms 599 out of 1,901 31.51%
Presence of hypernyms (direct) 583 out of 1,901 30.67%
Presence of hypernyms (2nd level) 996 out of 1,901 52.39%
Presence of hypernyms (3rd level) 1,254 out of 1,901 65,97%
Presence of hypernyms (all) 1,685 out of 1,901 88,63%
Presence of meronyms 201 out of 1,901 10.57%
Presence of purpose-relations 207 out of 1,901 10.89%
from different sources about 300 concrete concepts
highlighted possible features for their automatic gen-
eration and extraction from large corpora.
REFERENCES
Baker, C. F., Fillmore, C. J., and Lowe, J. B. (1998). The
berkeley framenet project. In Proc. of ACL, pages 86–
90. Association for Computational Linguistics.
Blei, D. M., Ng, A. Y., and Jordan, M. I. (2003). Latent
dirichlet allocation. Journal of machine Learning re-
search, 3(Jan):993–1022.
Bojanowski, P., Grave, E., Joulin, A., and Mikolov, T.
(2016). Enriching word vectors with subword infor-
mation. arXiv preprint arXiv:1607.04606.
Bond, F. and Foster, R. (2013). Linking and extending an
open multilingual wordnet. In ACL (1), pages 1352–
1362.
Brown, R. and McNeill, D. (1966). The “tip of the tongue”
phenomenon. Journal of verbal learning and verbal
behavior, 5(4):325–337.
Burgess, J. A. (2007). When is circularity in definitions
benign? The Philosophical Quarterly, 58(231):214–
233.
Carlson, A., Betteridge, J., Kisiel, B., Settles, B., H. Jr,
E. R., and Mitchell, T. M. (2010). Toward an archi-
tecture for never-ending language learning. In AAAI,
volume 5, page 3.
Devereux, B. J., Tyler, L. K., Geertzen, J., and Randall, B.
(2014). The cslb concept property norms. Behavior
research methods, 46(4):1119–1127.
Dumais, S. T. (2004). Latent semantic analysis. An-
nual review of information science and technology,
38(1):188–230.
Fillmore, C. J. (1977). Scenes-and-frames semantics. Lin-
guistic structures processing, 59:55–88.
G
¨
ardenfors, P. (2004). Conceptual spaces: The geometry of
thought. MIT press.
Harris, Z. S. (1954). Distributional structure. Word, 10(2-
3):146–162.
Huang, E. H., Socher, R., Manning, C. D., and Ng, A. Y.
(2012). Improving word representations via global
context and multiple word prototypes. In Proc. of
ACL, pages 873–882.
Iacobacci, I., Pilehvar, M. T., and Navigli, R. (2015).
Sensembed: learning sense embeddings for word and
relational similarity. In Proceedings of ACL, pages
95–105.
Klavans, J. L., Popper, S., and Passonneau, R. (2003). Tack-
ling the internet glossary glut: Automatic extraction
and evaluation of genus phrases. In Proceedings of
Semantic Web Workshop, SIGIR.
Leone, V., Siragusa, G., Di Caro, L., and Navigli, R. (2020).
Building semantic grams of human knowledge. In
Proceedings of The 12th Language Resources and
Evaluation Conference, pages 2991–3000.
McNamara, D. S., Crossley, S. A., and McCarthy, P. M.
(2010). Linguistic features of writing quality. Written
communication, 27(1):57–86.
McRae, K., Cree, G. S., Seidenberg, M. S., and McNorgan,
C. (2005). Semantic feature production norms for a
large set of living and nonliving things. Behav. r. m.,
37(4):547–559.
Mikolov, T., Chen, K., Corrado, G., and Dean, J. (2013).
Efficient estimation of word representations in vector
space. arXiv preprint arXiv:1301.3781.
Miller, G. A. (1995). Wordnet: a lexical database for en-
glish. Communications of the ACM, 38(11):39–41.
Moerdijk, F., Tiberius, C., and Niestadt, J. (2008). Access-
ing the anw dictionary. In Proc. of the workshop on
Cognitive Aspects of the Lexicon, pages 18–24.
Muresan, S. and Klavans, J. (2002). A method for automat-
ically building and evaluating dictionary resources. In
LREC, volume 2, pages 231–234.
Navigli, R. and Ponzetto, S. P. (2010). Babelnet: Building a
very large multilingual semantic network. In Proc. of
ACL, pages 216–225. Association for Computational
Linguistics.
Pennington, J., Socher, R., and Manning, C. D. (2014).
Glove: Global vectors for word representation. In
EMNLP, volume 14, pages 1532–43.
Silberer, C., Ferrari, V., and Lapata, M. (2013). Models
of semantic representation with visual attributes. In
Proceedings of the 51st Annual Meeting of the Associ-
ation for Computational Linguistics (Volume 1: Long
Papers), volume 1, pages 572–582.
Speer, R., Chin, J., and Havasi, C. (2016). Conceptnet
5.5: An open multilingual graph of general knowl-
edge. arXiv preprint arXiv:1612.03975.
Speer, R. and Havasi, C. (2012). Representing general re-
lational knowledge in ConceptNet 5. In LREC, pages
3679–3686.
Strehlow, R. A. (1983). Terminology and the well-formed
What’s in a Definition? An Investigation of Semantic Features in Lexical Dictionaries
231
definition. In Standardization of Technical Terminol-
ogy: Principles and Practices. ASTM International.
Thorat, S. and Choudhari, V. (2016). Implementing a re-
verse dictionary, based on word definitions, using a
node-graph architecture. In Proceedings of COLING
2016, the 26th International Conference on Compu-
tational Linguistics: Technical Papers, pages 2797–
2806.
Witte, S. P. and Faigley, L. (1981). Coherence, cohesion,
and writing quality. College composition and commu-
nication, 32(2):189–204.
WEBIST 2020 - 16th International Conference on Web Information Systems and Technologies
232