Bastian Bunzeck
@bbunzeck.bsky.social
Computational linguist trying to understand how humans and computers learn and use language 👶🧠🗣️🖥️💬
The work is mysterious and important. See https://bbunzeck.github.io
PhDing at @clausebielefeld.bsky.social
The work is mysterious and important. See https://bbunzeck.github.io
PhDing at @clausebielefeld.bsky.social
Also, the dialogue pairs taken from real data provide a much better reward signal than those synthetically generated. Here, real data beats synthetic data quite drastically!
October 28, 2025 at 12:56 PM
Also, the dialogue pairs taken from real data provide a much better reward signal than those synthetically generated. Here, real data beats synthetic data quite drastically!
While performance on most benchmarks also decreases, it actually increases on our own dialogue minimal pairs (real vs. randomly sampled adjacency pairs), from 64% for the pretrained model to 68% after reinforcement learning, even outperforming the BabyLM baseline by 10%.
October 28, 2025 at 12:55 PM
While performance on most benchmarks also decreases, it actually increases on our own dialogue minimal pairs (real vs. randomly sampled adjacency pairs), from 64% for the pretrained model to 68% after reinforcement learning, even outperforming the BabyLM baseline by 10%.
We replaced the standard BabyLM corpus with 10M tokens of dialogue triplets from CHILDES and trained an autoregressive model that we call llamalogue.
October 28, 2025 at 12:54 PM
We replaced the standard BabyLM corpus with 10M tokens of dialogue triplets from CHILDES and trained an autoregressive model that we call llamalogue.
As part of this year's BabyLM challenge, we (researchers from @gronlp.bsky.social and @clausebielefeld.bsky.social diverged from established pretraining paradigm by training only on dialogue data from CHILDES.
October 28, 2025 at 12:53 PM
As part of this year's BabyLM challenge, we (researchers from @gronlp.bsky.social and @clausebielefeld.bsky.social diverged from established pretraining paradigm by training only on dialogue data from CHILDES.
From conference to conference: September ends with a trip to #IWCS in beautiful Düsseldorf. Hyped for two days of semantics (and two more days of construction grammar and NLP). 🥳
September 22, 2025 at 7:51 AM
From conference to conference: September ends with a trip to #IWCS in beautiful Düsseldorf. Hyped for two days of semantics (and two more days of construction grammar and NLP). 🥳
From conference to conference — after last week’s #semdial I am at #konvens in Hildesheim this week. I will be presenting out German BabyLM Corpus (with @simphon.bsky.social) and our PI Sina Zarrieß will give a Keynote on BabyLMs tomorrow. 🥳
September 10, 2025 at 11:08 AM
From conference to conference — after last week’s #semdial I am at #konvens in Hildesheim this week. I will be presenting out German BabyLM Corpus (with @simphon.bsky.social) and our PI Sina Zarrieß will give a Keynote on BabyLMs tomorrow. 🥳
Now coming up: session 1 on naturalistic dialogue 👌
September 3, 2025 at 8:34 AM
Now coming up: session 1 on naturalistic dialogue 👌
There are similarities and fundamental differences. 🤷♂️
September 3, 2025 at 8:10 AM
There are similarities and fundamental differences. 🤷♂️
By testing LMs and human on the same stimuli…
September 3, 2025 at 8:03 AM
By testing LMs and human on the same stimuli…
The answer is: it depends, but generally yes. 😇
September 3, 2025 at 8:00 AM
The answer is: it depends, but generally yes. 😇
Finally, language models also display structural priming effects.
September 3, 2025 at 7:54 AM
Finally, language models also display structural priming effects.
What drives LLMs to generate repetitions? Nearby tokens in the context, and the previous repetition of constructions.
September 3, 2025 at 7:49 AM
What drives LLMs to generate repetitions? Nearby tokens in the context, and the previous repetition of constructions.
Their repetition patterns are quite similar to humans', but not consistently and as locally as humans' patterns.
September 3, 2025 at 7:46 AM
Their repetition patterns are quite similar to humans', but not consistently and as locally as humans' patterns.
For example in the map task:
September 3, 2025 at 7:39 AM
For example in the map task:
The frequency of different kinds of repetition changes across development, suggesting an implicit curriculum in human language acquisition.
September 3, 2025 at 7:34 AM
The frequency of different kinds of repetition changes across development, suggesting an implicit curriculum in human language acquisition.
Repetitions increase in complexity as language users progress, and they are also initiated more by learners.
September 3, 2025 at 7:30 AM
Repetitions increase in complexity as language users progress, and they are also initiated more by learners.
First keynote by Arabella Sinclair from the University of Aberdeen on “The many reasons for repetition in Dialogue”.
September 3, 2025 at 7:23 AM
First keynote by Arabella Sinclair from the University of Aberdeen on “The many reasons for repetition in Dialogue”.
We are very happy to welcome so many participants from so many different places 😇
September 3, 2025 at 7:19 AM
We are very happy to welcome so many participants from so many different places 😇
…which has continued to the present day! 😮💨
September 3, 2025 at 7:13 AM
…which has continued to the present day! 😮💨
Bielefeld University has a long, interdisciplinary history of research on dialogue and communication…
September 3, 2025 at 7:10 AM
Bielefeld University has a long, interdisciplinary history of research on dialogue and communication…
After 24 years, SemDial is back in the heart of Ostwestfalen 😎
September 3, 2025 at 7:06 AM
After 24 years, SemDial is back in the heart of Ostwestfalen 😎
Super cool talk by Charlotte Pouw on intonational phrasing, pauses and lengthening in TTS systems. They struggle especially with garden-path sentences, so apparently text-to-speech is not yet solved! @conll-conf.bsky.social #ACL2025NLP
July 31, 2025 at 9:19 AM
Super cool talk by Charlotte Pouw on intonational phrasing, pauses and lengthening in TTS systems. They struggle especially with garden-path sentences, so apparently text-to-speech is not yet solved! @conll-conf.bsky.social #ACL2025NLP