1.1 Introducing CLASPnet
CLASPnet (i.e. the CLAuse SPotting net) is the outcome of a project which was aimed at developing a computer simulation of how high-level linguistic properties of English clauses and sentences can be detected. For example, the sentence 'the cat, which was sleeping in the bag, can not be found any longer' is a statement (as opposed to a question or an order), and contains two separate clauses. The first one, 'the cat can not be found any longer', is the autonomous matrix clause, whereas the second one, 'which was sleeping in the bag' is a dependent relative clause. Two other differences are that the first one has a passive verb, and negative polarity through the presence of 'not' and 'any longer', while the second clause has an active verb, and positive polarity. Reliably detecting such properties of clauses is, of course, essential to understanding them correctly: for example, the difference between 'shake bottle before use' and 'never shake bottle before use' is just one word, but it is an important one when this bottle contains nitroglycerin. Similarly, 'Kim is eating' and 'Kim is being eaten' differ substantially in their meanings.
In real life, sentences and clauses are also presented to us as sequences of words, so when we read a text or hear someone speaking, we have to reconstruct the entire sentence and its clauses by joining all the right words together. People do not postpone trying to understand a sentence until they know all the words -- rather, as each word becomes available, it is processed immediately to update our hypothesis about what the writer or speaker is trying to tell us. Part of that processing deals with connecting individual words as we try to determine the separate clauses and their properties. It is only when the information carried by the separate words (e.g. 'cat', 'bottle', 'not') is combined with the information about the clauses (e.g. passive, negative) and the sentence (e.g. question) that 'Was the cat not put in the bottle?' can be comprehended.
CLASPnet is a computer model of how information about individual words can be used to construct hypotheses about the properties of the clause and sentence which the words belong to. CLASPnet is also a connectionist model (i.e. an artificial neural network): it consists of layers of units ( neurons) which are connected by modifiable links ( synapses). By changing the values of the links, the network can store information about how strongly certain units at one layer are associated with units at the next layer. (Note 1) In the case of CLASPnet, the network is 'shown' one word at a time on its input layer, and has to show on its output layer what it thinks the clausal and sentential properties for the word are. The words are actually presented to the network by means of an orthographic and semantic representation, and the clausal properties involved are the following:
Because CLASPnet shows its current hypotheses about these properties while it is processing the words of a sentence, its behavior can be compared to that of real people so as to check whether they parse sentences in a similar manner. Hence, its cognitive scientific character lies in the integration of elements from linguistics, computer science and psychology.
The remainder of Chapter 1 describes the motivations for CLASPnet, and a few methodological concerns. Chapter 2 takes a look at related work by summarizing the results of four other connectionist models -- it can probably be skipped by readers familiar with the field of connectionist language modeling. In Chapter 3, the precise implementation of CLASPnet is discussed, followed by the results of the model on the training and test corpora. The chapter ends with the presentation of a number of experiments carried out to throw more light on how CLASPnet works. Finally, a summary of all the results together with a general discussion can be found in Chapter 4.
Each link has a weight attached to it which can be modified to represent how strongly the unit at one side of the link is associated with the other unit: large positive weights are excitatory, while negative weights are inhibitory. Each unit, except for those at the input layer, has an activation value which is determined by the sum of the weights leading to it -- the units at the input layer have an activation value which is determined by the researcher.