TGA14: comments from OUCS

1. The representativeness issue
   We are not convinced by the argument in 2.3 concerning the
statistical validity of the sampling procedure proposed. We see
no reason to believe that `demographic representativeness'
(whatever that is) is transferable from the recruits to the
participants: on the contrary, our intuition is that most people
habitually converse with people just like themselves, so that
counting all participants (as opposed to recruits) as part of the
sampling vector actually decreases the variety of speech sampled
and thus its `representativeness'. We share the Advisory
Council's concern that the sampling procedure used should not be
regarded as any more significant than it actually is -- a random
sampling with no claim to anything but maximal variety.

2. Regional issues 

We would like to have more information about the regional distribution
within both the `demographic' and the `context- governed' samples. Are
the same sampling areas used in each? How are they defined? Does 'Great
Britain' include 'Northern Ireland', to take the obvious example?  

We note that so far all transcription has been done on samples from one
region, and moreover a region presumably fairly easy for Longmans'
transcribers to understand. The throughput on samples from other regions
is likely to be much worse, unless local transcribers are found (in
which case there will be an additional cost in training and recruiting
them)  

We do not understand why trades union activities should not be sampled
regionally in the same way as other political speeches, business
meetings or other local organisations. 

3. Categories in the `context-governed' samples
	We are a bit puzzled by some of the categories listed on pp
9-11, some of which seem to overlap considerably (e.g.
parliamentary proceedings come under two distinct headings;
`talks to WI, clubs etc', are hard to distinguish from  'club
meetings'). We think there should be a clear distinction between
'setting' (i.e. where the language is produced) and 'type' (i.e. what
kind of language it is), and strongly favour basing the sampling
explicitly on the former.  We propose a more balanced consideration of
situational parameters: the current list seems to have been
determined entirely by what kinds of institutions are likely to
allow material to be gathered from them. A (very preliminary)
alternative approach is sketched out in section 9 below.

4. Definition of conversation
	We are concerned at the notion of a `conversation' as an
independent unit. It's not clear whether the sequence in which
`conversations' occur will be recoverable from the corpus. This
seems essential when, as implied in 3.4 stage 2, pieces of
language have been chopped up during post editing: when this
happens, we think some indication of the length of omitted chunks
of silence should be included.  Likewise, we would like to see
some indication of whether the recording is coterminous with the
`conversation' e.g. whether the recorder was switched on after
the conversation began, whether the tape ran out while it was in
full swing etc.

5. Participant information 
	We think it's important to provide more information about
the participants in a conversation. We appreciate the difficulty
of obtaining full details analogous to those provided by the
recruits about themselves, but it should at the very least be
possible to know that person labelled <2> in conversation 234 is
the same as the person labelled <3> in conversation  233. We also
think it should be possible to know when participants enter and
leave a conversation. 
	
6. Information provided by recruits
	We still think that some sort of free form description of
the conversation provided by the recruit would be very useful.
This could be elicited during the post recording interview
mentioned on p.18, using the log itself as a prompt. It seems to
us to be of at least as much use to know what the recruits think
is important to understand the conversations they have recorded
as to know what their reading habits or lifestyle variables might
be. We would suggest recording and transcribing this interview as
well. 
	We dont think recruits are likely to characterise dialects
very helpfully (especially if they are judged incapable of
deciding race). First language is enough. In general, the questions
seem of more use for market research purpose than for linguistic
research.

7. Telephone conversation

Is this excluded from the `demographic' corpus on purely pragmatic
grounds? It seems a serious omission if `representativeness' is a
serious concern. Even if only half of the conversation is transcribable,
that's still better than nothing. Why not issue recruits with a
telephone answering machine (of the kind that can record both sides of a
conversation) as well as a taperecorder? Telephone talk should certainly
be included in the `context- governed' corpus.

8. Transcription and encoding
	There are several points of considerable controversy which
need to be discussed at more length. These include
- notions of 'turn' and `sentence' need better definition
- use of punctuation to represent intonational units (the current
proposal falls straight between two stools)
- use of controlled vocabulary for nonstandard orthography (we think
this is a real minefield which will need a lot more discussion)
- treatment of truncation false starts and repetition
	(the hyphen on its own is not enough)
- treatment of back-channels
	(are they regarded as turns or not?)
- paralinguistic features:
    (need to distinguish events that happen during an utterance
from those that happen between utterances; need to show duration
for some of these -- a `laugh' is not the same as 'laughing';
need to distinguish events produced by participants within their
own utterance from events happening externally --e.g. <passing
truck>-- or produced by other participants: these may be
overlapped separately. )
- voice quality, performance etc
	(these need to be thought about more carefully and we should
at least provide a set of suggestions for specific things worth
noting e.g. pitch, volume, tempo, rather than relying on the sort
of 'stage direction' approach implied here)

	We think these and related points should be taken up in more
detail either in a meeting of TGC or less formally between OUCS and
Longmans as soon as possible. 

9. Situtional parameters for spoken texts

The following list of situational parameters is a subset of that
proposed at the TEI work group on corpora chaired by Douglas
Biber during its recent meeting in Stockholm. Clearly, balanced
sampling across all these parameters is neither desirable or
possible: however it should be possible to characterise each
conversation along most of these dimensions. We leave it to TGA
to determine which are the most important from the point of view
of sampling.

Mode: (spoken,  written-to-be-spoken and spoken-to-be-written)
Channel: (radio, telephone, face to face etc.)
Language: (languages, dialects, sublanguages etc used)
Participant: (This parameter is subdivided as follows:
	- role (addressor, addressee, both)
	- number (single, multiple, corporate)
	- demographic characteristics (name, age, sex, place of
     residence, education, occupation, socio-economic status,
     parent-tongue, ethnic group, dialect etc.)
	- awareness of recording
	- relation to other participants
Setting: (i.e. physical location)
Factuality: (imaginative, non-imaginative (that is, meant to be
taken as factual")
Preparedness: (apparent degree of revision or polishing. Possible
values: high, medium, low).
Domain of use: This is probably the single parameter of most help
in sampling. We think it is of great importance that language
produced in each of the following situations should be equally
well represented in the corpus:
	- religious
	- art/ entertainment, 
	- business / workplace,
	- education
	- domestic 
	- government
	- other public, 
	- mass media (i.e. tv, radio, cinema, press)
Note that, with the exception of the last, which could be
combined with any of the others, these are more or less
independent of each other. 
Purpose: Three axes seem possible, as follows:
	- persuade or sell
	- transfer information
	- entertain
Topic: (subject matter)
Perceived value: (a culture-dependent subjective assessment, with
values such as important, beautiful, unmarked,  high, medium,
low, artistic trash, popular, highbrow, ephemeral )
Originality: (Possible values: translation, adaptation, 
condensation, abridgment, transcription)