Natural Language Processing · free preview

Conditional Random Fields

Why generative taggers strain under unknown words and rich features, the CRF as a log-linear sequence model, local and global features, weight learning, and entity-level evaluation.

1. Conditional Random Fields

2. Features, decoding, training, and evaluation

Features, decoding, training, and evaluation

3. A CRF scorer and entity F1 in NumPy

A CRF scorer and entity F1 in NumPy

4. The generative straitjacket

What is the core limitation of the HMM tagger that the CRF removes?

5. Summing local into global

How do the global features $F_k$ of a CRF relate to the local features $f_k$ ?

6. Computing a word shape

Map the word 'DC-10' to its word-shape fingerprint, upper-case letters to X, lower-case to x, digits to d, punctuation kept as itself. Answer with the exact shape string.

7. Entity-level precision and recall

On a test set an NER system finds $6$ spans correctly, produces $7$ spans in total, and the gold annotation contains $8$ spans. Compute precision, rounding to three decimals.

8. Viterbi with a different arithmetic

What changes when Viterbi runs on a linear-chain CRF instead of an HMM?

Practice this interactively

Free account · instant grading · spaced review that schedules itself.

Start this course for free