/
RELEASE-NOTES
55 lines (47 loc) · 2.34 KB
/
RELEASE-NOTES
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
Relase 1.3
----------
The third release includes:
- BSegment/BSegmentCRF having implementation of ICML 2006 paper 'Efficient inference on sequence segmentation models' by Sunita Sarawagi
- Bug fixes, lots of performance improvements
- New features in Model package
Release 1.2
-----------
The second release mostly includes a bunch of changes for improving performance,
and a set of new features.
The important changes are:
a) A new version of semi-CRF that allows dataset to specify candidate segments
instead of always evaluating all possible segment sizes from 1 to M. Refer
SegmentCRF.java this change.
b) Support for a large number of sparse labels (classes) along with a sparse
set of features (refer SparseCRF.java).
c) Support for pairwise constraints on labels assigned during inference by Viterbi.
d) Caching of logSumExp calculations to remove the bottlenecks of the massive
number of log/exp functions.
e) Constructor change in FeatureTypes to allow more than one CRF object at a time.
Instead of an object of the Model class, the FeatureTypes constructor now
accepts a FeatureGenerator object. This is a major design change, as all
the Features created by extending FeatureTypes class has to be modified for
this version.
f) Support for serilizability for learned models.
g) A set of new features
1) ClassPriorFeature.java
Encodes prior probability of class labels.
2) FeatureTypesConcat.java
Concatenates token features over a segment (subsequence of tokens).
3) ConcatRegexFeatures.java
Regular expression based features for a segment.
4) RegexCountFeatures.java
Encodes count of a pattern in a segment of tokens.
5) FeatureTypesEachLabel.java
A wrapper for features on input token sequence. Such features are
fired for all class labels once they are wrapped inside this class.
6) WindowFeatures.java
A wrapper for token features, which can specify a window of tokens
in or around the current token (segment) to be used to generate
features for the current token (segment).
7) FeatureTypesPosition.java
To encode relative position of labels within a sequence.
8) FeatureTypesSegmentLength.java & FeatureTypesSegmentLengthPoly2.java
To encode lengths of segments with a small number of new feature ids.
9) FeatureTypesWrapper.java
A wrapper for another feature types for backward compatibility.