Applying Text Generation Models to Scalable Customer Support at Airbnb

Modern AI has seen rapid growth in text generation models, which can produce natural language outputs after large‑scale pre‑training. Unlike traditional discriminative NLP models, generative models can encode domain knowledge and be prompted to perform a variety of tasks.

Airbnb has applied these models to its community support product in three main use cases: content recommendation, a real‑time support‑agent assistant, and a chatbot paraphrase model.

Content recommendation model : The previous ranking system used an XLM‑RoBERTa classifier. It was replaced by an MT5 encoder‑decoder model that receives a prompt‑augmented input consisting of the user query and candidate document representations (title, summary, keywords). The model is asked to generate a “yes”/“no” answer indicating whether the document resolves the query. Offline evaluation and online A/B tests showed significant improvements in relevance metrics and user experience.

Real‑time support‑agent assistant : The assistant provides suggested reply templates and performs intent detection via a QA model built on the same generative architecture. Prompts and annotations guide the model to produce context‑aware suggestions. Experiments with t5‑base, Narrativa, and other bases on mixed labeled and log data demonstrated that combining high‑precision labeled data with large‑scale noisy logs yields the best performance. Training leveraged DeepSpeed across multiple GPUs to reduce training time.

Chatbot paraphrase model : To increase user confidence, a paraphrase model re‑expresses the user’s problem at the start of a conversation. Various seq2‑seq transformers (BART, PEGASUS, T5) and autoregressive models (GPT‑2) were evaluated, with T5 performing best. Common failure modes such as overly generic replies were addressed by filtering clustered generic responses using Sentence‑Transformers similarity and by re‑ranking with a reverse‑probability model.

The overall conclusion is that large‑scale pre‑trained transformer models can encode domain knowledge, enable unsupervised learning, and be prompted to solve real‑world customer‑support challenges, leading to higher user satisfaction and more efficient agent workflows.