Paper 1

Helbig, Hermann and Marion Schulz (1997). Knowledge representation with MESNET: A multilayered extended semantic network. In Proceedings of the AAAI Spring Symposium on Ontological Engineering, pp. 64-72. Stanford, California.

see also his research webpage.

What is MESNET?

A multilayered extended semantic network. In which nodes are concepts and edges are relations/functions among concepts.

Goal of Mesnet: KRL for NLU (uses the term interlingua like in machine translation)

Criteria to compare KRLs:

  • universality domain independent
  • homogeneity usable to describe the sense of words, sentences and dialogues.
  • cognitive adequacy "concept or object centered representation"
  • interoperability it can be used to represent info in all components of a NLU system (lexicon, inference, grammar, generation)
  • communicability easy to understand (without confuzion)
  • automatizability feasibility of automatic knwoledge acquisition techniques.

Structure of Mesnet: figure 1 from the paper.

  • sort/feature:
    • sort: hierarchy, must have category. not combinable (exclusive?).
    • feature: polyhierarchy, typical category. combinable.
  • multidimensionality of each node:
    • intensional/extensional: few attempts? (compared to what? modal logic seems to be based on that distinction, no?). Called preextensional because an extensional concept is normally not fully expressed but expressed only by few examples. Interesting part, talking about the example "nearly all girls in the class love a boy". Intensional layer: condensed info. Extensional layer: all info. Here intensional/extensional is not the same as intension/extension as used in logic.
    • determinate/indeterminate: attributes REF={DET, INDET} and VARI={VAR, CO}. For example, a determiner that has an indepedent reference (i.e. a constant, even Skolem) will have REF=DET and VARI=CO. Something depending on something else (i.e. a skolem function) will have REF=INDET and VARI=VAR. (examples of intermediate situations?)
    • generic/non-generic: attribute GENER={GE,SP}. properties of a generic (GENER=GE) are inherited by default (typical or non-categorical, non-monotonic) by individuals of that class.
    • virtual/real: attribute "FACT", real for concepts that exists, virtual for hipotesized concepts. See figure 6.
    • order: attribute ORDER, Classifies different types of collective nouns. s^(0)=single entities. s^(1)=sets of entities, s^(2)=sets of sets of entities (on the extensional layer).
  • edges and subnets (See figure 5 for an example using relations/functions and shells):
    • relations/functions: various types, see the paper. To each relation/function is associated a set of inference rules in FOL. About 110 total number of relations/functions. Every thing is a concept, no need for explicit reification: Ax Ay ((x CAUS y) -> (x ANTE y)). Actually in this x and y could be what in EL are called episodes, so there is no need for reification.
    • shells (subnets): See figure 7. interesting example of what knowledge to use for answering the questions: "what is an house?" and "what did Peter buy?".

Comparison

Claims of beeing better than DRT (discourse representation theory) and GQT (generalized quantifier theory) because:
  • homogeneity
  • cognitive adequacy

Also, better than DRT and FCT (file change theory) because of shells.

In the conclusions it says that it can express donkey sentences (what is "combined quantification"?)

The comparison doesn't seem completely fair because DRT, GQT and FCT weren't done for knowledge representation and inference.

Observations (starting points for discussion)

the initial list of properties of a knowledge representation language are good, but: what is the cognitive adequacy thing? It seems to be the property of being able to retrieve information based on the concepts involved in it (propagation in a graph where the nodes are concepts).

But how do you produce that representation from normal NL? Example: "the lion is dangerous", "the lion is a mammal". First generic, second no. This difference is based on the property of the predicates involved.

Example in figure 3. How do you express "John gave Mary several dollars", "It was more than John gave to another girl" and "Two of them were counterfeits". One way is to existentially quantify over the set of several dollars and that is the reference for 'it'.

Interesting classification of KRLs for NLU in the conclusion.

Further reading necessary to see what they are able to do now, there are implemented systems (apparently, still to check).

Further readings

Ronald J. Brachman. On the epistemological status of semantic networks. In N. V. Findler, editor, Associative Networks: Representation and Use of Knowledge by Computers, pages 3--50. Academic Press, New York, NY, 1979.

Something on DRT, from this paper: Kamp, Hans (1981). "A Theory of Truth and Semantic Representation" in Formal Methods in the Study of Language, J. Groenedijk, J. Janssen, M. Stokhof, eds. Amsterdam: Mathematical Center Tracts.

Carnap: meaning and necessity (ref 5 in this paper)

What is MULTINET?

Multinet is the evolution of Mesnet, it is not clear what exactly was added/removed to Mesnet. But now Multinet is a complete system:

Overview of InSicht

InSicht is the question answering system based on Multinet that partecipated to QA@CLEF. I report here the 6 steps that fomr it as described in: Hartrumpf, Sven (2005b). University of Hagen at QA@CLEF 2005: Extending knowledge and deepening linguistic processing for question answering. In Results of the CLEF 2005 Cross-Language System Evaluation Campaign, Working Notes for the CLEF 2005 Workshop (edited by Peters, Carol). Wien, Österreich: Centromedia.
  • document processing step: adding an XML word, sentence and paragraph separator to the input document. Then this preprocessed document is parsed using WOCADI and Multinet is generated.
  • query processing: the same step above is applied to the user question.
  • query expansion: a lexicon (HAGenLex) and GermaNet are used to produce equivalent queries (multinets equivalent to or implied by the multinet generated from the user query).
  • semantic network matching: retrieve matching multinets.
  • answer generation: the retrieved multinets are converted to NL, not all of them succeed. This steps has the secondary effect of filtering some of the retrived multinets.
  • answer selection: dditinal filtering based on a preference for frequent and complex answers is done.

It obtained a 51% of correct answers of 400 questions (in 2005).

-- FabrizioMorbini - 24 Oct 2006

Topic attachments
I Attachment History Action SizeSorted ascending Date Who Comment
JPEGjpeg fig6.jpeg r1 manage 53.5 K 2006-10-24 - 17:17 FabrizioMorbini  
JPEGjpeg fig5.jpeg r1 manage 75.0 K 2006-10-24 - 18:14 FabrizioMorbini  
JPEGjpeg fig3.jpeg r1 manage 78.3 K 2006-10-24 - 16:47 FabrizioMorbini  
JPEGjpeg fig1.jpeg r1 manage 95.0 K 2006-10-24 - 15:59 FabrizioMorbini  
JPEGjpeg fig7.jpeg r1 manage 105.5 K 2006-10-24 - 18:23 FabrizioMorbini  
Edit | Attach | Watch | Print version | History: r5 < r4 < r3 < r2 < r1 | Backlinks | Raw View | Raw edit | More topic actions
Topic revision: r5 - 2007-03-12 - FabrizioMorbini
 
This site is powered by the TWiki collaboration platform Powered by PerlCopyright © 2008-2017 by the contributing authors. All material on this collaboration platform is the property of the contributing authors.
Ideas, requests, problems regarding URCS? Send feedback