Where NLP can solve your problems
We develop NLP models to solve your challenges
Chatbot For Databases
With our natural language understanding technology we developed a chatbot for databases. This enables non-technical users to get data from anywhere: databases, APIs, ERP-Systems and many more.
Natural language interface for databases
AskBy Data is a human friendly natural language interface for databases.
The core technology, AskBy NL Query, is able to translate natural language into any kind of structured format. For example the user's intent, but also API requests or SQL. That means everyone, even non-technical people, can simply ask for data through their preferred interface and it will be instantly delivered back to the user.
Our NLP AI Tech Stack
At AskBy.ai, we are using the latest insights from Natural Language Processing (NLP) to create and define our models. Especially the field of Natural Language Understanding (NLU) is rapidly developing in recent years due to breakthroughs in the design of task specific neural networks - in particular, these are currently recurrent neural networks in sequence-to-sequence scenarios, as well as attention models. Every year brings new developments that outperform old benchmarks, for example the ones in the GLUE benchmark - a new de-facto standard in the NLP community. This is why it is so important to be up-to-date in this field.
For training language models, we use an internally developed architecture inspired by BERT, an unsupervised transfer learning approach based on the Transformer that outperformed current state-of-the-art models on various tasks in late 2018. From our experience, those models require only little fine tuning to work very well on tasks like document representation, document classification and semantic search. In general, we argue that these models are the right choice for any language task that requires a deeper understanding of the content of texts.
Another scenario and one of the core technologies of AskBy is natural language to formal language translation. For that purpose, we developed a new kind of recurrent neural network architecture called Nefisto (Neural finite state output). Translating to formal language has its own challenges - due to the strict syntactical requirements on the output. For this reason, Nefistos allow to restrict their output using simple grammatical rules. This allows to do translation not only to complex sequences, but even to complex hierarchical calculation trees. Have a look on our calendar demo to see a simple example what it can do.
Often times, especially (but not only) when translating to formal languages, it is difficult to acquire the right amount of data for the task. Classically, training data is generated by humans. But if the prediction space is combinatorially big, it is intractable for a human to write down samples line by line, since the prediction space might grow faster than exponential. For this purpose, we developed a small programming language called Larala (Language randomization language). It gives its programmer a more expressive tool to create training data and capture larger parts of the prediction space with (asymptotically) much fewer lines of code. It was originally developed for AskBy NL Query, a product translating natural language to database requests. In this scenario, there was initially no training data at all - but using Larala we were able to build a system that works extremely well.
So if you think you don’t have enough training data or no training data at all - we might still be able to help you. Get in contact with us.