How AI Text Generation Models Are Reshaping Customer Support at Airbnb

April 10th 2024

DigitalDrip

Apr 13, 2024

From : AirBnb Blogs

Leveraging text generation models to build more effective, scalable customer support products.

One of the fastest-growing areas in modern Artificial Intelligence (AI) is AI text generation models

Problem Statement

Applying AI models in large-scale industrial applications like Airbnb customer support is not an easy challenge. Real-life applications have many long-tail corner cases, can be hard to scale, and often become costly to label the training data.

The first attractive trait is the capability to encode domain knowledge into the language models

The second trait of the text generation model we’ve found attractive is its “unsupervised” nature

Finally, text generation models transcend the traditional boundaries of ML problem formulations Over the past few years

In order to make the language model unification effective, a new but essential role is introduced: the prompt.

For content recommendation use case: fine-tuned the MT5 model
For Knowledge assistant use case : developed a mastermind Question-Answering (QA) model
For Real-Time Agent Assistant : developed a mastermind Question-Answering (QA) model

Summary

Accurate intent detection, slot filling, and effective solutions are not sufficient for building a successful AI chatbot. Users often choose not to engage with the chatbot, no matter how good the ML model is. Building a paraphrase model, which first rephrases the problem a user describes, can give users some confidence and confirm that the bot’s understanding is correct, and has significantly improved the bot’s engagement rate.

Best solution was to tune the training data. To do this, we ran text clustering on the training target data based on pre-trained similarity models from Sentence-Transformers.

With the fast growth of large-scale pre-training-based transformer models, the text generation models can now encode domain knowledge

Read More : Airbnb Blogs

Finding AI-generated (deepfake) faces in the wild

From : LinkedIn Engineering Blog

LinkedIn uses cutting-edge models to detect and remove fake accounts and AI-generated profile images to ensure authentic interactions on their platform.
They collaborated with academic experts like Professor Hany Farid through the LinkedIn Scholars program to co-create industry's first large-scale detector for AI-generated images in 2022.
The new research introduces a novel concept for a model that can detect AI-generated faces produced by various generative algorithms like GANs (StyleGAN, EG3D), generated.photos, and diffusion models (Stable Diffusion, DALL-E 2, Midjourney).
The model was trained on a diverse dataset of 120,000 real LinkedIn photos and 105,900 AI-generated faces across 10 different synthesis engines.
It achieved 98% recall in detecting in-engine AI faces and 84.5% for out-of-engine faces at a controlled 0.5% false positive rate.
The model is robust to JPEG compression, resolution changes and was able to identify the core property of an AI-generated face rather than minor artifacts.
Analysis showed the model focused on facial regions and skin areas to make its predictions, leveraging semantic facial features.
This research allows LinkedIn to enhance their automated fake account detection and improve platform authenticity at a large scale.

In summary, this novel AI-generated face detection model exhibits state-of-the-art performance by focusing on semantic facial features across diverse generative engines. Its robustness and ability to generalize enable LinkedIn to effectively combat misuse while ensuring authentic experiences for their members.

Beyond Bots: What Else Can AI Do For Business?

From: Forbes

This article deep dives into the complementary architectures of RAG (Retrieval Augmented Generation) and STAG (Stream-Triggered Augmented Generation) for leveraging AI systems. Here are the key points I took away:

RAG systems like ChatGPT work by taking a user query, retrieving relevant knowledge/context, and generating a response using a large language model (LLM). This reactive query-response pattern is very useful but limited to what the user asks about.
STAG systems, on the other hand, are proactive. They continuously monitor data streams for predefined queries/patterns and surface important insights to users before being explicitly asked.
While RAG amplifies user knowledge through querying, STAG amplifies user vigilance by autonomously analyzing data streams for relevant trends, issues, and opportunities.
A major gap RAG has is only surfacing what users think to ask about, missing emerging trends they aren't aware of. STAG fills this gap through autonomous monitoring.
STAG unlocks value from the vast unstructured data (99%+) that goes unanalyzed by utilizing LLMs to understand nuances, contextualize insights, and communicate them in an actionable way.
RAG and STAG are complementary - RAG provides interactive knowledge access, while STAG provides proactive insights. The most effective AI strategies will combine both frameworks synergistically.
Example synergies include using STAG to improve RAG performance, STAG surfacing opportunities that RAG assists in executing, and creating AI teams that balance query responsiveness with strategic vigilance.

In summary, while recent RAG models like ChatGPT have catalyzed AI adoption, the article advocates for a balanced strategy leveraging both reactive RAG and proactive STAG systems to comprehensively augment human intelligence. The future belongs to organizations that thoughtfully combine these complementary AI architectures.