Inductive Learning of Pronunciation Rules by hypothesis Testing and Correction

1. No conflict (type 4)

When a new rule enters the system it is compared with each rule in its letter list. This comparison starts at the end of the list (with the default rule) and finishes at the top. If no conflict condition occurs then the new rule is inserted at the top of the list.

2. Proper conflict (type 0)

If during the above process conflict is discovered then the count of the new rule is compared with that of the conflicting rule. The rule with the highest count is retained, and the other rule is broken into its constituents and input to the system without trying to combine them with other rules in the list.

3. Partial conflict (type 1 or 2)

If a type 1 conflict occurs as the comparison operation is taking place, then the incoming rule is inserted into the rule list immediately following the rule it conflicts with (but see Part IV). If however it is a type 2 conflict the comparison passes over (and therefore above) the conflicting rule and continues on until conflict occurs again, or the top of the list is found and the rule inserted there. This partial conflict handling therefore causes the rules to be ordered with the more general following the less general.

4. Ambiguity (type 3)

While this type is reported by the system, no special action is taken. No adverse effects seem to occur by ignoring it in this way, as stabilisation (see Part IV) corrects any resultant rule misplacement.

TV RULE STABILISATION

The system so far described reads in a set of words, creates basic rules to translate these words and attempts to form general rules from examination of these basic rules.

Occasionally, rules produced towards the end of reading a batch of words can negate the effect of rules produced earlier. To overcome such errors and check the rules more generally, the batch of words are again input to the system. All the words should be pronounced correctly but if this i s not the case then the rule set must be incorrect. The following example shows this.

Consider the two following ordered "S" rules (where the indented rules listed below them are the rules that were combined to form the general rule):-

When "WAS" is reinput, it is translated by rule rather than rule 2 (since 1 precedes 2 in the rule set), giving the wrong phoneme for the "S". This requires the following (specific) rule to be created:-

and this is checked against the current rule set in the usual way to look for possible generalisations. It of course finds rule 2, and attempts to combine the two.

At this point, the system checks to see if the rule being combined with rule 2 has already been combined with other rules to produce rule 2. Finding that it has (as 2a), rule 2 is broken into its (immediate) lower-level components, and these are re-entered individually into the system (as if they had just been created).

This in itself is insufficient, in that if no other action was taken these components would merely be recombined to give the original (over-general) rule. To stop this, during recombination the system will only combine those rules which have exactly the same pattern.

Using the above example, then, rule 2 is broken With t he to give the four separate rules 2a to 2d. exact matches now needed for generalisation, "WAS" and "HAS" rules can be combined to give:-

and then the "IS" and "AS" rules give:-

There are and the same occasions when this is insufficient rule is still formed. This situation is handled by breaking the one that interfered with the have been used rule 1 in the the other rule (ie rule that should example).

If this also recombined to form the same rule (even under this stricter control) then the least used rule is broken and its components re-entered into the rule set without allowing any combination.

This process of re-inputting repeated until all the words correctly and the system is stable. takes 2 or 3 iterations. The complete in that it will correctly the words given to it. the words is are translated This usually system is then translate all

V RESULTS

The system has three modes of operation. The first is the normal reading of words and creation of rules; the second is the stabilisation procedure; and the third is a rule evaluation mode. This third mode enables the system to read words, show their pronunciation using its rule set and compare this with the actual pronunciation to enable statistics to be produced. These show the number of words the system has pronounc~d correctly and the number of phonemes correctly produced. During this process the rule set is not changed in anyway.

If all the words that created a set of rules were run in this way, then the statistics would always show 100% correctness of pronunciation. To obtain more meaningful results all the basic rules are deleted before evaluation. The resultant statistics give an indication of how good the general rules produced are. It may also be the case that some of the basic rules may not be needed anyway as the general rules may have developed far enough to make them redundant.

The usage counts of all the rules are zeroised before an evaluation run and incremented during it, so the figures for usage are correct at the end of a run. Any unused rules (usage count of zero) can be deleted.

figures showing the and words correctly In addition, the together with an of the usefulness The result is a set of proportion of phonemes translated by the rule set. list of rules used is given indication (via the usage count) of individual rules.

Evaluatio~ runs have been carried out on the rules produced from reading the frequency based words from the Ladybird key word books (5), and the dictionary entries (4) for the first two letters of the alphabet (run separately). To act as a comparison, the same data have been run through a manually produced set of rules. This was obtained by Anglicising the rules produced by the American group (2). The results are shown in the following table:- The original American rule set was claimed to have a success rate of 90% words correct and 97% phonemes correct, but it should be stressed that these figures were derived from listening tests, not comparison with dictionary definitions. Their figures are also frequency weighted i.e. correct translation of a frequent word was more highly rated than an infrequent word.

This is shown in the table, where the best results were obtained from the Ladybird words (the most frequent 500 words of juvenile reading). Here the percentage correct produced by the manual system approaches the figures quoted. The deficiency can be accounted for by the different method of defining correctness and also because the Anglicising may have led to deterioration of the rules.

It can be seen that there is a large discrepancy between the results for the A's and B's for both the automatic and the manual systems. This i s caused by the difficulty of finding a rule to define the pronunciation of initial A's. The automatic system spent nearly 4o minutes trying to come up with a suitable rule to define which word should start with AE and which with AX. The problem is basically one of stress and ., it is interesting that no set of rules appears to exist to handle this situation.

The "B" run is a better test as it is not so affected by the stress problem. The automatic system took only 10 minutes to come up with a set of rules that perform much better than the "A" set.

In both the A's and the B's, the automatic and manual system are close in their results and only in the Ladybird set do they differ significantly. Frequent words often have unusual pronunciation (i.e. they do not follow set rules) so in most systems they are usually handled by a small exceptions dictionary. This is similar to the system's basic rules , so by deleting them the system's performance has been seriously affected for frequency based words input to it. This is also reflected in the rules used, where the manual system has used three times as many as the automatic system.

However, the system does successfully learn rules governing the pronunciation of words from examples. A number of rules produced are exactly the same as some in the manually produced set, while others show considerable similarity. And the rule set produced from a section of dictionary performs approximately as well as the manually produced set.

APPENDIX I: Constants used in the system

DELIM word delimiter (space, full stop, etc.)

VOWEL A, E, I, 0, U

CONS1 B,C,D,F,G,H,J,K,L,M,N,P,Q,R,S,T,V,W,X,Y,Z

INITIAL RULE SET: