📄 Aged to Perfection: Machine-Learning Maps of Age in Conversational English

Summary

This study asks how spoken British English varies with speaker age, and whether a speaker's age group can be predicted from what they say. It uses the spoken BNC2014, which contains about 1.25k informal conversations recorded in 2012–2014 with speakers aged 2 to 91. Speakers are sorted into five cohorts: Kids, Teens, Young Adults, Adults and Seniors.

The descriptive side combines ANOVA on utterance length and lexical diversity, a chi-squared search for words typical of each cohort, and BERT embeddings with UMAP and K-means.

The predictive side trains five scikit-learn classifiers on speaker-level text. SMOTE handles class imbalance, and four feature families are compared.

A Linear SVM on TF-IDF features reaches about 0.73 accuracy. Adding handcrafted lexical, syntactic and discourse features does not improve on that. Errors cluster between neighbouring cohorts, and t-SNE shows a smooth age gradient.

The takeaway is that age shapes conversation gradually rather than in discrete generational steps, and that simple text representations already carry most of the demographic signal.

Key points

• Data: BNC2014 spoken component with 1,251 dialogues, a mean of ≈941 tokens and ≈3 speakers per dialogue. Turns are short (mean 8.76 tokens, median 5). Higher social grades and graduates are over-represented, and the median speaker age is 34.
• Mean utterance length grows with age, from roughly 6 words in children to about 10 for speakers over 20 (F(4,518)=2.97, p=.019). Gender has no significant effect.
• Vocabulary overlap between interlocutors drops as their age gap widens (r = −0.219). The mean age of a dialogue's speakers shows no such link.
• Classification setup:
  • TF-IDF capped at 15,000 features, minimum document frequency of 3, stop words removed.
  • Seed fixed at 42.
• Classification results:
  • The best model is Linear SVM on text features, with accuracy 0.7264 and F1 0.708.
  • Combining all feature families gives 0.7075.
  • Lexical-only features score 0.30–0.40.
  • Discourse features reach up to 0.62 with the MLP.
• Embedding clusters (two clusters found via silhouette score) do not split by age and are dominated by markers like like, yeah, erm. Cosine similarity is nevertheless higher between adjacent cohorts than between distant ones.

Interesting ideas

• I1. Run ablations by feature family against a plain bag-of-terms baseline, so you can check whether engineered features add any signal the simple representation lacks. Here they added none.
• I2. Identify the characteristic words of each group in two independent ways (chi-squared tests and logistic-regression weights) and trust only what both methods agree on.
• I3. Treat age as an ordered, continuous variable rather than unordered bins, and read the confusion matrix for errors at neighbouring boundaries versus distant ones.
• I4. Measure lexical alignment between conversation partners and relate it to how socially distant they are, a cheap dialogue-level signal.
• I5. Spoken data is saturated with hesitations and discourse markers that blur semantic clustering.

For Mova

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

• I1 (learn): Any new hand-built feature for the perceptron parser or tagger could be required to beat a trivial baseline in a controlled ablation, which works as a negative control against feature bloat.
• I2 (learn): This mirrors witness agreement for silver data, and it could be used to build register or speaker-style lexicons that need consensus from separate statistical witnesses.
• I3 (learn): Evaluation of graded labels (formality, certainty, intensity) could score near-misses differently from gross errors, giving more honest error analysis.
• I4 (frames): Overlap of vocabulary between speakers could serve as an explainable feature when modelling dialogue participants, for example in tracking who is accommodating to whom.
• I5 (tokens): Tagging fillers and discourse markers explicitly at the token level would keep them from distorting the UD tree and downstream frames, while still preserving them as pragmatic evidence.