Some Frightening But Exciting PD-0332991 Guidelines
g., Miller, 1956). Yet speakers seem to apply (a variation of) the aforementioned decision rule in contexts with an atypically colored target. Such contexts are not more complex or visually cluttered than the typical ones. So, the decision rule that we propose above would not be one that merely applies when the (limited) processing capacity of speakers is exceeded, but one that is universally available whenever the content of a referring expression is determined. Implications for (Computational) Models of Reference Production Being able to refer to objects in a human-like manner is an important goal for NLG models of reference production (REG algorithms), and for the field of NLG (a subfield of Artificial Intelligence) in general (Dale and Viethen, 2009; Frank and Goodman, 2012; Van Deemter et al., 2012b). Our findings pose a new challenge Megestrol Acetate for current REG algorithms. In the light of our findings, models can PD-332991 be enhanced by incorporating general object knowledge, because without access to such information they are unable to distinguish between typical and atypical objects when determining the content of a referring expression. Moreover, in our data, the decision to include color in a referring expression appears not to be taken independently of the target object��s type. For example, speakers decide to mention redness when they describe a lemon, but not when they describe a tomato. This is something that a model should be able to take into consideration. Popular NLG models predict color use irrespective of the typicality and diagnosticity of the target��s color. In the Incremental Algorithm (IA; Dale and Reiter, 1995), attributes like color, size, and orientation are included in a referring expression on the basis of how informative they are, and they are considered one by one (i.e., incrementally). More salient attributes, like color, are considered early, because they are highly ranked in a predefined preference order GSK2118436 chemical structure (which is typically determined on the basis of empirical data). Type is likely to be included anyway, because it is necessary to create a proper noun phrase, and this would yield fully distinguishing referring expressions in all conditions in our experiments. The IA would therefore generate no color adjectives. If the IA was to be able to make the decision to mention the color of a yellow tomato, for example, and not for a red tomato, it would need a ranking (preference order) of certain colors for tomatoes (e.g., red, green, orange, yellow, blue), instead of a mere ranking of certain attributes (e.g., color, size, orientation). The model of pragmatic reasoning by Frank and Goodman (2012) allows salience of objects to be modeled for each visual context individually (instead of in a predefined preference order).
