he/they, scienting computational linguist, photographer, ex radiation worker
The Netherlands
Joined March 2013
- Tweets119
- Following593
- Followers93
- Likes1.1K
Excited to be presenting my paper, "Beyond Decodability: Reconstructing Language Model Representations with an Encoding Probe," at #EMNLP2026 this weekend!
arxiv.org/abs/2605.00607
In our two case studies, we show that features decoded from model representations don't necessarily contribute much to reconstructing them. We also find that syntactic features contribute information beyond what correlated lexical features explain.
Gaofei Shen retweeted
🎤 Just days before #Interspeech2025!
We are excited to announce Prof. 𝗞𝗮𝗿𝗲𝗻 𝗟𝗶𝘃𝗲𝘀𝗰𝘂 as our keynote speaker for the special session on Speech Interpretability!
📍 "What can interpretability do for us (and what can it not)?"
📋 More info lnkd.in/eUmR5xbZ
I am excited to announce that my paper "On the reliability of feature attribution methods for speech classification" has been accepted to #Interspeech2025!
Co-authors: @hmohebbi75 @AriannaBisazza @afraalishahi @gchrupala
Find the preprint here: arxiv.org/abs/2505.16406
We show that we need to be careful when applying existing feature attribution methods to modern speech classification models based on wav2vec 2.0, since the feature attribution results might not be as reliable as we think.
if only Dutch people bought better looking fireworks instead of whatever ugly ones they always have....
In fact the Netherlands has the worst protection for pedestrian out of all places I've lived for an extended period of time and I keep getting surprised by how they do things here
I find it incredibly weird that dutch drivers just do not yield to a pedestrian on the street (except a few days ago when I was carrying a long metal stick in my hand back from Gamma)
Visited the Museo Nacional de Antropología in CDMX thanks to #naacl2024 and found this piece of artifact. Is this kiki or bouba or something in between?
Gaofei Shen retweeted
If you're at #NAACL2024, check out @linguisticshen's talk tomorrow (June 19). The talk is about lexical tone in spoken language models aclanthology.org/2024.naacl-…. Session "Speech recognition, text-to-speech and SLU", Don Alberto 2, 11:00-12:30.
Sharing some great news today. My paper titled "Encoding of lexical tone in self-supervised models of spoken language" has been accepted to the main NAACL 2024 conference!
Many thanks to my co-authors @Phonologician @afraalishahi @AriannaBisazza @gchrupala
A brief 🧵below:
Additionally, we find SLMs don't follow the same developmental trajectory as human language acquisition in their pre-training. But the SLMs do show patterns similar to that of human listeners in tone and consonant perception experiments. [5/n]