Michael Collins

Ideas Mova took (2)

From the ideas ledger: an idea tried, kept or queued, or measured with no gain, and what it gave.

• kept — en/src/parse.rs: add a distance feature for each candidate arc (direction, binned distance, and a count of verbs or punctuation between the two words); measure cv LAS on the train folds, then EWT test once per release. (Head-Driven Statistical Models for Natural Language Parsing · verbs/punctuation between the pair, on top of conj+subcat: 79.80 -> 79.90 (+0.10, noise level), gold tags 84.41 -> 84.77 (3 shuffles), construction set 62.7…)
• kept — en/src/parse.rs: add a subcategorization feature for a verb head, namely which of its obligatory dependents (nsubj, obj, iobj, ccomp, xcomp) are already attached; measure cv LAS and the LAS of verb-headed arcs with en ud-eval. (Head-Driven Statistical Models for Natural Language Parsing · subcat bit set of attached core dependents, on top of conj: 79.37 -> 79.80 (+0.43, 3 shuffles), gold tags 84.55; commit 1c729b451)

Papers

1. Incremental Parsing with the Perceptron Algorithm (2004) measured
   The paper asks whether an unnormalised discriminative model can work well together with heuristic search in parsing.
2. Head-Driven Statistical Models for Natural Language Parsing (2003) 2 ideas kept
   Collins gives three generative models for lexicalized parsing of the Penn Treebank. A tree is generated top-down from its head: at each step the model chooses the head's modifiers on each side, with distance, subcategorization frames, and for Model 3 traces for wh-movement. Model 1 reaches 87.7% labelled precision…
3. Discriminative Training Methods for Hidden Markov Models: Theory and Experiments with Perceptron Algorithms (2002) measured
   The paper proposes training sequence-labelling models, such as part-of-speech taggers and base-NP chunkers, with a perceptron instead of maximum-entropy models or CRFs.
4. Discriminative Reranking for Natural Language Parsing (2000) measured
   The work splits parsing into two stages. A base probabilistic parser proposes a short list of candidate trees for each sentence together with their probabilities, and a second, discriminatively trained model reorders the list. Because the second model only scores complete candidates, it can use arbitrary, overlapping…
5. Prepositional Phrase Attachment through a Backed-off Model (1995)
   The paper treats the choice between noun and verb attachment of a prepositional phrase as a language-modeling problem.

Co-authors (3)

• Brian Roark
• James Brooks
• Terry Koo

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/authors/michael-collins/