name: Leon Derczynski
email: ld@itu.dk
twitter: @leonderczynski
telephone: +45 5157 4948
post: ITU Copenhagen, Rued Langgaards Vej 7, 2300 Copenhagen, Denmark
name: Leon Derczynski
email: ld@itu.dk
twitter: @leonderczynski
telephone: +45 5157 4948
post: ITU Copenhagen, Rued Langgaards Vej 7, 2300 Copenhagen, Denmark
NLPL: 2017-2020, NordForsk.
www.nlpl.eu.
Nordic Language Processing Laboratory. A cross-Nordic collaboration of high-performance computing resources and natural language processing resources, between universities and e-infrastructure organisations.
Role: PI for ITU.
COMRADES: 2016-2018, EC H2020 IA €2.0M.
www.comrades-project.eu.
Collective platform for community resilience & social innovation during crises.
Role: Co-I for U.Sheffield.
PHEME: 2014-2017, EC FP7 CP €4.3M.
www.pheme.eu.
Computing Veracity – the Fourth Challenge of Big Data. Pheme builds technology for finding how true claims made online are. This timely project was rated "excellent" at final evaluation.
Role: co-author, scientific co-ordinator.
uComp: 2013-2016, EC CHIST-ERA €1.25M.
www.ucomp.eu.
Embedded Human Computation for Knowledge Extraction and Evaluation. uComp built extensive resources for crowd-sourcing and social media processing, including an easy corpus construction tool integrated with GATE.
Role: named researcher.
TrendMiner: 2011-2014, EC FP7 CP €3.7M. Trendminer on CORDIS. Large-scale, Cross-lingual Trend Mining and Summarisation of Real-time Media Streams.
personal: github.com/leondz
ITU NLP: github.com/ITUnlp
GATE: github.com/GateNLP
Sheffield NLP: github.com/sheffieldnlp
Leon Derczynski
also: Leon Strømberg-Derczynski
View CV
Google Scholar: user=d8iwqa8AAAAJ
ORCID: 0000-0002-8656-3431
DBLP: Derczynski:Leon
1881 citations
18 h-index
29 i10-index
2018 10 01 at date
2018.08.2x: Program chair at COLING 2018, Santa Fe
2018.11.01: Dimensions of Variation in User-generated Text at the Workshop on Noisy User-generated Text (W-NUT), Brussels
2018.11.08: Fake News and Troll Detection at SLTC, Stockholm
2019.03.xx: OpenITU evening on Clinical AI, Copenhagen
2019.Q1: Guest lecturing ML & NLP at Innopolis University, Kazan, Russian Federation
2018
COLING: . 2018. Proceedings of the 27th International Conference on Computational Linguistics
ISCRAM: . 2018. Helping Crisis Responders Find the Informative Needle in the Tweet Haystack
2017
book: . 2017. Automatically ordering events and times in text
CSL: . 2017. Generalisation in Named Entity Recognition: A Quantitative Analysis
2016
COLING: . 2016. Broad Twitter Corpus: A Diverse Named Entity Recognition Resource
COLING: . 2016. Representation and Learning of Temporal Relations
AAAI: . 2016. Generalised Brown Clustering and Roll-up Feature Generation
2015
IPM: . 2015. Analysis of Named Entity Recognition and Linking for Tweets
SemEval: . 2015. SemEval-2015 Task 6: Clinical TempEval
WEBIST: . 2015. Enhanced Information Access to Social Streams through Word Clouds with Entity Grouping - best paper
2014
EACL: . 2014. Passive-Aggressive Sequence Labeling with Discriminative Post-Editing for Recognising Person Entities in Tweets
EACL: . 2014. DKIE: Open Source Information Extraction for Danish
EACL: . 2014. The GATE Crowdsourcing Plugin: Crowdsourcing Annotated Corpora Made Easy
2013
ICTIR: . 2013. Information Retrieval for Temporal Bounding
ACL: . 2013. Temporal Signals Help Label Temporal Relations
SemEval: . 2013. SemEval-2013 Task 1: TempEval-3: Evaluating Events, Time Expressions, and Temporal Relations.
HT: . 2013. Microblog-Genre Noise and Impact on Semantic Annotation Accuracy
EDBT: . 2013. Towards Context-Aware Search and Analysis on Social Media Data
2012
LREC: . 2012. TIMEN: An Open Temporal Expression Normalization Resource
Relative position of Starman (live data)
EARTH:km :mi VEL:km/s :km/h :mph (origin)
NLP: Polysynthetic morphology extraction. Greenlandic is a difficult language to work with because it forms words by adding many parts (called morphemes) together. This project builds a tool for identifying where those morphemes in a word in a sentence. It might use rules, or a deep learning based method, perhaps using character-level convolutional neural nets, or neural attention.
NLP: Clinical event recognition. Lots of information about patients and their health is kept in clinical notes. Clinical note technology is advanced for English but not so for Danish - a shame, because we have great digitalisation. This mini-project helps close that gap by building tools for automatically extracting clinical events (like surgeries, heart attacks, medication changes) from Danish clinical notes.
ML: Gemstone valuation. Gemstones are tough to value, due to their hugh variance. This project involves collecting a lot of data on one or two gem types, and building a model of gem stone value based on various features of the stone - e.g. color richness, cut depth, cut quality, mine of origin, weight, and so on. A working model will perform automatic estimation of gem value, an otherwise expensive service.
2018.07 Read an interview with me in "Alt om Data" (Danish): Nye sandheder om falske nyheder
2018.08 I was program co-chair for COLING 2018 (we had 1018 full paper submissions)
2018.10 Dagbladenes Bureau interviewed me on using AI to hire people (Danish): Ansat af en maskine
_ █
NLP: Time recognition. We should be able to connect dates and day mentions to calendars. This is a tougher task than it looks to automate (e.g. on what day is pinse 2014?), and needs to be done for each language. Here we'll use ISO-TimeML to process for Danish (or another non-English language) and then build a neural network for recognizing times automatically.
NLP: Stance detection. Fake news detection currently relies on knowing the attitude that people talking on social media are expressing towards and idea. Figuring this out is called stance detection. This project builds a stance detection system for Danish or a non-English language of your choice, over social media data.
NLP: Neural Eliza. There's an old program, Eliza, that lets you talk to the computer. It was far ahead of its time. This project builds a neural network that functions as Eliza does. Meet Eliza in javascript
NLP: Offensive & hate speech detection. Catching offensive speech helps us measure the tone of a dialog. This is useful in many contexts, assisting content moderation, thus avoiding legal action and maintining uptime. This mini-project involves building an offensiveness (or politeness) management tool.
brown (generalised)
\[ MI(C_i,C_j)= p(\left< C_i,C_j\right>)\ \log_2{\frac{p(\left<\ C_i,C_j\right>)}{p(\left< C_i,*\right>)\ p(\left<*,C_j\right>)}} \]
\[ AMI(C) = \sum_{C_i,C_j\in C}{MI(C_i,C_j)} \]
\[ C_{i\leftarrow j} = \left( C \setminus \left\{C_i,C_j\right\} \right) \cup \left\{C_i \cup C_j \right\} \]
\[ 0 < a \leq ||C|| \]
\[ i,j \in [1..a] \]
\[ \DeclareMathOperator*{\argmax}{arg\,max} \hat{\pi}(C) = \argmax_{C_i,C_j\in C,i\neq j}{\ AMI(C) - AMI(C_{i\leftarrow j})}. \]
NLP group at the ITU University of Copenhagen
Four faculty
Group leads the Data Science program
For details and research specialisations, visit the webpage: nlp.itu.dk
ITU is an agile, non-traditional, expanding university. I think of it as the startup of unis. Projects are easy to take on and novel ideas have plenty of instituional support. It achieves one of the highest rates of external funding per faculty member in the country, and the highest rate of female:male student applicants (29% of Bachelor in Software Development course applicants were female in 2018). The building is situated just between the national broadcaster, DR, and the KUA campus, and has a professional, light and spacious feel.
Natural Language Processing / Text Mining
Fake news, veracity & stance - continuing work from the PHEME project: how do we determin veracity of claims made on the web? What behaviours exist around news? How do we know that the data your system is processing is geniune and accurate?
Noisy text processing - including social media text, and clinical text (medics' notoriously bad handwriting does, in aggregate, sometimes map also to keyboard skills)
Clinical text mining - and pre-clinical public health. Continuing previous associations with Mayo Clinic, NHS SLaM, and Harvard Children's
Information extraction - who did what to whom, and where and when? This is explained in the text but harder to automatically extract and reason about, even though we can do a lot with that ability.
Spatio-temporal language - extracting timelines from text
keywords: natural language processing, machine learning, social media, time, semantics, distributionality, artificial intelligence