,
13 tweets,
8 min read
Read on Twitter
📺 THE VIDEOS FROM #spaCyIRL ARE NOW LIVE! And we don't mind saying, they turned out great. Here's 12 talks about NLP research, development and applications. Summaries and links in the thread 👇 youtube.com/playlist?list=…
Transfer learning has been the big topic in NLP for 2018 and 2019. @seb_ruder opened the conference with how the field has been changing, what it means for OSS, and what could be improved.
In promising preliminary work, @giannis_daras presents a new way to reason about attention architectures: information flow graphs. The graphs show problems with @OpenAI's sparse transformers, and suggest improvements.
When we invited @pmbaumgartner, we knew he'd steal the show. In the most tweeted talk of the conference, Peter discusses NLP's distinct project management problems, and how to give your applied NLP projects the best chance of success.
Chatbots are notoriously hard to build well. @juste_petr talks about some of the most common problems, and how @Rasa_HQ is giving developers tools to build these challenging NLP systems in-house.
Deep learning has brought big accuracy improvements on NLP benchmarks, but also introduced a growing gap between research and application. @yoavgo talks about what's missing, and presents a sketch of what a solution might look like.
Instead of defining words with other words, Entity Linking gives you grounded knowledge. @OxyKodit presents her ongoing work to bring this key NLP technology to spaCy. There are few production-ready EL systems, so this will have a big impact.
Without lemmatization you can't even build a decent word cloud. And spaCy's lemmatizer is pretty lacking. @_guadiromero describes a practical hybrid approach: a statistical system will predict rich morphological features enabling precise rule-engineering.
On lots of genres, default models perform poorly. @MarkNeumannnn shows an end-to-end example of how this can be solved. scispaCy is far faster than other biomedical NLP systems. We particularly liked the trick to make the model less domain-specific.
What gets done in the world depends on how markets allocate resources. Those decisions depend on data, much of which is produced by @SPGlobal. Patrick Harrison shows how NLP systems using spaCy fit into processes refined over decades for 100% reliability.
If you squint a little, active fund management is an information processing task. So NLP should be a clear win, right? McKenzie Marshall discusses why it isn't so easy & explains how @Barings are using spaCy & Prodigy to improve their investment research.
As time passes, old reporting can take on a new significance. Most publications struggle to surface the right information from their archives. As a young publication committed to doing "digital" right from day one, @qz has set up an innovative solution.
At spaCyIRL, the same questions kept coming up. What's the story behind spaCy? What's coming next? How do you guys make money? @_inesmontani and @honnibal got on stage to answer these questions, and look back on the spaCy story so far.
