A glance into how Airbnb makes use of speech recognition, intent detection, and language fashions to grasp customers and help brokers extra successfully.
By Yuanpei Cao, Heng Ji, Elaine Liu, Peng Wang, and Tiantian Zhang
At Airbnb, we goal to supply a clean, intuitive, and useful group help expertise, whether or not it’s serving to a visitor navigate a reserving change or serving to a number with a list challenge. Whereas our Assist Middle and buyer help chatbot helps resolve many inquiries effectively, some customers want the immediacy of a voice dialog with a help consultant. To make these interactions quicker and more practical, we’ve considerably improved our Interactive Voice Response (IVR) system through machine studying.
Over time, Airbnb has invested in conversational AI to boost buyer help. In our earlier weblog posts Activity-Oriented Conversational AI in Airbnb Buyer Help and Utilizing Chatbots to Present Quicker COVID-19 Neighborhood Help, we explored how AI-driven chatbots streamline visitor and host interactions via automated messaging. This put up explains how we prolong that work to voice-based help, leveraging machine studying to enhance real-time telephone interactions with our clever IVR system.
We’ll take you thru the end-to-end IVR journey, the important thing machine studying parts that energy it, and the way we designed a system that delivers quicker, extra human-like, and extra intuitive voice help for our group.
Conventional IVR programs usually depend on inflexible menu timber, requiring callers to press buttons and navigate pre-set paths. As a substitute, we designed an adaptive, conversational IVR that listens, understands, and responds in actual time. Right here’s usually what occurs when a caller reaches out to Airbnb help:
- Name and greeting: IVR picks up and prompts, “In a couple of sentences, please inform us why you’re calling at this time.”
- Automated speech recognition (ASR): The caller’s response is transcribed with Airbnb-specific ASR. For instance, if a caller says, “I have to request a refund for my reservation,” ASR precisely converts this speech into textual content, preserving key domain-specific phrases.
- Understanding intent: A Contact Purpose Detection mannequin classifies the difficulty right into a class like cancellations, refunds, account points, and so on.
- Determination-making: If self-service is feasible, the system retrieves and sends a related assist article or an clever workflow through SMS or app notification. If the caller explicitly requests agent help or the difficulty requires human intervention, the decision is routed to a buyer help agent with related particulars connected.
- Clarifying response: A Paraphrasing mannequin generates a abstract of the person intent, which IVR shares with the person earlier than delivering the answer. This ensures that customers perceive the context of the useful resource they obtain. Persevering with our instance, the system would reply, “I perceive your challenge is concerning a refund request. We now have despatched you a hyperlink to assets about this matter. Observe the directions to search out solutions. If you have to communicate with an agent, press 0 to be related to our customer support consultant.” The underscored Paraphrasing element enhances engagement by bridging the hole between system-generated responses and person comprehension, making the self-service expertise extra intuitive.
- Decision or escalation: The caller receives an SMS or app notification with a direct hyperlink to a related Airbnb Assist Middle article. If additional help is required, they will press 0 to attach with a customer support consultant.
By shifting away from inflexible menus to pure language understanding, we enable friends and hosts to specific their points in their very own phrases, serving to to extend satisfaction and backbone effectivity.
1. Automated speech recognition (ASR): transcribing with precision
In a voice-driven help system, reaching excessive transcription accuracy is important, significantly in noisy telephone environments the place speech will be unclear. Common speech recognition fashions usually battle with Airbnb-specific terminology, resulting in errors like misinterpreting “itemizing” as “lifting” or “assist with my keep” as “comfortable Christmas Day.” These inaccuracies create challenges in understanding person intent and affect downstream processes.
To reinforce ASR accuracy, we transitioned from a generic high-quality pretrained mannequin to 1 particularly tailored for noisy telephone audio. Moreover, we launched a domain-specific phrase checklist optimization that ensures Airbnb phrases are correctly acknowledged. Primarily based on a pattern of lots of of clips, this considerably diminished the phrase error fee (WER) from 33% to roughly 10%. The diminished WER considerably enhanced the accuracy of downstream assist article suggestions, rising person engagement, enhancing buyer NPS amongst customers who interacted with the ASR menu, whereas decreasing reliance on human brokers and decreasing customer support dealing with time.
2. Contact Purpose prediction: understanding the why
After transcribing the caller’s statements, the following step entails figuring out their intent. We achieved this by creating an in depth Contact Purpose taxonomy that categorizes all potential Airbnb inquiries, as elaborated in “T-LEAF: Taxonomy Studying and EvaluAtion Framework.” We then use an intent detection mannequin to categorise calls right into a Contact Purpose class, guaranteeing every inquiry is dealt with appropriately. For instance, if a caller mentions “I haven’t obtained my refund but,” the mannequin predicts the Contact Purpose as Lacking Refund and forwards it to the related downstream parts.
In manufacturing, we deploy the Difficulty Detection Service to host the intent detection fashions, operating them in parallel to realize optimum scalability, flexibility, and effectivity. Parallel computing ensures that intent detection latency stays below 50ms on common, making the method imperceptible to IVR customers and guaranteeing a seamless real-time expertise. The detected intent is then analyzed throughout the IVR workflow to find out the following motion, whether or not it’s guiding the person via a self-service decision or escalating on to a human agent.
Often, callers want to talk straight with a human agent as a substitute of describing their points, utilizing phrases like “agent” or “escalation.” For such eventualities, we use a special intent detection mannequin to acknowledge when a caller needs to escalate to a human agent. If this intent is detected, the IVR system honors the caller’s request and routes the decision to the appropriate help group.
3. Assist article retrieval: delivering the proper data
Many frequent Airbnb points will be shortly resolved by offering clear and related academic data. To assist present helpful data to customers and decrease the necessity for human buyer help, we use the Assist Article Retrieval and Rating system. This superior system robotically identifies the difficulty in a person’s inquiry and delivers probably the most related assist article hyperlink through SMS textual content message and Airbnb app notification. Our course of incorporates two machine studying levels.
Semantic retrieval and rating: We index Airbnb Assist Article embeddings right into a vector database, enabling environment friendly retrieval of as much as 30 related articles per person question utilizing cosine similarity, sometimes inside 60ms. An LLM-based rating mannequin then re-ranks these retrieved articles, with the top-ranked article straight offered to customers through IVR channels. This dual-stage system not solely powers IVR interactions but in addition helps our buyer help chatbot and Assist Middle search. Throughout these platforms, its effectiveness is constantly evaluated utilizing metrics like Precision@N, facilitating ongoing enhancements and refinements.
4. Paraphrasing mannequin: enhancing person understanding
A key problem in IVR-based buyer help is guaranteeing customers clearly perceive the decision earlier than receiving assist article hyperlinks, as they sometimes lack visibility into the article’s contents or title. To deal with this, we applied a light-weight paraphrasing method leveraging a curated set of standardized summaries.
UX writers created concise and clear paraphrases for frequent Airbnb eventualities. Throughout on-line serving, person inquiries are mapped to those curated summaries through nearest-neighbor matching based mostly on textual content embedding similarity. We calibrated a similarity threshold to make sure high-quality matches. Guide analysis of end-to-end mannequin outputs confirmed precision exceeding 90%.
The end result was a finite-state answer delivering probably the most applicable paraphrased IVR immediate earlier than presenting a assist article hyperlink. For instance, if a caller states, “I have to cancel my reservation and request a refund,” the mannequin generates a response like “I perceive your challenge is a few refund request” earlier than sending the retrieved assist article hyperlink.
Integrating this mannequin ensures customers obtain clear, contextually related summaries previous to accessing assist articles. In an experiment focusing on English hosts who contacted buyer help, we discovered that presenting a paraphrased abstract earlier than sending the article hyperlink will increase person engagement with article content material, leading to enchancment in self-resolution charges, serving to to cut back the necessity for direct buyer help help.
By combining Automated Speech Recognition and Contact Purpose Detection programs with a assist article retrieval system, and a paraphrasing mannequin, we’ve created an IVR system that streamlines help interactions and improves person satisfaction. Our answer allows callers to explain points naturally, reduces dependency on human brokers for frequent inquiries, and supplies immediate, related help via self-service. When human help is important, the system ensures a clean transition by routing customers to the proper agent with important context.
Fascinated by working at Airbnb? Try our open roles.
Because of Zhenyu Zhao, Mia Zhao, Wayne Zhang, Lucca Siaudzionis, Lulu Chen, Sukey Xu, Floria Wan, Michael Zhou, Can Yang, Yaolin Chen, Shuaihu Wang, Huifan Qu, Ming Shang,Yu Jiang, Wanting Chen, Elena Zhao, Shanna Su, Cassie Cao, Hao Wang, Haoran Zhu, Xirui Liu, Ying Tan, Xiaohan Zeng, Xiaoyu Meng, Gavin Li, Gaurav Rai, Hemanth Kolla, Ihor Hordiienko, Matheus Scharf, and Stepan Sydoruk who helped carry this imaginative and prescient to life. Additionally due to Paige Schwartz, Stephanie Chu, Neal Cohen, Becky Ajuonuma, Iman Saleh, Dani Normanm, Javier Salido, and Lauren Mackevich for the evaluation and enhancing.
Because of Jeremy Werner, Pleasure Zhang, Claire Cheng, Yashar Mehdad, Shuohao Zhang, Shawn Yan, Kelvin Xiong, Michael Lubavin, Teng Wang, Wei Ji, and Chenhao Yang’s management help on constructing conversational AI merchandise at Airbnb.
