2010/10/12 Jorge Llambías
<jjllambias@gmail.com>
So the conversation generator gets a sequence of "text"s as input and
outputs one "conversation", right? All it does is surround each text
with fa'ai/fa'o (when they are not already present, I suppose) and
then outputs their concatenation. That seems simple enough.
That is one possible conversationer, yes.
The problem I see is that we must already have the string of texts to
feed the conversationer. You can't feed it the string of phonemes "do
klama ma lo zarci" and expect it to split it into two texts, you must
feed it first chunk "do klama ma", second chunk "lo zarci". In other
words, the hard work will be done somewhere else, not by the
conversationer.
Right. If *all* you have is a stream of phonemes without knowledge of who spoke which, then the situation is hopeless (unless everyone is very cooperative). Fortunately, situations where that actually happens are exceedingly rare. At least, I can't think of any. In speech, we recognize who is saying what as easily (and often with much more accuracy) than the individual phonemes. In IRC, every PRIVMSG comes in separately, with its own source. In an email conversation, we have lots of headers on each mail (so starting the actual message with {fa'ai} and ending with {fa'o} might be a good practice; an entire mbox file could be fed to a de-conversationer). The idea is not that the information can be recovered from nowhere, but that there is a format in which it *can* be preserved, if we need to.