📄 Unsupervised Dialogue Act Induction using Gaussian Mixtures

Summary

The paper tackles dialogue act (DA) recognition without labelled training data. Annotated dialogue corpora are costly to build, especially across languages and channels such as chats, forums and social media. The authors treat each dialogue as a sequence of hidden states, one per DA, using a Hidden Markov model. Each utterance is observed as a dense real-valued vector, and that vector is modelled by a multivariate Gaussian specific to its state. Conjugate priors are used: Normal-inverse-Wishart for the Gaussians and a symmetric Dirichlet for the transitions. This makes collapsed Gibbs sampling straightforward. Utterance vectors are built by summing pre-trained GloVe word embeddings weighted by IDF.

On Switchboard-DAMSL, the context-aware model clearly outperforms earlier unsupervised approaches (BHMM, M4) as well as K-means. It still trails a supervised Maximum Entropy classifier. The design matters because the feature vector can be changed freely without altering the model. The results also show that modelling which act tends to follow which is consistently beneficial.

Key points

• Data: Switchboard-DAMSL with 42 DA labels and the Stolcke et al. split. Training uses 1115 dialogues (196,258 utterances) and testing uses 19 dialogues (4186 utterances).
• Setup: K = 42 clusters, 1000 Gibbs iterations and α = 50/K. GloVe vectors trained on 6B tokens (Wikipedia 2014 + Gigaword 5) were tested at M = 50–300.
• Best unsupervised result: ctx GMM GloVe (M = 300) reaches F1 65.7% and v-measure 41.2%. BHMM scores F1 41.1% / V1 34.7%, and M4 scores F1 45.1% / V1 18.0%.
• Supervised reference: ctx ME BoW + GloVe (M = 300) reaches accuracy 72.9%, F1 74.5% and V1 58.6%. This is about 1 point under the best published supervised system (73.9%, Kalchbrenner and Blunsom).
• Evaluation covers purity, collocation and their F1, plus homogeneity, completeness and v-measure. Degenerate baselines (random, majority, one-cluster-per-utterance) are reported alongside.

Interesting ideas

• I1. Utterances are emitted as continuous vectors from per-state Gaussians, so the feature representation can be swapped or extended without touching the inference code.
• I2. Adding transition structure between neighbouring acts turned a plain GMM into a much better clusterer, which shows that sequence context carries much of the signal.
• I3. Degenerate baselines are reported explicitly, and the paper notes that one-cluster-per-utterance gets a deceptively high v-measure (42.4%) that collapses as test data grows.
• I4. The supervised context model uses two rounds: it first classifies without context, then reclassifies using the first-round predictions of the neighbours as features.
• I5. The authors chose conjugate priors precisely so that sampling reduces to closed-form predictive distributions: smoothed counts for transitions and a Student-t for emissions.

For Mova

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

• I1 (frames): A speech-act inducer in prag could emit explainable feature vectors built from the UD tree and frames (mood, root verb class, polarity) instead of words, then cluster acts without labels.
• I2 (frames): Speech-act labelling could score each act together with the act before and after it, for example question → answer, instead of classifying sentences in isolation.
• I3 (learn): Evaluation with negative controls should always include such trivial clusterings, so that metric artefacts are not mistaken for real gains in induced categories.
• I4 (learn): A cheap two-pass scheme like this could let silver-data taggers or speech-act classifiers use neighbour predictions without full joint inference, keeping each pass inspectable.
• I5 (world): Unsupervised induction of latent categories, such as entity state types or verb-class refinements, could reuse this conjugate-prior Gibbs recipe to stay simple, fast and fully traceable.