Natural Language Generation homepages自然语言生成主页.abdn.ac.uk.pptVIP

Generation of Referring Expressions: the State of the Art SELLC Winter School, Guangzhou 2010 Kees van Deemter Computing Science University of Aberdeen Guangzhou, Dec 2010 Reference to sets: taking stock Guangzhou, Dec 2010 Reference to sets: taking stock Remarkably little computational or psycholinguistic work has been done on reference to sets. This stands in contrast to a significant amount of formal semantics work, mostly focussing on collectivity (e.g., Scha 1981, Link 1983) and intensionality (Landman 1989). Problems with computational complexity are even greater than with reference to single objects (but only if brevity is an issue). – Most people agree that getting “the right” descriptions is a more important problem. Guangzhou, Dec 2010 Reference to sets: taking stock 3. Gardent (2002, ACL) criticised Van Deemter (2002) for producing overly elaborate descriptions. This can be partly prevented by logical optimisation. [Project: try this, for example using Quine/McCluskey optimisation.] Gardent’s own proposal re-instates Full Brevity. Constraint satisfaction techniques are used to combine declarative problem statement with (relatively!) fast search. Horacek (2004) has proposed alternative techniques for weeding out infelicitous descriptions. Also Gatt 2007 (PhD thesis). Guangzhou, Dec 2010 Reference to sets: taking stock Evaluation of algorithms that refer to sets lags behind evaluation in GRE more generally. TUNA evaluation (Van der Sluis et al 2007, Gatt et al 2007) limited this to references to sets of 2. All algoritms performed badly. STEC invited people to submit algorithms that refer to sets, but none was received. Guangzhou, Dec 2010 Reference to sets: taking stock 5. Gatt’s work has put the spotlight on intra-NP coherence, e.g. “the Greek and the Italian” is better than “the Greek and the cook” Gatt’s research method is exemplary, and coherence might prove to be an important aspect of text quality more generally (also outside GRE). E.g., C