ACL 2023 findings paper - LEDA: a Large-Organization Email-Based Decision-Dialogue-Act Analysis Dataset

by Mladen Karan • Friday 2 June 2023 • Permalink

We have published a new paper in the Findings of the ACL!


Collaboration increasingly happens online. This is especially true for large groups working on global tasks, with collaborators all around the world. The size and distributed nature of such groups make decision-making challenging. In our new paper we propose a set of dialog acts for the study of decision-making mechanisms in such groups, and provide a new annotated dataset based on real-world data from the Internet Engineering Task Force (IETF). We also provide some interesting results of an initial exploratory data analysis


After experimenting with many variants of the label taxonomy in collaboration with annotators who were trained linguists we converged to the set of dialogue-act (DA) labels given in the table below:
The table containing the labels.
For annotation, we’ve split emails into segments and had at least two annotators independently label each segment. For this purpose we used a custom tool developed specifically for this task. We allow the same instance to be labeled with an arbitrary number of labels at the same time, making the task multilabel.


We divided the lifecycle of the draft into five periods (T1 … T5). Below we give the distribution of each dialogue act across the periods.
Label frequencies in different periods of the draft lifecycle.
We found that Answer and Question are more common in the early phases, likely due to more new issues being raised and unresolved issues discussed. ContextSetting and Extension are very frequent, increasingly so towards the end phases; we conjecture this is because those phases cover more complex issues requiring more background description. The frequency of ProposeAction is stable throughout the cycle and noticeably higher than StateDecision. This may imply that participants prefer to discuss actionable options rather than explicitly deciding on a single one.

We then divided the participants in terms of their role within the organisation – draft authors, working group chairs, and influencers (participants with high centrality in the email communication network).
Label frequencies for different types of participants.
Authors vs. non-Authors Authors are more social, give more answers, and ask fewer questions. Also, they use fewer NeutralResponse, Extension, and ContextSetting, indicating shorter, more focused messages. These trends imply they take a more reactive role in the discussion. Finally, they make the most decisions in the discussion, as would be expected, since they are in charge of the writing process.
Influential vs. non-Influential Influential people use Answer, Agreement, and NeutralResponse more, making them generally more responsive. They use less Extension, ContextSetting and Thanking, implying a concise, focused communication style. As expected, they make more decisions and propose slightly more actions.
Chairs vs. non-Chairs Similar to influential participants, chairs use NeutralResponse more than non-Chairs. However, they use more ContextSetting and Extension, and do more Thanking. We find this is because chairs send a lot of emails initiating and managing discussions and review assignments. Such emails are often composed of many small segments and contain a lot of these labels.

Prediction model

Finally, we made a BERT based prediction model for these DAs, which can serve as a baseline for more advanced models in future work. Results are given in the table below.
Prediction model scores.