StatNLP Heidelberg

StatNLP group at @UniHeidelberg. Our research is on the intersection of #ML and #NLProc, with a special focus on interactive statistical learning techniques.

[Raphael Schumann @RaphiRaph_]

In the recent days I gave talks about "Grounding Natural Language Processing in Geodata" at @UCSC ERIC lab, @UscNlp and Lyft Munich. The slides are available here: https://t.co/NzcO4g4srx

@xwang_lk @_jessethomason_ @cmarschner

[Raphael Schumann @RaphiRaph_]

Making geodata more accessible for non-experts 🙋‍♂️ and LLM tools 🤖.

🌍Text-to-OverpassQL: A Natural Language Interface for Complex Geodata Querying of OpenStreetMap🌏

Preprint📜: https://t.co/lqYs23ASbN
Demo🖱️: https://t.co/wU8J3X4EVm
Repo💻: https://t.co/Nn3p3l2ET8

[Raphael Schumann @RaphiRaph_]

What happens if you embody🚶GPT in Street View and provide it with navigation instructions?

✨🤖✨
VELMA: Verbalization Embodiment of LLM Agents for Vision and Language Navigation in Street View
✨🤖✨

💻 https://t.co/ZpWSHkVGra
📜 https://t.co/uZLtU2LXcD

[StatNLP Heidelberg @StatNLP_HD]

Last week we had Robert Baldock from @Aleph__Alpha as an invited speaker. He talked about recent developments and challenges of doing industry research on LLMs.

[Michael Hagmann @MHagmann1]

Interested in boosting the analysis of your ML experiments? 🪄

Read how you can gain more exciting findings (nearly) for free in our recently accepted #ICLR2023 paper

✨Towards Inferential Reproducibility of Machine Learning Research✨

📜: https://t.co/kaxenYiiAz

[StatNLP Heidelberg @StatNLP_HD]

Thank you Khanh for the good thread about the history of RLHF and the contributions of our Alumni @KreutzerJulia and Artem Sokolov.

[Khanh Nguyen @khanhxuannguyen]

The RLHF page of HuggingFace (https://t.co/piAJlycrWX) misses many important citations. Here are some classical RLHF papers that you should cite and why.

[Raphael Schumann @RaphiRaph_]

Updating ML models can introduce unseen errors, such as a virtual assistant 🤖 suddenly not understanding your often used command. How to avoid this?

✨
Backward Compatibility During Data Updates by Weight Interpolation
✨

📜 https://t.co/MX3ug3I7ph
⌨️ https://t.co/5y973CupOM

[Raphael Schumann @RaphiRaph_]

New paper to appear at #ACL2022nlp. We analyze Vision and Language Navigation in unseen outdoor areas. Spoiler: Vision isn't that important. Paper also includes strong baseline that achieves SOTA on Touchdown and introduces additional data for outdoor VLN. https://t.co/HTBerKZgSg

[StatNLP Heidelberg @StatNLP_HD]

We are happy that our book on Empirical Methods is now available. Data and code to replicate experiments presented in the book are available at https://t.co/pmPSK6rA1Y

[Morgan & Claypool Publishers @MorganClaypool]

New! Validity, Reliability, and Significance is self-contained, with an appendix on the mathematical background on GAMs and LMEMs, and with an accompanying webpage including R code to replicate experiments presented in the book. https://t.co/IzKOmj9T7f #NLP #DataScience

[StatNLP Heidelberg @StatNLP_HD]

Our paper "Don't Search for a Search Method" https://t.co/4S7pIS36Vr will be presented as a poster during #EMNLP2021's virtual poster session. We find that large fractions of adversarial text attacks' success can be achieved by simple baselines based on nearest neighbors

[StatNLP Heidelberg @StatNLP_HD]

Using high entropy in the language model as a proxy for uncertainty, they're able to ask the user clarification questions to determine what's actually wanted

[StatNLP Heidelberg @StatNLP_HD]

In semantic parsing without unique correct answers, such as understanding a user's query on #openstreetmaps, Michi Staniek and Stefan Riezler find that bringing the user in the loop for interactive parsing improves results. See the preprint here https://t.co/PrqenpzW5B #NLProc

[StatNLP Heidelberg @StatNLP_HD]

Additionally, the authors find that randomly selecting replacement tokens for the augmentation *outperforms* using a language model in terms of word error rate. This method can be combined with other augmentation techniques like SpecAugment

[StatNLP Heidelberg @StatNLP_HD]

We'll have a paper by Tsz Kin Lam, Mayumi Ohta, Shigehiko Schamoni, and Stefan Riezler at #Interspeech2021 -- On-the-Fly Aligned Data Augmentation for Sequence-to-Sequence ASR https://t.co/D5WvRFTGcp #NLProc

[StatNLP Heidelberg @StatNLP_HD]

The main idea is that, with text to audio alignment information and an audio dictionary, you can swap out both source audio and target tokens to augment your speech data. Replacement tokens can be selected from a language model or randomly.

[StatNLP Heidelberg @StatNLP_HD]

Every green line is a route in the training set and every red line is a route in the test set, covering Manhattan

[StatNLP Heidelberg @StatNLP_HD]

PhD student @RaphiRaph_'s paper "Generating Landmark Navigation Instructions from Maps as a Graph-to-Text Problem" was recently accepted to ACL-IJCNLP 2021 #NLProc #NLP #ACL2021 https://t.co/foTNN0esXp

[StatNLP Heidelberg @StatNLP_HD]

Check out the dataset he made along with some demos of the interface and model here: https://t.co/xfHyjnolZW

[StatNLP Heidelberg @StatNLP_HD]

New research from the StatNLP group on Cascaded Models With Cyclic Feedback For Direct Speech Translation has been accepted at the IEEE ICASSP 2021. The paper is available on arxiv: https://t.co/YkjQjRj3NW