|Application domain:||Learning Transfer Rules|
|Source:||Stephen Pulman and David Milward, SRI, Cambridge|
|Dataset size:||1.3 MB|
|Data format:||Prolog facts|
The transfer rules are normally hand-crafted through inspection of a set of non-transferrable QLF pairs, which is a tedious and time-consuming task. The main problem adressed in here is how to use ILP techniques in order to automatically learn transfer rules from examples.
540 QLF pairs for English and French have been obtained from SRI, Cambridge.
These were formed by running the SLT system which has accuracy of over 95%
on the ATIS 2 corpus. The pairs are given on the form
translation(Id,Source,SourceQLF,TargetQLF,Target), where Id gives a record giving the identifier of the utterance in the corpus, Source gives a list of words (the source utterance), SourceQLF gives the QLF corresponding to the source, TargetQLF gives the QLF corresponding to the target and Target gives the translation or translation_failed (in addition to the 540 QLF-pairs corresponding to the non-failed translations, 340 facts were given corresponding to failed translations).
This kind of corpus hopefully should be fine as a first step - if we can't learn transfer rules from the more uniform examples created by automated translation, it is unlikely we can learn them from human translations. Of course, if learning was to be done in earnest for a new language pair or domain, you might expect to take human produced translation pairs (consisting of English/French sentences) and use the individual grammars for English and French to give you the best QLFs for each sentence.
Examples were all of full ATIS utterances. Some of these are not traditional sentences e.g. "flights to Boston please". The danger of using parts of utterances is that it might bias the results - the difficult cases are those where there is some context dependence.
Just one target QLF was generated for each source QLF. The system uses statistical methods to choose the best QLF which is both a good French sentence and a good translation of the original (according to weighted transfer rules).
Three main features of the transfer rule learning problem is that i) often more than one clause should be produced from each example, ii) only positive examples are provided, and iii) the produced hypothesis should be recursive. Most previous ILP systems produce at most one clause from each positive example, and this is a significant problem when learning transfer rules, since it is not practically feasible to provide examples of each clause to be induced, but rather a set of clauses should be produced from each example. Initial experiments with this dataset is currently undertaken using the system TRL (deliverable STO 1.a from the first project year).