M. Bisson. Learning in fol with a similarity measure. In Proceedings of the 10th American Association for Artificial Intelligence conference, San-Jose (CA US), pages 8287, 1992.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....how we can extend our algorithm to handle relational data natively. There are two research questions that arise. First, how does one de ne a distance metric to compare objects which may have a variable number of linked objects There has been some work on de ning metrics for relational data [6, 9, 15]. The central idea is to apply a recursive distance measure. That is, to compare two objects one starts by comparing their features directly, and then moves on to compare linked objects and so on. Second, how does one eciently retrieve an object and its related objects to compare them in the ....

G. Bisson. Learning in FOL with a similarity measure. In Proceedings of the Tenth National Conference on Arti cial Intelligence, pages 82-87, 1992.

....how we can extend our algorithm to handle relational data natively. There are two research questions that arise. First, how does one define a distance metric to compare objects which may have a variable number of linked objects There has been some work on defining metrics for relational data [6, 9, 15]. The central idea is to apply a recursive distance measure. That is, to compare two objects one starts by comparing their features directly, and then moves on to compare linked objects and so on. Second, how does one e#ciently retrieve an object and it s related objects to compare them in the ....

G. Bisson. Learning in FOL with a similarity measure. In Proceedings of the Tenth National Conference on Artificial Intelligence, pages 82--87, 1992.

....and AI (e.g. 8, 1, 4, 3] Though this research covers many wide areas and application possibilities, most of it has restricted its attention to the determination of similarity of lexicon, concepts, and relations within one ontology. The nearest to our comparison between two ontologies come [2, 3] and [13] 2] introduces several similarity measures in order to locate a new complex concept into an existing ontology by similarity rather than by logic subsumption. Bisson restricts the attention to the conceptual comparison level. In contrast to our work the new concept is described in terms ....

....[8, 1, 4, 3] Though this research covers many wide areas and application possibilities, most of it has restricted its attention to the determination of similarity of lexicon, concepts, and relations within one ontology. The nearest to our comparison between two ontologies come [2, 3] and [13] [2] introduces several similarity measures in order to locate a new complex concept into an existing ontology by similarity rather than by logic subsumption. Bisson restricts the attention to the conceptual comparison level. In contrast to our work the new concept is described in terms of the ....

G. Bisson. Learning in FOL with a similarity measure. In Proc. of AAAI-1992, pages 82--87, 1992.

....1996; Hovy, 1998] Though this research covers many wide areas and application possibilities, all of it has restricted its attention to the determination of similarity of lexical entries, concepts, and template slots within one ontology. The nearest to our comparison between two ontologies come [Bisson, 1992] and [Weinstein and Birmingham, 1999 ] Bisson, 1992] introduces several similarity measures in order to locate a new complex concept into an existing ontology by similarity rather than by logic subsumption. Bisson restricts the attention to the semantic comparison level. In contrast to our ....

....areas and application possibilities, all of it has restricted its attention to the determination of similarity of lexical entries, concepts, and template slots within one ontology. The nearest to our comparison between two ontologies come [Bisson, 1992] and [Weinstein and Birmingham, 1999 ] [Bisson, 1992] introduces several similarity measures in order to locate a new complex concept into an existing ontology by similarity rather than by logic subsumption. Bisson restricts the attention to the semantic comparison level. In contrast to our work the new concept is described in terms of the existing ....

G. Bisson. Learning in FOL with a similarity measure. In Proc. of AAAI-1992, pages 82--87, 1992.

....1996; Hovy, 1998] Though this research covers many wide areas and application possibilities, all of it has restricted its attention to the determination of similarity of lexical entries, concepts, and template slots within one ontology. The nearest to our comparison between two ontologies come [Bisson, 1992] and [Weinstein and Birmingham, 1999 ] Bisson, 1992] introduces several similarity measures in order to locate a new complex concept into an existing ontology by similarity rather than by logic subsumption. Bisson restricts the attention to the semantic comparison level. In contrast to our ....

....areas and application possibilities, all of it has restricted its attention to the determination of similarity of lexical entries, concepts, and template slots within one ontology. The nearest to our comparison between two ontologies come [Bisson, 1992] and [Weinstein and Birmingham, 1999 ] [Bisson, 1992] introduces several similarity measures in order to locate a new complex concept into an existing ontology by similarity rather than by logic subsumption. Bisson restricts the attention to the semantic comparison level. In contrast to our work the new concept is described in terms of the existing ....

G. Bisson. Learning in FOL with a similarity measure. In Proc. of AAAI-1992, pages 82--87, 1992.

....problem itself. This is clearly undesirable. Case assessment has therefore to reduce the matching complexity. For this means, one classic approach is predicate typing [14] Combinatorial optimisation problems have in general both, symbolic and numeric characteristics. Predicate typing in KBG [5] copes with such numeric features in first order inductive learning. In this representation, literals have the form (predicateName [attribute] starting with the predicate name and followed by a variable number of attributes. Attributes have to belong to one of the two classes: object or value, ....

Bisson G., Learning in FOL with a Similarity Measure, AAAI, 1992, p. 82-7

....e : V E are mappings which assign a label to every node or edge of the graph. During the last years, different measures of the distance or the similarity between graphs have been introduced. The majority of them are based on the computation of a best mapping between the graphs, see for instance [25, 8, 22, 9, 3, 5, 24]. Only a few of them show metric properties. We suggest the following family of distance measures for graphs with metric properties. Let G = NG ; VG ; l G ; e G ; LG ; EG ) and H = NH ; VH ; l H ; e H ; LH ; EH ) be two graphs and h a one to one mapping from NG [h] ae NG to NH [h] ae NH . Assume ....

G. Bisson. Learning in FOL with a similarity measure. In Proc. of the 10th AAAI-92, pages 82--87. AAAI Press/The MIT Press, Menlo Park Cambridge London, 1992.

....they introduce a similarity measure for relational descriptions [14] that is a metric in the special case of descriptions with the same number of variables. A similar distance is introduced by Eshera and Fu [15] Other distance measures for graphs and other relational descriptions are suggested in [16, 17, 18, 19, 20]. In [21] graph distances are investigated considering their metric properties. As a result, some families of graph metrics which can be easily expressed in terms of connectionist computation are introduced. In this paper, one of these distances between graphs and its computation by a ....

....[73] and developed, for instance, by Lebowitz [74] Fisher et al. 75] and others. Because it is impossible to compute all possible generalizations, many authors use a similarity measure for objects as a guideline for generalization, for example, Lebowitz for attribute value descriptions, Bisson [19, 76] for logical representations, and the distance guided generalization for graphs, using MatchBox s results, which is described below. Another problem is that often a class of objects cannot be described by a single prototype because the class consists of several subclasses. A prototype has to be ....

G. Bisson. Learning in FOL with a similarity measure. In Proc. of the Tenth National Conference on Artificial Intelligence AAAI-92, pages 82--87. AAAI Press/The MIT Press, Menlo Park Cambridge London, 1992.

....show that some first order is necessary in order to deal with the relations among features of the image. Generalization to first order is well known for being highly combinatory, and most of the efforts in ML are done in order to reduce the combinatorics, as explained for instance in [30] and in [7]. The same can be said from the vision algorithms, as pointed out by Grimson [20] In principle, first order knowledge is well suited to the recognition of relations among objects 2 . On the other hand, attribute value representations are very efficient to describe the objects in terms of their ....

Bisson G., "Learning in FOL with a similarity measure", Proceeding of AAAI92, San-Jos, to appear.

....of food, the fact that any corn is a seed. As is emphasized by several studies, accounting for such knowledge may improve the relevance of the nodes of the GS [26] 4] 11] In machine learning, accounting for domain knowledge is often performed by using a preliminary stage of saturation [15] [1]. The saturation consists in an exhaustive application of a set of rules onto the objects descriptions. These rules allow to improve objects descriptions by adding all the knowledge which may be deduced from domain knowledge and their descriptions. A rule which is frequently used is the rule ....

....Space construction. This method consists in anticipating the filtering step (3 Fig. 1) by generalizing the descriptions by layers . 3. 2 The generalization by layers The method which we propose for accounting for knowledge about the types in the generalization relies, as in CHARADE [9] or KBG [1], upon a saturation stage. Nevertheless, we propose a way to counter the problem of the size of the required memory space. Our method consists in anticipating the filtering step (3 Fig. 1) by eliminating as soon as possible the useless arcs added (i.e. those which are not maximally specific) This ....

G. Bisson. Learning in fol with a similarity measure. In Proc. of AAAI, pages 82--87, San Jose, CA, 1992.

....in order for knowledge to be conquered. Inductive logic programming (ILP) 15] can benefit from clustering, too: e.g. KBG uses a similarity function specifically designed for first order languages, and gradually constructs hypotheses by generalizing the most similar examples and or hypotheses [2]. ffl A distance allows the retrieval of the examples or hypotheses most similar to the instance at hand. In case based reasoning (CBR) the retrieval stage commands the success of the whole process; hence much attention has been paid in CBR to developing flexible distances or similarity ....

....criterion most often refers to a relational or structural distance [10, 3] In this paper, we first compare the respective advantages and weaknesses of rules and distances in regard to supervised learning. We then discuss previous work devoted to constructing distances on first order languages [2, 7]. Section 3 presents an alternative to distances based on syntax and weights, namely hypothesis driven distances (HDD) We show that a set of d hypotheses induces a mapping from the problem domain L h onto the space of vectors of integers IN d . A distance on L h then follows, by defining the ....

[Article contains additional citation context not shown here]

G. Bisson. Learning in FOL with a similarity measure. In Proceedings of 10 th AAAI, 1992.

....can be compared to that of RIBL [5] which is also based on nearest neighbors. The essential difference is the following: in RIBL, the similarity between E and a training example Ex only depends on E and Ex (this is true also for the even more sophisticated first order similarity used in KBG [2]) But here, the neighborhood of Ex (and the fact that E is neighbor of Ex or not) depends on E, Ex and the counter examples Ce 1 ; Cen to Ex: the underlying similarity is driven by discrimination. 2.7 Complexity Under the standard assumption that the domain of any variable is explored ....

....in the propositional case [32, 27] Unfortunately, building the set of such minimal substitutions turns out to be intractable too. Another possibility would be to consider the best substitution oe, defined as minimizing some distance to in the line of the structural similarity developed in [2]. For instance, the best substitution in Sigma Ex;Ce would minimize the sum of the distances between atom i in Ex and atom oe(i) in Ce. As noted in [33] the description of an atom can be handled as a single treestructured feature since the element of an atom commands its atom type (e.g. the ....

G. Bisson. Learning in FOL with a similarity measure. In Proceedings of 10 th AAAI, 1992.

....effective and or more efficient, e.g. enables a system to infer missing information for partially described objects. Some conceptual clustering programs (e.g. Cluster [Michalski Stepp 83] or Cobweb [Fisher 87] construct a hierarchy of classes, other programs (e.g. Unimem [Lebowitz 87] Kbg [Bisson 92b] construct a directed acyclic graph of classes, which means that the classes are not necessarily disjoint. Cola uses the conceptual clustering program Sprite to perform the conceptual clustering step. With respect to the topic of this paper Sprite can be regarded as a reimplementation of Kbg 2 ....

....construct a directed acyclic graph of classes, which means that the classes are not necessarily disjoint. Cola uses the conceptual clustering program Sprite to perform the conceptual clustering step. With respect to the topic of this paper Sprite can be regarded as a reimplementation of Kbg 2 [Bisson 92b] The advantage of the conceptual clustering approach developed with Kbg 2 compared to systems like Unimem or Cobweb lies in the fact that Kbg s knowledge representation language is based on first order logic (without negation and function symbols) with some extensions, e.g. for numerical ....

[Article contains additional citation context not shown here]

Gilles Bisson. Learning in FOL with a Similarity Measure. In AAAI92, pp. 82--87. AAAI Press, 1992.

....effective and or more efficient, e.g. enables a system to infer missing information for partially described objects. Some conceptual clustering programs (e.g. Cluster [Michalski Stepp 83] or Cobweb [Fisher 87] construct a hierarchy of classes, other programs (e.g. Unimem [Lebowitz 87] Kbg [Bisson 92b] construct a directed acyclic graph of classes, which means that the classes are not necessarily disjoint. An example of a directed acyclic graph of classes constructed by the conceptual clustering system Kbg is shown in figure 1 below. It is a sub graph of a graph constructed from descriptions ....

....one positive and one negative example were necessary to build the description. on table(X4) part of(X1,X2) ne(X4,X1) brick(X4) supports(X4,X1) supports(X3,X1) ne(X3,X4) not(touches(X3,X4) on table(X3) is an arch(X2) 4. 1 Kbg 2 Cola uses the conceptual clustering program Kbg 2 [Bisson 92b] to perform the conceptual clustering step. The advantage of Kbg 2 compared to systems like Unimem or Cobweb lies in the fact that Kbg s knowledge representation language is based on first order logic (without negation and function symbols) with some extensions, e.g. for numerical values. ....

[Article contains additional citation context not shown here]

Gilles Bisson. Learning in FOL with a Similarity Measure. In AAAI92, pp. 82--87. AAAI Press, 1992.

....representing a whole dictionary would be required (about 2000 words in our experimentation) and most of their values would be equal to zero. AUTOCLASS [Cheeseman et al..88] CLASSIT [Gennari et al..88] or ADECLU [Decaestecker91] have the same drawback. FOL based clustering method such as KBG [Bisson92] or RIBL [Emde et al..96] use a more powerful representation than we need but those methods have the following limitations for our approach. They learn strict hierarchies of concepts although to express different viewpoints in an ontology a semantic class may have more than one super class. ....

....= 2 4 . P FC1 = 4 12 = 16, and P FC2 = 8. thus, the distance is equal to: 1 Gamma [ 16 2 3 ) 8 2 4 ) 4 12 2) 5 3 6 2) 0:57 Dist is metric (the triangular inequality is still to be proved) It is inspired by the Hamming distance, and the O SIM distance of KBG [Bisson92] except that the frequency of words in contexts is taken into account. Dist has been adjusted with the recipe corpora on which it gives promising results (section 5) 3.4. Subcategorization frames learning In parallel with concept formation, subcategorization frames are generalized so that new ....

[Article contains additional citation context not shown here]

Bisson (G.). -- Learning in fol with a similarity measure. In : Tenth National Conference on Artificial Intelligence. -- San Jose, California, 1992.

....in all vectors, very large vectors representing a whole dictionary would be required (about 2000 words in our experimentation) and most of their values would be equal to zero. autoclass [Che88] classit [Gen88] or adeclu [Dec91] have the same drawback. FOL based clustering method such as kbg [Bis92] or RiBl [Emd96] use a more powerful representation than we need but those methods have the following limitations for our approach. They learn strict hierarchies of concepts although to express different viewpoints in an ontology a semantic class may have more than one super class. Moreover their ....

....) 2 4 . P FC 1 = 4 12 = 16, and P FC 2 = 8. thus, the distance is equal to: 1 Gamma [ 16 2 3 ) 8 2 4 ) 4 12 2) 5 3 6 2) 57 Dist is metric (the triangular inequality is still to be proved) It is inspired by the Hamming distance, and the O SIM distance of kbg [Bis92], except that the frequency of words in contexts is taken into account. Dist has been adjusted with the recipe corpus on which it gives promising results (section 5) 3.4 Subcategorization frames learning In parallel with concept formation, subcategorization frames are generalized so that new ....

[Article contains additional citation context not shown here]

Bisson G. Learning in fol with a similarity measure. In Tenth National Conference on Artificial Intelligence, San Jose, California, 1992.

No context found.

BISSON G. 1992b. Learning in FOL with a similarity measure. In Proceeding of 10th AAAI Conference. San-Jose.

No context found.

BISSON G. 1992a. Learning in FOL with a similarity measure. In Proceedings of 10th National Conference on Artificial Intelligence (AAAI) . San-Jose. 82-87.

No context found.

M. Bisson. Learning in fol with a similarity measure. In Proceedings of the 10th American Association for Artificial Intelligence conference, San-Jose (CA US), pages 8287, 1992.

No context found.

Bisson, G. 1992b. Learning in FOL with a similarity measure. In Proc. AAAI-92. The MIT Press. 82--87.

No context found.

Bisson, M.: Learning in FOL with a similarity measure. In: Proceedings of the Tenth National Conference on Artificial Intelligence. (1992) 8287

No context found.

Gilles Bisson. Learning in FOL with similarity measure. In Proc. 10th American Association for Artificial Intelligence conference, San-Jose (CA US), pages 82--87, 1992.

No context found.

Gilles Bisson. Learning in FOL with similarity measure. In Proc. 10th American Association for Artificial Intelligence conference, San-Jose (CA US), pages 82-- 87, 1992.

No context found.

Bisson G., "Learning in FOL with a Similarity Measure", in Proceedings of AAAI, pp. 82-87, San Jose, 1992.

No context found.

G. Bisson. Learning in fol with a similarity measure. In AAAI-92 Proc. Tenth Natl. ConferenceonArtif. Intelligence, 1992.

First 50 documents

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC