Improving Robustness and Flexibility of Concept Taxonomy Learning from Text

Improving Robustness and Flexibility of
Concept Taxonomy Learning from Text

F. Leuzzi1 , S. Ferilli1,2 , C. Taranto1 , and F. Rotella1
1
Dipartimento di Informatica – Universit` di Bari
a
{fabio.leuzzi, fulvio.rotella}@uniba.it
{ferilli, claudio.taranto}@di.uniba.it
2
Centro Interdipartimentale per la Logica e sue Applicazioni – Universit` di Bari
a

Abstract. The spread and abundance of electronic documents requires
automatic techniques for extracting useful information from the text they
contain. The availability of conceptual taxonomies can be of great help,
but manually building them is a complex and costly task. Building on
previous work, we propose a technique to automatically extract concep-
tual graphs from text and reason with them. Since automated learning of
taxonomies needs to be robust with respect to missing or partial knowl-
edge and flexible with respect to noise, this work proposes a way to deal
with these problems. The case of poor data/sparse concepts is tackled
by finding generalizations among disjoint pieces of knowledge. Noise is
handled by introducing soft relationships among concepts rather than
hard ones, and applying a probabilistic inferential setting. In particular,
we propose to reason on the extracted graph using different kinds of re-
lationships among concepts, where each arc/relationship is associated to
a number that represents its likelihood among all possible worlds, and
to face the problem of sparse knowledge by using generalizations among
distant concepts as bridges between disjoint portions of knowledge.

1 Introduction

The spread and abundance of electronic documents requires automatic tech-
niques for extracting useful information from the text they contain. The avail-
ability of conceptual taxonomies can be of great help, but manually building
them is a complex and costly task. Obtaining automatically Full Text Under-
standing is not trivial, due to the intrinsic ambiguity of natural language and to
the huge amount of common sense and linguistic/conceptual background knowl-
edge needed to switch from a purely syntactic representation to the underlying
semantics. Nevertheless, even small portions of such knowledge may significantly
improve understanding performance, at least in limited domains. Although stan-
dard tools, techniques and representation formalisms are still missing, lexical
and/or conceptual taxonomies can provide a useful support to many NLP tasks,
allowing automatic systems to exploit different kinds of relationships that are
implicit in the text but required to correctly understand it. Building on previous
work, we propose a technique to automatically extract conceptual graphs from

text and reason with them. Since automated learning of taxonomies needs to be
robust with respect to missing or partial knowledge and flexible with respect to
noise and to the need of shifting the representation, this work proposes a way
to deal with these problems. Data poorness, concept sparsity and representation
shift are tackled by finding generalizations among disjoint pieces of knowledge
in order to build a bridge between them. Noise is handled by introducing soft
relationships among concepts rather than hard ones, and applying a probabilis-
tic inferential setting to reason on the extracted graph using different kinds of
relationships, where each arc/relationship is associated to its likelihood of being
true in all possible worlds.
This work is organized as follows: the next section describes related works;
Section 3 outlines the proposed approach; then we present an evaluation of our
solution; lastly we conclude with some considerations and future works.

2 Related Work

Many approaches have been attempted to build taxonomies and ontologies from
text. [1] builds concept hierarchies using Formal Concept Analysis by grouping
objects with their attributes, which are determined from text linking terms with
verbs. A few examples of other approaches: [10, 9] build ontologies by labelling
taxonomical relations only, while we label also non-taxonomical ones with actions
(verbs); [12] builds a taxonomy considering only concepts that are present in a
domain but do not appear in others, while we are interested in all recognized
concepts independently of their being generic or domain-specific; [11] defines a
language to build formal ontologies, while we are interested in the lexical level.
As regards our proposal, for the syntactic analysis of the input text we ex-
ploit the Stanford Parser and Stanford Dependencies [8, 2], two very effective
tools that can identify the most likely syntactic structure of sentences (including
active/passive and positive/negative forms), and specifically ‘subject’ or ‘(di-
rect/indirect) object’ components. They also normalize the words in the input
text using lemmatization instead of stemming, which allows to distinguish their
grammatical role and is more comfortable to read by humans.
We also apply probabilistic reasoning on the extracted knownledge using
ProbLog [13]. It is essentially Prolog where all clauses are labelled with the prob-
ability that they are true, that in turn can be extracted from large databases
by various techniques. A ProbLog program T = {p1 : c1, ..., pn : cn} specifies
a probability distribution over all its possible non-probabilistic subprograms ac-
cording to the theoretical basis in [14]. The semantics of ProbLog is then defined
by the success probability of a query, which corresponds to the probability that
the query succeeds in a randomly sampled program. Indeed, the program can be
split into a set of labelled facts pi :: fi , meaning that fi is a fact with probability
of occurrence pi , and a Prolog program using those facts, which encodes the back-
ground knowledge (BK ). Probabilistic facts correspond to mutually independent
random variables (RV s), which together define a probability distribution over

all ground logic programs L ⊆ LT (where LT is the set of all fi ’s):

P (L|T ) = pi (1 − pi )
fi ∈L fi ∈LT L

In this setting we will use the term possible world to denote the least Herbrand
model of a subprogram L together with BK, and we will denote by L both the
set of sampled facts and the corresponding world.
Lastly, we need in some steps of our technique to assess the similarity among
concepts in a given conceptual taxonomy. A classical, general measure, is the
Hamming distance [6], that works on pairs of equal-lenght vectorial descriptions
and counts the number of changes required to turn one into the other. Other
measures, speciﬁc for conceptual taxonomies, are mathrmsf F a [4] (that adopts
a global approach based on the whole set of hypernyms) and sf W P [15] (that
focuses on a particular path between the nodes to be compared).

3 Proposed Approach

This proposal relies on a previous work [5], in which we assume that each noun
in the text corresponds to an underlying concept (phrases can be preliminarily
extracted using suitable techniques, and handled as single terms). A concept
is described by a set of characterizing attributes and/or by the concepts that
interact with it in the world described by the corpus. The outcome is a graph,
where nodes are the concepts/nouns recognized in the text, and edges represent
the relationships among these nodes, expressed by verbs in the text (whose
direction denotes their role in the relationship). In this work, we handle noise
by weighting the relationships among concepts, this way obtaining a semantic
network with soft links between nodes.

3.1 Graph Construction

Natural language texts are processed by the Stanford Parser in order to extract
triples subject, verb, complement that will represent the concepts (the subjects
and complements) and attributes (verbs) for the graph. We have adopted some
representational tricks: indirect complements are treated as direct ones by em-
bedding the corresponding preposition into the verb; sentences involving verb
‘to be’ or nouns with adjectives contributed in building the sub-class structure
of the taxonomy (e.g., “the penguin is a bird” yields is a(penguin,bird )). In this
way we have two kind of edges among nodes in the graph: verbal ones, labelled
with the verb linking the two concepts, and taxonomic (is a) ones. In order to
enrich the representation formalism previously deﬁned, we analyzed the syntatic
tree to seize the sentence positive or negative form based on the absence or pres-
ence (respectively) of a negation modifer for the verb. Moreover we decided to
take into account separately the frequency of each arc between the concepts in
positive and negative sentences.

This setting allowed us to give robusteness to our solution through a statis-
tical approach. In fact, the obtained taxonomy could be inspected and used by
filtering out all portions that do not pass a given level of reliability. This could
be useful for the identification of relevant concept, as shown in [5], or for other
applications that will be explained in the next two subsections.

3.2 Probabilistic Reasoning ‘by association’
Reasoning ‘by association’ means finding a path of pairwise related concepts that
establishes an indirect interaction between two concepts c and c in the seman-
tic network. However, since real world data are typically noisy and uncertain,
there is a need for strategies that soften the classical rigid logical reasoning.
In particular, when knowledge is learned from text, as in our case, we might
run into the above problems, which require an inference engine that allows to
perform several kinds of probabilistic queries, like getting the minimal path, or
choosing the best (i.e., the most likely) path, rather than computing the exact
probability of all possibile ones.
In this work we propose two reasoning stategies: the former aims at obtaining
the minimal path between concepts together with all involved relations, the latter
relies on ProbLog in order to allow probabilistic queries on the conceptual graph.
In the former strategy, we look for a minimal path using a Breadth-First
Search (BFS) technique, applied to both concepts under consideration. The ex-
pansion steps of the two processes are interleaved, checking at each step whether
the new set of concepts just introduced has a non-empty intersection with the
set of concepts of the other process. When this happens, all the concepts in such
an intersection identify one or more shortest paths connecting c and c , that
can be retrieved by tracing back the parent nodes at each level in both directions
up to the roots c and c . Since this path is made up of concepts only, to obtain
a more sensible ‘reasoning’ it must be filled with the specific kind of interaction
represented by the labels of edges (verbs) that connect adjacent concepts in the
chain. We also provide the number of positive/negative instances, and the corre-
sponding ratios over the total, to help understanding different gradations (such
as permitted, prohibited, typical, rare, etc.) of actions between two objects.
The latter strategy requires the definition of a formalism based on ProbLog
language. We used pi :: fi where fi is a ground literal of the form link (subject,
verb, complement) and pi is ratio between the sum of all examples for which
fi holds and the sum of all possibile links between subject and complement.
Figure 1 shows an example where many links are present between farmer and
plant, expressing different kinds of interactions between these two concepts on
which the above kinds of reasoning can be applied. One might be interested in
getting minimal paths, regardless of their probability, from farmer to trasporter,
reporting all traversed concepts and verbs. Or, one might want to compute the
explanation probability among all possibile path in the SLD-tree between these
two nodes, obtained as the product of all traversed links together with the proof
of the query (most likely explanation). One of the problems of these approaches
is the tight connection between the quality of the reasoning results and that

transporter

1.0::carry

f armer cow

0.3::prune 0.25::cultivate 0.35::eat 0.3::eat 0.6::trample

0.1::own plant

0.3::grow on

ground

Fig. 1. Example of ProbLog Network

of the network, in turn depending on the processed texts. Indeed, if two nodes
belong to disjoint regions of the graph, reasoning cannot succeed. We tackle this
problem by defining a generalization operator as follows.

3.3 Generalization operator

Let us first provide a more precise account of our generalization.
Definition. (Generalization)
Given two concepts G and C, G generalizes C if anything that can be labelled
as C can be labelled as G as well, but not vice-versa.
The use of generalizations provides many opportunities of enrichment and/or
manipulations on the graph:

1. building taxonomic structures by (repeatedly) applying this operator, also
after the addition of new text (possibly causing the presence of new nodes
in the graph);
2. shifting the representation, by removing the generalized nodes from the
graph and leaving just their generalization (that inherits all their relation-
ships to other concepts);
3. extending the amount of relationships between concepts belonging to the
same connected component of the graph, or building bridges between disjoint
components that open new (previously impossible) reasoning paths.

In particular, as regards the third case, it can be viewed as a tool that aids
reasoning ‘by association’ to reach its objectives, and hence as a multi-strategy
setting in which induction and deduction cooperate.
More formally, we provide the following definition.
Definition. (Bridge)
Given a piece of knowledge K represented as a concept graph made up of different
connected components, a bridge in K is a modification of the graph that connects
some of such components, allowing to reach any node of either from any node of
another. A bridge that unifies n isolated sub-graphs is named n-way bridge.

Regardless of the task, three steps can be outlined to describe the general
procedure:
1. Concept Grouping, in which all concepts are grossly partitioned to obtain
subsets of concepts: we group similar concepts if the aim is to enrich the
relationships, or dissimilar ones in the bridging perspective (Algorithm 1);
2. Word Sense Disambiguation, that associates a single meaning to each term
by solving possible ambiguities using the domain of discourse;
3. Computation of taxonomic similarity, in which WordNet [3] is exploited in
order to further filter with an external source the groups found in step 1,
and to choose an appropriate subsumer (Algorithm 2).
To generalize two or more concepts, we use as their description their direct
neighbor concepts and the verbs (in positive or negative form) used to connect
them. Thus, considering the set Attributes of all verbs in their positive and
negative forms, we build a Concepts × (Concepts ∪ Attributes) matrix C where:
Ci,j = 1 if j denotes a concept column and there is at least a relationship between
concepts i and j; Ci,j = 1 if j denotes an attribute column and there is at least
a relationship between concept i and verb j; Ci,j = 0 otherwise. Each row is a
feature vector describing the concept, and hence two vectors can be compared
according to the Hamming distance. Pairwise clustering under the complete link
assumption is applied to these descriptions: initially, each non-null row becomes a
singleton cluster; then, clusters are merged while a merging condition is fulfilled.
In its standard view, complete link states that the distance of the farthest items
of the involved clusters must be less than a given threshold.
As stated earlier, generalizations can be carried out for different objectives,
affecting the way complete link is applied. In particular, if the objective is the
enrichment of relationships within connected components, it is applied in the
standard way, otherwise, if the objective is to build bridges, the distance of
the closer items of the involved clusters must be greater than a given threshold.
When the condition is satisfied, the average score between all pairs of items
in the two clusters is saved, and only the pair of clusters corresponding to the
smallest (respectively, greatest) average is merged. We define more formally the
clustering of dissimilar objects as follows.
Definition. (Inverse clustering)
Given a set of objects and a distance measure, the Inverse clustering is obtained
grouping iteratively most dissimilar objects until a given threshold of distance
is satisfied.
Now, the clusters contain similar (resp., dissimilar) concepts that can be
generalized in order to create new relationships (resp, to merge nodes) for the
enrichment (resp., bridging), but this procedure alone might not be reliable, be-
cause terms that occur seldom in the corpus have few connections (which would
affect their cluster assignment due to underspecification) and because the ex-
pressive power of this formalism is too low to represent complex contexts (which
would affect even more important concepts). Note that the bridging setting is
less affected by the underspecification problem, because it tends to group dissim-
ilar concepts. Since underspecification corresponds to almost zero vectors, taking

dissimilar vectors we tend to group the most specified distant vectors. Indeed,
since there cannot be any 1 in the same positions in both descriptions (because
this would mean that there exists a path between them), the more 1’s overall in
the two descriptions, the larger their distance, which means that the bridge is
merging two hub (i.e., highly connected) nodes. This clearly improves the quality
of the bridge. This solution allows to limit the shortest medium length among all
possible paths built between the sub-graphs that the bridge connects. However,
the support of an external resource might be desirable. We consider WordNet
as a sensible candidate for this, and try to map each concept in the network to
the corresponding synset (a non trivial problem due to the typical polysemy of
many words) using the one domain per discourse assumption as a simple crite-
rion for Word Sense Disambiguation, whose algorithm is described in [5]. Thus,
WordNet allows to check and confirm/reject the similarity of concepts belonging
to the same cluster, by considering all possible pairs of concepts whose similarity
is above a given threshold. Let us state this more formally.
Definition. (Star)
Given a set of unordered pairs of items S and another unordered pair P = {A, B}
s.t. P ∈ S, its Star is a set of pairs S s.t. ∀P ∈ S , (A ⊂ P ∨ B ⊂ P ) ∧ P = P
(i.e., the set of pairs of items that contain A or B).
Definition. (Generalization set)
Given a cluster of concepts C, a similarity function sf (), a similarity threshold T ,
and a pair P = {A, B} s.t. P ∈ C, P is a Generalization set if it can be covered by
a most specific common subsumer G, sf (P ) > T , and ∀P ∈ C, sf (P ) > sf (P )
(i.e., has the largest similarity value among all possible pairs). Moreover, each
pair P in the Star of P s.t. P > T that can be covered from G, is used to
extend the Generalization set.
Where most specific common subsumer is the first common node found ex-
ploring the hypernym relations in WordNet. Similarity is determined using a mix
of the measures proposed in [4] and in [15], to consider both a global perspective
which avoids the choice of a single path and the actual generalizing path:

sf (A, B) = sfF a (A, B) · sfW P (A, B)

The underlying idea is that the former general measure can be weighted with
the latter real measure that represent the single path from which the subsumer
is taken, and vice-versa.

4 Evaluation

The proposed approach was evaluated using ad-hoc tests that may indicate its
strengths and weaknesses. Due to lack of space, only a few selected outcomes
will be reported here. Although preliminary, these results seem enough to suggest
that the approach is promising. We exploited a dataset made up of documents
concerning social networks on socio-political and economic topic. It was chosen
deliberately small in order to tackle problems as noise and poor knowledge.

Algorithm 1 Pair-wise clustering of all concepts in the network.
Input: matrix C × C+A (where C is the set of objects/concepts, and A is the set of positive
and negative verbs), that for each cell has the value 1 if at least one link exist, 0 otherwise;
THRESHOLD for Hamming distance.
Output: set of clusters.
pairs ← empty
averages ← empty
for all Oi | Oi = zero vector ∧ i ∈ C do
newCluster ← Oi
clusters.add(newCluster)
end for
for all pair(Ck , Cz ) | C ∈ clusters ∧ k, z ∈ [0, clusters.size] do
if completeLink(Ck , Cz ) then
pairs.add(Ck , Cz )
averages.add(getScoreAverage(Ck , Cz ))
end if
end for
pair ← getP airW ithM axM in(pairs, averages)
merge(pair)
completeLink → check the complete link assumption for the passed clusters.
getPairWithMaxMin → get the pair with the maximum or minimum average depending on task.

4.1 Probabilistic Reasoning ‘by association’

Table 1 shows a sample of outcomes of reasoning ‘by association’. In this ex-
periment we want to investigate the minimum path (if any) between two nodes
in the graph. In order to show the strength of this kind of reasoning, each verb
is labelled with the frequency with which it occurs in the paths subject, verb,
complement. Looking at case 1, we wanted to explore the relationship between
young and schoolwork. This chain includes verb look, that occurs only in posi-
tive sentences with probability 1.0; this means that our knowledge is coherently
indicating that young “always” look television. In the same case there is the re-
lationship between facebook and schoolwork, in fact the verb help appears with
probability 0.25 in positive and 0.75 in negative sentences. We can interpret this
behaviour as facebook “may” help schoolwork, or with the speciﬁc associated
probability. We need to point out that this reasoning strategy shows all possibile
verbs between each pair of adjacent nodes in the path.
On the other hand, Table 2 shows a sample of probabilistic queries executed
on the same paths. We have computed the exact success probability, the most
likely explanation probability and an approximate probability according to the
MonteCarlo method implemented in [7] keeping a 95% of conﬁdence interval.
The last type of query has been provided because exact inference can be in-
tractable even for small networks, and so sometimes it can be reasonable to
make approximate inference. The probability of each sentence was computed
according to the criteria in section 3.2. For instance, in case 2, the explanation
proof entails the set {(young, look, television), (television, talk about, facebook),
(facebook, not help, schoolwork)}, respectively with probability {1.0, 0.75, 0.5},
whose product is 0.375.

Algorithm 2 Effective generalization research.
Input: the set of C clusters returned by pair-wise clustering; T similarity threshold; max the
max number of generalizations to try to extract from a cluster.
Output: set of candidate generalizations.
generalizations ← empty set
for all c ∈ C do
good pairs ← empty set
for all pair(Oi , Oj ) | i, j ∈ c do
if similarity score(pair(Oi , Oj )) > T then
good pairs.add(pair(Oi , Oj ), wordnet hypernym(pair(Oi , Oj )))
end if
end for
for all i ∈ [0, max] do
if good pairs = empty set then
new set ← {good pairs.getBestP air, good pairs.getStar}
generalizations.add(new set)
good pairs.remove(new set)
end if
end for
end for
good pairs → contains a list of pairs that satisfy T , with their relative subsumer.
good pairs.getBestPair → get the pair that has the best similarity score.
good pairs.getStar → get the Star of the pair
good pairs.remove → remove all pairs in the passed set.
wordnet hypernym → get the subsumer discovered in WordNet for the passed pair.

4.2 Evaluation of generalization operator
Two toy experiments are reported for concept generalization, the former aimed
to the enrichment of relationships, the latter with the bridging perspective. The
maximum threshold for the Hamming distance was set to 0.001, while the mini-
mum threshold for taxonomic similarity was fixed at 0.45 in both. The exploited
dataset included 669 objects (subjects and/or complements) and 727 verbs.
In Table 3 we can see that, coherently with the enrichment-only perspective,
no bridges are built. Conversely, applying Inverse clustering we obtain as ex-
pected also bridges among two or more disjoint graph regions. Analyzing the
two conceptual similarity measures in both experimental settings, they both re-
turn very high values for almost all pairs, leading to final scores that neatly pass

Table 1. Examples of smooth reasoning ‘by association’ throught BFS (start and
target nodes in emphasis).

# Subject Verb Complement
young look [Pos: 3/3] television
1 television talk about [Pos: 3/3], critic [Pos: 1/1] facebook
facebook help [Pos: 1/4, Neg: 3/4], distract [Pos: 1/1] schoolwork
people be in [Pos: 1/1] group
2 group invite [Pos: 2/2] facebook
facebook help [Pos: 1/1] individual
everyone want [Pos: 5/5] occupation
3 occupation maintain [Pos: 1/1] lifestyle
lifestyle see in [Pos: 1/1], change in [Pos: 1/1], media

Table 2. Examples of probabilistic reasoning ‘by association’ throught ProbLog.

# Query Probability
1 problog exact(path(young - schoolwork )) 0.530
2 problog max(path(young - schoolwork )) 0.375
3 problog approx(path(young - schoolwork )) 0.555
4 problog exact(path(people - individual)) 0.167
5 problog max(path(people - individual)) 0.167
6 problog approx(path(people - individual)) 0.162
7 problog exact(path(occupation - media)) 0.750
8 problog max(path(occupation - media)) 0.500
9 problog approx(path(occupation - media)) 0.744

Table 3. Generalizations for pairwise clustering of similar concepts, and corresponding
conceptual similarity scores (bottom).

# Bridge Subsumer Subs. Domain Concepts Conc. Domain
variable [105857459] mathematics
1 No variable [105857459] mathematics
factor [105858317] mathematics
type [109909060] person
2 No person [100007846] biology, person
collegian [109937056] factotum
type [109909060] person
3 No person [100007846] biology, person
model [110324851] person
nineteen [113747989] number
4 No integer [113728499] mathematics
forty [113749527] number
scholar [110251779] school
5 No person [100007846] biology, person job [110222949] person
name [110344443] person
# Pairs F a score W P score Score
1 variable, factor 0.7 0.857 0.6
2 type, collegian 0.659 0.737 0.486
3 type, model 0.692 0.778 0.538
4 nineteen, forty 0.75 0.75 0.562
scholar, job 0.711 0.823 0.585
5
scholar, name 0.678 0.778 0.528

the 0.45 threshold. Another very interesting result is that sf W P is always greater
than sf F a when the generalizations are enrichments. Since the former is more
related to a specific path, and hence to the goodness of the chosen subsumer, this
confirms the previous outcomes (suggesting that the chosen subsumer is close
to the generalized concepts). This regularity is lost in Table 4, when Inverse
clustering was used to build bridges, validating the motivations for which sf F a
is used: the identification of similar concepts when this is not so evident based
only on the single subsuming path. Observing the outcome, three aspects can
be emphasized: the effectiveness of the searching for bridges in case 5, in which
a three-way bridge has been built; the overall quality of the generalizations ob-
tained; and the opportunity to perform not only representation shifts, but also
an alignment as in case 1 of Table 3. Note that, after enriching the network, we
are now able to reason ‘by association’ also between the new paths opened by
generalization. We have decided to assign probability 1.0 to each generalization
because an oracle, in our case WordNet, has suggested the common subsumer.
For example, if we search for a path between serenity and food, we have to pass
through verbs {find [Pos: 2/2], want [Pos: 3/3] } for reaching type and then fol-

Table 4. Generalizations for pairwise Inverse clustering, and corresponding conceptual
similarity scores (bottom).

# Bridge Subsumer Subs. Domain Concepts Conc. Domain
teaching talk [100893243] pedagogy
1 Yes pedagogy
[100887081] lecture [100892861] pedagogy
person scientist [110560637] person
2 Yes biology, person
[100007846] youth [110804406] person
music introduction [106396930] literature, music
3 Yes music
[107020895] end [106398401] literature, music
figure two [113743269] number
4 Yes number
[113741022] pair [113743605] number
viewer [110633450] person
person
5 Yes biology, person model [110324851] person
[100007846]
volunteer [110759151] person
theory basis [105793554] factotum
6 No factotum
[105888929] theory [105888929] factotum
administration, hometown [108671644] geography
municipality
7 No geography, literature, administration,
[108626283] city [108524735]
town planning geography, town planning
territorial dominion administration, state [108544813] geography
8 No
[108552138] town planning department [108548733] geography
structure wall [104546855] buildings
9 No factotum
[104341686] level [103365991] buildings
photography, racing,
representation scene [106614729]
10 No factotum sociology, telecommunication
[104076846]
photo [103925226] photography
# Pairs F a score W P score Score
1 talk , lecture 0.739 0.869 0.643
2 scientist, youth 0.724 0.727 0.526
3 introduction, end 0.75 0.714 0.536
4 two, pair 0.72 0.823 0.593
viewer, model 0.71 0.666 0.474
5
viewer, volunteer 0.71 0.666 0.474
6 basis, theory 0.694 0.842 0.584
7 hometown, city 0.738 0.842 0.621
8 state, department 0.75 0.75 0.562
9 wall, level 0.733 0.8 0.586
10 scene, photo 0.735 0.823 0.605

low the new arc to reach model ; then, through the verbs {eat [Pos: 1/4 Neg:3/4],
buy [Pos: 2/2] } we finally reach food.

5 Conclusions
This work proposes a technique to automatically extract conceptual graphs from
text and reason with them. In particular a way to deal with missing or partial
knowledge, noise and to the need of shifting the representation is presented,
facing these problems through reasoning ‘by association’ and the generalization
operator. Preliminary experiments confirm the results of previous works and
show that this approach can be viable. Further extensions and refinements are
needed to improve the effectiveness of this approach, as the definition of an
Anaphora Resolution strategy including also handling the Named Entities, or a
study on how to set automatically suitable thresholds for searching generaliza-
tions, or lastly a way for freeing from WordNet in the choise of the appropiate
subsumer.

References
[1] P. Cimiano, A. Hotho, and S. Staab. Learning concept hierarchies from text
corpora using formal concept analysis. J. Artif. Int. Res., 24(1):305–339, August
2005.
[2] M.C. de Marneffe, B. MacCartney, and C. D. Manning. Generating typed depen-
dency parses from phrase structure trees. In LREC, 2006.
[3] C. Fellbaum, editor. WordNet: An Electronic Lexical Database. MIT Press, Cam-
bridge, MA, 1998.
[4] S. Ferilli, M. Biba, N. Di Mauro, T.M. Basile, and F. Esposito. Plugging taxonomic
similarity in first-order logic horn clauses comparison. In Emergent Perspectives
in Artificial Intelligence, Lecture Notes in Artificial Intelligence, pages 131–140.
Springer, 2009.
[5] S. Ferilli, F. Leuzzi, and F. Rotella. Cooperating techniques for extracting concep-
tual taxonomies from text. In Proceedings of The Workshop on Mining Complex
Patterns at AI*IA XIIth Conference, 2011.
[6] R.W. Hamming. Error detecting and error correcting codes. Bell System Technical
Journal, 29(2):147–160, 1950.
[7] A. Kimmig, V. S. Costa, R. Rocha, B. Demoen, and L. De Raedt. On the efficient
execution of problog programs. In Maria Garcia de la Banda and Enrico Pontelli,
editors, ICLP, volume 5366 of Lecture Notes in Computer Science, pages 175–189.
Springer, 2008.
[8] D. Klein and C. D. Manning. Fast exact inference with a factored model for
natural language parsing. In Advances in Neural Information Processing Systems,
volume 15. MIT Press, 2003.
[9] A. Maedche and S. Staab. Mining ontologies from text. In EKAW, pages 189–202,
2000.
[10] A. Maedche and S. Staab. The text-to-onto ontology learning environment. In
ICCS-2000 - Eight International Conference on Conceptual Structures, Software
Demonstration, 2000.
[11] N. Ogata. A formal ontology discovery from web documents. In Web Intelligence:
Research and Development, First Asia-Pacific Conference (WI 2001), number
2198 in Lecture Notes on Artificial Intelligence, pages 514–519. Springer-Verlag,
2001.
[12] A. Cucchiarelli P. Velardi, R. Navigli and F. Neri. Evaluation of OntoLearn, a
methodology for automatic population of domain ontologies. In Ontology Learning
from Text: Methods, Applications and Evaluation. IOS Press, 2006.
[13] L. De Raedt, A. Kimmig, and H. Toivonen. Problog: a probabilistic prolog and its
application in link discovery. In In Proceedings of 20th IJCAI, pages 2468–2473.
AAAI Press, 2007.
[14] T. Sato. A statistical learning method for logic programs with distribution se-
mantics. In Proceedings of the 12th ICLP 1995, pages 715–729. MIT Press, 1995.
[15] Z. Wu and M. Palmer. Verbs semantics and lexical selection. In Proceedings
of the 32nd annual meeting on Association for Computational Linguistics, pages
133–138, Morristown, NJ, USA, 1994. Association for Computational Linguistics.

Improving Robustness and Flexibility of Concept Taxonomy Learning from Text

More Related Content

What's hot (17)

Viewers also liked (9)

Similar to Improving Robustness and Flexibility of Concept Taxonomy Learning from Text (20)

Recently uploaded (20)

Improving Robustness and Flexibility of Concept Taxonomy Learning from Text