TRF top level design

1.Features:

See module Feature

• Class Feat：store all the features that has occurred. Each feature is indexed with an integer. The class Feat has an member array of class: FeatTable.
  • Class FeatTable：Store features of a set of similar feature templates. For instance, a FeatTable with constructing parameter w[1:3] will store the features of the following templates:（w_0, w_-1w_0, w_-2w_-1w_0）. A FeatTable object use a member object Trie<VocabID, int> to store those similar fetures. The key of the Trie is IDs of the words, and the value of the Trie is the index for the corresponding feature.
  • class FeatStyle：A specific feature temple. Represented by a struct Field.
• Class Seq：A word sequence which represent a sentence. The class Seq has a member variable of type Mat, with index of each word as the first row, classes of each word as second row.

2.The model and sample algorithms：

See module Model

• class Model：Define and update the parameters. Define the sample algorithms：
  • LocalJump(Seq &seq) Propose and accept the length of next sentence.
  • MarkovMove(Seq &seq) Gibbs sampling each word of a fixed-length sentence.
    Class Model has member objects of following classes:
    • class Corpus：read the training/validation set and build a corpus
    • class AlgNode： implementation of forward-backward algorithm

3.Training of the model:

See module SA Training

• class SAfunc：Implementation of the SA training algorithm.
  For each iteration, commit a TRF sampling and obtain a set of sentences which obey the joint distribution of the pair (l, xl) calculated by current model parameters. Use the difference of E[f] with respect to empirical distribution and joint distribution as the update vector.
  class SAfunc has member objects of following classes：
  • class Model
• class SAtrain：Carry out the SA training
  class SAtrain has member objects of following classes：
  • class SAfunc
  • class LearningRate：adjust learning rate in respect to iteration times

4.Word Clustering：

See module Word Cluster

• class WordCluster：Cluster the words into several classes. The objective to minimize is dRes = dSumClassGram - 2 * dSumClass + m_dWordLogSum. Greedy method is used in each iteration.

5.Data Structure：

See mudule Data Strcture Used

6.System Tools：

See module System Tools