Open innovation at

Flexudy Pipe

Improve Education and Research through Artificial Intelligence

Open Innovation at Flexudy

In short Flexudy Pipe is Flexudy Education’s research group. Here, we brainstorm, design and implement innovative solutions to improve the quality of education and research. Moreover, we also work on open source tools so that we can give back to the community. Everyone is welcome to join us by sending us a request on social media or via mail. We distinguish between proprietary and open source projects. In principle, anyone can work on any project. Ownership is public when it comes to open source projects, meanwhile Flexudy Education retains sole ownership of proprietary projects. Of course, you can directly contribute on open source projects without sending a request. Whether you are a designer, AI specialist, Enthusiast, Programmer, Mathematician or even a Musician, we will be happy to find something for you to work on and improve your portfolio.

A multi-lingual dataset for FORMULA, TITLE and CITATION detection in text (en, de, fr).

Text quality can greatly influence the predictions of neural networks. We generated a dataset that can be used to train a model to do minor cleaning tasks like math, citation and title detection, which can then be masked before any further processing.

Sentence-Doctor: T5 Sequence to Sequence Transformer Model For Sentence Reconstruction

Many Natural Language Processing tasks rely on sentence boundary detection (SBD). Although amazing libraries like spacy provide state of the art SBD, they often depend on text extractors (e.g pdf text extractors or OCR). The quality of these extractors greatly influence the quality of SBD libraries and as a consequence, the performance of downstream models as well. To help address this problem, we fine-tuned a Google AI‘s T5 model from the hugging face hub that attempts to reconstruct “broken sentences” (sentences that were wrongfully detected due to noise in text). The model is open-sourced here.

Competitive Analysis Trees (CATs): An Alternative to Regular Charts

Often , startups have to create pitch decks and a lot of documents to present their value proposition. This also holds for established companies that want to introduce a new service or process. In the end, competitive analysis tools can provide a visual explanation as to why the new service might be better than existing ones. Commonly, competitive analysis radars and 2D charts are selected for this purpose. At Flexudy Pipe, we argue that using trees drastically improve explainability while enabling multiple the support for multiple dimensions. Radars are very good when it comes to depict multiple dimensions but become quite messy and unreadable with increasing competitors. CATs provide a clear “happy” path of the customers decisions. Check out this article for more.

Labely: Data Labelling Tool For Data Scientists

To train Machine Learning models, you often have to label your data. Data labelling is often tedious and the existing tools are often either expensive or unintuitive. That is why we decided to develop an open source tool for data labelling called Labely. We all those who worked on this project for their huge contribution. The tool can be found here.

T5 for Punctuation Insertion. A post-processor for Facebook AI's Wav2vec 2.0 Speech to Text model.

A no-frills post-processor for facebook AI’s amazing Wav2Vec 2.0, here is one. We fine-tuned a T5 model from the Hugging Face Hub using about ~120K paragraphs extracted from popular QA datasets.

Contributors

Join us now and get your own Flexudy avatar.