Sayan Ghosh

PhD student at @uncnlp. IIT KGP CSE 2018

[Kerem Zaman @KeremZaman3]

❓ Did you know you can fuse models to reduce unwanted knowledge like shortcuts and biases?

💡 Fused models keep shared knowledge and forget unshared.

🚨Fuse to Forget: Bias Reduction and Selective Memorization through Model Fusion

📄https://t.co/hTUv9QRdsr
@LChoshen @shsriva

[Rakesh R Menon @rrmenon10]

Can we learn to classify purely from language explanations?

Check out our #ACL2022 @aclmeeting paper: "CLUES: A Benchmark for Learning Classifiers from Language Explanations"

https://t.co/SwYxH1heNr

Joint work w/ @sayandgp and adv. @shsriva (1/6)

[Sayan Ghosh @sayandgp]

We present three tasks - proverb & alignment prediction, narrative generation, and identifying narratives with similar motifs. We benchmark pretrained lang. models on these tasks and show they struggle on abstract language understanding and finding analogies in narratives.
(3/3)

[Sayan Ghosh @sayandgp]

Check out our #ACL2022 @aclmeeting paper: "ePiC: Employing Proverbs in Context as a Benchmark for Abstract Language Understanding"

Paper: https://t.co/K1EF3WOlmX
Website: https://t.co/rYgW0j4ZqZ

Joint work with advisor @shsriva
@uncnlp
(1/3)

[Sayan Ghosh @sayandgp]

ePiC is a crowdsourced dataset of narratives paired with English proverbs. ePiC also provides fine-grained annotation of aligned spans between proverbs and narratives, and is designed to minimize lexical overlap between narratives and proverbs.
(2/3)

[UNC AI @unc_ai_group]

Here is the full program from @UNCNLP+collaborators at #EMNLP2021 this week ➡️ https://t.co/OcoKgHx0Q6

We're also 𝗵𝗶𝗿𝗶𝗻𝗴 𝗣𝗼𝘀𝘁𝗱𝗼𝗰𝘀+𝗣𝗵𝗗 𝘀𝘁𝘂𝗱𝗲𝗻𝘁𝘀! Plz spread the word & ping faculty @mohitban47 @snigdhac25 @colinraffel @shsriva +students for any questions😀

[UNC AI @unc_ai_group]

Congrats to @uncnlp+collaborators on these exciting #EMNLP2021 papers on explainability/fairness, composition/generalization, generation/summarization, story understanding+visualizn, groundedNLP. Join us for PhD/postdoc positions!

cc @mohitban47 @snigdhac25 @colinraffel @shsriva

[Swarnadeep Saha @swarnaNLP]

Our new #emnlp2020 paper on PRover: An interpretable transformer-based model that emulates formal reasoning by jointly answering questions over *explicit knowledge* as facts+rules (in natural language) & generating corresp proof graphs.

with @sayandgp, @shsriva & @mohitban47
1/4

[Lisa Lee @rl_agent]

I wrote a Colab tutorial on MaxEnt RL: https://t.co/ABiLuAL8Od It implements the graphical model from @svlevine's "RL as Inference" tutorial for a simple chain environment. Play around with the reward function to learn different policies using the forward-backward algorithm!