We propose in this paper an approach to learn term to concept mapping with the joint utilization of verb relations and an existing ontology. This is a non-supervised solution that can be applied to any field for which an ontology modeling verbs as relations holding between the concepts was already created. Conceptual graphs, representing a set of verb relations, are learned from a natural language corpus by using part-of-speech information and statistic measures. Labeling strategies are proposed to assign terms of the corpus to concepts of the ontology by taking into account the structure of the ontology and the extracted conceptual graphs. Results of this assignation could be used to automatically create semantic annotations of documents. A first experimentation in the field of accidentology was done and its results are also presented.
The rapid evolution in the production of documents in natural language requires the definition of efficient automated approaches allowing finding relevant information in those documents. This paper presents an approach that uses verb relations and a domain ontology to assign terms of a given corpus to concepts of the field. Those assignations can be used thereafter for various exploitation scenarios, that is to say: semantic annotation of documents, estimating similarities between documents, etc.
This approach is based on an entirely automatic and nonsupervised process, unless the use of a domain ontology to support the process.
The task to achieve could be described as follows: let o be a domain ontology and c a collection of domain-specific texts.
Sylvie Desprès LIPN UMR CNRS 7030 - University of Paris 13 99 avenue Jean Baptiste Clément
93430 Villetaneuse, France sylvie.despres@lipn.univ-paris13.fr For this work, we assume that the ontology takes into account the linguistic level of entities. Thus, concepts and roles are labeled by terms, which are linguistic manifestation of ontology entities in a specific language (French, English, etc.). Therefore, ontology considered for this work has two levels: a conceptual level, describing domain specific entities (concepts and roles) and a linguistic level, providing expressions of those entities in a given language.
The goal of this approach is to identify within c terms t representing linguistic expression of concepts of o ontology. Thus, we can label terms identified in the corpus by concepts of ontology. We propose a three steps approach to carry out this labeling process: (1) in a first stage, verb relations are extracted from the corpus. Each verb relation is composed of a verb, be that a general one or a field specific one, and a pair of terms connected by this verb. (2) in a second phase, statistical processing is performed to structure verb relations as conceptual graphs. As the verb is considered to be the key element of a verb relation, it is placed at the top of the conceptual graph. Terms occurring as arguments of the verb are connected to this verb through links representing theirs syntactic function which could be subject or object. (3) the last phase is based on the assumption that the domain ontology models verbs of the field as relations holding between the concepts. If this is the case, labeling strategies are using both the ontology and extracted conceptual graphs to assign field specific terms to field specific concepts.
We shall approach that topic by answering a number of questions: which method should be used to extract verb relations from corpus? How to learn conceptual graphs from the extracted verb relations? Those questions are analyzed in sections 2 and 3. Given a domain ontology and a set of conceptual graphs, which strategies will be used to assign terms to concepts? The solution is discussed in section 4. A first experimentation in the field of accidentology is described and its results are presented in section 5. Related work is presented in section 6. Conclusions and perspectives end this paper.
To extract verb relations from corpus, we adopted an approach based on pattern recognition. This approach is using part-ofspeech information and consists in seeking within the corpus for particular associations of lexical categories. Such an association represents a lexical pattern. For example Verb, Noun or Verb, Preposition, Noun are lexical patterns.
We manually crafted a set of lexical patterns including a verb
(among other categories). Associations of words matching
patterns of this set are identified by a pattern recognition
algorithm, described in [4]. The algorithm takes as input the
corpus tagged by TreeTagger, see [
The goal of this phase is to learn conceptual graphs from the results of pattern recognition algorithm.
A conceptual graph represents a hierarchy having as a top a verb
and, on a second layer, arguments connected to the verb by their
grammatical function, subject or object. We use the term
conceptual graph as it was introduced by [
As many terms could be the subject or object of the same verb, a conceptual graph corresponds to a set of verb relations generated by the same verb. To learn conceptual graphs, the chain of treatments based on lexical similarity measures presented bellow is performed.
A lexical similarity measure associates a real number r to a pair of strings s, t . Important values of r indicate a significant Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. similarity of strings s, t . In a similar way, a lexical distance measure associates a real number to a pair of strings but the interpretation is different: important values of r indicates minor similarity of strings s, t .
Many coefficients were proposed to calculate similarities or
distances between strings. A number of them are presented in
[
Jaccard coefficient calculates the similarity between two strings s and t by considering a string composed of several substrings. Jaccard coefficient is given by:
Jaccard (s, t ) s t s t This measure is takes into account the number of sub-strings common to s and t and the number of all sub-strings of s and t . If we consider characters as sub-strings, the coefficient expresses the similarity by taking into account the number of common characters of s and t only.
Jaro and Jaro-Winkler coefficients, introduced below, express the distance between two strings s, t by taking into account the number and the position of characters shared by s and t. Let s s1...sk and t t1...t k be two strings. A character si in s is considered common to both strings if there exists t j in t such as: si t j and i h j i h , where h min(s, t ) 2
.
Let s1 s11 ...s1k be characters of s common to t and t 1 t11
...t k1 characters of t common to s . We define a transposition between s and t as an index i such as: si1 ti1 . If Ts,t is the number of transpositions from s i to t i the Jaro coefficient calculates the lexical distance between s and t as follows:
s1
1
Jaro(s, t)
3 s
t 1
t
s1 Ts,t
s1
[
...tl1 by performing two steps. First, the two strings are divided into sub-strings; then the similarity is given by: Monge Elkan(s, t )
max lj1sim(si , t j ) k i1 where sim(si , t j ) are given by some similarity function, for instance one of those previously presented. Such a function is called a level 2 function.
Conceptual graphs are learned from the set of lexical pattern instances extracted according to section 2. An iterative solution is proposed, performing a number of steps, each of them adding a new layer to the graphs. (1) The first step identifies verb classes which represent the set of verb relations generated by the same verb, see Table 1. For each verb class, instances of patterns “Verb” and “Verb, Preposition, are added to the set of roots. We argue that for verbs accepting prepositions, each “verb, preposition” pattern accepts specific arguments and for this reason conceptual graphs are created for each instance of those patterns. This step creates a number of conceptual graphs having one level, which is to say the root (see Figure 1). (2) For each root, its arguments are identified: terms that are subjects and objects. As each relation accepts many terms as subject or object, lists of arguments are obtained. This step is adding a second layer to each conceptual graph. (3) We observe that, for a given verb, arguments can have different levels of granularity, as we can see in table 2: partie (side ) partie gauche (left side) partie droite (right side) rétroviseur (rear view mirror) rétroviseur extérieur (external rear view mirror) Hence, a new layer can be added to each conceptual graph by clustering those arguments.
A cluster is a group of similar terms, having a central term c called centroid and its k nearest neighbors. Based on the heuristic that the greater number of words in a word regrouping there are, the more specific his meaning is, an algorithm is proposed to cluster arguments of verb relations. The clustering algorithm is written as follows: (1) for each list of arguments, create the list L of centroids, composed of all one-word arguments; (2) for each centroid c , calculate the lexical similarity with other terms of the list by using Monge-Elkan coefficient; (3) add to cluster c terms having a similarity value greater than a given threshold. An expert intervention allows us to chose the value of this trehsolds.
At that stage, Monge-Elkan function is used because it carries out recursive comparisons between sub-strings. Consequently, it has the capacity to agglomerate around a word (as centroids of clusters are one-word terms, which is to say words), terms derived from this word.
We chose one-word terms as centroids as they have the most general meaning, and, by consequence, will be able to attract into a cluster terms that are similar from a lexical point of view and that have more specific meanings. Figures 1 and 2 show the iterative construction of conceptual graphs. We can see onelevel conceptual graphs learned from diriger (to direct) class and two-level conceptual graphs learned from circuler (to circulate) class. given conceptual graph, the verb R representing its root node is already modeled by the ontology. If this is the case, let r be the corresponding relation and concepts of the ontology connected by r . Those concepts and theirs descendants will be used to label arguments of the verbs. As arguments are connected to verb by links corresponding to the syntactic function, Domainr will be used to label subject Ranger , Domainr arguments, while
A first strategy ignores the hierarchical organization of arguments. Thus, similarities between each argument and terms naming concepts of the ontology are calculated using one of presented similarity measures. The argument is assigned to the concept maximizing this similarity, if the value of this similarity is greater than a pre-defined threshold. If similarity values are below the threshold, the term will be labelled as inconnu (unknown). This is a non-oriented strategy because all arguments are considered at the same level.
Further on, we present two strategies (the second and the third) which take into account the hierarchical structure of arguments. Therefore, each cluster of arguments is considered as a hierarchy having on its first level the centroid and on its second level terms that are specializations of centroid.
The second strategy we propose is a top-down strategy. In the first phase, it identifies concepts of ontology which label the centroid of the cluster. If the centroid of a cluster is labeled as unknown, the same label is assigned to each term of the cluster. If the centroid of a cluster is labeled by a concept c of ontology, labels for other terms of the cluster are searched only in the set of sub-concepts of c . In this way, the top-down labelling strategy reduces the search space.
A third strategy is based on a bottom-up approach. For each cluster, the similarities between its terms and the concepts of ontology are calculated by using one of presented coefficients. If values of similarities are higher than a threshold, the concept labels the term. If this is not the case, the term will be labeled as inconnu (unknown). Based on the assignments of each term of cluster to ontology concepts, the similarity between the centroid and a concept of ontology is given by: sim(Centroid , c) 1 k
sim(ti , c) , where ti is a term k i1 of the cluster, c is a concept of ontology, sim(ti , c) is the similarity between ti and c , and k is the number of terms labeled by c .
Those three labelling strategies are used in a first experimentation in the field of accidentology which is described in the next section.
Results of our approach can be affected by different parameters: the corpus we use, that is to say its size and its nature (a domain specific corpus or a general one) and the ontology. Our first experimentation was performed in accidentology and aimed to points out how different ontologies affect the outcome. For this experimentation, we used a corpus and we aimed to assign terms extracted from this corpus to concepts of two different ontologies. Here after we describe those resources. The corpus we used is composed of about 250 accident reports of accidents which occurred in and around Lille region (130 KO, 205 000 words). Accident reports are documents created by the police describing road accidents. They are written by policemen, according to declarations of people involved in accident and testimonies of witnesses.
A first case study was done by using an ontology created from
accident reports O , see [4]. The ontology was created with
1
Terminae, see [3], and it is expressed in OWL, see [
Our second case study was done by using an ontology O ,
2
created from accident scenarios, see [7]. Accident scenarios are
documents created by researchers in road safety which describe
prototypes of road accidents. The ontology was created with
Protégé see [
Each ontology models concepts (see figure 3) and roles. number of entities of O2 is less important, see table 3. ontology; Ttotal is the number of terms extracted from corpus (arguments of verb relations); Ca is the number of concepts to whom arguments of verb relation are assigned, and Ctotal is the , where Ta is the number number of concepts of the ontology.
This definition is based on relative measures which enables us to compare results obtained by using different corpus and ontologies.
Values of assignation degree rank from 0 (all terms extracted from corpus are labelled as unknown) to 1 (each term extracted is assigned to a concept and each concept of the ontology labels at least one term).
For each case study, terms are assigned to concepts by using a labelling strategy. Results obtained are presented below:
O1 Result analysis is two-fold: for each case study, we compare the results provided by each labelling strategy; for the same case study (which is to say the same ontology), we compare the results provided by each labelling strategy.
As tables above show, the assignation degree has more
important values if the corpus and the ontology share the same
community. This could be explained by the similarity between
the linguistic level of the ontology (terms designing its entities)
and the corpus. By using a corpus and an ontology belonging to
different communities, the assignation degree decrease
drastically. In order to overcome this problem, we can use
lexical resources such as WordNet, see [
Among the labelling strategies, the bottom-up one shows lows values of assignation degrees in both case studies. This is because the most clusters we have obtained have less than 10 words, and this strategy fails in case of small sized clusters. The non oriented strategy and the top-down strategy provide similar values of assignation degree. Nevertheless, the top-down strategy performs faster, as it reduces the search space.
Approaches proposed in different application fields, such as ontology learning or word-sense disambiguation are at the origin of this work.
Among them, [
We have presented a non supervised approach developed to automatically assign terms of a corpus to concepts of ontology. This approach is using jointly verb relations and a domain ontology. Results provided could be used to semantically annotate or index documents.
A first experimentation in the accidentology domain was done in order to point out how different ontologies affect the outcome. In order to evaluate the results of this evaluation we have defined a new measure, called assignation degree. This evaluation shows that the approach provide better results if the corpus and the ontology belong to the same community. If they belong to different communities, values of assignation degree decrease. This experimentation shows that our approach is sensitive to lexical level (changing the vocabulary, by passing from a community to another, affects values of assignation degree).
As a future work, new evaluation scenarios have to be proposed in order to study how other factors (namely the corpus: its size and its nature) affect the results.
Another perspective concerns the exploitation of lexical resources such as WordNet, in order to take into account the synonymy between terms. Namely, this will allow us to overcome the problem of lexical variation between different communities.
As a continuation of this work, a feedback could be added in order to enrich the domain ontology by integrating new concepts.
[1] Alfonseca, E., Manandhar, S.: Improving an ontology refinement method with hyponymy patterns. In proceedings of the Third International Conference on Language Resources and Evaluation, 2001. [2] Aussenac-Gilles, N., Seguela, P.: Les relations sémantiques: du linguistique au formel. Cahiers de grammaire 25 (175), 2000. [3] Biébow, B., Szulman, S.: A linguistic-based tool for the building of a domain ontology. In proceedings of the International Conference on Knowledge Engineering and Knowledge Management, 1999. [4] Ceausu, V., Desprè, S.: Towards a text mining driven approach for terminology construction. In proceedings of the 7th International conference on Terminology and Knowledge Engineering, TKE 2005, 2005. International Conference on Knowledge Discovery and Data Mining, 1996.