NLIs to ontologies

In recent years, the popularity of NLIDBs and even open-domain Question-Answering systems is replaced to some extent by the new kind of NLIs – those which are finding answers in an ontology or a set of ontologies. As this is quite a young research topic which has been around for less than a decade, it can also be seen as a continuation of the work which has been researched for more than five decades now. There are many similarities between various NLI systems (which have been discussed here) mainly related to ways of solving the language complexity problem. The advantages brought by the NLIs to ontologies are related to the possibility to link the word meanings, inherit the relationships based on the existing structure and deal with ambiguities more effectively. Moreover, in comparison to NLIDBs, these new systems have been promoting the benefits of reasoning over structured data, and portability – extracting the lexicon from the ontology directly, without any need for customisation.

In other words, as noted by Grosz et al. [1987] who developed the TEAM system, one feature which is missing in NLIDBs is what makes NLIs to ontologies attractive:

[TEAM] shares such constraints of customized interfaces as being restricted to single queries and being able only to retrieve the facts from a database, not to reason about them. [Grosz et al., 1987][p. 237]

A good example to demonstrate the reasoning was given by Professor Daniel Weld from the University of Washington, during his invited talk at K-CAP 2009 [Weld, 2009]:

what vegetable prevents osteoporosis?

If we enter this query into Google, there will be no answer (or rather, no documents which contain the answer). The answer can be found in the documents available on the Web, however, Information Retrieval engines can not locate it as they do not implement reasoning. Namely, kale is a vegetable which prevents osteoporosis – but no such documents exist on the Web which mention this, however, there are documents which mention the following:

kale is a vegetable (1)

kale contains calcium (2)

calcuim prevents osteoporosis (3)

NLI systems which interface ontologies are built to answer these and similar kinds of questions.

Another important advantage of NLIs to ontologies is interoperability - the possibility to easily combine and merge resources from various locations on the Web. For the example above, statements 1, 2 and 3, could or could not be contained within one single resource on the Web. Therefore, the knowledge which has been collected for decades can now be merged in order to successfully accomplish what has been a great challenge for such a long time: answer the questions automatically using the distributed sources available on the Web. This has not been possible with databases as they are distributed over the Web and not interoperable, while Question-Answering systems have to process large amounts of unstructured text and use techniques such as Information Retrieval to locate the documents in which the answer may appear. This step can be misleading as Information Retrieval methods although scale well, do not often capture enough semantics — documents with the answer could be easily disregarded if the answer was hidden in a form which is not in-line with the patterns expected by the QA systems. Finally, NLIs to ontologies can also use the techniques applied in interactive NLI systems in order improve the user’s experience. Instead of a single question session, they can move towards conversational systems which can give answers simulating the human, but not being restricted to a topic or a domain.

References

  • Barbara J. Grosz, Douglas E. Appelt, Paul A. Martin, and Fernando C. N. Pereira. TEAM: An experiment in the design of transportable natural language interfaces. Artificial Intelligence, 32(2):173 – 243, 1987.
  • Daniel Weld. Invited Talk: “Machine Reading: from Wikipedia to the Web”, 2009. K-CAP 2009, Redondo Beach, California.