There is a notable lack of large scale, easy to use, labeled data sets for information retrieval in most specific domains. We propose a method for generating them.

Semantic search has generated a lot of hype recently by promising search engines that can understand the meaning behind a search, rather than just looking for keywords. However, the only people using and building these models are Information Retrieval (IR) researchers and their code is not easily transferrable to a production environment with real users. The best of the models that these researchers have developed are capable of impressive language understanding, topping the leaderboard of IR competition such as MS MARCO and TREC-CAR. These semantic search methods represent a significant improvement over traditional keyword search, as much as doubling standard search performance.

Machine learning models are excellent specializers; they’re fantastic at being trained for a particular task and doing it really well.

Fake news, political deception, and rumor mills are not new phenomena, but in recent years the internet has fueled an explosion of deliberate disinformation.