<?xml version="1.0" standalone="yes"?>
<Paper uid="W99-0906">
  <Title>A Computational Approach to Deciphering Unknown Scripts</Title>
  <Section position="8" start_page="42" end_page="42" type="concl">
    <SectionTitle>
7 Discussion
</SectionTitle>
    <Paragraph position="0"> We have presented and tested a computational approach to phonetically deciphering written scripts. We cast decipherment as a special kind of text-to-speech conversion in which we have no rules or data that directly connect speech sounds with written characters. We set up general finite-state transducers for turning sounds into writing, and use the EM algorithm to estimate their parameters. The whole process is driven by knowledge about the spoken language, which may include frequency information about sounds, sound sequences, words, grammar, meaning, etc. An interesting result is that decipherment is possible using limited knowledge of the spoken language, e.g., sound-triple frequencies. This is encouraging, because it may provide robustness against language evolution, a fixture of archaeological deciphering. However, our experiments have been targeted a bit narrowly. We were able to re-use the Spanish decoder on Chinese, but it could not work for Japanese kana. Even our Japanese decoder would fail on an alternative syllabic script for Japanese which employed a single symbol for the sound KAO, instead of separate kana symbols for KA and O. One ambitious line of research would be to examine writing systems in an effort to invent a single, generic &amp;quot;mother of all writing systems,&amp;quot; whose specializations include a large fraction of actual ones. To cover Spanish and Japanese, for example, we could set up a scheme in which each sound produces zero or more characters, where the sound is potentially influenced by the two sounds immediately preceding and following it. This gets tricky: the &amp;quot;mother Of all&amp;quot; has to be general, but it also has to be narrow enough to support decipherment through automatic training. (Sproat, forthcoming) suggests the class of finite-state transducers as one candidate. This narrows things down significantly from the class of Turing machines, but not far enough for the direct application of known training algorithms.</Paragraph>
    <Paragraph position="1"> In the future, we would like to attack an* cient scripts. We would start with scripts that have already been roughly deciphered by archaeologists. Computer decipherments could be checked by humans, and published human decipherments could be checked by computer. We would subsequently like to attack ancient scripts that yet to be deciphered. High-speed computers are not very intelligent, but they display a patience that exceeds even the most thorough human linguist.</Paragraph>
    <Paragraph position="2"> It will be important to consider text-layout questions when dealing with real scripts. For example, Mayan glyphs may run from top to bottom, right to left, or they may run differently.</Paragraph>
    <Paragraph position="3"> Furthermore, each glyph contain sub-parts representing up to ten sounds, and these may be organized in a spiral pattern.</Paragraph>
    <Paragraph position="4"> Another intriguing possibility is to do language identification at the same time as decipherment. Such identification would need to be driven by online sound sets and spoken corpora that span a very wide range of languages.</Paragraph>
    <Paragraph position="5"> Whether a document represents a given language could then be estimated quantitatively.</Paragraph>
    <Paragraph position="6"> In case language identification fails, we may be faced with a completely extinct language.</Paragraph>
    <Paragraph position="7"> Current computational techniques demonstrate that it is theoretically possible to figure out where nouns, verbs, and adjectives from raw text, but actual translation into English is another matter. Archaeologists have sometimes succeeded in such cases by leveraging bilingual documents and loan words from related languages. Only a truly optimistic cryptanalyst would believe that progress could be made even without these resources; but see (AI-Onaizan and Knight, 1999) for initial results on Arabic-English translation using only monolingual resources. null Finally, we note that the application of source-channel models to the text-to-speech problem is promising. This kind of statistical modeling is prevalent in speech recognition, but ours is one of the few applications in speech synthesis. It may be possible to use uncorrelated streams of speech and text data to learn mappings that go beyond character pronunciation, to pitch, duration, stress, and so on.</Paragraph>
  </Section>
class="xml-element"></Paper>