CodeMixESC cross-lingual retriever
paraphrase-multilingual-mpnet-base-v2 fine-tuned with the Multiple Negatives Ranking (in-batch
contrastive) loss on 16,951 Hinglish-English sentence pairs (PHINC + Roman-script Hinglish rewrites
of ESConv training utterances), so that a Roman-script Hinglish query retrieves the same English
ESConv cases as its English version. Part of CodeMixESC
(course project, IIT Bhilai); used by its live demo.
| Dev retrieval (151 turns, k = 10) | Overlap@10 Light | Overlap@10 Heavy |
|---|---|---|
| all-roberta-large-v1 (MultiAgentESC) | 56.0 | 22.1 |
| multilingual mpnet (base) | 69.9 | 29.6 |
| this model | 72.9 | 54.0 |
Overlap@10 = share of the cases retrieved for a Hinglish post that are also retrieved for its English
original. Training: 1 epoch, batch 32, lr 2e-5, scale 20, word embeddings frozen; best checkpoint by
dev retrieval. esconv_bank_embeddings.npy holds this model's embeddings of the 13,484 ESConv
case-bank posts (no text). For research use.
from sentence_transformers import SentenceTransformer
m = SentenceTransformer("roshan9136/codemix-retriever")
m.encode(["yaar bahut tension ho rahi hai", "I am very stressed"])
- Downloads last month
- 14