#15: MLDublin meets @ Linkedin

In our Big Data world, the amount of text being gathered is ever expanding. This presentation identifies an approach to text classification using a graph-of-words keyword extraction and a bipartite co-clustered graph of documents and terms. The keyword extraction can scale well on a large corpus and the resultant model can be visually demonstrated to users.

Faceted search is the predominantly used interface in vertical search engines such as people and product searches. This traditional interface is cumbersome in complex queries, particularly on smaller screen devices. On the other hand, natural language is efficient in conveying complex queries in a short form. Moreover, a dialogue conversational interface facilitates navigational searches. In this talk, we explain how to turn a faceted search engine to a conversational search. In particular, we explain in more details how to create a natural language understanding (NLU) unit for a conversation engine.

Training data is always key to solving supervised machine learning problems. In many cases the labeled training data will be available in one domain (e.g. news/sports..etc) but our solution needs to work for other domains. There are many approaches to solve this issue by adjusting the training data so it can be useful and work better for the target domain. In this talk we will go through few of the baseline approaches for domain adaptation and present one specifically simple and interesting approach based on the paper “Frustratingly Easy Domain Adaptation By Hal Daumé III”