8.1.3 Homograph Disambiguation

HOMOGRAPHS

The goal of our NSW algorithms in the previous section was to determine which sequence of standard words to pronounce for each NSW. But sometimes determining how to pronounce even standard words is difficult. This is particularly true for homographs, which are words with the same spelling but different pronunciations. Here are some examples of the English homographs use, live, and bass:

(8.8) It's no use (/y uw s/) to ask to use (/y uw z/) the telephone.

(8.9) Do you live (/l ih v/) near a zoo with live (/l ay v/) animals?

(8.10) I prefer bass (/b æ s/) fishing to playing the bass (/b ey s/) guitar.

French homographs include fils (which has two pronunciations [fis] ‘son’ versus [fil] ‘thread]), or the multiple pronunciations for fier ('proud' or 'to trust'), and est ('is' or 'East') (Divay and Vitale, 1997).

Luckily for the task of homograph disambiguation, the two forms of homographs in English (as well as in similar languages like French and German) tend to have different parts of speech. For example, the two forms of use above are (respectively) a noun and a verb, while the two forms of live are (respectively) a verb and a noun. Fig. 8.5 shows some interesting systematic relations between the pronunciation of some noun-verb and adj-verb homographs.

Indeed, Liberman and Church (1992) showed that many of the most frequent homographs in 44 million words of AP newswire are disambiguatable just by using part-of-speech (the most frequent 15 homographs in order are: use, increase, close, record, house, contract, lead, live, lives, protest, survey, project, separate, present, read).

Thus because knowledge of part-of-speech is sufficient to disambiguate many homographs, in practice we perform homograph disambiguation by storing distinct pronunciations for these homographs labeled by part-of-speech, and then running a part-of-speech tagger to choose the pronunciation for a given homograph in context.

There are a number of homographs, however, where both pronunciations have the same part-of-speech. We saw two pronunciations for bass (fish versus instrument) above. Other examples of these include lead (because there are two noun pronunciations, /i/y/ (a leash or restraint) and /l/eh/ (a metal)). We can also think of the task of disambiguating certain abbreviations (mentioned early as NSW disambiguation) as homograph disambiguation. For example, Dr. is ambiguous between doctor and drive, and St. between Saint or street. Finally, there are some words that dif

fer in capitalizations like polish/Polish, which are homographs only in situations like sentence beginnings or all-capitalized text.

In practice, these latter classes of homographs that cannot be resolved using part-of-speech are often ignored in TTS systems. Alternatively, we can attempt to resolve them using the word sense disambiguation algorithms that we will introduce in Ch. 20, like the decision-list algorithm of Yarowsky (1997).

8.2 PHONETIC ANALYSIS

The next stage in synthesis is to take the normalized word strings from text analysis and produce a pronunciation for each word. The most important component here is a large pronunciation dictionary. Dictionaries alone turn out to be insufficient, because running text always contains words that don't appear in the dictionary. For example Black et al. (1998) used a British English dictionary, the OALD lexicon on the first section of the Penn Wall Street Journal Treebank. Of the 39923 words (tokens) in this section, 1775 word tokens (4.6%) were not in the dictionary, of which 943 are unique (i.e. 943 types). The distributions of these unseen word tokens was as follows:

Thus the two main areas where dictionaries need to be augmented is in dealing with names and with other unknown words. We'll discuss dictionaries in the next section, followed by names, and then turn to grapheme-to-phoneme rules for dealing with other unknown words.