Amir Salehi

Thesis: “A Prototype-Based Prompt Learning Approach for Out-of-Distribution Intent Detection,” advised by Dr. Behrooz Koohestani.

My thesis asks how a dialogue system can tell when a request falls outside what it knows, instead of confidently misclassifying it as something familiar. I replace the fixed, vocabulary-derived answer space used in prior prompt-based methods with learnable prototype representations refined jointly with the encoder, so the model's notion of "known" and "unknown" intents is learned rather than hard-coded. To keep closely related known intents from collapsing into each other, I introduce a fine-grained contrastive loss that uses label-token overlap and proximity in prototype space to pull confusable intent families apart. I'm now extending the approach to a cross-lingual, low-resource setting with a translated Persian Banking77 dataset, to test how well prompt-based out-of-distribution detection transfers to languages like Farsi.

• Designed a prototype-based prompt-learning framework for out-of-distribution intent detection in task-oriented dialogue, replacing the fixed, vocabulary-derived answer space of prior prompt-based methods with learnable prototype representations refined jointly with the encoder.
• Extending the framework to a cross-lingual setting with a translated Persian Banking77 dataset, evaluating the transferability of prompt-based OOD detection to low-resource languages like Farsi.

• Built the first fine-grained, 77-class Persian intent-detection benchmark by translating and adapting the Banking77 dataset (13,083 samples) for the Persian banking and fintech domain, following TEXTOIR's standard train/dev/test splits.
• Designed an LLM-based machine translation pipeline with intent-aware prompting, followed by a structured human review process with documented translation guidelines and terminology standards for reproducibility.

• Shipped the QA and shopping-assistant features now used by users (around 10k chatbot requests a day).
• Developed end-to-end pipelines for building and deploying chatbot features — covering data ingestion, retrieval indexing, response generation, and error handling — cutting the time from feature idea to production release so new capabilities reached users faster.