Our QA section is all about knowledge – we seek to gather and maintain knowledge, develop new ways for computers to understand the knowledge, and make the assimilated knowledge useful – mainly by building great user interfaces, often using voice. This is as much an academic as a startup endeavor. Even when we do basic research, we always have practical applications in our cross-hairs.
Unlike other sections, next we list concrete project ideas rather than general topics, but we are quite flexible and will love it if you present us your own goals and ideas. Some projects have a clear path to success, while “(X)” denotes more explorative (but exciting!) projects.
We want to turn computers into real partners in conversation – not just having impressive demos of our existing technology ailao.eu/yodaqa, but something that you’ll actually want to use every day.
- Let’s make a very polished version of ailao.eu/yodaqa
- Mobile / web app improvements – do you like eye-candy too?
- Tweaks related to the movies domain to always make it answer perfectly
- YodaQA+ – a simple 1-month project for improving live.ailao.eulive.ailao.eu
- Integrate YodaQA + Wolfram Alpha + other sources to answer a much wider scope of questions; a lot of this work is already done
- In-car assistant – your hands and eyes are busy, let’s talk!
- Let’s put traffic info, mapy.cz, Wikipedia, weather and news behind a single voice interface
- Czech language interface for YodaQA
- We have a question answering system that can rival IBM Watson (in some aspects), can we teach it Czech?
- Uh-oh-eeh dialog (X)
- When you want to ask the computer something, you probably don’t know the exact wording in your head. So you make a mistake or two and put a lot of pauses and fill words into your speech. We need to make the computer cope, by filtering all this in the speech recognition phase.
Deep Learning for Natural Language
Excited by neural networks and machine learning? Our team is reaching world-class results when processing and analyzing written text, and looking for help!
- Reach state-of-art results with our open source neural network framework with existing and new datasets; neural networks are actually easy, we’ll teach you!
- Pleasant book reader (X) – let’s train a computer to talk like an actor on a series of audiobooks (Harry Potter, The Song of Ice and Fire, Discworld, …). Current programs sound like depressed robots because they don’t understand what’s going on in the story and the emotions in the dialogue, but our NN models could!
- Someone just asked a two-paragraph question – let’s train a model that will find the most similar question already asked before (“question-to-question”)
- (X) We are trying to train a single model that can do many NLP-related tasks, hoping to breathe more general “text comprehension” skills into a machine
Teaching Computers Human Knowledge
We want to feed any text, PDF manual or datasheet to our existing question answering system, so that anyone can just ask what they need to know, and in time computers can themselves ponder and reason on all the human knowledge. Here are the blank spots we need to fill first.
- Yes/No on Knowledge Graph – a simple self-contained project (maybe you only have 1 month?). When presented with “is michelle obama president of the US?” try to verify that assertion (and similar ones) in our knowledge graph and just say yes or no.
- Question Manager – a pretty app that takes a page or two of text, tries to generate all possible questions about its contents, guess the answers from the sentence structures and let the operator fine-tune the result.
- Fast Knowledge Graph – enjoying down-to-metal programming? Try to build a graph database that can retrieve all facts related to an entity – out of 3 billion facts (graph edges), in a split-second! There are database engines that are powerful and general, but too slow for the one thing we really need.
- PDF Information Extractor (X) – let’s put together existing tools and more to process and store knowledge from PDFs with tables, pictures and diagrams.
- Zero-shot Knowledge Absorption (X) – a peek or two is enough for a human to learn the basic info from an eshop page or housing property advertisement, no matter how they are presented or stored. We have all the components ready to give our software the same power by combining vision and NLP neural networks.