The success of neural word embedding models like word2vec and GloVe motivated research on representing sentences in an n-dimensional space. Michael Manukyan and Hrayr Harutyunyan reviewed several sentence representation algorithms and their applications in state-of-the-art automated question answering systems during a talk at the Armenian NLP meetup. The slides of the talk are below. Follow us on SlideShare to get the latest slides from YerevaNN.
Many languages have their own non-Latin alphabets but the web is full of content in those languages written in Latin letters, which makes it inaccessible to various NLP tools (e.g. automatic translation). Transliteration is the process of converting the romanized text back to the original writing system. In theory every language has a strict set of romanization rules, but in practice people do not follow the rules and most of the romanized content is hard to transliterate using rule based algorithms. We believe this problem is solvable using the state of the art NLP tools, and we demonstrate a high quality solution for Armenian based on recurrent neural networks. We invite everyone to adapt our system for more languages.
Last year Hrayr used convolutional networks to identify spoken language from short audio recordings for a TopCoder contest and got 95% accuracy. After the end of the contest we decided to try recurrent neural networks and their combinations with CNNs on the same task. The best combination allowed to reach 99.24% and an ensemble of 33 models reached 99.67%. This work became Hrayr’s bachelor’s thesis.
Recently we have implemented Dynamic memory networks in Theano and trained it on Facebook’s bAbI tasks which are designed for testing basic reasoning abilities. Our implementation now solves 8 out of 20 bAbI tasks which is still behind state-of-the-art. Today we release a web application for testing and comparing several network architectures and pretrained models.
The Allen Institute for Artificial Intelligence has organized a 4 month contest in Kaggle on question answering. The aim is to create a system which can correctly answer the questions from the 8th grade science exams of US schools (biology, chemistry, physics etc.). DeepHack Lab organized a scientific school + hackathon devoted to this contest in Moscow. Our team decided to use this opportunity to explore the deep learning techniques on question answering (although they seem to be far behind traditional systems). We tried to implement Dynamic memory networks described in a paper by A. Kumar et al. Here we report some preliminary results. In the next blog post we will describe the techniques we used to get to top 5% in the contest.