Semantic Entropy Confidence Flags for LLM Answers in Urdu and English

The problem: confident answers that are wrong

Banks, telecom companies, universities and clinics in Pakistan are adding language model assistants to their websites and WhatsApp channels, often in Urdu as well as English. The models answer fluently, which is the danger: a wrong answer about a fee, a deadline or a policy sounds exactly as sure as a right one. Customers act on it.

A useful assistant needs to know when it does not know, and say so, or hand over to a person. Asking a model "how confident are you?" does not work well. A better signal is consistency: if you ask the same question several times and get answers that mean different things, the model is probably making it up.

What already exists and the gap

• Evaluation tools such as promptfoo and DeepEval test model outputs before release, but they do not decide, live, whether to trust one answer.
• Guardrail products check answers for policy breaches and leaks, not for whether the content is likely invented.
• Research on uncertainty for language models is active, almost entirely in English.

The gap is a working, measured confidence layer for Urdu and English that a team can put behind its assistant, with honest numbers on how many wrong answers it catches and what it costs in extra calls.

Inspiration

"Detecting hallucinations in large language models using semantic entropy" by Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn and Yarin Gal, published in Nature in June 2024, samples several answers, groups those that entail each other into meaning clusters and measures the entropy over clusters. High entropy signals confabulation. The method needs no labelled examples of hallucination. Your project tests whether it holds up in Urdu, where the entailment step is the weak link.

What you build

• Teams building assistants call your service with a question and get back an answer plus a flag: confident, uncertain or abstain.
• End users see "I am not sure about this, here is how to reach a person" instead of an invented answer.
• Researchers get an Urdu and English evaluation of a published method, with code.

The technically interesting part

• Comparing meaning in Urdu. Two Urdu answers can say the same thing with different word order, spelling variants and borrowed English words. You need a multilingual entailment model, and you must check by hand how often it gets Urdu pairs right, because every error there flows into the flag.
• Cost. Sampling five or ten answers multiplies cost and delay. Study how few samples you can get away with and whether a cheap first pass can decide when the full check is needed.
• Short versus long answers. The method works best on short factual answers. Long answers need splitting into claims first; decide whether that is in scope.
• Thresholds. A flag that abstains too often is useless; one that rarely abstains misses errors. Show the trade-off curve and let the team pick.

How to evaluate it

• Separation of right and wrong answers: area under the ROC curve for each method, per language.
• Selective accuracy: accuracy of the answers the system chooses to give, at different abstention rates.
• Entailment quality: agreement of the entailment model with human judgement on a sample of Urdu answer pairs.
• Cost and latency: extra calls and seconds per question for each sample count.
• Demo test: volunteers ask the support bot real questions; count wrong answers given with and without the flag.

Suggested plan

• Weeks 1 to 4: read the paper, reproduce a small English result with an open model.
• Weeks 5 to 9: build the Urdu evaluation set from public question answering datasets such as UQA and UQuAD, and check gold answers.
• Weeks 10 to 14: Urdu entailment checks, the full pipeline and baselines.
• Second semester, weeks 1 to 5: cost study, thresholds, and the service.
• Weeks 6 to 10: demo support bot and volunteer test.
• Weeks 11 to 14: report and release.

Path to a product

Any company running an assistant in Urdu needs a way to reduce wrong answers it cannot catch by hand. A hosted confidence service or an open library with paid integration is realistic, and the evaluation set is a contribution in its own right. The project is strong preparation for AI engineering and evaluation roles.

Scope warning

Do not build your own language model or your own entailment model from scratch. Keep to short factual answers in the first semester. Report results honestly even when the method works worse in Urdu; that finding is useful. Watch API costs: cap the number of questions per run.

Learn what you need on Sefism

• AI engineering roadmap: working with model APIs, evaluation and guardrails.
• LLM APIs cheat sheet: sampling, temperature and cost control.
• Hugging Face Transformers Documentation: running the open model and the entailment model.
• AI Evals: Everything You Need to Know: building an evaluation you can trust.
• Speech and Language Processing, 3rd Edition: background on entailment and question answering.
• CS224N: Natural Language Processing with Deep Learning: the modelling background.

Get help with this idea

The risk is spending the year on infrastructure and never reaching the Urdu results. Tauseef can help you scope the evaluation set, plan the cost study and keep the work focused on the research question. If you are new here, create a free account first.

• Discuss this idea with a mentor
• Send a query to Tauseef
• Book a 1:1 session