📄 Modeling Long-Range Context for Concurrent Dialogue Acts Recognition

Summary

The paper addresses Concurrent Dialogue Acts (CDA) recognition, a setting in which one forum post or dialogue turn can carry several dialogue acts at once, such as thanking someone and confirming that a fix worked. Earlier work fell short in one of two ways. Some models handled whole dialogue sequences but assumed a single act per utterance, typically through a CRF output layer. Others allowed multiple labels but considered only a small fixed window of neighbouring turns.

The authors adapt a Convolutional Recurrent Neural Network (CRNN) to cover both needs:

• a CNN encodes each utterance;
• a bidirectional LSTM/GRU runs across the whole dialogue to collect distant context;
• independent sigmoid outputs with binary cross-entropy produce the multi-label decisions.

Two refinements are added. A highway link carries the utterance's own CNN features straight to the output layer, and dynamic k-max pooling handles long posts.

On MSDialog-Intent, a tech-support forum corpus, the full model beats the previous best system (CNN-CR) by about 4.7 points of accuracy and 3.1 points of micro-F1. This shows that modelling long-range context matters when utterances are long and serve several functions.

Key points

• MSDialog-Intent has 10,020 utterances labelled with a subset of 12 dialogue acts (e.g., GG, PA, FD, OQ, PF, IR). Utterances average about 72 tokens and 1.83 acts each, and roughly 60% have more than one label.
• The authors kept every label combination. Qu et al. had dropped some labels and collapsed rare combinations into the 32 most frequent, so the authors re-ran CNN-CR on the unsimplified data for a fair comparison.
• The best model, CRNN (v3) with LSTM, reached 0.6822 accuracy and 0.7337 micro-F1, against 0.6354 and 0.7029 for CNN-CR. The GRU variant of v3 scored the highest precision (0.7358).
• Setup:
  • GloVe embeddings (Wikipedia + Gigaword 5);
  • 8:1:1 train/validation/test split;
  • 100 CNN filters, 2 recurrent layers of 900 units, k=2;
  • Adam with learning rate 0.001.
• Evaluation and ablation:
  • The main metrics are Hamming-style label accuracy and micro-F1.
  • Paired t-tests are run only on accuracy, because it is the only metric computed per utterance.
  • Each addition (highway connection, then dynamic k-max pooling) improved results step by step.

Interesting ideas

• I1. Treat each dialogue act as its own yes/no decision, using sigmoid outputs and binary cross-entropy instead of one softmax or CRF choice, so a single turn can be both a greeting and an answer.
• I2. A highway path sends an utterance's local features past the context layer, so surrounding turns add information without washing out what the utterance itself says.
• I3. Dynamic k-max pooling splits a long text into p segments and keeps the strongest signal from each, rather than a single global maximum, so a long post can express several functions in different places.
• I4. Context is not limited to a fixed window, because a reply in a thread may target any earlier turn, as when a potential answer addresses the very first question.
• I5. The evaluation refuses to simplify the label space and breaks accuracy down by how many gold labels an item has, which shows whether a model really handles the multi-label cases.

For Mova

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

• I1 (frames): the prag speech-act layer could assign a set of acts per utterance, each backed by its own explainable evidence, instead of forcing one winning act.
• I2 (frames): when discourse context refines a speech-act or implicature reading, keep the sentence-internal cues as a separate, inspectable input that the final decision can always fall back on.
• I3 (answer): when reading long fables or test passages, collect evidence segment by segment so each answer can point to where in the text its support came from.
• I4 (world): the world-of-text model should keep the whole dialogue or story history available for linking states and events, instead of looking back only a few sentences.
• I5 (learn): evaluation of acts or frames could be stratified by item complexity (one versus several gold labels), with baselines re-run on the same unsimplified data so gains are not artefacts of an easier test.