3.1.1.2 The Semantic Representation

As has been mentioned above, all the words in the training and testing corpora have also been given a semantic representation. As it has been my aim to follow cognitive linguistic tenets, these representations are of a somewhat unusual nature: rather than using 'standard' semantic features like [+HUMAN] or [+ABSTRACT], I have tried to limit myself to strictly sensory and perceptual features. A personal selection of 25 semantic features has been made. Each feature is represented by one input unit, and the unit can have continuous activation values between 0 and 1 (actually 0.1 intervals have been used).

There are 14 features which are directly based on sensory perceptions. They have a value of 0 if I consider the feature to be irrelevant for the semantic definition of a word, or a continuous value between 0.1 and 1 if I think it is relevant. As one might expect, there are more units for those senses which I expect to be of more relevance for most words. It is also worth mentioning that I have paid little attention to 'objective' comparisons, which I deem unlikely to be of much importance in most concepts: e.g. both 'barn' and 'Berlin' have received very high values for the unit representing visual size, because both are much larger than humans -- the latter is much larger than the former, but this knowledge is not readily available from visual perception alone, if at all. Similarly, both 'aircraft' and 'shout' are considered to be very loud from an auditory point of view, despite the fact that aircraft make more noise than shouting people when measured in decibels.

• Visual perception:
  • Primary color: 0.1 (black) to 1 (white) with dark hues (e.g. gray) closer to black and lighter hues (e.g. pink) closer to white;
  • Secondary color: idem;
  • Size: 0.1 (very small) through 0.5 (adult human) to 1 (very large);
  • Rate of movement: 0.1 (very slow) through .5 (walking pace) to 1 (very fast);
  • Fuzziness: 0.1 (unclear edges, e.g. fog) through 0.5 (e.g. the sun) to 1 (clear edges); large physical structures usually have a lower value on this unit than small ones, because the latter are difficult to see completely at a single glance;
  • Has_Face: 0.1 (no face at all) through 0.5 (some animals) to 1 (e.g. humans, puppets);

• Auditory perception:
  • Pitch: 0.1 (very low pitched sounds) through 0.5 (adult human voice talking) to 1 (very high pitch);
  • Loudness: 0.1 (very soft sounds) through 0.5 (normal conversation) to 1 (very loud);
  • Understandability: 0.1 (white noise) through 0.5 (e.g. sobbing, crying) to 1 (humans talking);

• Tactile perception:
  • Softness: 0.1 (very hard) through 0.5 (human skin) to 1 (very soft);
  • Smoothness: 0.1 (very rough) through 0.5 (human skin) to 1 (very smooth);
  • Graspability: 0.1 (e.g. water) through 0.5 (e.g. a car) to 1 (e.g. a spoon) -- another way of looking at this unit is that it codes how easily one can literally manipulate the object;

• Olfactory perception:
  • Strength of smell: 0.1 (very weak) through 0.5 to 1 (very strong);

• Gustatory perception:
  • Strength of taste: 0.1 (very weak) through 0.5 to 1 (very strong).

• Gestalt-like properties:
  • Spatio-temporal range: this unit expresses the extent to which it is possible to physically stay in or experience the concept, either because the concept is large (and allows movement inside) or because it lasts -- values range from 0.1 (e.g. a flash) through 0.5 (e.g. a human) to 1 (e.g. cities);
  • Proximity: this unit indicates whether the concept is typically physically close to the conceptualizer, or whether it is typically linked to another concept -- hence, 0.1 is appropriate for fairly autonomous concepts, while 1 suits many adjectives and prepositions. (Note 5);
  • Up-down axis: 0.1 for concepts which are much closer to the ground than 0.5 (adult humans), and up to 1 for concepts which are typically very high;
  • Front-back axis: 0.1 for concepts which happen behind one's back, through 0.5 (no special orientation) to 1 (typically in front of oneself);
  • Homogeneity: 0.1 for concepts which are internally not uniform at all, through 0.5 to 1 -- all plurals, for example, had lower values on this unit than their corresponding singulars;
  • Duration: this unit expresses whether the concept is subjectively experienced as typically lasting for a very short (0.1), normal (0.5) or very long (1) time -- hence, the objective difference between a century and a millennium is not relevant for this unit;
  • Repetition: 0.1 for concepts which never repeat, through 0.5 for those who do so irregularly, to 1 for concepts which always include repetition -- plurals, for example, had higher values for this unit than their corresponding singulars.

• Instinctive properties:
  • Danger: from 0.1 for concepts which are not associated with physical (or mental) harm, to 1 for concepts which are;
  • Pleasure: a similar range from 0.1 to 1 for physical (and mental) joy;
  • Excitation: from 0.1 for concepts which are dull, to 1 for those which cause a lot of adrenaline to be released -- obviously, there is some overlap with the previous two units;
  • Food: from 0.1 for concepts which are not related to the feelings of hunger and thirst to 1 for those which are strongly related;

Note 5
Because of the ambiguous nature of this unit, and because it carries obvious word class information, a network has also been trained in which this unit and the other Gestalt-like units had been removed. (See 3.4.4 for the results.)