summaryrefslogtreecommitdiff
path: root/gi/posterior-regularisation/prjava/src/phrase/C2F.java
blob: 31fd4fda73ff38f5f9c0a890e82ec12bba08df4a (plain)
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
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
package phrase;

import gnu.trove.TIntArrayList;

import io.FileUtil;

import java.io.File;
import java.io.IOException;
import java.io.PrintStream;
import java.util.Arrays;
import java.util.List;

import phrase.Corpus.Edge;

/**
 * @brief context generates phrase
 * @author desaic
 *
 */
public class C2F {
	public int K;
	private int n_words, n_contexts, n_positions;
	public Corpus c;
	
	/**@brief
	 *  emit[tag][position][word] = p(word | tag, position in phrase)
	 */
	public double emit[][][];
	/**@brief
	 *  pi[context][tag] = p(tag | context)
	 */
	public double pi[][];
	
	public C2F(int numCluster, Corpus corpus){
		K=numCluster;
		c=corpus;
		n_words=c.getNumWords();
		n_contexts=c.getNumContexts();
		
		//number of words in a phrase to be considered
		//currently the first and last word in source and target
		//if the phrase has length 1 in either dimension then
		//we use the same word for two positions
		n_positions=c.phraseEdges(c.getEdges().get(0).getPhrase()).size();
		
		emit=new double [K][n_positions][n_words];
		pi=new double[n_contexts][K];
		
		for(double [][]i:emit){
			for(double []j:i){
				arr.F.randomise(j);
			}
		}
		
		for(double []j:pi){
			arr.F.randomise(j);
		}
	}
	
	/**@brief test
	 * 
	 */
	public static void main(String args[]){
		String in="../pdata/canned.con";
		String out="../pdata/posterior.out";
		int numCluster=25;
		Corpus corpus = null;
		File infile = new File(in);
		try {
			System.out.println("Reading concordance from " + infile);
			corpus = Corpus.readFromFile(FileUtil.reader(infile));
			corpus.printStats(System.out);
		} catch (IOException e) {
			System.err.println("Failed to open input file: " + infile);
			e.printStackTrace();
			System.exit(1);
		}
		
		C2F c2f=new C2F(numCluster,corpus);
		int iter=20;
		double llh=0;
		for(int i=0;i<iter;i++){
			llh=c2f.EM();
			System.out.println("Iter"+i+", llh: "+llh);
		}
		
		File outfile = new File (out);
		try {
			PrintStream ps = FileUtil.printstream(outfile);
			c2f.displayPosterior(ps);
		//	ps.println();
		//	c2f.displayModelParam(ps);
			ps.close();
		} catch (IOException e) {
			System.err.println("Failed to open output file: " + outfile);
			e.printStackTrace();
			System.exit(1);
		}
		
	}
	
	public double EM(){
		double [][][]exp_emit=new double [K][n_positions][n_words];
		double [][]exp_pi=new double[n_contexts][K];
		
		double loglikelihood=0;
		
		//E
		for(int context=0; context< n_contexts; context++){
			
			List<Edge> contexts = c.getEdgesForContext(context);

			for (int ctx=0; ctx<contexts.size(); ctx++){
				Edge edge = contexts.get(ctx);
				double p[]=posterior(edge);
				double z = arr.F.l1norm(p);
				assert z > 0;
				loglikelihood += edge.getCount() * Math.log(z);
				arr.F.l1normalize(p);
				
				int count = edge.getCount();
				//increment expected count
				TIntArrayList phrase= edge.getPhrase();
				for(int tag=0;tag<K;tag++){

					exp_emit[tag][0][phrase.get(0)]+=p[tag]*count;
					exp_emit[tag][1][phrase.get(phrase.size()-1)]+=p[tag]*count;
					
					exp_pi[context][tag]+=p[tag]*count;
				}
			}
		}
		
		//System.out.println("Log likelihood: "+loglikelihood);
		
		//M
		for(double [][]i:exp_emit){
			for(double []j:i){
				arr.F.l1normalize(j);
			}
		}
		
		emit=exp_emit;
		
		for(double []j:exp_pi){
			arr.F.l1normalize(j);
		}
		
		pi=exp_pi;
		
		return loglikelihood;
	}

	public double[] posterior(Corpus.Edge edge) 
	{
		double[] prob=Arrays.copyOf(pi[edge.getContextId()], K);
		
		TIntArrayList phrase = edge.getPhrase();
		TIntArrayList offsets = c.phraseEdges(phrase);
		for(int tag=0;tag<K;tag++)
		{
			for (int i=0; i < offsets.size(); ++i)
				prob[tag]*=emit[tag][i][phrase.get(offsets.get(i))];
		}
			
		return prob;
	}

	public void displayPosterior(PrintStream ps)
	{	
		for (Edge edge : c.getEdges())
		{
			double probs[] = posterior(edge);
			arr.F.l1normalize(probs);

			// emit phrase
			ps.print(edge.getPhraseString());
			ps.print("\t");
			ps.print(edge.getContextString(true));
			int t=arr.F.argmax(probs);
			ps.println(" ||| C=" + t);
		}
	}
	
	public void displayModelParam(PrintStream ps)
	{
		final double EPS = 1e-6;
		
		ps.println("P(tag|context)");
		for (int i = 0; i < n_contexts; ++i)
		{
			ps.print(c.getContext(i));
			for(int j=0;j<pi[i].length;j++){
				if (pi[i][j] > EPS)
					ps.print("\t" + j + ": " + pi[i][j]);
			}
			ps.println();
		}
		
		ps.println("P(word|tag,position)");
		for (int i = 0; i < K; ++i)
		{
			for(int position=0;position<n_positions;position++){
				ps.println("tag " + i + " position " + position);
				for(int word=0;word<emit[i][position].length;word++){
					if (emit[i][position][word] > EPS)
						ps.print(c.getWord(word)+"="+emit[i][position][word]+"\t");
				}
				ps.println();
			}
			ps.println();
		}
		
	}
	
}