Senior Lecturer in NLP for AI, Institute for @PeopleCentedAI | University of Surrey | #nlproc
United Kingdom
Joined June 2008
- Tweets969
- Following1.5K
- Followers777
- Likes3.6K
Diptesh Kanojia retweeted
Last chance (3 days!) to participate in the refreshed WMT evaluation shared task. Pick any subtask and submit an automatic metric that's aligned with human annotations of translation quality across many languages.
www2.statmt.org/wmt26/mteval…
Diptesh Kanojia retweeted
| 27 October: World Day for Audiovisual Heritage 2025 🎬 |
Today we celebrate the sounds and images that tell humanity’s story and the professionals who ensure those stories transcend language and culture
#WorldDayForAudiovisualHeritage #AudiovisualTranslation #Accessibility
Diptesh Kanojia retweeted
💡 I'm happy to share our new article on how language & communication barriers impact mental healthcare for migrants across Europe — based on survey responses from over 600 health & social care professionals in 9 countries buff.ly/9nyU73r #1nt #health #mentalhealth
Diptesh Kanojia retweeted
Organizers are happy to help with any questions. 🙂
Website with all details and contacts: www2.statmt.org/wmt25/mteval…
Diptesh Kanojia retweeted
📐Task 3: Quality-informed segment-level error correction
Automatically post-edit machine-translated text using quality annotations to generate minimal and accurate corrections.
Description: www2.statmt.org/wmt25/mteval…
Submission platform: codabench.org/competitions/8…
Diptesh Kanojia retweeted
📐Task 2: Span-level error detection
Identify and locate translation errors within each segment (start/end indices) and classify their severity.
Description: www2.statmt.org/wmt25/mteval…
Submission platform: codabench.org/competitions/9…
Diptesh Kanojia retweeted
📐Task 1: Segment-level quality score prediction
Predict a quality score for each source–target segment pair, using document-level context and either ESA or MQM annotations.
Description: www2.statmt.org/wmt25/mteval…
Submission platform: codabench.org/competitions/9…
Diptesh Kanojia retweeted
The 2025 MT Evaluation shared task brings together the strengths of the previous Metrics and Quality Estimation tasks under a single, unified evaluation framework.
The following tasks are now open (deadline July 31st but participation has never been easier 🙂)
Diptesh Kanojia retweeted
Machine Translation is my first and final love. Every single work I do has some flavor of Machine Translation to it. Machine Translation is the best test bed for any sequence to sequence neural architecture. So it's best you read the book on NMT by the OG MT teacher Prof Philipp Koehn of @JHUCompSci. I can't recommend this enough.
arxiv.org/abs/1709.07809
Diptesh Kanojia retweeted
📢 Presenting IndicSeamless: A Speech Translation Model for Indian Languages 🎙️🌍
IndicSeamless is a speech translation model fine-tuned from SeamlessM4Tv2-large on 13 Indian languages. Trained on a curated subset of BhasaAnuvaad, the largest open-source Speech Translation dataset for Indian languages and English, this model enhances translation quality across diverse linguistic contexts.
🔹 Why IndicSeamless?
Indian languages pose unique challenges in speech translation due to their linguistic diversity, resource limitations, and speech variations. IndicSeamless addresses these challenges by improving accuracy, fluency, and contextual relevance across multiple languages.
🔹 Key Features:
✅ Fine-tuned on a diverse subset of BhasaAnuvaad 🗂️
✅ Handles spontaneous and read speech effectively 🗣️📖
✅ Robust to noise variations and diverse accents 🎧
✅ Optimized for real-world deployment in speech applications
🔹 Resources:
🔗 Demo: huggingface.co/spaces/ai4bha…
🔗 Model: huggingface.co/ai4bharat/ind…
🔗 Paper: arxiv.org/abs/2411.04699
This represents a significant step forward in multilingual speech translation for the Indian subcontinent.
Please try out the demo hosted on @huggingface who have kindly allotted us an A100 GPU! Thanks, @reach_vb for the support!
Diptesh Kanojia retweeted
Prof. Pushpak Bhattacharyya, in conversation with @EconomicTimes, advocates for trinity models—smaller, cost-effective AI models tailored to India’s diverse languages, domains, and tasks.
Link: economictimes.indiatimes.com…
#CFILT #NLP #AI #LLM
Diptesh Kanojia retweeted
📣 Exciting #Hiring opportunity! Lecturer in Natural Language Processing #NLP. Apply 👉 tinyurl.com/4u2b8xyc. Closes 12 Mar. [email protected].
Our School of Computer Science & Electronic Engineering seeks Lecturer in Natural Language Processing to grow AI research 💡
Diptesh Kanojia retweeted
Summary in case you missed any LLM research from the past month:
* RL on math datasets improves math ability v1
* RL on math datasets improves math ability v2
* RL on math datasets improves math ability v3
* RL on math datasets improves math ability v4
* RL on math datasets...
Diptesh Kanojia retweeted
Created a quick summary of DeepSeek-R1: anoopkunchukuttan.gitlab.io/…
Diptesh Kanojia retweeted
Nice paper on LLMs struggle to solve simple reasoning tasks. I love comment "What becomes quite clear through our study is the failure of current standardized benchmarks to reflect true model reasoning capabilities and to reveal their weaknesses."
arxiv.org/abs/2406.02061
Diptesh Kanojia retweeted
After COLING 2012, a top-tier NLP conference is back to IIT Bombay - my alma mater
🚨 IJCNLP-AACL 2025 will be held in @iitbombay, India!
📅 Date: 20th - 24th December 2025
🏢 Venue: Victor Menezes Convention Centre (VMCC), IIT Bombay, Mumbai, India
🗺️ Map Location: maps.app.goo.gl/Cd9JPMaiWr7T…
In the following days, more information will be made available!
Diptesh Kanojia retweeted
HALOGEN is a comprehensive benchmark with automated verifiers that decomposes and analyzes LLM outputs into atomic facts to detect and classify hallucinations across diverse tasks.
Methods in this Paper 🔧:
→ HALOGEN tests LLMs on 9 different domains like coding, summarization, and scientific citations.
→ For each domain, it breaks down model outputs into atomic units (like package names or factual statements).
→ These units are then automatically verified against trusted knowledge sources.
→ Hallucinations are classified as Type A (incorrect recall), Type B (bad training data), or Type C (fabrication).
Key Insights from this Paper:
→ Even the best LLMs hallucinate 4-86% of facts depending on domain
→ No single domain predicts hallucination behavior in other domains
→ Larger models generally hallucinate less on response tasks
→ Open-source models lag behind closed models in factual accuracy
Results 📊:
→ GPT-4 hallucination rates: 4% (coding) to 86% (scientific citations)
→ 91% verification accuracy for summarization
→ 92% accuracy for text simplification
→ 83% accuracy for historical events
Diptesh Kanojia retweeted
I appreciate DeepSeek providing examples of failure, especially since these are ideas that have been widely discussed for achieving o1-style models. This is very rare to see in AI papers.
