ACOPOST: User manual
Contents
1. PROMPT tbt r l train lex train rules lt test raw gt test tbt Transformation based Tagger c Ingo Schrder ingo nats informatik uni hamburg de done PROMPT evaluate pl test cooked test tbt 2061 sentences test tbt 34430 1244 96 513 5 13 tt2cooked pl 5 13 1 Purpose Convert a corpus in a format Bra97 used by the TnT tagger package Bra00 to a corpus in cooked format 5 13 2 Usage tt2cooked pl h lt in tt gt out cooked h display a short help text and exit 5 13 3 Example PROMPT tt2cooked pl lt negra tt gt negra cooked 396309 lines read PROMPT 5 14 wsj2cooked pl 5 14 1 Purpose Convert a corpus in Wall Street Journal format to cooked format 5 14 2 Usage wsj2cooked pl lt in wsj gt out cooked 15 5 14 3 Example PROMPT wsj2cooked pl lt corpus wsj gt negra cooked PROMPT References Bra97 Thorsten Brants The negra export format for annotated corpora version 3 1997 Bra00 Thorsten Brants TnT a statistical part of speech tagger In Proceedings of the Sixth Applied Natural Language Processing Conference ANLP 2000 Seattle WA USA 2000 BW98 Eric Brill and Jun Wu Classifier combination for improved lexical disambiguation In Proc Joint Conference COLING ACL 798 pages 191 195 Montral Canada 1998 DvdBW97 Walter Daelemans Antal van den Bosch and Ton Weijters Igtree Using trees for compression and classification in lazy learning algorithms Artificial Intelligence2. ref 0 t3 O tnt 1002 sentences 0 t3 16651 654 96 221 O tnt 16732 573 96 689 PROMPT 5 9 majority voter pl 5 9 1 Purpose Report how often different numbers of different taggers have tagged words correctly See author Sch02 This immediately tells one how efficient a parallel combination of different taggers can be Four numbers are given in each line The number of taggers that were correct the percentage of words the accumulated percentage of words and the mean ambiguity of tags if all emitted tags are counted 5 9 2 Usage majority voter pl h ref t1 t2 h display short help text and exit ref reference corpus in cooked format ti first tagged corpus in cooked format t2 second tagged corpus in cooked format 5 9 3 Example PROMPT majority voter pl O ref 0 t3 O tbt O et O met 2061 sentences 35674 words 4 92 928 92 928 0 937658 3 3 493 96 420 0 983041 2 1 343 97 763 1 010988 1 1 090 98 854 1 068313 0 1 146 100 000 PROMPT 5 10 met 5 10 1 Purpose Nothing yet 11 5 10 2 Usage met OPTIONS modelfile inputfile where modelfile is a trained or a new model file and inputfile is either a corpus in cooked format for training or in raw format for tagging OPTIONS can be one or more of the following bb beam factor default 1000 for viterbi search or n best width default 5 for n best search c command mode tag train or test d d dictionary file f threshold fo
3. to the right are not yet known CLASS relpos Use the ambiguity class at the relative position relpos as a criterion For example CLASS 1 considers the ambiguity class of the word to the right of the current word WORD Lrelpos Use the word form at the relative position relpos as a criterion Note that only frequent words see options r and w are used For rare words the artifical token RARE is substituted LETTER relpos index Use the letter at position index of the word at the relative position relpos as a criterion Negative values of index count from the end of the word backwards CAP Lrelpos Use the binary answer whether the word at the relative position relpos is capitatized as a criterion HYPHEN relpos Use the binary answer whether the word at the relative position relpos contains a hyphen as a criterion NUMBER repos Use the binary answer whether the word at the relative position relpos contains a digit as a criterion INTER relpos Use the binary answer whether the word at the relative position relpos contains an interpunctuation mark as a criterion The directory examples et contains example feature files 5 6 4 Example PROMPT PROMPT 5 7 5 7 1 et Purpose Assign tags to a natural language text in raw format using the example based paradigm Sch02 Section 5 4 Note that the learning phase is done by the Perl script cooked2wtree pl cf Section 5 6 5 7 2 Usage et OPTIONS know
4. 6 221 16651 654 0 accuracy B 96 689 16732 573 comp A B 22 783 505 654 comp B A 11 867 505 573 PROMPT 5 2 cooked2lex pl 5 2 1 Purpose Convert a corpus in cooked format to a lexicon 5 2 2 Usage cooked2lex pl h c lt in cooked gt out lex h display a short help text and exit c output deprecated word count after the word form cf Section 3 5 2 3 Example PROMPT cooked2lex pl lt negra cooked gt negra lex 20602 sentences 55 tags 51272 types 355096 tokens 1 49189 95 937 238545 67 178 2 1884 3 675 45586 12 838 3 164 0 320 46789 13 176 4 32 0 062 20090 5 658 5 1 0 002 2715 0 765 6 1 0 002 1363 0 384 7 1 0 002 8 0 002 Mean ambiguity A 1 611544 Entropy H p 4 273873 PROMPT 5 3 cooked2ngram pl 5 3 1 Purpose Convert a corpus in cooked format to a file containing counts for tag n grams 5 3 2 Usage cooked2ngram pl h lt in cooked gt out ngram h display a short help text and exit 5 3 3 Example PROMPT cooked2ngram pl lt corpus cooked gt corpus ngram 5 4 cooked2tt pl 5 4 1 Purpose Convert a corpus in cooked format to a corpus in the format Bra97 used by the TnT tagger package Bra00 5 4 2 Usage cooked2tt pl h lt in cooked gt out tt h display a short help text and exit 5 4 3 Example PROMPT cooked2tt pl lt negra cooked gt negra tt 20602 sentences PROMPT 5 5 cooked2wsj pl 5 5 1 Purpose Convert a corpus in cooked format to a corpu
5. ACOPOST User manual Contents Contents 1 Introduction 2 Installation 3 File formats 4 Tutorial Version 1 8 4 Ingo Schrder ixs users sourceforge net 5 Program references 5 1 complementary rate pl X ee 5 1 1 Purpose saaa aaa sas sososnsossnis is is nsnsi nsn is os nis it a 5 1 2 Usage 5 1 3 Example ___ _ _ ea ee 5 2 cooked2lex pl 5 2 1 Purpose c rs ss sas sos sssi isis nisnsnisn nsnis is os ons os a 5 2 2 Usage 5 2 3 Example 2 2 ee 5 3 cooked2ngram pl ooa ee 5 3 1 Purpose cia es ras sas sosisini nis is ninss ns nis is os ons s a 5 3 2 Usage 5 3 3 Example __ _ ee 5 4 cooked2tt pl 5 4 1 Purpose i ss sas sas is sssi is is ins nsn nis is os nis it a 5 4 2 Usage NN NN NN TD DBA DB RRR eH 5 5 5 6 5 7 5 8 5 9 5 10 5 11 5 4 3 Example 0 2 ee 7 cooked2wsj pl 0 ee 8 5 5 1 Purpose i as sss sas s sss oss is is nsnsi nsn ns i os nis nis a 8 5 5 2 Usage naaa aaa is sns nisi is nsn isis nis is ns nis nis nis 8 5 5 3 Example ____ _ a ee 8 cooked2wtree pl _ _ r ee 8 5 6 1 Purpose i as sss ss sosisini isis oisnsnsn nsn nis is os nis nis 8 5 6 2 Usage sis iss os isis isis is sns nis sns is nsn isis nis nis ns is i s 8 5 6 38 Features sas ss issenensis is is nsnsi nis nis is os nis nis a 9 5 6 4 Example __ __ _ 2 s
6. Review 11 407 423 1997 MSM93 M Marcus B Santorini and M Marcinkiewicz Building a large annotated corpus of English the Penn Treebank Computational Linguistics 19 2 1993 Sch02 Ingo Schrder A case study in part of speech tagging using the ICOPOST toolkit Tech nical report Computer Science University of Hamburg 2002 16
7. a ss ee 9 CIS S 9 5 7 1 Purpose i i s sss ss sos iss nsnsi is is ninss nis nis is is nis nit a 9 5 7 2 Usage i sis e 10 5 7 3 Example ___ _ ee s s sss ee 10 evaluate pl 0 ee 10 5 8 1 Purpose 2 ee 10 5 8 2 Usage 2 is sssi isis isis is sns nis sns nis nsn nis nsn nis is ns nis ni nes 10 5 83 Example ___ _ _ e ee 11 majority voter pl X T 2 a 2 a s azscz ee 11 5 9 1 Purpose i as sss ss sosisini is is ninss ns nis nis is is nis nis 11 5 9 2 Usage _ sis sassa iss ss is sns nis snn nsnsi is nis nis nis nis nis nis is 11 5 9 3 Example ___ 2 ss s sas ee 11 met i s sisi issenensis ee 11 5 10 1 Purpose r ca sr s ss sas sas sas sss sos nis ins nis nis os nis it ta 11 5 10 2 Usage a r s s rss s ss sis sss sss is sns nsn nis nis ns nis is is 12 5 10 8 Example ____ _ a ee 12 t3 cis sissnsisnisnis isis nis isian nis is nis nis nsn nis nis nis nis nis nis ns nis ns ns it sa 12 5 11 1 Purpose 2 ee 12 5 11 2 Usage ee 12 5 11 8 Example 0 ee 13 tht aoaaa ee 13 5 12 1 Purpose r ca sr s ss sas esas esas sss sss is nis os os nis nis e 13 5 12 2 Usage a ss css e 14 5 12 3 Templates ___ __ _ a ee sa s ee 14 5 12 4 Example ____ _ ee 15 5 13 tt2cooked pl ee 15 5 13 1 Purpose r ca sr s ss sas sas sas s
8. can be 12 t u v v x y z 5 11 3 PROMPT PROMPT smoothing parameters for transitional probabilities see author Sch02 Section 5 1 1 and author Bra00 for the default beam factor default 1000 states that are worse by this factor or more than the best state at this time point are discarded debug mode display short help and exit maximum suffix length for estimating output probability for unknown words default 10 mode of operation default 0 0 means tagging 1 testing quiet mode of operation rare word count default 1 for output probabilities theta for suffix backoff default SD of tag probabilities see author Sch02 Sec tion 5 1 1 and author Bra00 test mode reads cooked input use line buffered IO for input default block buffered on files verbosity default 1 case insensitive suffix tries default sensitive case insensitive when branching in suffix trie default sensitive use zero probability for unseen transition probabilities default 1 tags Example cooked2lex pl lt train cooked gt train lex cooked2ngram pl lt train cooked gt train ngram t3 train ngram train lex lt test raw gt test t3 O ms 1 O ms 1 Trigram POS Tagger c Ingo Schrder schroeder informatik uni hamburg de O ms 1 80 ms 1 model generated from 18541 sentences thereof 491 one word 80 ms 1 found 55623 uni 74164 bi and 92214 trigram counts for the bounda
9. he Perl programming language which you want to have installed anyway Find a convenient place in your directory tree and unzip the archive which unpacks into a new directory acopost 2 y z PROMPT gunzip c acopost 1 8 4 tar gz tar fxv acopost 1 8 4 acopost 1 8 4 src acopost 1 8 4 src Makefile acopost 1 8 4 src array c The fresh directory contains at least the following files and directories e Text file README with a short intro and latest changes e Directory bin which contains the Perl scripts and where the binaries are installed after compi lation e Directory src which contains the C files e Directory docs which contains the documentation this user guide and a technical report Sch02 e Directory examples which contains some example files To compile change to the src directory and type make If everything works out ok issue the command make install which installs the binaries into the directory bin Congratulations You re done If something goes wrong try to fix it by adpating the Makefile or the source code Don t forget to tell me about your problems so that I can provide a better solution with the next release You can now chose to add the bin directory as a full path to your PATH variable to move copy all binaries from the bin directory to a directory already in your PATH variable or simply decide to always use the full path to an ACOPOST program 3 File formats I tried to keep everything as
10. is s sis nsn isnis ons os nis it a 15 5 13 2 Usage a cia rss s ss sis rss sss is sns osit nis ns nis is nis 15 5 13 8 Example ____ _ a ee 15 5 14 wsj2cooked pl 2 ee 15 5 14 1 Purpose r ca sr s ss ss ease esas sissoo os sns is os nis it a 15 5 14 2 Usage 2 a e 15 5 14 3 Example __ _ _ _ ee 16 References 16 1 Introduction This document describes how to use the ACOPOST program suite ACOPOST is a collection of part of speech tagging algorithms each originating from a different ma chine learning paradigm t3 is a trigram tagger based on Markov models met is a maximum entropy inspired tagger tbt is an error driven learner of transformation rules et is an example based tagger An evaluation of the individual part of speech taggers and of novel combination techniques can be found in an accompanying technical report Sch02 2 Installation ACOPOST is available under the GNU public license from the project homepage hosted at http www sourceforge net See http www gnu org licenses gpl html ACOPOST comes as a gzipped tar archive of the source code named acopost 2 y z tar gz where x y z is the version number No pre compiiled binaries are available but don t worry Compiling is easy You only need a C compiler gcc is recommended and the make program which are both most probably already installed on your machine if you re using UNIX Some scripts use t
11. nder consideration and the conditions are prerequisites for the application of the rule All conditions must be fulfilled for a rule to trigger The following types of conditions are allowed tag relpos tag The current tag of the word at relative position relpos is tag bos relpos Begin of sentence marker at relative position relpos eos relpos End of sentence marker at relative position relpos wordL relpos word The word at relative position relpos is word rare relpos The word at relative position relpos is rare prefixLlength prefix The prefix of length length of the current word is prefix suffixLlength suffizx The suffix of length length of the current word is suffix cap Lrelpos mode The capitilization of the word at relative position relpos is as mode which can be no No character is capitilized some Some characters are capitilized all All characters are capitilized digit Lrelpos mode The word at relative position relpos contains digits according to mode which can be no some or all see cap The placeholders tag word prefix suffix and mode can also be the wildcard symbol in templates A typical rule template which takes the two preceding tags into account would then be tag 2 tag 1 The examples tbt directory contains example template files 14 5 12 4 Example PROMPT cat train rules rare 0 NN rare 0 digit 0 no ADJA rare 0 tag 0 NN cap 0 no VVPP rare 0 tag 0 ADJA suffix 0 t
12. nwtree unknownwtree lexiconfile in raw gt out cooked where knowntree is a weighted tree file generated by cooked2wtree pl cf Section 5 6 for known words unknowntree is a weighted tree file for unknown words and lexiconfile is a lexicon file generated by cooked2lex pl cf Section 5 2 If the input file in raw is omitted standard input is used OPTIONS can be v v verbosity default 1 5 7 3 Example PROMPT cooked2lex pl lt train cooked gt train lex PROMPT cooked2wtree pl a 3 known etf lt train cooked gt known wtree PROMPT cooked2wtree pl b 2 unknown etf e closed class tags lt train cooked gt unknown wtree PROMPT et known wtree unknown wtree train lex lt test raw gt test et O ms 1 O ms 1 Example based Tagger c Ingo Schrder schroeder informatik uni hamburg de O ms 1 2240 ms 1 read wtree with 156173 nodes from known wtree 3580 ms 1 read wtree with 116334 nodes from unknown wtree 3590 ms 1 done PROMPT evaluate pl test cooked test et 2060 sentences test et 33990 1434 95 952 L L L L L L 5 8 evaluate pl 5 8 1 Purpose Report tagging accuracy on sentence level for unknown known and all words 5 8 2 Usage evaluate pl h H 1 1 v ref t1 h display short help text and exit i use case insensitive lexicon 1 1 use lexicon 1 v be verbose ref reference corpus in cooked format t1 tagged corpus in cooked format 10 5 8 3 Example PROMPT evaluate pl O
13. pts which convert from and into different formats e g ws j2cooked pl cf Section 5 14 tt2cooked pl cf Section 5 13 and cooked2tt p1 cf Section 5 4 The individual taggers use additional data formats to store the model information These formats have been chosen to be human readable but completely understanding them requires deep insights into the tagging algorithms The formats of model file might change between releases The format of the lexicon files is also line based Each line lists the word form and the possible tags including the tag counts WORDFORM TAG1 TAGCOUNT1 TAG2 TAGCOUNT2 An older format allowed for an optional word count after the word form but since this information is redundant it is deprecated 4 Tutorial Nothing yet 5 Program references Note that not all programs in the bin directory are described here This may be the case due to one of the following reasons e The program is considered to be of marginal importance e It hasn t reached a stable state e It s obsolete 5 1 complementary rate pl 5 1 1 Purpose Report the complementary error rate BW98 of two versions of a tagged corpus 5 1 2 Usage complementary rate pl h ref ab h display short help text and exit ref reference corpus in cooked format a first tagged corpus in cooked format b second tagged corpus in cooked format 5 1 3 Example PROMPT acopost bin complementary rate pl O ref 0 t3 O tnt accuracy A 9
14. r feature count default 5 h display short help and exit maximum number of iterations default 100 training only m m probability threshold default 1 0 n use n best instead of viterbi p p UNIX priority class default 19 r r rare word threshold defualt 5 s case sensitive dictionary t t minimum accuracy improvement per iteration default 0 0 training only v v verbosity default 1 I P H 5 10 3 Example PROMPT met c test d train lex train model met lt test cooked O ms 1 running as test O ms 1 using test lex as dictionary file 1390 ms 1 read 54 tags 40690 predicates and 83343 features 2090 ms 1 read 45779 lexicon entries discarded 2237 entries 24620 ms 1 35674 35257 pos 417 neg words tagged accuracy 98 831 P 5 11 1 Purpose Assign tags to a natural language text in raw format using the Viterbi algorithm based on a hidden Markov model HMM The model information is extracted from a tag trigram file and a lexicon file Note that the learning phase is very easy for HMMs For that reason the training phase is done by the Perl script cooked2ngramn p1 cf Section 5 3 5 11 2 Usage t3 OPTIONS modelfile lexiconfile in raw gt out cooked where modelfile is a tag trigram file generated by cooked2ngram pl cf Section 5 3 and lexiconfile is a lexicon file generated by cooked2lex p1 cf Section 5 2 If the input file in raw is omitted stan dard input is used OPTIONS
15. ry tag 210 ms 1 computed smoothed transition probabilities 1940 ms 1 built suffix tries with 32602 lowercase and 74242 uppercase nodes 1970 ms 1 leaves single total LC 8628 20073 32603 2040 ms 1 leaves single total UC 18627 47180 74243 4420 ms 1 suffix probabilities smoothing done theta 7 489e 02 21690 ms 1 done PROMPT evaluate pl test cooked test t3 2061 sentences 5 12 5 12 1 test t3 34547 1127 96 841 tbt Purpose Nothing yet 13 5 12 2 Usage tbt OPTIONS rulefile inputfile i i maximum number of training iterations default unlimited training only 1 1 lexicon file default none m m minimum improvement per training iteration default 1 training only n n rare wore threshold default 0 o o mode of operation default 0 0 tagging 1 testing 2 training p p preload file default lexically most probable tag start from a different initial tagging r assume raw format for input default cooked format tagging only t t template file default none training only see below u u unknown word default tag default most probable tag from lexicon v v verbosity defualt 1 5 12 3 Templates Templates are patterns for rules The file format is line based i e one rule per line empty lines and everything after a hash sign is ignored The format for a rule or template is as follows TARGETTAG CONDITION1 CONDITION2 where TARGETTAG is the new tag for the word u
16. s in the format used by the Wall Street Journal corpus MSM93 5 5 2 Usage cooked2wsj pl h lt in cooked gt out wsj h display a short help text and exit 5 5 3 Example PROMPT PROMPT 5 6 cooked2wtree pl 5 6 1 Purpose Convert a corpus in cooked format to a weighted tree DvdBW97 Sch02 for use in example based disambiguation Warning the current implementation is far from efficient Training on the Wall Street Journal corpus requires large amounts of main memory Be careful 5 6 2 Usage cooked2wtree pl OPTIONS f file lt in cooked gt out wsj where f file is a feature file see below and OPTIONS can be one or more of a a is the minimal word count that a word must have to be considered default unlimited b b is the maximal word count that a word must have to be considered default unlimited d debug flag e e file with tags to be excluded default exclude none i i file with tags to be explicitely included default include all h display a short help text and exit r r rare word count threshold w w word rank threshold default 100 5 6 3 Features Features describe characteristics of tagging context that can be used for the tagging decision The following features are allowed TAGL relpos Include the tag at the relative position relpos as a criterion for the decision For example TAG 1 means the tag of the word immediately to the left Of course relpos must be negative since the tags
17. simple as possible in order to be able to use other tools on the corpora e g UNIX tools like grep sed wc etc or Perl Therefore I chose line based formats for the corpora i e each line of texts separated by the newline character n holds exactly one sentence The items 21 have not tried to compile ACOPOST on MS Windows but I am interested in reports from Windows users See http www perl org and http www perl com in a sentence are separated by one or more white space characters i e tabular Nt or space characters Punctuation marks should be separated from preceding words ACOPOST uses two file formats for text raw and cooked e Raw text follows the line based format described above but doesn t contain any additional in formation Here s an example from the Wall Street Journal corpus MSM93 The rest went to investors from France and Hong Kong e Cooked text contains the part of speech tags for the words The tag immediately follows the word and the two are separated by one or more white space characters i e in the same way ajacent words are separated Of course a line of cooked text must always contain an even number of items Here s the same example as above as cooked text The DT rest NN went VBD to TO investors NNS from IN France NNP and CC Hong NNP Kong NNP Note that the period functions as both a word and a tag symbol in the Wall Street Journal corpus The ACOPOST program suite contains Perl scri
Download Pdf Manuals
Related Search