---
title: "Dialogue-Act Prediction of Future Responses Based on"
description: "Neural chatbots generate replies end-to-end, so it is hard to see why a particular reply… 69.7% accuracy. 9 references, cited by 6 papers here."
authors: ["Koji Tanaka", "Junya Takayama", "Yuki Arase"]
year: 2019
venue: "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop"
source: "https://aclanthology.org/P19-2027/"
doi: "10.18653/v1/p19-2027"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/tanaka-2019-dialogue-act-prediction-future-responses/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/tanaka-2019-dialogue-act-prediction-future-responses/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2019](https://md.research.elamur.ai/free/en/papers/2019/page/6/) / Dialogue-Act Prediction of Future Responses Based on Conversation History

# Dialogue-Act Prediction of Future Responses Based on Conversation History

**Authors:** Koji Tanaka, [Junya Takayama](https://md.research.elamur.ai/free/en/authors/junya-takayama/) and Yuki Arase · **Year:** 2019 · **Venue:** Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/P19-2027/) · [PDF](https://aclanthology.org/P19-2027.pdf) · [DOI 10.18653/v1/p19-2027](https://doi.org/10.18653/v1/p19-2027) · [OpenAlex](https://openalex.org/W2952100657)

**Citations:** 16 citations (OpenAlex, 2026-10-14) · [6](https://md.research.elamur.ai/free/en/papers/tanaka-2019-dialogue-act-prediction-future-responses/cited-by/) in our library, of the 141,840 papers whose bibliographies we read

## Summary
Neural chatbots generate replies end-to-end, so it is hard to see why a particular reply was chosen. As a step toward more transparent generation, the authors forecast the dialogue-act (DA) of the *next* response before it is produced, using only the preceding conversation. The response text itself is never seen. Their model has three GRU encoders:

- an Utterance Encoder for each sentence,
- a Context Encoder over the sequence of sentence vectors, with a speaker-change flag added,
- a DA Encoder over the one-hot DAs of earlier turns.

A softmax classifier receives the joined vectors produced by the Context Encoder and the DA Encoder. On the SwDA corpus, with tags collapsed to 9 classes and conversations cut into 5-utterance windows, the full model reaches 69.7% accuracy and 32.4% macro-F1. This is 10.8 F1 points and 3.0 accuracy points above an utterance-only baseline. The ablations point to a division of labour: the DA history mainly helps rare acts, while the text history mainly lifts overall accuracy. The predicted act could later serve as an interpretable control signal for a response generator.

## Key points
- **Data.** SwDA has 1,155 telephone conversations and 219,297 utterances. Its DAMSL tags were merged into 9 classes following easy damsl. A sliding window of 5 produced 212,367 sequences, split 80/10/10 for training, development and test.
- **Architecture.** All encoders are GRUs:
  - word embeddings have 300 dimensions;
  - the Utterance Encoder has 512 hidden units, the Context Encoder 513 (one extra for the speaker-change tag), and the DA Encoder 128;
  - the classifier has one fully-connected layer of 100 units;
  - training used Adam with a learning rate of 5e-5 for 30 epochs.
- **Main results.**
  - The proposed model scores 52.7 precision, 32.5 recall, 32.4 macro-F1 and 69.7 accuracy.
  - Utterance-only scores 21.6 F1 and 66.7 accuracy.
  - A non-neural Max-Probability transition baseline scores only 16.9 F1 and 54.8 accuracy.
  - The best existing system for labelling the *current* utterance (Khanpour et al., 80.1%) is treated as an upper bound; the forecast reaches about 87% of it.
- **Rare classes.** Against Utterance-seq, per-class F1 improves on every DA.
  - Greeting rises from 46.7 to 81.3.
  - Directive and Other go from 0.0 to 2.7 and 3.6.
  - Uninterpretable and Question remain below 10 F1.
- **Caveat.** Gold previous DAs are fed in during testing as well as training, so the reported scores assume an error-free act history.

## Interesting ideas
- Keeping the history of symbolic labels in a separate small-vocabulary channel next to the raw text gave robustness on rare classes, because the label channel suffers far less from data sparseness than word sequences do.
- A one-bit speaker-change flag lets the model handle a single speaker producing several sentences in a row, without forcing one sentence per turn.
- A trivial transition-table baseline (the most probable next label given the previous one) exposed how skewed the act transitions are, and showed that real context is needed.
- The authors report macro-F1 next to accuracy because frequent tags dominate. They also say openly that test inputs used gold history, which marks the scores as oracle-condition results.
- Choosing the act of a reply before writing its words turns the predicted act into an inspectable plan for generation.

## For Mova

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

- **I1** (frames): crate prag could keep the running sequence of recognised speech acts as its own explicit feature when classifying or anticipating the next act, rather than relying only on the current sentence's tree.
- **I2** (frames): marking turn boundaries in dialogue texts would help resolve "I/you" coreference and assign speech acts to the right speaker in tales with dialogue.
- **I3** (learn): such a majority-transition predictor makes a cheap negative control that any learned act or event-order model must clearly beat before its gains are trusted.
- **I4** (learn): evaluations should always report macro scores on skewed label sets and run both gold-history and predicted-history conditions, so that error propagation through the pipeline stays visible.
- **I5** (back): the generator could first commit to a speech act from the meaning graph (answer, agree, ask) and then realise text for it, making each output's intent explainable.

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

## Institutions

- [National Institute of Advanced Industrial Science and Technology](https://md.research.elamur.ai/free/en/labs/national-institute-of-advanced-industrial-science-and-technology/) — Research institute, Tsukuba, Japan · 13 works · 2,747 citations
- [Artificial Intelligence Research Center](https://md.research.elamur.ai/free/en/labs/artificial-intelligence-research-center/) — Research institute, Tokyo, Japan · 8 works · 1,556 citations
- [The University of Osaka](https://md.research.elamur.ai/free/en/universities/the-university-of-osaka/) — University, Osaka, Japan · 6 works · 495 citations

📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 9,086 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 8,889 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 9,296 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 9,206 works

🧭 Topics (2): [🗨️ Dialogue systems](https://md.research.elamur.ai/free/en/topics/dialogue-systems/) · 401 works · [💬 Speech act theory](https://md.research.elamur.ai/free/en/topics/speech-act-theory/) · 160 works

## Cite

Cite as: Koji Tanaka, Junya Takayama and Yuki Arase. 2019. Dialogue-Act Prediction of Future Responses Based on Conversation History. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop.

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@inproceedings{tanaka2019dialogue,
  title = {{Dialogue-Act Prediction of Future Responses Based on Conversation History}},
  author = {Koji Tanaka, Junya Takayama and Yuki Arase},
  booktitle = {Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop},
  year = {2019},
  doi = {10.18653/v1/p19-2027},
  url = {https://doi.org/10.18653/v1/p19-2027}
}
```

- [← previous Are we there yet? Encoder-decoder neural networks as cognitive models of English past tense inflection](https://md.research.elamur.ai/free/en/papers/corkery-2019-are-we-there-yet-encoder/)
- [next → Natural Language Inference with Monotonicity](https://md.research.elamur.ai/free/en/papers/hu-2019-natural-language-inference-monotonicity/)
- [↑ back to papers of 2019](https://md.research.elamur.ai/free/en/papers/2019/page/6/)

## References

9 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Alessandra Cervone, Evgeny Stepanov, and Giuseppe Riccardi. 2018. Coherence models for dialogue. In Proceedings of the Interspeech, pages 1011–1015. [find](https://scholar.google.com/scholar?q=Alessandra+Cervone%2C+Evgeny+Stepanov%2C+and+Giuseppe+Riccardi%2E+2018%2E+Coherence+models+for+dialogue%2E+In+Proceedings+of+the+Interspeech%2C+pages+1011%E2%80%931015%2E)
2. Kyunghyun Cho, Bart Van Merriënboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation. In Proceedings of the EMNLP, pages 1724–1734. [↗](https://doi.org/10.48550/arxiv.1406.1078)
3. Naoki Isomura, Fujio Toriumi, and Kenichiro Ishii. 2009. Evaluation method of non-task-oriented dialogue system by HMM. The IEICE transactions on information and systems (Japanese edition), 92(4):542–551. [find](https://scholar.google.com/scholar?q=Naoki+Isomura%2C+Fujio+Toriumi%2C+and+Kenichiro+Ishii%2E+2009%2E+Evaluation+method+of+non%2Dtask%2Doriented+dialogue+system+by+HMM%2E+The+IEICE+transactions+on+information+and+systems+%28Japanese+edition%29%2C+92%284%29%3A542%E2%80%93)
4. Nal Kalchbrenner and Phil Blunsom. 2013. [Recurrent convolutional neural networks for discourse compositionality](https://md.research.elamur.ai/free/en/papers/kalchbrenner-2013-recurrent-convolutional-neural-networks-discourse/). In Proceedings of the Workshop on CVSC, pages 119–126.
5. Hamed Khanpour, Nishitha Guntakandla, and Rodney Nielsen. 2016. [Dialogue act classification in domain-independent conversations using a deep recurrent neural network](https://md.research.elamur.ai/free/en/papers/khanpour-2016-dialogue-act-classification-domain-independent/). In Proceedings of the COL-ING, pages 2012–2021.
6. Diederik Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. In Proceedings of the ICLR. [↗](https://doi.org/10.48550/arxiv.1412.6980)
7. Norbert Reithinger, Ralf Engel, Michael Kipp, and Martin Klesen. 1996. Predicting dialogue acts for a speech-to-speech translation system. In Proceeding of the ICSLP, pages 654–657 vol.2. [↗](https://doi.org/10.22028/d291-25293)
8. Oriol Vinyals and Quoc V Le. 2015. [A Neural Conversational Model](https://md.research.elamur.ai/free/en/papers/vinyals-2015-neural-conversational-model/). In Proceedings of the ICML. [↗](https://doi.org/10.48550/arxiv.1506.05869)
9. Tiancheng Zhao, Kyusong Lee, and Maxine Eskenazi. 2018. Unsupervised discrete sentence representation learning for interpretable neural dialog generation. In Proceedings of the ACL, pages 1098–1107. [find](https://scholar.google.com/scholar?q=Tiancheng+Zhao%2C+Kyusong+Lee%2C+and+Maxine+Eskenazi%2E+2018%2E+Unsupervised+discrete+sentence+representation+learning+for+interpretable+neural+dialog+generation%2E+In+Proceedings+of+the+ACL%2C+pages+1098%E2%80%931107%2E)

*Updated 24 October 2026*