💻 PhD Student at @dh-fbk.bsky.social @mobs-fbk.bsky.social @land-fbk.bsky.social
🇮🇹 FBK, University of Trento
🇪🇺 @ellis.eu
☕ NLP, CSS and coffee
https://nicolopenzo.github.io/
Nicolò Penzo
Loading...
Interested? Come by our Poster @ Poster Session (NLP) 1 on January 22nd, 12–2 PM, Hall 2!
See you in Singapore! 🇸🇬
🧵9/9
@dh-fbk.bsky.social @land-fbk.bsky.social @mobs-fbk.bsky.social
For current LLMs, multi-party conversations (MPCs) potentially represent a clear distributional shift, since they are mostly tested and optimized for two-party (human–assistant) interactions.
🧵2/9
By constraining LLMs, we can generate MPCs that: 1) exhibit diverse interaction structures and 2) capture rich speaker–addressee relationships, including multi-addressee turns
🧵4/9
One way to address the lack of structural diversity in social media–derived MPCs is to synthesize multi-party conversations.
🧵3/9
Interactional analysis and qualitative evaluation further demonstrate that LLMs are suitable for synthesizing large-scale, diverse Written MPC datasets, which can be potentially used to fine-tune smaller models for tasks like next speaker or addressee prediction.
🧵8/9
We also propose a novel evaluation framework combining quantitative and qualitative metrics.
Such metrics measure how well generated MPCs adhere to content and structural constraints and whether they capture the complexity of real-world multi-party interactions.
🧵6/9
In this paper, we propose synthetic MPC generation testing four different LLMs guided by explicit constraints.
We explore two strategies:
1 - One-Long (OL): generate the entire conversation in one step
2 - Turn-by-Turn (TT): generate sequentially, one turn at a time
🧵5/9
Social media platforms typically enforce a one-to-one reply structure, ignoring implicit addressees and oversimplifying interaction dynamics.
In natural conversations, turns are often directed to multiple participants, generating much richer structures.
🧵1/9
🎶One more “song”?
Come see our poster next week at #AAAI2026 on our paper:
📣 “Don’t Stop the Multi-Party! On Generating Synthetic Written Multi-Party Conversations with Constraints”
With @marcoguerini.bsky.social Bruno Lepri @gglavas.bsky.social @satonelli.bsky.social
🧵
arxiv.org/abs/2502.13592
We found that some LLMs (i.e., Llama3.1, Qwen2.5) can generate well-structured multi-party conversations that conform to constraints.
Comparing generation strategies, Turn-by-Turn adheres more closely to constraints and also produces greater linguistic variability.
🧵7/9
