104 AI chatbot using ElasticSearch and deep learning

摘要：

In this talk, I will share the experience of 104 hackathon contest which includes how to build a chatbot system and I will also go through each solution for solving three kinds of question types (true-or-false, multiple choices and fill-in-the-blank).

In this contest, our team (Tobacco AI) build information retrieval system (ElasticSearch) for fetching the source text and do the similarity between source and question first. Then, for the different question types, we use the sentence embedding SIF (Smooth Inverse Frequency) for the true-or-false question and present a neural network model (context2vec) for learning a context embedding vector from 104 text corpus, using bidirectional LSTM. Therefore, we could use three types of similarity metrics in embedding space: target-to-context (t2c), context-to-context (c2c) and target-to-target (t2t) for the fill-in-blank question, As a result, we are the first place and win the prize in this contest.

講者簡介：

Chris Lin (Tsung-Fu Lin) is Sr. software engineer of TrendMicro. He is an enthusiast in big data and cloud computing technologies, such as Spark and Hadoop. Recently, he is responsible for graph mining and deep learning on threat research, such as email importance identification, fraud message detection, and advisor, etc.

• ← Metadata Store: Generalized Entity Database for Intelligence Services and Machine Learning
• Building-up Centralized Data Catalog Service on AWS among various data/schema file formats, analytic tools →