Probing classifier performance comparison between early and late checkpoint across layers. While the early checkpoint shows uniformly high performance, the later checkpoint exhibits relatively high variance across layers.

How and when do multilingual LMs achieve cross-lingual generalization during pre-training? And why do later, supposedly more advanced checkpoints, lose some language identification abilities in the process? Our #ACL2025 paper investigates.

07.06.2025 10:11 — 👍 3    🔁 2    💬 1    📌 1

Many thanks to Philipp Heinisch, Janosch Gehring, Philipp Cimiano and @anettemfrank.bsky.social (@hd-nlp.bsky.social) for their great contributions.

Please reach out if you have any questions, and see you in Albuquerque at @naaclmeeting.bsky.social 2025 🥳

21.02.2025 16:08 — 👍 2    🔁 0    💬 0    📌 0

For example, we find signature perspectives that authors tend to agree, disagree, or be orthogonal on. Many more details, considering e.g., different stakeholder groups are in the paper.

21.02.2025 16:08 — 👍 0    🔁 0    💬 1    📌 0

We propose methods for all three subtasks and evaluate them individually using human annotations and automated metrics.
After validating our approach, we conduct a case study on real-world data, demonstrating how PSVs analyze debates with a focus on deliberative resolutions.

21.02.2025 16:08 — 👍 0    🔁 0    💬 1    📌 0

3️⃣ Finally, we aggregate PSVs to compute (dis)agreement scores, both for individual perspectives and overall argument alignment. E.g., the two example arguments agree on “trophy hunting” but disagree on “hunting for food”. For “eating meat” the arguments are orthogonal.

21.02.2025 16:08 — 👍 0    🔁 0    💬 1    📌 0

2️⃣ Each argument is then mapped to a Perspectivized Stance Vector (PSV), a structured representation that captures the stance (✅❔❌) toward each perspective.

21.02.2025 16:08 — 👍 0    🔁 0    💬 1    📌 0

Our approach consists of three steps:

1️⃣ For a given debate topic, we identify signature perspectives that represent key viewpoints.

Consider the image above. Hunting for food, trophy hunting, eating meat and sustainability are the signature perspectives for discussed topic.

21.02.2025 16:08 — 👍 0    🔁 0    💬 1    📌 0

Debates aren’t always black and white—opposing sides often share common ground. These partial agreements are key for meaningful compromises
Presenting “Perspectivized Stance Vectors” (PSVs) — an interpretable method to identify nuanced (dis)agreements

📜 arxiv.org/abs/2502.09644
🧵 More details below

21.02.2025 16:08 — 👍 4    🔁 3    💬 1    📌 0