Institute of Formal and Applied Linguistics Wiki

Comments by : Loganathan

The objective of the paper is to make the statistical parsers adapting to new domains. Best parsing model for a particular testing data is identified by combining training data(source mixture) from different domains. This source mixture is learned from a regression model which will identify the appropriate parsing model.

Some of the comments and doubts arised during the discussion

• It was asked how the data was collected, mainly due to the size of the data used in training.
• Training and testing were reported in the development set not on the parsing models.
• It was noted that the parser has been tested across various domains.
• Entropy feature was not clear.
• The idea was to successfully adapt to new domains than to achieve very good accuracy for a particular domain.

• The multiple source adaptation method can identify the factors which affect the parsing accuracy for texts from different domains.
• They successfully included methods for domain detection compared to previous works.
• Inclusion of self trained corpora helped avoiding data sparsity in small corporas.

• Results (just before section 7) could have been better explained.