Linguistically Informed Question Answering

ion over Parts of Speech. Lexical units are grouped into frames irrespective of their parts of speech. This allows to easily map, e. g., two text fragments onto each other that carry essentially the same meaning, but where one is headed by a verb and the other by a noun, such as ‘A bought B’ vs. ‘(the) acquisition of B by A’. In GermaNet, this mapping requires additional knowledge in the form of derivation relations (see above). Semantic Role Labelling. By semantic role labelling (‘frame elements’), syntactic variations are abstracted over: As such syntactic variants are mapped onto the same FrameNet representations, no additional relabelling mechanism is required. Frame-to-Frame Relations. Frame-to-frame relations, as recorded in the FrameNet database, list correspondences between frame elements. ‘A sold B to C.’ can be directly mapped onto ‘C bought B from A.’: The frames COMMERCE_SELL and COMMERCE_BUY are properly related, as are the participating frame elements BUYER, SELLER and GOODS. For every subtree for which a FrameNet representation can be found (based on the lemma of the node and the argument realisation), the corresponding FrameNet labels will be added: The name of the frame is added as a supplementary label to the node corresponding to the frame evoking element; the edges are labelled with the corresponding frame elements. Thus, the subtree is annotated as representing an instance of the respective frame. An example is shown in fig. 5.11. Note that frame structures are not fully disambiguated. Syntactic differences are used for disambiguation. For example, the reflexive use of a verb may be associated with a different frame than the intransitive one. In these cases, disambiguation is done. In other cases no disambiguation is performed, for example, where the correct frame can only be identified through sortal preferences on arguments. 5.2. THE LINGUISTIC KNOWLEDGE-BASE 213 Frame information provides an additional level of normalisation: syntactically different realisations, as, e. g., occasioned by dative shift, will receive the same FrameNet representation. For ‘John gave the book to Mary.’ and ‘John gave Mary the book.’, a GIVING frame with the same frame elements is derived. In particular, ‘Mary’ is identified as the RECIPIENT in both cases. Frame Relations as Sources of Inferences. The FrameNet lexical database not only defines frames as abstract semantic predicates and frame elements as abstract semantics role labels, but also a hierarchy based upon different frameto-frame relations defined both between frames and frame elements. We translate frame-to-frame relations directly into relabelling relations with corresponding relevance values. We currently use all available FrameNet frame-to-frame relations, except for the SEE_ALSO relation, even though some of them are only rather vaguely defined (cf. 4.3). It is therefore not always possible to foresee whether or not using a frame-to-frame relation will or will not result in a valid inference relation. For example, when two words evoke the same frame, this does not mean that they stand in the classical synonymy relation: ‘Good’ and ‘bad’ both evoke the DESIRABILITY frame, even though they would be considered antonyms in terms of classical lexical relations. We have decided to exploit all frame-to-frame relations as sources of inferences. From the definition of the relations (cf. 4.3), we considered that this would in most cases produce interesting, if possibly sometimes unlikely inferences. We considered, however, that the additional step of answer checking should detect and properly mark those cases. In our experiments, we did not observe any serious problems with this approach (cf. 7.2.2.3). This may to a large extent be due to the limited overall current coverage of the frame lexicon that we use (cf. 4.3.3). We expect clearer definitions of the relations to emerge together with growing coverage; at some point, it may turn out to be advisable to remove all but some core relations from consideration. This is how the relations are currently utilised for inferences: ‘Same Frame’. This is not strictly a frame-to-frame relation: Two subtrees labelled with the same frame (and frame elements) match during direct answer matching, simply because the labels are identical. No additional inference rules are required. Inheritance. We treat inheritance like a classical hyponymy relation, that is, we use it to introduce inferences in both directions (cf. 3.5.2.3). 214 CHAPTER 5. MATCHING STRUCTURED REPRESENTATIONS