We present HotelMatch-LLM, a multimodal dense retrieval model for the travel domain that enables natural language property search, addressing the limitations of traditional travel search engines which require users to start with a destination and editing search parameters. HotelMatch-LLM features three key innovations: (1) Domain-specific multi-task optimization with three novel retrieval, visual, and language modeling objectives; (2) Asymmetrical dense retrieval architecture combining a small language model (SLM) for efficient online query processing and a large language model (LLM) for embedding hotel data; and (3) Extensive image processing to handle all property image galleries. Experiments on four diverse test sets show HotelMatch-LLM significantly outperforms state-of-the-art models, including VISTA and MARVEL. Specifically, on the test set—main query type—we achieve 0.681 for HotelMatch-LLM compared to 0.603 for the most effective baseline, MARVEL. Our analysis highlights the impact of our multi-task optimization, the generalizability of HotelMatch-LLM across LLM architectures, and its scalability for processing large image galleries.

In high-stakes industrial NLP applications, balancing generation quality with speed and efficiency presents significant challenges. We address them by investigating two complementary optimization approaches: Medusa for speculative decoding and knowledge distillation (KD) for model compression. We demonstrate the practical application of these techniques in real-world travel domain tasks, including trip planning, smart filters, and generating accommodation descriptions. We introduce modifications to the Medusa implementation, starting with base pre-trained models rather than conversational fine-tuned ones, and developing a simplified single-stage training process for Medusa-2 that maintains performance while reducing computational requirements. Lastly, we present a novel framework that combines Medusa with knowledge distillation, achieving compounded benefits in both model size and inference speed. Our experiments with TinyLlama-1.1B as the student model and Llama-3.1-70B as the teacher show that the combined approach maintains the teacher’s performance quality while reducing inference latency by 10-20x.

Production NMT systems typically need to serve niche domains that are not covered by adequately large and readily available parallel corpora. As a result, practitioners often fine-tune general purpose models to each of the domains their organisation caters to. The number of domains however can often become large, which in combination with the number of languages that need serving can lead to an unscalable fleet of models to be developed and maintained. We propose Multi Dimensional Tagging, a method for fine-tuning a single NMT model on several domains simultaneously, thus drastically reducing development and maintenance costs. We run experiments where a single MDT model compares favourably to a set of SOTA specialist models, even when evaluated on the domain those baselines have been fine-tuned on. Besides BLEU, we report human evaluation results. MDT models are now live at Booking.com, powering an MT engine that serves millions of translations a day in over 40 different languages.