Inductive Learning for the Semantic Web: What does it buy?

Paper Title: 
Inductive Learning for the Semantic Web: What does it buy?
Authors: 
Claudia d'Amato, Nicola Fanizzi, Floriana Esposito
Abstract: 
Nowadays, building ontologies is a time consuming task since they are mainly manually built. This makes hard the full realization of the Semantic Web view. In order to overcome this issue, machine learning techniques, and specifically inductive learning methods, could be fruitfully exploited for learning models from existing Web data. In this paper we survey methods for (semi-)automatically building and enriching ontologies from existing source of information such as linked data, tagged data, social networks, ontologies. In this way, a large amount of ontologies could be quickly available and possibly only refined by the knowledge engineers. Furthermore, inductive incremental learning techniques could be adopted to perform reasoning at large scale, for which the deductive approach has showed its limitations. Indeed, incremental methods allow to learn models from samples of data and then to refine/enrich the model when new (samples of) data are available. If on one hand this means to abandon sound and complete reasoning procedures for the advantage of uncertain conclusions, on the other hand this could allow to reason on the entire Web. Besides, the adoption of inductive learning methods could make also possible to face with the intrinsic uncertainty characterizing the Web, that, for its nature, could have incomplete and/or contradictory information.
Full PDF Version: 
Submission type: 
Other
Responsible editor: 
Decision/Status: 
Accept
Reviews: 

Review 1 by Philipp Cimiano:

The paper proposes that inductive reasoning techniques should be a crucial part of the SW, in particular for the task of ontology mining.
The authors discuss how inductive learning can help to help to create ontologies by mining folksonomies and linked data, support ontology evolution or refinement as well as support the retrieval of instances.
I agree that inductive approaches are very important in the context of ontologies.

The article could be improved in several ways:

1) The authors mention that inductive reasoning could make the task of "reasoning on the entire Web" more feasible. It would be nice to elaborate on this, in particular making clear what the characteristics of Inductive Reasoning are that can make it scale to the Web. Scaling to the Web means having sub-linear algorithms to achieve a certain task. Are the techniques the authors proposed linear or sublinear (in the size of semantic data etc.) It would also be nice to have some indication of the complexity of the inductive reasoning approaches that the authors have in mind.

2) It would be nice to see some concrete indication of the methods that are suited to deal with the different problems that the authors mention. An example would help to understand what the input and output could be. The description as it stands is rather abstract. A reader not familiar with the authors' previous work might have difficulties getting a concrete idea of the methods at stake.

3) The application of inductive learning techniques could lead to a non-monotonic behaviour in the sense that predictions of a inductive learning approach might in fact have to be retracted later as new evidence comes in. It would be interesting to see some comment on this.

4) The authors mention on page 3, 2nd column that negative and unlabeled examples are introduced "to cope with the OWA". This aspect is not clear to me. I am not aware of any inductive method that can do OWA-style reasoning. I would welcome some explanations/clarifications here.

5) There are many errors in the paper that need to be corrected (see my detailed comments below)

Some minor comments:

Abstract, last sentence: to my knowlege face is transitive and can not be used together with "with", so I suggest to rephrase as: "also possible to deal with the intrinsic uncertainty characterizing the Web"

Introduction, 1st column:

"thus resulting a highly time consuming task that could sap" ???
What is "sap" here? Further, "result" can not be used as a transitive verb to my knowledge. It should be "result in" I think.

last sentence of first paragraph: "which results to be an even". Again, I think that result can either be used as in "result from" or "result in".

Introduction, 2nd column

"treat another problem that is emerged in the last few year" -> "deal with another problem that ** emerged in the last few year*s* ?

"some proposals" ... "have been done" sounds a bit odd.

"already existis" -> "exists"

last sentence of this column, "occur" sounds strange here, I would say "holds"

In the footnote: "intended" -> "considered/meant/taken into account here"

2nd page, section 3.1

"folksonoimies" -> "folksonomies"

"so popular to constitute a non trivial source of knowledge" -> non-trivial in which way ???

"Hierarchical methods returns" -> "Hierarchical methods return**"

Page 3, Section 3.3.

"Ontology refinement could result a very" -> "Ontology refinement could turn out to represent a very complex task" ???

Page 3, second column

The following is not a complete sentence: "A Terminological Decision Tree from which a new concept definition is derived".

Section 3.4.

The following sentence is odd (syntactically): "If the overall cluster evaluation is lower that the global decision than it is as existing concept otherwise...". As a result, the meaning is not clear.

Section 3.5

"Indeed, this methods" -> "Indeed, *these* methods"

Section 3.6

(for instance the classmemberships) -> "(for instance the probability of belonging to a certain class"

Conclusion:

"is considered one of the most challenging and interesting problem*s*."

References:

Reference [4] Should be "B. Ganter" and not "G. Bernhard"

Reference [16]: There are some strange characters in the page indication

Reference [41] International Symposium** (no "s" at the end)???

Review 2 by Bernardo Cuenca Grau:

Comments for improvement:

- It would be useful if you could explain in a couple of sentences in the introduction what inductive learning is.

- It doesn't come across clearly in the first two sections what is the difference between "ontology mining" and "ontology learning". A clarification would be in order.

- In the introduction, you suggest that inductive learning could be helpful for different aspects of learning ontologies. You seem to focus almost exclusively in the generation of uncertain knowledge and data. Is this the only application? I think it would be useful if you replaced the last paragraph of the introduction with an itemised list enumerating all the applications of inductive learning that you are envisioning.

- Section 2 is a bit too cluttered with technical jargon: SVN, FCA, kernel function, NN, inductive concept retrieval, inductive query answering, OWA, CWA, ALCQ, and so on. I would make this section a bit less technical and more focused on the main ideas and intuitions.

- In the beginning of Section 3, you suddenly start mentioning deductive reasoning
and its limitations. This comes at a surprise to the reader, and seems rather disconnected with the content of the previous sections. It seems that your goal is to learn ontological content. What does automatic deduction have to do with that? Please, explain.

- Again, Section 3.1 is too cluttered with technical jargon and it is sometimes not clear which are your main points.

- Also, in Section 3.1 you seem to survey existing inductive learning methods> Are these methods exclusive to folksonomies and linked data? Do inductive methods work in a fundamentally different way for other types of data? If so, why?

- In Section 3.1, what do you mean exactly by "linked data"? Is it RDF data?

- Why is the availability of negative examples problematic with the OWA?

- Section 3.2: What does concept retrieval have to do with enriching the assertional part of the ontology? At least at first sight, it seems that concept retrieval should be much more related to taxonomy building. Please, clarify what do you mean by concept retrieval in this context.

- Section 3.2 is again cluttered with technical jargon, which blurs the main ideas.

- Is imbalance learning a type of inductive learning? Please, clarify.

- What is a "balanced dataset"?

- What is a terminological decision tree?

- In the conclusions you mention for the first time ontology matching. This again comes at a surprise. I would either comment on this potentially interesting application in the main text or I wouldn't mention it anywhere.

Tags: 

Comments

It is very nice that the paper provides a list of many aspects where Semantic Web can benefit from machine learning. As a comment, I'd like to add a few thoughts on the relationship between inductive and deductive reasoning which the authors may want to contemplate in preparing the final version.

I entirely agree that it is a fascinating venture to explore machine learning as an alternative to deduction - e.g. by understanding deductive reasoning as a classification problem (although I personally wouldn't talk about inductive reasoning in this context; in my opinion it's conceptually not "inductive") - and this also relates your paper nicely with our vision paper (A Reasonable Semantic Web).

However, there is more to the deductive/learning interaction story, in particular since you also discuss ontology learning and ontology refinement (sections 3.1 and 3.3) in your paper. In this context, I see two major (and related) challenges, where it is currently not clear how to address them.

1) Learning complex axioms (beyond taxonomies). With this I mean that, in order to be true to the original semantic web vision, one would have to go beyond the learning of subclass relationships, and learn, e.g., OWL axioms involving complex classes. There is some work on this, e.g. [A,B] (and in fact also [24] cited in your paper, but this won't work from texts and other raw data), but strong results are still missing.

2) Learning for reasoning. With this I mean that, in the end, it would be necessary to close the cycle in the sense that ontologies acquired using machine learning techniques could be used in (deductive) reasoning systems. This is, in fact, a much more general problem than only for the Semantic Web, where it is also an (obvious?) bottleneck. There exist some attempts in this direction (although I'm not aware of any in the Semantic Web realm) but they are very limited, and don't really go beyond toy examples.

[A] Johanna Völker, Denny Vrandecic, York Sure, Andreas Hotho: Learning Disjointness. ESWC 2007: 175-189

[B] Johanna Völker, Pascal Hitzler, Philipp Cimiano: Acquisition of OWL DL Axioms from Lexical Resources. ESWC 2007: 670-685

Thanks Pascal for your very interesting comments.