Trip Report Language Resources and Evaluation Conference 2016 (LREC 2016)

lrec2016

From 23 until 28 May the biannual Language Resources and Evaluation Conference (LREC) took place in Portorož, Slovenia. LREC is a large conference in our field covering all aspects of language technology. About 1200 people attended (who were all quite happy that the WiFi worked!) and nearly 750 papers were presented (4 parallel oral sessions and 5 poster sessions throughout the conference). So plenty for everyone out there, and naturally this post can only reflect the papers that caught my attention and what I think might be of interest to you. 

First all: CLARIAH and CLARIN ERIC were well represented:

 

 

Language

Besides a fair amount of attention to sign language (sessions P15 and O30) and less-resourced languages (session P42), there was also attention for historical language use, such as POS-tagging for Historical Dutch by Dieuwke Hupkes and Rens Bod. What I found really nifty is that they use word alignments between contemporary Dutch (for which we have lots of language tools) and historical Dutch to assign the correct POS-tag. 

There was also a poster presentation by Maria Sukhareva and Christian Chiarcos on Combining Ontologies and Neural Networks for Analyzing Historical Language Varieties. A Case Study in Middle Low German. Again projections are used (I guess I never had to worry about that working on contemporary text) and I like that it combines machine learning with background information from an ontology to improve performance.

Resources   

There were lots of interesting resources and frameworks for publishing linguistic resources presented. One where we can learn (and tag onto) our colleagues from the Semantic Web is the Linguistic Linked Open Data Cloud, where linguistic resources can be stored in a uniform format which enables easier (not yet entirely painless) reuse. 

Corpus building is a time-consuming task, so I also really liked the The Royal Society corpus: From Uncharted Data to Corpus poster. Whilst the Royal Society dataset interests me anyway, they adopted an approach to build the corpus based on agile software development. Whilst this may not be suitable to every corpus building effort, it may be worthwhile to take notice of and see where we can make our approaches more flexible to publish data faster and use feedback loops to improve it. 

Then there were also several datasets convering non-english languages such as the Royal Library 1 Million Captioned Dutch Newspaper Images by Desmond Elliott and Martijn Kleppe,  An Open Corpus for Named Entity Recognition in Historic Newspapers by Clemens Neudecker, containing Dutch, French and German newspaper text including historical spellings and Publishing the Trove Newspaper Corpus by Steve Cassidy on the corpus derived from the National Library of Australia's digital archive of newspaper text.

Here, I should also mention the 2nd keynote by Ryan McDonald from Google on "The Language Resource Spectrum: A perspective from Google". In his talk he presented some experiments done at Google on different NLP tasks to figure out whether to put more effort (=money) into annotated data or fancier language models. Whilst some of the results were not that surprising I think it's an interesting to question to ask and we don't always ask ourselves this are researchers because we are 'used to using method X or Y" (at least in my limited experience).

Evaluation

Unfortunately, the poster didn't make it to Slovenia, but the paper on Complementarity, F-score, and NLP Evaluation by Leon Derczynski raises some interesting issues on how we compare systems; when two systems reach the same F-score for example it doesn't mean they perform the same on all aspects of the problem. 

<shameless plug>I also got to present our paper on Evaluating Entity Linking: An Analysis of Current Benchmark Datasets and a Roadmap for Doing a Better Job where we looked at the different characteristics of different entity linking benchmark datasets and found that there is still a fair bit of work to do before we are testing different dimensions of the problem.</shameless plug>

 

Concluding remarks:

All in, LREC was yet again a great, varied three day whirlwind of what's hot and happening in language technology in Europe (and a little bit beyond). After having gotten some sleep and catching up on the papers I didn't get to see, I'm looking forward to LREC 2018!

Marieke van Erp