GIZA++输出缺少*.ti.final和*actual.ti.final文件
在理解如何运行GIZA++的基础上，我遇到了一些问题。
我在StackOverflow (Is there a tutorial about giza++?)上和人们在那里提供的链接上进行了讨论。我已经从摩西-SMT Github下载并编译了最新的giza。
git clone https://github.com/moses-smt/giza-pp.git
cd giza-pp
make在成功编译之后，我编写了一个简单的脚本用于测试。
#!/bin/bash
SRC=french
TRG=english
PREFIX=out
GIZA=../giza-pp
# Cleaning from previous run ...
rm -f *.log
rm -f *.vcb
rm -f *.snt
rm -f *.vcb.classes
rm -f *.vcb.classes.cats
rm -f *.gizacfg
rm -f *.cooc
rm -f ${PREFIX}*
# Converting plain text into sentence format using the "plain2snt.out" tool ...
${GIZA}/GIZA++-v2/plain2snt.out ${SRC} ${TRG}
# Generating word clusters using the "mkcls" tool ...
${GIZA}/mkcls-v2/mkcls -p${SRC} -V${SRC}.vcb.classes
${GIZA}/mkcls-v2/mkcls -p${TRG} -V${TRG}.vcb.classes
# Generating coocurrence using the "snt2cooc" tool ...
${GIZA}/GIZA++-v2/snt2cooc.out ${SRC}.vcb ${TRG}.vcb ${SRC}_${TRG}.snt > ${SRC}_${TRG}.cooc
# Running "GIZA++" ...
${GIZA}/GIZA++-v2/GIZA++ -S ${SRC}.vcb -T ${TRG}.vcb -C ${SRC}_${TRG}.snt -CoocurrenceFile ${SRC}_${TRG}.cooc -o ${PREFIX} >> giza.log 2>&1现在，这是我运行脚本后目录的内容。
jakub@jakub-virtual-machine:~/Master/giza-pp_test$ ls
english french_english.snt out.d3.final out.perp
english_french.snt french.vcb out.d4.final out.t3.final
english.vcb french.vcb.classes out.D4.final out.trn.src.vcb
english.vcb.classes french.vcb.classes.cats out.Decoder.config out.trn.trg.vcb
english.vcb.classes.cats giza.log out.gizacfg out.tst.src.vcb
french out.a3.final out.n3.final out.tst.trg.vcb
french_english.cooc out.A3.final out.p0_3.final run_test.sh重点是输出缺少(对我来说很重要)下面列出的文件。
out.ti.final
out.actual.ti.final现在，我一直在查看GIZA的Main.cpp (第260-273行)，可以看到应该创建这些文件的行。
cerr << "writing Final tables to Disk \n";
string t_inv_file = Prefix + ".ti.final" ;
if( !FEWDUMPS)
m1.getTTable().printProbTableInverse(t_inv_file.c_str(), m1.getEnglishVocabList(),
m1.getFrenchVocabList(),
m1.getETotalWCount(),
m1.getFTotalWCount());
t_inv_file = Prefix + ".actual.ti.final" ;
if( !FEWDUMPS )
m1.getTTable().printProbTableInverse(t_inv_file.c_str(),
eTrainVcbList.getVocabList(),
fTrainVcbList.getVocabList(),
m1.getETotalWCount(),
m1.getFTotalWCount(), true);我还在日志中打印了"cerr“行，但我就是找不到为什么输出中没有这些文件。
jakub@jakub-virtual-machine:~/Master/giza-pp_test$ cat giza.log | tail -n 25
p0_count is 4.0073 and p1 is 5.99635; p0 is 0.400584 p1: 0.599416
Model4: TRAIN CROSS-ENTROPY 0.80096 PERPLEXITY 1.74226
Model4: (10) TRAIN VITERBI CROSS-ENTROPY 0.801289 PERPLEXITY 1.74266
Dumping alignment table (a) to file:out.a3.final
Dumping distortion table (d) to file:out.d3.final
Dumping nTable to: out.n3.final
Model4 Viterbi Iteration : 10 took: 0 seconds
H3333344444 Training Finished at: Fri Oct 23 16:24:44 2015
Entire Viterbi H3333344444 Training took: 0 seconds
==========================================================
writing Final tables to Disk
Writing PERPLEXITY report to: out.perp
Writing source vocabulary list to : out.trn.src.vcb
Writing source vocabulary list to : out.trn.trg.vcb
Writing source vocabulary list to : out.tst.src.vcb
Writing source vocabulary list to : out.tst.trg.vcb
writing decoder configuration file to out.Decoder.config
Entire Training took: 0 seconds
Program Finished at: Fri Oct 23 16:24:44 2015
==========================================================有没有人遇到过类似的问题？这是某种bug，还是我做错了什么？
编辑：
现在，我在Makefile中的CFLAGS中重新编译了整个GIZA++，但没有使用-DBINARY_SEARCH_FOR_TTABLE选项。并更改了脚本，使其不会生成并发文件并将其提供给GIZA++。在我重新运行脚本之后，输出确实包含out.actual.ti.final和out.ti.final。有人知道如何解释这种行为吗？我教过我使用coocurrence文件可以得到更好的对齐和概率估计，有什么需要吗？或者仅仅是为了提高性能？
回答 1
Stack Overflow用户
发布于 2015-12-06 02:04:50
我以前也遇到过同样的问题。我认为缺少的步骤在位于.\giza-pp\GIZA++-v2\的生成文件中，代之以行: CFLAGS_OPT = $(CFLAGS) -O3 -funroll-loops -DNDEBUG -DWORDINDEX_WITH_4_BYTE -DBINARY_SEARCH_FOR_TTABLE -DWORDINDEX_WITH_4_BYTE
命令行: CFLAGS_OPT = $(CFLAGS) -O3 -funroll-loops -DNDEBUG -DWORDINDEX_WITH_4_BYTE -DWORDINDEX_WITH_4_BYTE
检查这个推杆，祝你好运
https://stackoverflow.com/questions/33305090
复制相似问题
腾讯云开发者
