A Low-budget Tagger for Old Czech Jirka Hana 1 Anna Feldman 2 - - PowerPoint PPT Presentation

▶
a low budget tagger for old czech
SMART_READER_LITE
LIVE PREVIEW

A Low-budget Tagger for Old Czech Jirka Hana 1 Anna Feldman 2 - - PowerPoint PPT Presentation

Mar 26, 2024 274 likes •696 views

A Low-budget Tagger for Old Czech Jirka Hana 1 Anna Feldman 2 Katsiaryna Aharodnik 2 1 Charles University, Prague 2 Montclair State University, NJ ACL 2011 LaTeCH Portland, OR, June 24, 2010 J. Hana et al. (Charles University & MSU) A

slide-1
SLIDE 1

A Low-budget Tagger for Old Czech

Jirka Hana1 Anna Feldman2 Katsiaryna Aharodnik2

1Charles University, Prague 2Montclair State University, NJ

ACL 2011 – LaTeCH Portland, OR, June 24, 2010

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 1 / 30

slide-2
SLIDE 2

Outline of the talk

1

Introduction

2

Czech

3

Corpora & Tagsets

4

Taggers Translation Model Resource-light Morphological Analysis Even Tagger Combining the Translation and Even Taggers

5

Conclusion

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 2 / 30

slide-3
SLIDE 3

Introduction

Outline of the talk

1

Introduction

2

Czech

3

Corpora & Tagsets

4

Taggers Translation Model Resource-light Morphological Analysis Even Tagger Combining the Translation and Even Taggers

5

Conclusion

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 3 / 30

slide-4
SLIDE 4

Introduction

Introduction

Creating morphosyntactic resources for Old Czech on the basis of Modern Czech

Two goals

1

Practical: Create morphologically annotated resources for Old Czech to investigate various morphosyntactic patterns underpinning the evolution of Czech

2

Theoretical: Test the resource-light cross-lingual method we have been developing on a source-target language pair divided by time

Difficulties

500+ years of language evolution at all layers, e.g., phonology, graphemics, syntax, vocabulary

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 4 / 30

slide-5
SLIDE 5

Introduction

Introduction

Creating morphosyntactic resources for Old Czech on the basis of Modern Czech

Two goals

1

Practical: Create morphologically annotated resources for Old Czech to investigate various morphosyntactic patterns underpinning the evolution of Czech

2

Theoretical: Test the resource-light cross-lingual method we have been developing on a source-target language pair divided by time

Difficulties

500+ years of language evolution at all layers, e.g., phonology, graphemics, syntax, vocabulary

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 4 / 30

slide-6
SLIDE 6

Introduction

Introduction

Creating morphosyntactic resources for Old Czech on the basis of Modern Czech

Two goals

1

Practical: Create morphologically annotated resources for Old Czech to investigate various morphosyntactic patterns underpinning the evolution of Czech

2

Theoretical: Test the resource-light cross-lingual method we have been developing on a source-target language pair divided by time

Difficulties

500+ years of language evolution at all layers, e.g., phonology, graphemics, syntax, vocabulary

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 4 / 30

slide-7
SLIDE 7

Czech

Outline of the talk

1

Introduction

2

Czech

3

Corpora & Tagsets

4

Taggers Translation Model Resource-light Morphological Analysis Even Tagger Combining the Translation and Even Taggers

5

Conclusion

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 5 / 30

slide-8
SLIDE 8

Czech

Czech

Basic info:

West Slavic language, significant influences from German, Latin and (in modern times) English, fusional (flective) language with rich morphology and, high degree of homonymy of endings

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 6 / 30

slide-9
SLIDE 9

Czech

Modern Czech

10M speakers Two variants with differences mainly in phonology, morphology, lexicon The official variant is based on the 19th-century resurrection of the 16th century Czech Writing system is mostly phonological.

Old Czech

1150-1500 AD No native speakers Amount of available texts limited (??10MW) Spelling not standardized

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 7 / 30

slide-10
SLIDE 10

Czech

Examples of sound/spelling changes from Old Czech to Modern Czech

change example ú > ou non-init. múka > mouka ‘flour’ sˇ e > se sˇ eno > seno ‘hay’ ó > uo > ˚ u kóˇ n > kuoˇ n > k˚ uˇ n ‘horse’ šˇ c > št’ šˇ cír > štír ‘scorpion’ ˇ cs > c ˇ cso > co ‘what’ (Mann 1977, Boris Leheˇ cka p.c.).

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 8 / 30

slide-11
SLIDE 11

Czech

Morphology

dual number virtually disappeared animacy distinction in masculine gender emerged many verbal forms disappeared (three simple past tenses, supinum), and some are archaic (verbal adverbs, plusquamperfectum). some forms have different meaning

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 9 / 30

slide-12
SLIDE 12

Czech

Old vs Modern Czech verbs

category Old Czech Modern Czech infinitive péc-i péc-t ‘bake’ present 1sg pek-u peˇ c-u 1du peˇ c-evˇ e – 1pl peˇ c-em(e/y) peˇ c-eme : imperfect 1sg peˇ c-iech – 1du peˇ c-iechovˇ e – 1pl peˇ c-iechom(e/y) – : imperative 2sg pec-i peˇ c 2du pec-ta – 2pl pec-te peˇ c-te : verbal noun peˇ c-enie peˇ c-ení

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 10 / 30

slide-13
SLIDE 13

Corpora & Tagsets

Outline of the talk

1

Introduction

2

Czech

3

Corpora & Tagsets

4

Taggers Translation Model Resource-light Morphological Analysis Even Tagger Combining the Translation and Even Taggers

5

Conclusion

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 11 / 30

slide-14
SLIDE 14

Corpora & Tagsets

Corpora needed

Annotated corpus of Modern Czech

PDT, 1.5M tokens. Daily newspapers, business and popular scientific magazines.

Plain corpus of Old Czech

STB; http://vokabular.ujc.cas.cz; 740K tokens. Much smaller than what we used before (e.g., 63M for Catalan). Chronicles, legends, poetry, fiction, letters, etc. Transliterated.

Annotated corpus of Old Czech – for testing

About 1000 words. Much less than we would wish for. Making a bigger one.

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 12 / 30

slide-15
SLIDE 15

Corpora & Tagsets

Corpora needed

Annotated corpus of Modern Czech

PDT, 1.5M tokens. Daily newspapers, business and popular scientific magazines.

Plain corpus of Old Czech

STB; http://vokabular.ujc.cas.cz; 740K tokens. Much smaller than what we used before (e.g., 63M for Catalan). Chronicles, legends, poetry, fiction, letters, etc. Transliterated.

Annotated corpus of Old Czech – for testing

About 1000 words. Much less than we would wish for. Making a bigger one.

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 12 / 30

slide-16
SLIDE 16

Corpora & Tagsets

Corpora needed

Annotated corpus of Modern Czech

PDT, 1.5M tokens. Daily newspapers, business and popular scientific magazines.

Plain corpus of Old Czech

STB; http://vokabular.ujc.cas.cz; 740K tokens. Much smaller than what we used before (e.g., 63M for Catalan). Chronicles, legends, poetry, fiction, letters, etc. Transliterated.

Annotated corpus of Old Czech – for testing

About 1000 words. Much less than we would wish for. Making a bigger one.

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 12 / 30

slide-17
SLIDE 17

Corpora & Tagsets

Tagset

Modern Czech

positional tagset (Hajiˇ c 2004) more than 4200 tags encodes categories like POS, detailed POS, gender, number, case, person, voice, etc.

Old Czech

based on the modern tagset roughly the same set of categories, but some values added (e.g. imperfect), some removed co-occurrence restrictions are different (e.g. dual number is not limited to few tags)

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 13 / 30

slide-18
SLIDE 18

Taggers

Outline of the talk

1

Introduction

2

Czech

3

Corpora & Tagsets

4

Taggers Translation Model Resource-light Morphological Analysis Even Tagger Combining the Translation and Even Taggers

5

Conclusion

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 14 / 30

slide-19
SLIDE 19

Taggers Translation Model

Modernizing OC and Aging MC

An idea:

◮ Translate an annotated MC corpus to OC; then train a tagger on

the result.

◮ Too costly and probably, not needed since we deal only with

morphology.

Another idea:

◮ Modify the MC corpus so that it looks more like the OC just in the

aspects relevant for morphological tagging.

◮ Still not easy (e.g. the opposite of what historical linguistics does)

One more idea:

◮ Age the MC corpus ◮ Modernize the OC corpus ◮ Train on the Aged MC, tag the Modernized OC

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 15 / 30

slide-20
SLIDE 20

Taggers Translation Model

Modernizing OC and Aging MC

An idea:

◮ Translate an annotated MC corpus to OC; then train a tagger on

the result.

◮ Too costly and probably, not needed since we deal only with

morphology.

Another idea:

◮ Modify the MC corpus so that it looks more like the OC just in the

aspects relevant for morphological tagging.

◮ Still not easy (e.g. the opposite of what historical linguistics does)

One more idea:

◮ Age the MC corpus ◮ Modernize the OC corpus ◮ Train on the Aged MC, tag the Modernized OC

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 15 / 30

slide-21
SLIDE 21

Taggers Translation Model

Modernizing OC and Aging MC

An idea:

◮ Translate an annotated MC corpus to OC; then train a tagger on

the result.

◮ Too costly and probably, not needed since we deal only with

morphology.

Another idea:

◮ Modify the MC corpus so that it looks more like the OC just in the

aspects relevant for morphological tagging.

◮ Still not easy (e.g. the opposite of what historical linguistics does)

One more idea:

◮ Age the MC corpus ◮ Modernize the OC corpus ◮ Train on the Aged MC, tag the Modernized OC

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 15 / 30

slide-22
SLIDE 22

Taggers Translation Model

Modernizing OC and Aging MC

An idea:

◮ Translate an annotated MC corpus to OC; then train a tagger on

the result.

◮ Too costly and probably, not needed since we deal only with

morphology.

Another idea:

◮ Modify the MC corpus so that it looks more like the OC just in the

aspects relevant for morphological tagging.

◮ Still not easy (e.g. the opposite of what historical linguistics does)

One more idea:

◮ Age the MC corpus ◮ Modernize the OC corpus ◮ Train on the Aged MC, tag the Modernized OC

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 15 / 30

slide-23
SLIDE 23

Taggers Translation Model

Modernizing OC and Aging MC

An idea:

◮ Translate an annotated MC corpus to OC; then train a tagger on

the result.

◮ Too costly and probably, not needed since we deal only with

morphology.

Another idea:

◮ Modify the MC corpus so that it looks more like the OC just in the

aspects relevant for morphological tagging.

◮ Still not easy (e.g. the opposite of what historical linguistics does)

One more idea:

◮ Age the MC corpus ◮ Modernize the OC corpus ◮ Train on the Aged MC, tag the Modernized OC

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 15 / 30

slide-24
SLIDE 24

Taggers Translation Model

Translation Tagger

STB

  • ld plain

O2M: form translation tag & form (back) translation STB' plain STB' tagged STB tagged tagging train Old Czech HMM tagger PDT corpus modern annotated M2O: tag & form translation PDT ' corpus annotated train HMM tagger 1 2 3 4 5 6

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 16 / 30

slide-25
SLIDE 25

Taggers Translation Model

Translation Model – Major POSs

All Full: 70.6 SubPOS 88.9 Nouns Full 63.1 SubPOS 99.3 Adjs Full: 60.3 SubPos 93.7 Verbs Full 47.8 SubPOS 62.2

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 17 / 30

slide-26
SLIDE 26

Taggers Translation Model

Translation Model – Individual Positions

Tags: 70.6 Position 0 (POS ): 91.5 Position 1 (SubPOS ): 88.9 Position 2 (Gender ): 87.4 Position 3 (Number ): 91.0 Position 4 (case ): 82.6 Position 5 (PossGen): 99.5 Position 6 (PossNr ): 99.5 Position 7 (person ): 93.2 Position 8 (tense ): 94.4 Position 9 (grade ): 98.0 Position 10 (negation): 94.4 Position 11 (voice ): 95.9

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 18 / 30

slide-27
SLIDE 27

Taggers Resource-light Morphological Analysis

Resource-light morphological analysis

Resource-light morphological analyzer (Hana 2008, Feldman & Hana 2010) Manually provided information:

◮ Direct analyses of frequent words ◮ Endings organized into paradigms

12h of language-specific work needed in total. Done by a non-linguist on the basis of (Važný 1964, Dostál 1967). A cascade of modules:

  • 1. Word list – 250 most frequent words with their analyses.
  • 2. Lexicon-based analyzer – the lexicon has been automatically

acquired from a plain corpus using the knowledge of manually provided information about paradigms.

  • 3a. Guesser – analyzes words based on their tails (string suffixes).
  • 3b. Modern Czech word list – a simple analyzer of Modern Czech;
  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 19 / 30

slide-28
SLIDE 28

Taggers Resource-light Morphological Analysis

Resource-light morphological analysis

Resource-light morphological analyzer (Hana 2008, Feldman & Hana 2010) Manually provided information:

◮ Direct analyses of frequent words ◮ Endings organized into paradigms

12h of language-specific work needed in total. Done by a non-linguist on the basis of (Važný 1964, Dostál 1967). A cascade of modules:

  • 1. Word list – 250 most frequent words with their analyses.
  • 2. Lexicon-based analyzer – the lexicon has been automatically

acquired from a plain corpus using the knowledge of manually provided information about paradigms.

  • 3a. Guesser – analyzes words based on their tails (string suffixes).
  • 3b. Modern Czech word list – a simple analyzer of Modern Czech;
  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 19 / 30

slide-29
SLIDE 29

Taggers Resource-light Morphological Analysis

Resource-light morphological analysis

Resource-light morphological analyzer (Hana 2008, Feldman & Hana 2010) Manually provided information:

◮ Direct analyses of frequent words ◮ Endings organized into paradigms

12h of language-specific work needed in total. Done by a non-linguist on the basis of (Važný 1964, Dostál 1967). A cascade of modules:

  • 1. Word list – 250 most frequent words with their analyses.
  • 2. Lexicon-based analyzer – the lexicon has been automatically

acquired from a plain corpus using the knowledge of manually provided information about paradigms.

  • 3a. Guesser – analyzes words based on their tails (string suffixes).
  • 3b. Modern Czech word list – a simple analyzer of Modern Czech;
  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 19 / 30

slide-30
SLIDE 30

Taggers Resource-light Morphological Analysis

Resource-light morphological analysis

Resource-light morphological analyzer (Hana 2008, Feldman & Hana 2010) Manually provided information:

◮ Direct analyses of frequent words ◮ Endings organized into paradigms

12h of language-specific work needed in total. Done by a non-linguist on the basis of (Važný 1964, Dostál 1967). A cascade of modules:

  • 1. Word list – 250 most frequent words with their analyses.
  • 2. Lexicon-based analyzer – the lexicon has been automatically

acquired from a plain corpus using the knowledge of manually provided information about paradigms.

  • 3a. Guesser – analyzes words based on their tails (string suffixes).
  • 3b. Modern Czech word list – a simple analyzer of Modern Czech;
  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 19 / 30

slide-31
SLIDE 31

Taggers Resource-light Morphological Analysis

Lexicon & leo no yes Recall Ambi Recall Ambi Overall 96.9 14.8 91.5 5.7 Nouns 99.9 26.1 83.9 10.1 Adjectives 96.8 26.5 96.8 8.8 Verbs 97.8 22.1 95.6 6.2

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 20 / 30

slide-32
SLIDE 32

Taggers Even Tagger

MA Based Even Tagger

Old Czech text MA analyzed Old Cz text tagged Old Cz text 3 tag translation 4 record of the

  • riginal tags

compiling tnt emissions 5 tnt tag back translation even OCz emissions Old Czech text tag translation 1 tag translation 2 Cz transitions Cz emissions MA Creation Frequent forms Lexicon + Paradigms Ending based Guesser Modern Czech Forms

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 21 / 30

slide-33
SLIDE 33

Taggers Even Tagger

Even Tagger on major POS categories

Transl All Full: 70.6 67.7 SubPOS 88.9 87.0 Nouns Full 63.1 44.3 SubPOS 99.3 88.6 Adjs Full: 60.3 50.8 SubPos 93.7 87.3 Verbs Full 47.8 74.4 SubPOS 62.2 78.9

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 22 / 30

slide-34
SLIDE 34

Taggers Even Tagger

Ending -e and noun cases in Old Czech

case form lemma gender gloss nom moˇ r-e moˇ re neuter sea gen

  • ráˇ

c-e

  • ráˇ

c masculine plowman dat vládyc-e vládyka masculine local ruler acc

  • ráˇ

c-e

  • ráˇ

c masculine plowman voc chlap-e chlap masculine guy loc vládyc-e vládyka masculine local ruler inst – –

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 23 / 30

slide-35
SLIDE 35

Taggers Even Tagger

Old Czech verbs

category Old Czech Modern Czech infinitive péc-i péc-t ‘bake’ present 1sg pek-u peˇ c-u 1du peˇ c-evˇ e – 1pl peˇ c-em(e/y) peˇ c-eme : imperfect 1sg peˇ c-iech – 1du peˇ c-iechovˇ e – 1pl peˇ c-iechom(e/y) – : imperative 2sg pec-i peˇ c 2du pec-ta – 2pl pec-te peˇ c-te : verbal noun peˇ c-enie peˇ c-ení

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 24 / 30

slide-36
SLIDE 36

Taggers Combining the Translation and Even Taggers

The Even model clearly performs better on the verbs (and pronouns, conjuctions, ...), The Translation model predicts other categories much better. Use Even for verbs etc, Translation for the rest.

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 25 / 30

slide-37
SLIDE 37

Taggers Combining the Translation and Even Taggers

Even Tagger on major POS categories

Transl All Full: 70.6 74.1 SubPOS 88.9 90.6 Nouns Full 63.1 57.0 SubPOS 99.3 91.3 Adjs Full: 60.3 60.3 SubPos 93.7 93.7 Verbs Full 47.8 80.0 SubPOS 62.2 86.7

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 26 / 30

slide-38
SLIDE 38

Taggers Combining the Translation and Even Taggers

Combined tagger on individual positions

Full tags: 74.1 Position 0 (POS ): 93.0 Position 1 (SubPOS ): 90.6 Position 2 (Gender ): 89.6 Position 3 (Number ): 92.5 Position 4 (case ): 83.6 Position 5 (PossGen): 99.5 Position 6 (PossNr ): 94.9 Position 7 (person ): 94.9 Position 8 (tense ): 95.6 Position 9 (grade ): 98.6 Position 10 (negation): 96.1 Position 11 (voice ): 96.4

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 27 / 30

slide-39
SLIDE 39

Conclusion

Outline of the talk

1

Introduction

2

Czech

3

Corpora & Tagsets

4

Taggers Translation Model Resource-light Morphological Analysis Even Tagger Combining the Translation and Even Taggers

5

Conclusion

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 28 / 30

slide-40
SLIDE 40

Conclusion

Conclusion

Traditional statistical taggers rely on large amounts of training data – There is no realistic prospect of annotation for Old Czech. Old Czech is an ideal candidate for testing our resource-light method – no native speakers, limited corpora and lexicons, limited funding Challenging: Old Czech and Modern Czech departed significantly over the 500+ years; Old Czech and Modern Czech corpora belong to different genres. Results: 74% accuracy on the whole tag, 90+% on detailed POS.

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 29 / 30

slide-41
SLIDE 41

Acknowledgments

Thanks to

the Grant Agency Czech Republic (project ID: P406/10/P328) the U.S. NSF grants #0916280, #1033275, and #1048406. Alena M. ˇ Cerná and Boris Leheˇ cka for annotating the testing corpus and for answering questions about Old Czech. Institute of Czech Language of the Czech Academy of Sciences for the plain text corpus of Old Czech Anonymous reviewers for their insightful comments

  • J. Hana et al. (Charles University & MSU)

A Low-budget Tagger for Old Czech ACL 2011 – Latech 30 / 30