Multidimensionality of concepts in multidisciplinary domains is a problem terminographers have to deal with. We apply Corpus Pattern Analysis (CPA; Pustejovsky, Hanks, & Rumshisky, 2004) to extract conceptual dimensions according to context. The dynamic nature of these concepts is exemplified with the case study of SAND. On the other hand, knowledge patterns (KPs) often convey different conceptual relations and are therefore polysemic structures. The development of pattern-based constraints can help to disambiguate them and at the same time avoid conceptual noise, which would be a first step towards the systematization of automatic knowledge extraction. Two KPs are analyzed in detail: rang* from, which conveys the conceptual relation is_a, and the polysemic KP formed by. Mots clés: Extraction de connaissances, Multidimensionalité, Corpus Pattern Analysis, Patron de connaissance.
Résumé: La multidimensionalité conceptuelle dans les domaines multidisciplinaires est un problème auquel les terminographes doivent faire face. On a appliqué le Corpus Pattern Analysis (CPA ; Pustejovsky et al., 2004) afin d’extraire les dimensions conceptuelles selon le contexte. La nature dynamique qui caractérise les concepts est illustrée avec l’exemple de SAND. Par ailleurs, les patrons de connaissance (KPs) expriment très souvent de différentes relations conceptuelles, étant ainsi des structures polysémiques. Dans ce sens, le développement de contraintes basées sur les KPs peut être une bonne approche pour contourner le problème de la polysémie et en même temps éviter le bruit conceptuel dans les concordances. Cela représenterait un premier pas vers la systématisation de l’extraction automatique de connaissances. Dans cet article, on présente l’approche suivi dans l’analyse de deux KPs: rang* from, exprimant la relation is_a, et formed by, un des KPs les plus polysémiques. 1
According to Kageura (1997: 120) the characteristics of a concept are frequently specified from different points of view or facets (function, material, shape...) and the set of characteristics that constitutes a concept is normally multidimensional. Moreover, when concepts are represented in different contexts (i.e. specialized subdomains) certain facets or dimensions become more or less salient. Corpus linguistics provides the clues to distinguish which dimensions are activated in each case and a sound methodology must be applied to study this contextual multidimensionality in a consistent way.
Corpus Pattern Analysis
We apply the CPA approach in a slightly different way. We analyse concordances
starting with a direct search of specialized terms. After that, they are classified
according to the dimensions they show, where different knowledge patterns (KPs;
On the other hand, a pattern-based search is conducted in order to find new relations among other concepts. The problem is that knowledge patterns are very often polysemic structures that convey different conceptual relations. As a result, the development of pattern-based constraints can help to disambiguate them and at the same time avoid conceptual noise. 2
Contextual multidimensionality is derived from the situated nature of concepts. It occurs especially in the case of concepts with a low degree of specificity. We call them versatile concepts because they are involved in a myriad of events and they are not always related to the same concepts or through the same relations, especially in interdisciplinary domains.
For example, the concept SAND is generally (or prototypically) defined as a kind of sediment located in the sea, rivers or soil layers. However, in real texts, SAND activates many other dimensions. In a more general domain, such as GEOLOGY, the concept is linked to others through: type, as a kind of SEDIMENT; attribute, related to grain size as a classification parameter; and material, linking the concept to the natural elements of which it is part (VALLEY, SOIL, AQUIFER, DESERT, etc):
However, in a COASTAL PROCESS domain (see Fig. 2), salient dimensions become: material, although values (natural elements) are restricted to coastal ones (SAND BARRIER, SAND BERM, SAND SPIT, BEACH, etc.); and patient, where the concept is involved in certain natural processes (WAVE ACTION, STORMS, LONGSHORE CURRENT, DEPOSITION). If context is again restricted to the COASTAL DEFENCE domain (see Fig. 3), dimensions are still the same, but values are focalized to artificial elements (FENCE, BERM, DIKE) or processes (TRAPPING, PUMPING, DUMPING). Furthermore, there is a new dimension, highlighting the functional nature of the concept in this specific context (SAND protects DUNE-BLUFFS, SAND BODY is used for BEACH NOURISHMENT, etc.). These three domains form a hierarchy (GEOLOGY COASTAL PROCESS COASTAL DEFENCE), but in a completely different domain, changes are more remarkable:
In the WATER TREATMENT domain, a new dimension is found, where SAND is linked to a particular instrument used in water treatment plants. The functional dimension now has a different value (FILTRATION) and patient and material are no longer representative conceptual dimensions.
Yeh and Barsalou (2006) claim that when situations are incorporated into a cognitive task, processing becomes more tractable than when situations are ignored, and the same can be applied to knowledge acquisition processes. As a result, a more believable representational system should account for re-conceptualization according to the situated nature of concepts.
We are developing context-based conceptual networks by dividing the interdisciplinary field of the environment2 in different contextual domains, but first we need to know which concepts are activated in each situation and how to extract this information. This is where CPA can help to accomplish our aim. In our approach, patterns expressing contextual information delimit conceptual dimensions, which at the same time delimit domain membership.
First of all, if a concept only activates a particular dimension in just one domain, the identification of any KP linking the concept to that dimension will be enough to associate the concept to a concrete domain. In the case of SAND, if a KP expresses the instrument dimension, the concept will be automatically assigned to the WATER TREATMENT domain.
Nevertheless, domain disambiguation is not that easy when different domains can activate the same dimensions and, as a result, KPs are usually the same. For example, if SAND is found next to KPs like consist of, comprising, formed from, containing, or composed of, the material dimension ascribes concepts to three possible domains: GEOLOGY, COASTAL PROCESSES and COASTAL DEFENCE. Sometimes, certain KPs express a particular dimension in just one domain, such as made of, where SAND is always related to COASTAL DEFENCE because the pattern needs the activation of an artificial concept. However, most of the time, this is not the case, and domain disambiguation requires a second step based on the kind of values associated to each dimension. At this stage, semantic annotation seems the only way to differentiate
domain membership. As mentioned above, in the GEOLOGY domain materials must be natural elements found in nature, in the COASTAL PROCESS domain, materials are restricted to those found in the coastal area, and in the COASTAL DEFENCE domain materials are no longer natural elements. Consequently, annotation should be conceptoriented, differentiating all concepts in the hierarchy and assigning each of them to particular contexts. 3
Many KPs can be found in the manual identification process. However, automatic extraction needs a certain level of reliability to be effective. In table 1 we show some of the most reliable patterns for seven conceptual relations in our specialized domain: Relation Is_a Part_of Made_of Located_at Result_of Has_function Effected_by
Most patterns are general language expressions and can be applied to many other domains. Domain-specific patterns are generally more reliable, such as built of/from and constructed of. However, even reliable ones show a certain degree of conceptual noise and polysemy. Thus it is necessary to discover certain rule constraints according to each KP’s specific needs. In the next sections we will deal with the KPs rang* from and formed by. 3.1
One of the most reliable KPs that activates the relation is_a in the environmental domain is rang* from. However, when rang* from is followed by a number (see Fig. 7), the relation that is expressed is always of magnitude. The same goes for the combination with a number written in full, or an adverb or an article and a number. For example, measurable concepts such as GROWTH TIMES and RECHARGE EVENTS are defined by certain time and amount combinations. Therefore, if our aim is to analyze the is_a relation, there should also be restrictions on words expressing duration such as minute, hour, month, day, week, etc. (see Fig. 5).
Furthermore, if the terms are too far away from the KP, it is very hard to extract useful information from the concordance (see Fig. 6). This noise could be partly solved with the implementation of a candidate term extractor. This way, terms which did not fall under the consideration of specialized terms would be excluded from the extraction process from the beginning. Anaphora would still be a problem, but the validation of conceptual propositions would be at least more efficient.
The knowledge pattern rang* from has proven to be a very reliable pattern if the above-mentioned restrictions are taken into account. It is not only very informative for the extraction of hyponymic relations, but also for relations of coordination, as the expression requires at least two coordinate concepts:
From figure 7, the following information can be extracted: FINE SAND and PEBBLE is_a BEACH SEDIMENTS; LONG BEACH and STRAIGHT BEACH is_a BEACH; SAND, MUD,
3.2
As in the case of rang* from, certain constraints must be applied in order to avoid noise. However, the main problem of the KP formed by is polysemy. Concordances in figure 8 show the way formed by works in the three different dimensions it can express, although the result dimension prevails over part and material (see Fig. 8).
The disambiguation of this polysemic KP requires different steps. If the KP is followed by a verb, it is definitely related to the result dimension. Instead, if the KP is followed by a noun, it can be linked to any of the three dimensions. Then the difference lies in two factors: if the noun is a process concept type, the concept still falls into the result dimension; if the noun is an object concept type, the dimension can be either part or material, but if the noun is uncountable, it will always refer to the material dimension, whereas countable nouns will always link wholes with parts.
In order to construct a knowledge resource based on the real behaviour of concepts in all the possible contexts of a domain, all the dimensions of multidimensional concepts must be taken into account. On the other hand, we have shown how constraints can be defined for KPs to facilitate knowledge extraction. Needless to say, there is still much work to be done before achieving an automatic knowledge extraction system. For example, even if all KPs were tightly constrained, we are still analysing terms, and synonymic values may seem different concepts. That could be solved if knowledge representation and extraction processes were complementary. In this way, an ontological system could inform the extraction system and semantic annotation could be based on an already defined hierarchy. In any case, manual validation would still be necessary. context. Proceedings of the 20th international conference on Computational Linguistics. RUMSHISKY, A., HANKS, P., HAVASI, C. & PUSTEJOVSKY, J. (2006). Constructing a Corpusbased Ontology using Model Bias. Proceedings of FLAIRS 2006, Melbourne Beach, Florida,