Hectoralos/GSOC 2020 work plan control

From Apertium
Jump to navigation Jump to search
Week Dates Goals Fulfilled
Bidix
(excluding
proper names)
Coverage WER Testvoc
---
Manual disamb. of frp texts
Frp monodix
(excl.
proper names)
Bidix
(excl.
proper names)
Non-WP
coverage
(%)[1]
WP
coverage
(%)[2]
WER
(%)
Testvoc
(clean %)
---
Manual disamb.
(words)
1 1 June -
7 June
7,006 1,213 fra > frp
64.5
frp > fra
74.7
fra > frp
61.3
frp > fra
47.9
2 8 June -
14 June
~1,500 13,699 6,739 fra > frp
77.6
frp > fra
82.5
fra > frp
74.1
frp > fra
57.9
3 15 June -
21 June
~3,000 14,863 8,922 fra > frp
-
frp > fra
86.3
fra > frp
77.7
frp > fra
61.0
4 22 June -
28 June
~4,500 15,805 10,747 fra > frp
85.2
frp > fra
90.7
fra > frp
81.8
frp > fra
72.0
5 29 June -
5 July
~6,000 fra > frp >80%
frp > fra >85%
16,746 13,637 fra > frp
87.6
frp > fra
92.6
fra > frp
84.3
frp > fra
75.3
6 6 July -
12 July
~7,500 17,392[3] 18,419
fra > frp
14,639
frp > fra
16,743
fra > frp
92.1
frp > fra
95.3
fra > frp
89.2
frp > fra
78.0
7 7 July -
19 July
~8,500 18,160 19,401
fra > frp
15,552
frp > fra
17,498
fra > frp
93.2
frp > fra
95.4
fra > frp
90.2
frp > fra
78.5
8 20 July -
26 July
~9,500 Disamb. of frp texts 19,411 20,844
fra > frp
16,915
frp > fra
18,744
fra > frp
94.5
frp > fra
95.6
fra > frp
91.6
frp > fra
80.1
0
9 27 July -
2 August
~10,500 fra-frp ~89%
frp > fra ~92%
fra-frp <25% Disamb. of frp texts 19,813 21,465
fra > frp
17,425
frp > fra
19,216
fra > frp
94.8
frp > fra
95.8
fra > frp
91.9
frp > fra
80.8
fra > frp
5.5[4]

frp > fra
10.7[5]

0[6]
10 3 August -
9 August
~11,500 21,193 23,086
fra > frp
18,902
frp > fra
20,502
fra > frp
95.3
frp > fra
95.9
fra > frp
92.6
frp > fra
81.2
11 10 August -
16 August
~12,500 Testvoc: closed categories, vblex 21,775 23,718
fra > frp
19,472
frp > fra
21,061
fra > frp
95.6
frp > fra
95.9
fra > frp
92.6
frp > fra
81.3
non-verbs:
fra > frp
100%
frp > fra
100%
12 17 August -
23 August
~12,750 Testvoc: adj, adv 22,407 24,459
fra > frp
20,142
frp > fra
21,694
fra > frp
95.7
frp > fra
96.0
fra > frp
92.7
frp > fra
81.4
verbs:
fra > frp
99,99%
frp > fra
100%
13 24 August -
30 August
~13,000 fra > frp ~90.0%
frp > fra ~93.0%
fra > frp <20%
frp > fra <25%
Testvoc: n 22,667 24,775
fra > frp
20,423
frp > fra
21,964
fra > frp
95.8
frp > fra
96.0
fra > frp
92.8
frp > fra
81.5
fra > frp
5.7[7]
frp > fra
15.5[8]
fra > frp
100%
frp > fra
100%


See also[edit]

Work plan in the original proposal

Notes[edit]

  1. The Arpitan non-Wikipedia corpus contains a few texts in ORB. At the beginning of the project it had 81,670 words. 80% were written by Dominique Stich, and the rest were several sociopolitical texts.
  2. The Arpitan Wikipedia corpus contains all the articles written in ORB, except the 366 for every day of year (2375 articles, 347,101 words). My impression is that 75+% of the content is made by bots, so it is very repetitive and little representative.
  3. + massive inclusion of np.ant and np.cog (thus the jump in the coverage)
  4. The test was conducted using three randomly selected texts (total: ~1000 words): two "good" Wikipedia articles (In My Life and Niger (cheval)) and the most outstanding article on lemonde.fr at the time of the test (Barbara Pompili : « C’est le bon moment pour gagner la bataille écologique »). The results are surprisingly good, even though the translator still needs work. The reasons for this may be several: the articles may have been easy; the revision (by a native speaker) may have been too acquiescent; the close proximity of the languages and the great variety of Arpitan may make it more difficult to consider the translator's choice wrong than in more distant languages with well-established standards; and, of course, it is a huge advantage to have been able to use an extensive electronic dictionary and to have the help of two great language experts who are very committed to the project.

    Test file: '200727_fra-frp.ini.txt'
    Reference file '200727_fra-frp.fin.txt'

    Statistics about input files

    Number of words in reference: 1180
    Number of words in test: 1180
    Number of unknown words (marked with a star) in test: 78
    Percentage of unknown words: 6.61 %

    Results when removing unknown-word marks (stars)


    Edit distance: 65
    Word error rate (WER): 5.51 %
    Number of position-independent correct words: 1122
    Position-independent word error rate (PER): 4.92 %

    Results when unknown-word marks (stars) are not removed


    Edit distance: 125
    Word Error Rate (WER): 10.59 %
    Number of position-independent correct words: 1062
    Position-independent word error rate (PER): 10.00 %

    Statistics about the translation of unknown words


    Number of unknown words which were free rides: 60
    Percentage of unknown words that were free rides: 76.92 %

  5. No he trobat (encara) la manera de fer una prova realment aleatòria de textos. No conec cap revista electrònica en arpità a la xarxa i la Viquipèdia arpitana està escrita amb diferents normes i, a més, acostuma a tenir articles que són simples plantilles. He agafat tres articles (France, Rimbaud i Toquio, ~ 1000 paraules), sobre els quals no havia treballat abans, i he obtingut uns resultats massa bons. És inimaginable que això passi en textos reals, especialment quan s'escriu amb una acceptació molt àmplia de les normes per a adaptar-les a la parla local (el que en diuen "grafia estreta").

    Test file: '200727_frp-fra.ini.txt'
    Reference file '200727_frp-fra.fin.txt'

    Statistics about input files

    Number of words in reference: 932
    Number of words in test: 939
    Number of unknown words (marked with a star) in test: 57
    Percentage of unknown words: 6.07 %

    Results when removing unknown-word marks (stars)


    Edit distance: 100
    Word error rate (WER): 10.73 %
    Number of position-independent correct words: 848
    Position-independent word error rate (PER): 9.76 %

    Results when unknown-word marks (stars) are not removed


    Edit distance: 115
    Word Error Rate (WER): 12.34 %
    Number of position-independent correct words: 833
    Position-independent word error rate (PER): 11.37 %

    Statistics about the translation of unknown words


    Number of unknown words which were free rides: 15
    Percentage of unknown words that were free rides: 26.32 %

  6. Disambiguation is working pretty well, using the French prob-file and ad-hoc CG rules , so I used the time for doing other things
  7. The test was conducted using four randomly selected texts (total: ~1600 words): three "good" Wikipedia articles (Cheval au Togo, Élisabeth de Bavière and Nyctalope) and the most outstanding article on lemonde.fr at the time of the test (Pour Emmanuel Macron, un été de crises diplomatiques).

    Test file: '200820_fra-frp.ini.txt'
    Reference file '200820_fra-frp.fin.txt'

    Statistics about input files

    Number of words in reference: 1635
    Number of words in test: 1629
    Number of unknown words (marked with a star) in test: 61
    Percentage of unknown words: 3.74 %

    Results when removing unknown-word marks (stars)


    Edit distance: 93
    Word error rate (WER): 5.69 %
    Number of position-independent correct words: 1548
    Position-independent word error rate (PER): 5.32 %

    Results when unknown-word marks (stars) are not removed


    Edit distance: 113
    Word Error Rate (WER): 6.91 %
    Number of position-independent correct words: 1528
    Position-independent word error rate (PER): 6.54 %

    Statistics about the translation of unknown words


    Number of unknown words which were free rides: 20
    Percentage of unknown words that were free rides: 32.79 %

  8. The test was conducted using randomly selected texts (total: ~1700 words): 30 (!) randomly selected Wikipedia articles (200826_WP.fra.txt -- there are no labelled "good articles" in the Arpitan Wikipedia and most of them have 1-3 sentences), adding up 1063 words, and the first three pages of the "manual" of the site arpitan.eu (200826_Manuel.frp.txt), 636 words.

    Test file: '200826_WP_frp-fra.ini.txt'
    Reference file '200826_WP_frp-fra.fin.txt'

    Statistics about input files

    Number of words in reference: 1699
    Number of words in test: 1703
    Number of unknown words (marked with a star) in test: 269
    Percentage of unknown words: 15.80 %

    Results when removing unknown-word marks (stars)


    Edit distance: 264
    Word error rate (WER): 15.54 %
    Number of position-independent correct words: 1480
    Position-independent word error rate (PER): 13.13 %

    Results when unknown-word marks (stars) are not removed


    Edit distance: 372
    Word Error Rate (WER): 21.90 %
    Number of position-independent correct words: 1373
    Position-independent word error rate (PER): 19.42 %

    Statistics about the translation of unknown words


    Number of unknown words which were free rides: 108
    Percentage of unknown words that were free rides: 40.15 %

    There is a noticeable difference in the results in the Wikipedia texts (WER: 13.4%) and the manual (WER: 19.0%). It seems significant that the author of the manual is from the Aosta Valley and uses constructions that are less similar to the French ones (which does not mean at all that his Arpitan is worse than any other variety, maybe the opposite could be sustained).