Nobody talks like that: a stylometric exercise

Back in March 2022, I was asked if I’d like to be interviewed as part of a research project concerning editing in the 21st century . What the hell I said, I have close to no real experience of digital editing (unless you count my lovely digital edition of “Through Beatnik Eyeballs”), though I have made a reasonably satisfactory career out of telling other people how they should do it. One thing they really do teach well at Oxford is the ability to sound as if you know what you’re talking about… Anyway, I signed up, and after some vicissitudes was duly interviewed via Skype, sitting in my birdsong-filled garden, some time in June. Some considerable time later (at the end of December to be precise)  I was invited to revise the transcript someone had made of my interview, and did so. I removed some of the more egregious hesitations and a few garden path sentences, but left the bullshit intact, collection of that being after all the object of the exercise. And later still, I learned, my transcribed interview joined a group of fifty or so others on a website, in a proper digital archive no less. This was all very satisfactory, though I was a little disappointed to find that the edited transcriptions were being made available in PDF or in RTF only. Are those really now considered to be appropriate long term preservation formats? I suppose in a world where Boris Johnson can become prime minister, nothing should be surprising. And what happened to the audio? The design goals of this project were firmly based in a part of the forest of the Digital Humanities somewhat removed from linguistic analysis, discourse semantics, problems of speech transcription, or the textual analysis of academic talk. The project was to deliver something readable by human beings, like a book. Quite enough for one grant.

But hoorah for the open-minded spirit of open access, which makes it possible for me (and anyone else so inclined) to play with the resulting resources and do at least some things not originally envisaged in the project design! Entirely unsurprisingly, I spent a happy few days last week downloading the RTF files and converting them to TEI (the scripts and the results are now available in a github repo ). TEI because that’s what I do, but also because I wanted to be able to do textual analysis properly.

My resulting TEI corpus contains 46 small documents, one for each interview, each consisting of a sequence of TEI <sp> elements, with a who attribute to indicate a unique code for the speaker Each element contains one or more paragraphs of text, preceded by the speaker code as given in the transcription (there are a few differences). Like this:

<sp who="#LB">
<p>I would have liked to have been better paid.</p></sp>
<sp who="#MK"><speaker>MK</speaker><p>Sure.</p></sp>

This was prefixed by a paragraph of background information about the interviewee, which I banished to the <front> of the document, as a source of metadata. I also created a rudimentary TEI Header for each document.

Importing my 36 documents into TXM, I found the corpus had a total of 237,271 words. I made a partition on the basis of the @who attribute, so that all the words for each distinct speaker were grouped together. Here’s the bar graph from TXM showing how many words were associated with each of the 51 distinct speakers. It shows that one speaker (JOS) talks a lot more than anyone else: but this is unsurprising, since he is one of the two interviewers, and I have simply aggregated his side of each discussion irrespective of participant. I did the same for the other interviewer, MK, who has fewer interviews (14 as opposed to 32 for JOS); at 6000 words, he actually talks less than the three most garrulous interviewees (JC, AG, and RR), all of whom manage more than 6500 words. At the other end of the scale, there are five interviewees who hover around 2000 words apiece. The bulk of respondents fall comfortably between these extremes.

What can I do with this data? Well, treating it just as data, it might be interesting to see whether the frequency with which words, or lemmas, or POS codes appear in each speaker’s chunk is much the same, or whether some stylometrical statistic can be used to group like-speaking speakers together. Does everyone talk in more or less the same way or (ex hypothese) do professors and old lags like me talk differently from early career researchers? It’s not quite the typical stylometric use-case (which tries to establish probable authorship on the basis of similarities) but close. Fortunately for the mathematically challenged, there exists a fairly well established range of tools designed to explore such matters. Unfortunately for the mathemagically challenged (amongst whom I unreservedly place myself), you do need to know what you’re doing with these really quite sharp edged tools. So please forgive any idiocies in what follows…

I played with TXM and with Stylo, both of which claim to be usable by the non specialist, and both of which have (interestingly different, but that’s another story) user interfaces. TXM has the great advantage of accepting TEI XML as input and treating it sensibly. Stylo requires me to pre-process my XML text into dumbed down plain ascii using arbitrary naming conventions to provide metadata. Both produce fancy graphics.

Here, for example, is a dendogram from Stylo, showing how my 50 locuteurs cluster together, if we look just at the highest frequency lemmas. If I interpret this correctly, it shows the interviewers (“allJOS” and “allDK”) grouped together and distinct from all other respondents, which seems reassuring.

This is even more evident in the Principle Component Analysis also produced with Stylo, which shows JOS and MK as complete outliers.

And TXM provides further confirmation of this from a lexical perspective:

I think what this is telling me is not only that JOS and MK are outliers, miles from all the other documents shown here only as a red splodge in the blue cloud, but also that the words they favour are characteristically to do with their role as interviewers (What, future,  question, projects, maybe etc.)  Or so I believe. But clearly I am going to have to do a lot more background reading before I can say anything really interesting about this little dataset…

Rechercher dans OpenEdition Search

Vous allez être redirigé vers OpenEdition Search