3.3.3.4 The Clausal Types

In the final group of output units each one stands for a different type of clause (see 3.1.3 for the details). Figure 3.16 shows how well the network manages to classify them -- as always the percentages are based on a tolerance value of 0.2. The performance of the network on the clausal types is remarkably uniform, ranging from 99% for the relative clauses in the training corpus to 93% for the declarative clauses in the test corpus. This is quite surprising if one takes into account that the number of occurrences of each clausal type in the training corpus varies wildly (see Table 3.5).

Clausal Type	Patterns
Declarative Order WH-Question YN-Question Relative Connector Complement	15,740 2,764 10,220 4,558 2,870 3,808 1,250

Table 3.5: Number of patterns in the training corpus for which the clausal type units had a desired activation value of 1.

There seems to be no linear correlation between the number of occurrences of a clausal type and the number of times the network has been trained on it -- what's more, the clausal type which occurs most often in the training corpus, the declarative clause, is apparently classified worst. Luckily, the percentages for the missed and spurious errors show a more complicated picture (see Figure 3.17):

The performance on declarative clauses actually is not that bad at all; only the WH-clauses show better results. This is more in line with the numbers in Table 3.5, as there are far more declarative clauses and WH-clauses in the corpus than any of the other types (Note 22). Another observation about Figure 3.17 is that, for the first time, there are significant differences in performance between the training corpus and the test corpus. As these differences are associated with the less frequent clausal types, it seems reasonable to hypothesize that there is a causal relation between the two. A factor which may also play a role in this regard is that the training corpus simply might not be large enough for all the different clause constructions to be represented sufficiently -- if at all. For example, if the training corpus contains no instances of the connector 'although' then an occurrence of this form in the test corpus should be expected to give the network some problems. (The experiment reported on in 3.4.2 is an attempt to check the validity of this second hypothesis: the network was trained on corpora of increasing size, but always tested on a fixed large corpus.)

The spurious errors in Figure 3.17 are almost exclusively the result of the Declarative and WH-clause units being highly active when they should not. Again, the most natural explanation is that the high frequency of these two clausal types has led the network to develop connection weights which easily excite their output units. Especially the Declarative unit seems to have become the default option for the network. As a consequence, many of the missed errors for the other clausal types end up being spurious errors for the declarative type. A good example of this type of error is the sentence initial 'the' in 'the more ..., the less ...' constructions: in these cases, the desired active unit for 'the' is the one for connector clauses. But because sentences like 'the man is too happy to be hungry.' are far more common, the network always activates the Declarative unit to a very high value (about 0.88) when it sees the first 'the', while the activation value of the Connector unit is only weakly active (about 0.14). As soon as the second word is processed, the network uses the presence or absence of 'more', 'less' or 'fewer' to make the final decision for that clause. (The network also 'remembers' its decision, because when the second clause-initial 'the' is processed, it is immediately classified as belonging to the connector type.)

There are also other cases in which two conflicting interpretations interfere:

• The Yes/No and the Order units both receive medium activation values of 0.4 to 0.5 for the sentence-initial 'do' in 'do the sharks chase the fishes?' and 'do not drink!' -- the initial hesitance seems perfectly justifiable, because more information is needed to make a correct decision;
• When a relative clause begins with 'that', the Declarative unit is sometimes also weakly activated -- perhaps because the determiner 'that' is mainly used in declarative clauses;
• Compressed relative clauses which are not introduced by a comma and relative pronoun are also problematic: for example, in 'the boys loved by me [...]' the compressed relative clause is classified throughout as if it were a declarative clause, but in 'the tigers liked by me [...]' the Relative unit is highly active, be it in addition to an also highly active Complement unit;
• The Complement and the Order units interfere in clauses like 'the man be told to hit a child.' Only the Order unit should be active, but because the complement of 'tell' is usually a complement clause (e.g. 'the tigers tell him that the cats be hungry.'), the Complement unit also has a high activation value. It is worth noting that there is no interference whatsoever with the connecter clause 'to' + infinitive forms in 'they be too interesting to eat.';
• Sentence-initial modals in Yes/No-Questions seem to have been learnt rather poorly, because the activation values of many different clausal type output units tend to reach levels of 0.15 when such a modal is encountered;
• Many of the missed errors in connector clauses (next to the ones in 'the more ..., the less ...' constructions referred to above) occur when the network apparently fails to recognize the crucial connector word itself: in clauses beginning with 'although' the weak activation of the Connector unit is soon swamped by the default Declarative unit, while connector clauses introduced by 'when' sometimes fall prey to the Relative unit.

There are also two types of errors which affect many clausal types: the coordinating conjunctions 'and' and 'or', which are both used to connect different clauses and different noun phrases, never receive a high activation value on any of the clausal output units. While this is the desired behavior when two different clauses are connected, it is considered an error when two noun phrases are linked. Naturally, noun phrases occur in all the different clausal constructions. The second error of this nature occurs with embedded relative clauses: in many cases, when the clause surrounding the relative clause continues, the information about the clausal type of the surrounding clause appears to have been lost completely: the activation values of a number of different output units reach 0.1, but there is no clear winner and the network has to await the beginning of the next sentence to get back on track (declarative clauses seem somewhat resistant to this effect, though).

In this section, I have spent quite some time on discussing which input patterns CLASPnet fails to analyze correctly. Before I move on to the next section, however, I would like to point out again that the great majority of input patterns is classified as desired. And this despite the fact that the network has to make use of strictly local input data in order to detect higher-level sentential and clausal properties.

Note 22
No attempt has been made to make the relative numbers for the different clause types correspond to their frequency in natural language corpora, because I have not found any frequency lists for a classification similar to the one used for CLASPnet. Manipulating the frequencies would be fairly simple: if rules in the grammar are duplicated then their chances of being picked during generation are also doubled. (From a developmental perspective, though, the abundance of declarative clauses and WH-Questions might be justifiable to some extent -- e.g. "Where is Daddy? [...] Daddy is in the kitchen.")