📄 The pragmatic annotation of a corpus of academic lectures

Summary

Alsop and Nesi describe a scheme for marking pragmatic functions in spoken academic discourse. Such functions cannot be recovered from lexical items or structural mark-up alone. Their data is the Engineering Lecture Corpus (ELC), which holds 76 transcribed English-medium engineering lectures of roughly an hour each, recorded in the UK, Malaysia and New Zealand. The transcripts were first encoded as TEI-compliant XML with structural information such as speakers, turns and laughter. Annotators then watched the videos and added inline XML tags for five functional categories: explaining, housekeeping, humour, storytelling and summarising, each refined by subtypes. Comparing the three national subcorpora shows that the same discourse functions are realised differently depending on cultural setting. Examples include how much lecturers summarise, whose experience their stories draw on, and which kinds of humour they use. The authors connect these patterns to syllabus rigidity, lecturers' career paths and differing views of what a lecture is for. The work matters for EAP/ESP teaching and as a source of authentic annotated examples of pedagogically important spoken language.

Key points

• The ELC contains 76 lectures from Coventry University (UK), Universiti Teknologi Malaysia (MS) and Auckland University of Technology (NZ), treated as three subcorpora.
• The annotation has five top-level elements with subcategories (the paper reports 22):
  • explaining: equating, defining, translating
  • humour: nine types, e.g. bawdy, ironic/sarcastic, self-deprecating, wordplay
  • storytelling: anecdote, exemplum, narrative, recount
  • summarising: review or preview of past, current or future lecture content
• Humour tags and structural laughter tags coincide in only about a quarter of cases. Laughter is therefore an unreliable signal of humour, since it can also express anxiety or relief.
• Summarising takes up the most text in NZ lectures (8.8% per lecture on average) and the least in UK lectures (2.3%). In MS lectures it is spread evenly across subtypes at around 1% each.
• Storytelling varies by country:
  • UK narratives mostly draw on the lecturer's own experience, MS narratives on other people's, and NZ narratives on both equally.
  • Anecdotes are significantly more frequent in the UK.
  • Exempla are frequent in MS and rare in NZ.

Interesting ideas

• I1. The scheme keeps objectively observable mark-up (turns, pauses, laughter) separate from interpretive functional labels, then cross-checks the two layers; this showed that an obvious surface cue (laughter) is a weak proxy for the function it seems to signal (humour).
• I2. Inline tags are used only while the transcript is still being corrected, and each pragmatic category is later exported as its own stand-off XML layer over stable text.
• I3. Summarising is subdivided by direction (review vs preview) and by target (past, current or future lecture), which turns meta-discourse into a small temporal grid.
• I4. Martin's storytelling genres separate tales told to illustrate a point (exemplum) from tales told to entertain or provoke emotion (anecdote), as well as from recounts and full narratives.
• I5. Frequencies are normalised by tokens and expressed as the proportion of text spanned by a function, so corpora of different sizes can be compared on equal terms.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (frames): In crate prag, irony and humour detection should hold surface cues and the inferred speech act in separate layers and report how often they agree, rather than treating a cue as proof of the act.
• I2 (frames): Speech-act, implicature and coreference annotations could each live in an independent offset-based layer over a frozen token stream, so that fixing one layer never invalidates the others.
• I3 (world): Phrases like "as we saw" or "later we will" could be modelled as references to earlier or later segments of the text itself, giving the event-ordering component a notion of discourse time alongside story time.
• I4 (answer): Classifying a fable or tale segment by genre would tell the QA reader when a moral or generalisation is expected, and when the story is only a sequence of events.
• I5 (learn): Evaluation reports on pragmatic labels could give span coverage per thousand tokens per domain, making shifts between fables, school texts and other corpora visible instead of being hidden in raw counts.