🧑‍🔬 Sam Clark

Sam Clark's papers here concern language processing for Twitter, where tools trained on newswire perform poorly because messages are short and noisy, with slang, irregular spelling and capitalization that cannot be relied on. The recurring question is how to adapt standard shallow-analysis components, chiefly part-of-speech tagging and named entity recognition, to such text when little annotated material exists. The methods typically combine a small amount of hand-labelled tweets with out-of-domain corpora, word clusters learned from unlabelled messages, and ways of reusing existing taggers.

The most cited work is Named Entity Recognition in Tweets (2011). It rebuilt the processing pipeline for Twitter step by step. A retrained tagger, a chunker and a classifier that judges whether a message's casing is informative supply features for entity segmentation, and entity types are assigned through distant supervision. Twitter Part-of-Speech Tagging for All (2013) evaluated several off-the-shelf taggers on tweets and classified their mistakes on unknown words. It then extended the Stanford tagger with treatment of rare and slang forms, priors for unambiguous tokens, and a bootstrapping scheme. In that scheme, an unlabelled tweet becomes training data only when two taggers with different tagsets agree on it.

The 2013 tagging paper is also the newest here. It shifts attention from entity extraction toward the quality of training data. It uses agreement over an entire tweet as a filter for automatically labelled material and argues for reporting whole-sequence accuracy alongside token scores. Clark's papers here list the University of Washington between 2011 and 2013. Alan Ritter appears on both works. The entity recognition study was written with Mausam and Oren Etzioni, and the tagging paper with Leon Derczynski and Kalina Bontcheva.

Sam Clark has 2 papers in our library, in 2011–2013, most of them on Part-of-speech tagging, Named entity recognition; often with Alan Ritter, Kalina Bontcheva, Leon Derczynski.