Stolcke
Stolcke (1990) used NN to map simple English language phrases to semantic representations. The input to the network was a series of words describing the spatial relationship between two geometrical shapes, presented one at a time, and a 22 bit semantic description of the sentence being presented. This semantic pattern was held active and constant for the duration of the entire sentence. The network was trained to repeat the semantic pattern on its output. After training, the words of a sentence were presented at the input, and the correct semantic pattern would appear at the output, little by little as the sentence progressed. For example, a sentence like the red circle below the triangle touches the blue square would be entered. The output nodes would activate to represent the nouns circle and square, and the relationship touches.
The average Hamming distance between the expected and the actual outputs was of .082 for the training set, 0 for simple sentences, 1.07 for sentences with one adjective, and .94 for sentences with two adjectives. Stolcke also found that he needed to use a different network topology to effectivelly process embedded sentences.
Figure 8: Sketch of the network used by Stolcke to process embedded sentences (Stolcke, 1990).
The results obtained by Stolcke are in accordance with that of linguists who seem to believe that semantical grounding provides useful constraints to the process of syntax acquisition (such as Langucker (1985) and Pinker (1984) ).
Back to the Table of Content

Back to the previous topic

To the next topic