David McClovsky, Eugene Charniak, Mark Johnson (ACL 2010)

Presented by: Nathan Green
Report by: Katerina Topilova

Idea – when parsing large data from diverse domains, it is useful for parsers to be able to generalize to a variety of domains.
The result is a system that proposes linear combinations of parsing models trained on the source corpora.

• Uses regression to predict f-scores
• Features – CosineTop50, UnkWords and Entropy
• Training data for the regressor consists of examples of source domain mixtures and their actual f-scores on target texts

Evaluation – 2 scenarios – out-of-domain evaluation, in-domain evaluation
Baselines – Uniform, Self-Trained Uniform, Fixed Set: WSJ, Best Single Corpus, Best Seen, Best Overall
Feature selection – round-robin tuning scenario

Results:

• Self-trained corpora are beneficial
• This model is the best non-oracle system for both scenarios
• Only 0.3% worse than Best Seen for out-of domain
• Within 0.6% of the Best Seen for in-domain
• 0.7% better than Best Overall model

• CosineTop50 feature is not sufficiently explained and leaves more possibilities of understanding
  1. we could take the top 50 words from first corpus and their frequencies and find frequencies of these words in the second corpus and then compare
  2. we could take the top 50 words from both corpora and compare only the frequencies, where these words overlap
  3. we could find the top 50 words that the two corpora have in common and then compare the frequencies
• Entropy feature could also have been more explained - we have agreed on RG, that this is the entropy of the distribution over mixtures of source domains, taking as better the mixtures, where the entropy value is closer to maximum
• Only Genia and SWBD are really different and the other perform well on WSJ, more domains would be better
• Overall, it would be very hard to implement this system based solely on this article

• Results are good