28/11/2008

Yet another change. Since November I work in the CS department at Stanford as a senior research associate. More specifically, I am in the Natural Language Processing group, where I work under prof. Chris Manning on NLP-ish applications for the legal domain.

7/7/2008

The CoNLL shared task was a success: more than 50 teams registered and 22 of these teams submitted results. Most groups implemented pipeline architectures, which suggests that joint learning is not trivial for this problem. Nevertheless, some teams ventured in the joint-learning territory with some interesting approaches. Read all about it here. For more details, please attend the shared task session at CoNLL.

7/12/2007

I am one of the organizers of the next CoNLL shared task evaluation, which this year is devoted to the joint parsing of syntactic and semantic dependencies. If you are in this field or are just interested in the problem, read more about the task here and participate!

22/11/2007

I have recently changed jobs. Previously I was with Technical University of Catalonia. Since mid October I am a research scientist at Barcelona Media-Innovation Center and visiting at Yahoo! Research Barcelona in the same time. The best way to contact me is to use the surdeanu DOT name address: this will always point to wherever I am currently.

27/07/2007

I have moved all my public software at sourceforge.net. With this occasion I released new versions for both SwiRL and Bios. There are several significant changes: I changed the inference strategy in SwiRL from greedy to approximate inference using global (i.e., frame level) information. In Bios I switched the default classifiers to Crammer's multiclass Perceptron. More importantly, I have separated all the machine learning software into a new project called MiLL. In case you're wondering this stands (more or less) for Machine Learning Library. MiLL includes SVM, Maximum Entropy and Perceptron classifiers under a unique and simple interface. MiLL includes the novel Perceptron algorithm with dynamic uneven margins I designed for my ACE Information Extraction system (see the publication page).

12/07/2006

I have released Bios, my suite of syntactico-semantic analyzers for English. Bios is implemented in Java and it includes a named-entity recognizer, a syntactic chunker, a POS tagger, and a smart tokenizer. All tools obtain state-of-the-art performance on existing corpora.

27/03/2006

I have released SwiRL, my semantic role labeling system. Coming soon: a named entity recognizer that achieves a F1 measure of over 90%. You can download all this code and more in my software page.

3/03/2006

I have decided to release some of the NLP software I created as GPL. For now, I have posted spear, which is a complete implementation of Michael Collins' parsing model 1 (using most of his code, actually). Coming soon: a named entity recognizer that achieves an F1 measure of over 90% and a semantic role labeling system that scores ~77% F1. You can download all this code in my software page.