- 1 https://www.sketchengine.eu/eurlex-corpus/
1Parallel text corpora have been firmly becoming an essential tool for natural language processing and linguistic studies in the domain of contrastive analysis, translation studies and lexicology (Borin 2002; Hansen-Schirra, Neumann & Čulo 2017; Doval & Sánchez Nieto 2019) Such corpora range from small-sized single-author or even single-text collections (Buk 2012; Sitchinava 2016) to large scale ones, like EUR-Lex Corpus1 based on the European Union legislation and other documents (Baisa et al. 2016). Resources for African languages remain under-represented and mostly focused on Swahili (De Pauw, Wagacha & de Schryver 2011; Wójtowicz 2018), Amharic (Rychlý & Suchomel 2016; Woldeyohannis, Besacier & Meshesha 2018) or languages of South Africa (Wallmach 2000; Moropa 2007). The parallel Bamana–French corpus, which is a part of a larger project, the Bamana Reference Corpus (BRC, see Vydrin 2013; Vydrin et al. 2011–2019) is the only example of the Mande languages.
2In the present work, six Bamana tales recorded by Umaru Ɲanankɔrɔ Jara (Oumar Nianankoro Diarra) are studied. Analysis is made using texts from the abovementioned Bamana–French parallel corpus. Distributions of parts of speech are obtained for both Bamana originals and French translations.
3The following texts have been analyzed:
-
“Dununba kumata” (“Le tam-tam qui parle” = “The talking tom-tom”)
-
“Juguya sara” (“Le prix de la méchanceté” = “The price of wickedness”)
-
“Juman nɔrɔla farakolo la” (“Diouman s’est collée à une pierre” = “Diouman stuck to a stone”)
-
“Ntalen” (“Ntalen” [Parabole : araignée = Parable: spider])
-
“Sigidankelen ka labanko juguya” (“La fin tragique de Sigidankelen” = “The tragic end of Sigidankelen”)
-
“Warabilenkɔrɔ ka walijuya” (“La sainteté du vieux singe rouge” = “The holiness of the old red monkey”)
4The first two texts were published in a children’s book entitled Dununba kumata : Mali nsiirinw (Diarra & Fenayon 2011a; Diarra & Fenayon 2011b), see Figure 1. Apart from this Bamana version, a French translation of the book (Le tam-tam qui parle : contes du Mali, translated by Umaru Ɲanankɔrɔ Jara and Antoine Fenayon) as well as a German one (Die sprechende Trommel: Geschichten aus Mali, translated by Tim Hentschel) also appeared in 2011. Four other tales were provided by Umaru Jara himself as handwritten notebooks.
Figure 1. Book cover of Dununba kumata : Mali nsiirinw (Diarra & Fenayon 2011a)
Image source: http://donniyakadi.over-blog.com
5The rest of the paper is organized as follows: Section 2 discusses details of autosemantic parts of speech (PoS) as well as PoS-tagging and lemmatization issues; Section 3 contains results about frequency data in the analyzed Bamana and French texts; Section 4 discusses some peculiarities of adjective functioning; Section 5 briefly describes an application of the theory of complex networks. Finally, conclusions are drawn in Section 6.
6The paper analyzes the distribution of autosemantic parts of speech in the texts of the tales. The term ‘autosemantic’ refers to meaningful parts of speech (also known as content words), like nouns, verbs, adjectives, adverbs, etc. (Popescu, Altmann & Köhler 2010). These are contrasted with synsemantic (auxiliary) PoS (also known as function words), like particles, conjunctions, prepositions, etc. As there are no strict approaches to defining a particular PoS across languages, especially when dealing with languages of different families, it is worth discussing briefly which parts of speech are considered autosemantic in this work for the two languages, Bamana and French.
7The problem of parts of speech in Bamana has been discussed in a number of works (see especially Vydrine 1999 and references therein). Using different approaches, the authors mostly agree on the core set of nominals, verbs, and adjectives (Creissels 1983; Kastenholz 1998; Dumestre 2003), even though their definitions and the respective PoS-attributions do not necessarily coincide. Bamana is also sometimes described as a language with flexible word classes (Rijkhoff & van Lier 2013). In the present work, I mostly adhere to the definitions of the parts of speech based on morphosyntactic criteria as described by Vydrin (2017a) and applied in the Bamana Reference Corpus.
8For Bamana, texts from the tagged and disambiguated part of the BRC are used. The tools for building this and related corpora are described in detail by Maslinsky (2014). With PoS tags at hand, the following PoS are considered autosemantic: adjective, adverb (including preverbial), copula, determinative, noun, numeral, participle, pronoun (personal and non-personal), qualitative verb, and verb. Copulas are treated as autosemantic due to their syntactic role close to that of verbs. A similar syntactic criterion is applied to determinatives behaving like adverbs and to pronouns, which can substitute nouns (or adjectives in certain contexts).
9French texts were lemmatized using the TreeTagger software (Schmid n. d.) yielding a set of tags (Stein 2003) corresponding to the following PoS treated as autosemantic: adjective, adverb, noun (including a separate NAM tag for proper names), numeral, pronoun (personal, possessive, etc.), and verb.
10The autosemantic parts of speech in both languages and the respective tags are summarized in Table 1 for convenience.
11Lemmatization in the Bamana texts was performed by cutting affixes corresponding to flexion2, namely: verbal progressive suffix ‑la/‑na (glossed as PROG), non-productive plural marker ‑lu/‑nu (PL2), perfective intransitive marker ‑ra/‑la/‑na (PFV.INTR), optative marker ‑ra/‑la/‑na (OPT2), and plural marker ‑w (PL). Some examples are as follows:
-
sèginna ‘revenir.PROG’ is lemmatized as sègin ‘revenir’;
-
mínnu ‘REL.PL2’ is lemmatized as mîn ‘REL’;
-
táara ‘aller.PFV.INTR’ is lemmatized as táa ‘aller’;
-
mɔ̀gɔw ‘homme.PL’ is lemmatized as mɔ̀gɔ ‘homme’.
12No optative morphemes have been attested in the analyzed texts.
13Here and below, glosses are given in French as they appear in the BRC. This facilitates comparisons with the French translations in the parallel texts. The free French translations taken from the French part of the parallel corpus are followed by ther English equivalents.
Table 1. Autosemantic parts of speech and respective tags
Part of speech
|
Bamadaba tags
|
French TreeTagger tags
|
noun
|
n
|
NOM, NAM
|
verb
|
v, ptcp
|
VER
|
qualitative verb
|
vq
|
—
|
copula
|
cop
|
—
|
adjective
|
adj
|
ADJ
|
determinative
|
dtm
|
—
|
adverb
|
adv
|
ADV
|
numeral
|
num
|
NUM
|
pronoun
|
prn, pers
|
PRO (including PRO:DEM, PRO:IND, PRO:PER, etc.), DET:POS
|
14Bamana lemmas obtained from the corpus underwent some normalization. First of all, contracted forms resulting from the vowel elision were lemmatized as full ones, e. g., copulas y’ ‘être’ as yé ‘être’, t’ ‘COP.NEG’ as tɛ́ ‘COP.NEG’, d’ ‘donner’ as dí ‘donner’, f’ ‘dire’ as fɔ́ ‘dire’, k’ ‘faire’ as kɛ́ ‘faire’, predicative markers k’ ‘INF’ as kà: ‘INF’, m’ ‘PFV.NEG’ as ma ‘PFV.NEG’, etc. Next, dialectal forms were replaced with primary ones according to the Bamadaba dictionary (Bailleul et al. 2011–2020), e. g., búbagatoo ‘termitière’ → búbaganton ‘termitière’, dímin ‘faire.souffrir’ → dími ‘faire.souffrir’, tága ‘aller’ → táa ‘aller’, tágama ‘voyage’ → táama ‘voyage’, etc. Finally, a few typos, mostly resulting from incorrect accent placement, were corrected.
15The lemmatized French texts were manually post-processed to remove ambiguities and make some corrections. In particular, the two most frequent ambiguous lemmatizations were (variants are separated by a vertical line “|”)
16suis VER suivre|être
17fils NOM fil|fils
18In both cases, only the second variant was found in the analyzed texts.
19The most frequent incorrect lemmatization was
20nouvelle(s) ADJ nouveau
21instead of
22nouvelle(s) NOM nouvelle
23There is also another problem, which cannot be solved automatically, namely, the lemmatization of the French ‘un/une’. It is not always clear whether such a word should be considered an indefinite article or a numeral. This problem is known to occur in the tagging of texts in Romance languages. Sometimes, a portmanteau tag is used, e.g., \ARTi:NUMc in the Portuguese corpus (Bacelar do Nascimento et al. 2005). In the collocation ‘une fois’, which is very frequent in the text of tales, the tag corresponding to an indefinite article is used for ‘une’ in an example quoted by Salamanca (2019). In the present work, only those instances of ‘un/une’ are tagged as numerals where the cardinality is clear, for instance, ‘Une des femmes…’ or ‘Un mois passe, deux mois, trois mois, quatre mois, cinq mois…’.
24To facilitate comparisons with the Bamana texts, all French personal pronoun lemmas were replaced with a person-number gloss, e.g.:
je PRO:PER 1SG
me PRO:PER 1SG
moi PRO:PER 1SG
25or
ils PRO:PER 3PL
eux PRO:PER 3PL
elles PRO:PER 3PL
26Table 2 shows statistics on parts of speech in the analyzed texts. To clarify the terms used below, consider the following examples. The number of tokens is the total number of running words, while the number of types is the number of different words (lemmas) in a given text. For instance, the sentence
(1a)
|
À
|
yɛ́lɛla
|
kà
|
yɛ́lɛ ,
|
fɔ́
|
kà
|
à
|
ɲɛ́ji
|
bɔ́ .
|
|
pers
|
v
|
pm
|
v
|
conj
|
pm
|
pers
|
n
|
v
|
|
3SG
|
rire.PFV.INTR
|
INF
|
rire
|
jusqu'à
|
INF
|
3SG
|
larme
|
sortir
|
|
‘Il rit beaucoup, il rit tellement qu’il en pleura.’ = ‘He laughed a lot, he laughed so much that he cried.’ [“Juguya…”]
|
27contains nine tokens, of which six are autosemantic PoS (given in boldface). The number of lemma types is six (à, yɛ́lɛ, kà, fɔ́, ɲɛ́ji, bɔ́), of those four are autosemantic (à, yɛ́lɛ, ɲɛ́ji, bɔ́). Note that the verb yɛ́lɛla ‘rire.PFV.INTR’ was lemmatized as yɛ́lɛ ‘rire’.
28Another example
(1b)
|
“Jalakɔro”
|
yé
|
dùgu
|
yé
|
dùgu
|
bèlebele .
|
|
n.prop
|
cop
|
n
|
pp
|
n
|
adj
|
|
TOP
|
EQU
|
terre
|
PP
|
terre
|
gros
|
|
‘Dialakoro est un village, un bien gros village.’ = ‘Dialakoro is a village, a pretty big village.’ [“Juman…”]
|
29contains six tokens (including five autosemantic) and five types, of which four are autosemantic. Note that the two occurrences of yé are counted as different instances (copula and postposition).
Table 2. Data about the number of tokens and types in the Bamana and French versions of tales
|
Dununba…
|
Juguya…
|
Juman…
|
Ntalen
|
Sigidankelen…
|
Warabilenkɔrɔ…
|
Whole collection
|
|
bam
|
fra
|
bam
|
fra
|
bam
|
fra
|
bam
|
fra
|
bam
|
fra
|
bam
|
fra
|
bam
|
fra
|
All lemma tokens
|
1491
|
1633
|
656
|
769
|
723
|
968
|
1723
|
2007
|
984
|
1202
|
1643
|
2017
|
7218
|
8596
|
Autosemantic lemma tokens
|
1055
|
1145
|
491
|
580
|
524
|
665
|
1264
|
1486
|
698
|
863
|
1180
|
1460
|
5212
|
6199
|
All lemma types
|
396
|
461
|
239
|
266
|
283
|
296
|
461
|
472
|
361
|
383
|
454
|
484
|
1178
|
1322
|
Autosemantic lemma types
|
335
|
426
|
195
|
233
|
235
|
268
|
396
|
437
|
307
|
351
|
386
|
448
|
1079
|
1270
|
30From Table 2 one can see that autosemantic parts of speech account for 69–75% of all words used in a text while they are constitute 82–93% of the vocabulary (list of types). Details about their distribution are presented in the next section.
31Frequency lists of lemmas in the Bamana and French texts were compiled and are available from the author upon request. The numbers and percentages of autosemantic parts of speech in both text and vocabulary are shown in Tables 3–6.
Table 3. Distributions of autosemantic parts of speech in Bamana tales (types)
|
Dununba…
|
Juguya…
|
Juman…
|
Ntalen
|
Sigidankelen…
|
Warabilenkɔrɔ…
|
Whole collection
|
verb
|
110
|
34,0%
|
57
|
29,8%
|
60
|
26,8%
|
106
|
28,1%
|
93
|
32,0%
|
102
|
28,2%
|
304
|
28,2%
|
noun
|
157
|
48,5%
|
87
|
45,5%
|
116
|
51,8%
|
205
|
54,4%
|
150
|
51,5%
|
200
|
55,2%
|
658
|
61,0%
|
adj
|
7
|
2,2%
|
5
|
2,6%
|
10
|
4,5%
|
14
|
3,7%
|
7
|
2,4%
|
13
|
3,6%
|
32
|
3,0%
|
adv
|
17
|
5,2%
|
10
|
5,2%
|
11
|
4,9%
|
12
|
3,2%
|
10
|
3,4%
|
11
|
3,0%
|
37
|
3,4%
|
prn
|
17
|
5,2%
|
14
|
7,3%
|
14
|
6,3%
|
19
|
5,0%
|
16
|
5,5%
|
19
|
5,2%
|
23
|
2,1%
|
num
|
2
|
0,6%
|
5
|
2,6%
|
2
|
0,9%
|
7
|
1,9%
|
1
|
0,3%
|
4
|
1,1%
|
8
|
0,7%
|
dtm
|
8
|
2,5%
|
8
|
4,2%
|
6
|
2,7%
|
9
|
2,4%
|
9
|
3,1%
|
8
|
2,2%
|
11
|
1,0%
|
cop
|
6
|
1,9%
|
5
|
2,6%
|
5
|
2,2%
|
5
|
1,3%
|
5
|
1,7%
|
5
|
1,4%
|
6
|
0,6%
|
|
324
|
100%
|
191
|
100%
|
224
|
100%
|
377
|
100%
|
291
|
100%
|
362
|
100%
|
1079
|
100%
|
32On average, nouns account for around 50% of the vocabulary and verbs account for about another 30%. Note that the proportion of nouns increases to over 60% when the whole collection is analyzed. The reason is quite clear: in every new text, new objects and concepts are more likely to appear than words belonging to other parts of speech.
33The category of verbs includes participles as well as qualitative verbs. The latter are not numerous: only 24 occurrences together in all six texts (seven in “Dununba…”, two in “Juguya…”, two in “Juman…”, seven in “Ntalen”, three in “Sigidankelen…”, and three “Warabilenkɔrɔ…”). Of those, dí ‘[être] agréable’ is found in all but one text (five occurrences), ɲì ‘bon’ is found in three texts (seven occurrences), and kán ‘égal’ is found in two texts (four occurrences). Depending on the approach used, qualitative verbs can be counted together with adjectives, so the above data allow for simple recalculations if required.
Table 4. Distributions of autosemantic parts of speech in Bamana tales (tokens)
|
Dununba…
|
Juguya…
|
Juman…
|
Ntalen
|
Sigidankelen…
|
Warabilenkɔrɔ…
|
Whole collection
|
verb
|
271
|
25,7%
|
108
|
22,0%
|
115
|
21,9%
|
276
|
21,8%
|
163
|
23,3%
|
276
|
23,4%
|
1208
|
23,2%
|
noun
|
383
|
36,3%
|
132
|
26,9%
|
218
|
41,6%
|
510
|
40,3%
|
252
|
36,0%
|
440
|
37,3%
|
1935
|
37,1%
|
adj
|
12
|
1,1%
|
6
|
1,2%
|
13
|
2,5%
|
17
|
1,3%
|
11
|
1,6%
|
15
|
1,3%
|
74
|
1,4%
|
adv
|
28
|
2,7%
|
16
|
3,3%
|
13
|
2,5%
|
21
|
1,7%
|
20
|
2,9%
|
18
|
1,5%
|
116
|
2,2%
|
prn
|
255
|
24,2%
|
160
|
32,6%
|
102
|
19,5%
|
283
|
22,4%
|
169
|
24,1%
|
299
|
25,4%
|
1268
|
24,3%
|
num
|
9
|
0,9%
|
9
|
1,8%
|
7
|
1,3%
|
28
|
2,2%
|
3
|
0,4%
|
10
|
0,8%
|
66
|
1,3%
|
dtm.
|
45
|
4,3%
|
20
|
4,1%
|
23
|
4,4%
|
62
|
4,9%
|
46
|
6,6%
|
55
|
4,7%
|
251
|
4,8%
|
cop
|
52
|
4,9%
|
40
|
8,1%
|
33
|
6,3%
|
67
|
5,3%
|
36
|
5,1%
|
66
|
5,6%
|
294
|
5,6%
|
|
1055
|
100%
|
491
|
100%
|
524
|
100%
|
1264
|
100%
|
700
|
100%
|
1179
|
100%
|
5212
|
100%
|
34When texts are analyzed, one counts tokens. Their proportion differs from the vocabulary (with types counted). On average, a half of a text is nearly equally split between verbs and pronouns, so words belonging to the respective parts of speech account for about 25% of autosemantic PoS in a text. Approximately another 40% of autosemantic words are nouns, see Table 4.
35French translations demonstrate somewhat different proportions, both in the vocabulary and in the text (see Tables 5 and 6). Nouns constitute about 40% of the vocabulary, followed by verbs with nearly 30%. Adjectives and adverbs are represented in the vocabulary of individual texts in almost equal parts, about 10% each. In the French texts of the tales, nouns constitute about 30%. They are followed by verbs and pronouns, with the proportions similar to Bamana’s (about 25% each).
36The most pronounced difference between Bamana and French is the relative frequencies of adjectives and adverbs, especially in texts. With an average share of about 11% (of both vocabulary and text), adverbs in French are five times more frequent compared to the Bamana text (2.4%) and almost three times more frequent compared to the Bamana vocabulary (4.2%). Adjectives in French appear also nearly five times more frequently in texts (6.9% versus 1.5%) and more than three times in the vocabulary (10.4% versus 3.2%). Note that the respective numbers are the values averaged over individual texts, not those given in the last columns of Tables 3–6 corresponding to the whole collection of six tales.
Table 5. Distributions of autosemantic parts of speech in the French versions of the tales (types)
|
Dununba…
|
Juguya…
|
Juman…
|
Ntalen
|
Sigidankelen…
|
Warabilenkɔrɔ…
|
Whole collection
|
verb
|
140
|
32,9%
|
71
|
30,5%
|
73
|
27,2%
|
122
|
28,0%
|
112
|
31,9%
|
131
|
29,2%
|
372
|
29,3%
|
noun
|
167
|
39,2%
|
77
|
33,0%
|
104
|
38,8%
|
197
|
45,2%
|
140
|
39,9%
|
197
|
44,0%
|
603
|
47,5%
|
adj
|
45
|
10,6%
|
27
|
11,6%
|
34
|
12,7%
|
41
|
9,4%
|
29
|
8,3%
|
44
|
9,8%
|
151
|
11,9%
|
adv
|
45
|
10,6%
|
32
|
13,7%
|
35
|
13,1%
|
41
|
9,4%
|
41
|
11,7%
|
40
|
8,9%
|
96
|
7,6%
|
pron
|
26
|
6,1%
|
21
|
9,0%
|
21
|
7,8%
|
27
|
6,2%
|
29
|
8,3%
|
32
|
7,1%
|
39
|
3,1%
|
num
|
3
|
0,7%
|
5
|
2,1%
|
1
|
0,4%
|
8
|
1,8%
|
0
|
0,0%
|
4
|
0,9%
|
9
|
0,7%
|
|
426
|
100%
|
233
|
100%
|
268
|
100%
|
436
|
100%
|
351
|
100%
|
448
|
100%
|
1270
|
100%
|
Table 6. Distributions of autosemantic parts of speech in French versions of the tales (tokens)
|
Dununba…
|
Juguya…
|
Juman…
|
Ntalen
|
Sigidankelen…
|
Warabilenkɔrɔ…
|
Whole collection
|
verb
|
309
|
27,0%
|
157
|
27,1%
|
163
|
24,5%
|
378
|
25,5%
|
230
|
26,7%
|
370
|
25,3%
|
1607
|
25,9%
|
noun
|
364
|
31,8%
|
122
|
21,0%
|
220
|
33,1%
|
497
|
33,5%
|
247
|
28,6%
|
450
|
30,8%
|
1900
|
30,7%
|
adj
|
70
|
6,1%
|
42
|
7,2%
|
61
|
9,2%
|
72
|
4,8%
|
58
|
6,7%
|
107
|
7,3%
|
410
|
6,6%
|
adv
|
125
|
10,9%
|
80
|
13,8%
|
72
|
10,8%
|
148
|
10,0%
|
111
|
12,9%
|
143
|
9,8%
|
679
|
11,0%
|
pron
|
273
|
23,8%
|
172
|
29,7%
|
148
|
22,3%
|
369
|
24,8%
|
217
|
25,1%
|
379
|
26,0%
|
1559
|
25,1%
|
num
|
4
|
0,3%
|
7
|
1,2%
|
1
|
0,2%
|
21
|
1,4%
|
0
|
0,0%
|
11
|
0,8%
|
44
|
0,7%
|
|
1145
|
100%
|
580
|
100%
|
665
|
100%
|
1485
|
100%
|
863
|
100%
|
1460
|
100%
|
6199
|
100%
|
- 3 Obviously, the approach to the definitions of parts of speech plays an important role here, as disc (...)
37With the frequency data obtained for each text, it is easy to compile a frequency dictionary of the entire text collection. Table 7 contains a complete list of lemmas corresponding to autosemantic parts of speech common to all six tales. There are 46 lemmas in Bamana and also 46 in French. In accordance with the observations made above, the Bamana part contains only one adverb (bì ‘aujourd’hui’). While the lack of adverbs and adjectives is not unexpected (cf. Creissels 2003; Segerer 2008)3 and can be partly compensated for by some other parts of speech, like determinatives or qualitative verbs, the absence of equivalents for French grand ‘big’ and petit ‘small’ in the Bamana list catches the eye immediately. Some reasons for this are discussed in the next Section.
38The proportion of nouns is much higher in the Bamana list of common words (12) versus the French one (4). There are a number of reasons for such a relation. For instance, dùgu glossed as ‘terre’ can also denote ‘village’, which is reflected in the French side. The occurrences of ‘village’ in the French texts are also due to compound words, such as dùgutigi ‘chef du village’. On the other hand, síra ‘chemin’ mostly appears in the French texts not as a physical path but rather as more abstract concepts, such as ‘relation’ or ‘link’. A more detailed analysis can be made for all the instances, which is beyond the intended scope of the present study.
Table 7. Words common to all six texts in Bamana and French
|
Bamana
|
|
French
|
Rank
|
Lemma
|
PoS
|
Gloss
|
Freq
|
Cover
|
Rank
|
Lemma
|
PoS*
|
Freq
|
Cover
|
1
|
à
|
pers
|
3SG
|
442
|
8,5%
|
1
|
3SG
|
pers
|
507
|
8,2%
|
2
|
ò
|
prn
|
ce
|
205
|
12,4%
|
2
|
être
|
v
|
229
|
11,9%
|
3
|
kɛ́
|
v
|
faire
|
124
|
14,8%
|
3
|
ne
|
adv
|
166
|
14,6%
|
4
|
í
|
pers
|
2SG
|
112
|
16,9%
|
4
|
avoir
|
v
|
139
|
16,8%
|
5
|
ù
|
pers
|
3PL
|
107
|
19,0%
|
5
|
ce
|
prn
|
136
|
19,0%
|
6
|
nê
|
pers
|
1SG.EMPH
|
91
|
20,7%
|
6
|
son
|
poss
|
127
|
21,0%
|
7
|
kó
|
cop
|
QUOT
|
88
|
22,4%
|
7
|
pas
|
adv
|
113
|
22,9%
|
8
|
bɛ́
|
cop
|
être
|
87
|
24,1%
|
8
|
1SG
|
pers
|
96
|
24,4%
|
9
|
mîn
|
dtm
|
REL
|
65
|
25,3%
|
9
|
2SG
|
pers
|
82
|
25,7%
|
10
|
yé
|
cop
|
EQU
|
55
|
26,4%
|
10
|
dire
|
v
|
74
|
26,9%
|
11
|
í
|
pers
|
REFL
|
51
|
27,4%
|
11
|
3PL
|
pers
|
72
|
28,1%
|
12
|
fɔ́
|
v
|
dire
|
48
|
28,3%
|
12
|
tout
|
prn
|
62
|
29,1%
|
13
|
sé
|
v
|
arriver
|
48
|
29,2%
|
13
|
que
|
prn
|
56
|
30,0%
|
14
|
tùma
|
n
|
moment
|
48
|
30,1%
|
14
|
faire
|
v
|
55
|
30,9%
|
15
|
dɔ́
|
dtm
|
certain
|
44
|
31,0%
|
15
|
qui
|
prn
|
49
|
31,7%
|
16
|
dùgu
|
n
|
terre
|
44
|
31,8%
|
16
|
mon
|
poss
|
48
|
32,4%
|
17
|
mɔ̀gɔ
|
n
|
homme
|
44
|
32,7%
|
17
|
village
|
n
|
43
|
33,1%
|
18
|
ìn
|
dtm
|
DEF
|
43
|
33,5%
|
18
|
aller
|
v
|
41
|
33,8%
|
19
|
mîn
|
prn
|
REL
|
43
|
34,3%
|
19
|
jour
|
n
|
41
|
34,5%
|
20
|
sɔ̀rɔ
|
v
|
obtenir
|
42
|
35,1%
|
20
|
cela
|
prn
|
38
|
35,1%
|
21
|
táa
|
v
|
aller
|
42
|
35,9%
|
21
|
1PL
|
pers
|
36
|
35,7%
|
22
|
ń
|
pers
|
1SG
|
42
|
36,7%
|
22
|
prendre
|
v
|
34
|
36,2%
|
23
|
bɔ́
|
v
|
sortir
|
39
|
37,5%
|
23
|
grand
|
adj
|
32
|
36,7%
|
24
|
bɛ́ɛ
|
dtm
|
tout
|
39
|
38,2%
|
24
|
petit
|
adj
|
31
|
37,2%
|
25
|
ê
|
pers
|
2SG.EMPH
|
37
|
38,9%
|
25
|
ton
|
poss
|
31
|
37,7%
|
26
|
tɛ́
|
cop
|
COP.NEG
|
35
|
39,6%
|
26
|
arriver
|
v
|
29
|
38,2%
|
27
|
kélen
|
num
|
un
|
35
|
40,3%
|
27
|
en
|
prn
|
29
|
38,7%
|
28
|
yé
|
v
|
voir
|
34
|
40,9%
|
28
|
autre
|
adj
|
24
|
39,0%
|
29
|
tó
|
v
|
rester
|
33
|
41,6%
|
29
|
leur
|
poss
|
22
|
39,4%
|
30
|
dón
|
n
|
jour
|
32
|
42,2%
|
30
|
alors
|
adv
|
21
|
39,7%
|
31
|
kó
|
n
|
affaire
|
31
|
42,8%
|
31
|
aujourd'hui
|
adv
|
21
|
40,1%
|
32
|
nà
|
v
|
venir
|
29
|
43,3%
|
32
|
mettre
|
v
|
19
|
40,4%
|
33
|
dòn
|
cop
|
ID
|
28
|
43,9%
|
33
|
bien
|
adv
|
18
|
40,7%
|
34
|
yɔ́rɔ
|
n
|
lieu
|
27
|
44,4%
|
34
|
2PL
|
pers
|
17
|
40,9%
|
35
|
bìla
|
v
|
mettre
|
26
|
44,9%
|
35
|
celui
|
prn
|
17
|
41,2%
|
36
|
sú
|
n
|
nuit
|
25
|
45,4%
|
36
|
tout
|
adv
|
17
|
41,5%
|
37
|
kánto
|
v
|
s'adresser
|
22
|
45,8%
|
37
|
lever
|
v
|
16
|
41,7%
|
38
|
dɔ́n
|
v
|
connaître
|
20
|
46,2%
|
38
|
savoir
|
v
|
16
|
42,0%
|
39
|
dòn
|
v
|
entrer
|
20
|
46,6%
|
39
|
comme
|
adv
|
14
|
42,2%
|
40
|
bì
|
adv
|
aujourd'hui
|
19
|
46,9%
|
40
|
œil
|
n
|
13
|
42,4%
|
41
|
sí
|
dtm
|
aucun
|
13
|
47,2%
|
41
|
rester
|
v
|
13
|
42,7%
|
42
|
tɔ́gɔ
|
n
|
nom
|
12
|
47,4%
|
42
|
gens
|
n
|
12
|
42,8%
|
43
|
síra
|
n
|
chemin
|
11
|
47,6%
|
43
|
là
|
adv
|
11
|
43,0%
|
44
|
tɔ̀
|
n
|
le.reste
|
10
|
47,8%
|
44
|
devoir
|
v
|
10
|
43,2%
|
45
|
tìle
|
n
|
soleil
|
10
|
48,0%
|
45
|
voilà
|
adv
|
10
|
43,3%
|
46
|
fàn
|
n
|
côté
|
7
|
48,1%
|
46
|
beaucoup
|
adv
|
7
|
43,5%
|
* PoS tags for French are given according to the abbreviations accepted in the Bamana corpus, with an additional notation poss for possessive pronouns.
39There is only one numeral, kélen ‘un’, in the Bamana list of common words but no numerals at all in the French one. The reason, at least partly, might be sought in the issue of the article/numeral ambiguity in tagging the French counterpart ‘un/une’ discussed in Section 2.
40The “Cover” column in Table 7 shows the proportion of text covered by the respective lemmas relative to the total number of autosemantic tokens in all the texts (5212). So, the 46 lemmas common to all six texts in Bamana account for 48.1% of all words. Seventy lemmas are common to at least five texts and cover 54.7% of tokens. The lemmas common to at least three texts count 167 and cover already 66.4% of tokens. The numbers for the French versions are slightly different. There are 6199 autosemantic tokens in all six French texts, of which 46 are common to all six texts; they cover 43.5% of text. Eighty lemmas are common to at least five texts and cover 51.7% of text, while 219 lemmas are common to at least three texts and cover 66.9% of text.
41Four reasons can be identified leading to a significantly smaller number of size adjectives in Bamana texts compared to their French translations. They are listed in subsections 4.1–4.4.
424.1. First of all, in Bamama diminutive suffix ‑nin and augmentative suffix ‑ba are used extensively in instances where one would expect ‘petit’ or ‘grand / gros’ in French. Several examples are shown in (2a–c).
(2a)
|
Tòro
|
bɔ́ra
|
à
|
ka
|
wònin
|
fɛ̀
|
|
rat.voleur
|
sortir.PFV.INTR
|
3SG
|
POSS
|
trou.DIM
|
par
|
43The respective French sentence reads ‘Toro sortit par son petit trou.’ = ‘Toro came out through his little hole.’ [“Dununba…”].
(2b)
|
– Cɛ̀nin
|
wó,
|
í
|
tɛ
|
sábali !
|
|
jeune.homme = mâle.DIM
|
hé
|
2SG
|
IPFV.NEG
|
être.patient
|
44‘– Petit garçon, tu n’exagères pas ?’ = ‘– Little boy, aren't you exaggerating?’ [“Dununba…”].
(2c)
|
Dúnuya
|
kó-ba
|
cáman
|
sún
|
bɛ́
|
mɔ̀ɔsow
|
lá :…
|
|
monde
|
affaire-AUGM
|
nombreux
|
tronc
|
être
|
homme.maison
|
à
|
45‘L’origine de bien des grandes œuvres de la vie, c’est les femmes : …’ = ‘The origin of many great works of life is women: …’ [“Sigidankelen…”].
464.2. Single Bamana words can be translated into French by lexical equivalents containing two words, one of which is a size adjective (3a–c).
(3a)
|
– Ń
|
kɔ̀rɔ
|
Kélennako
|
nê
|
séra
|
í
|
fɛ̀
|
|
1SG
|
aîné
|
NOM.M
|
1SG.EMPH
|
arriver.PFV.INTR
|
2SG
|
par
|
47‘– [Mon] Grand frère Kélénako, me voilà aujourd’hui devant toi.’ = ‘[My] Big brother Kélénako, here I am today in front of you.’ [“Juguya…”].
(3b)
|
Nê
|
dɔ́gɔmuso
|
nàna
|
kó
|
ń
|
ka
|
|
1SG.EMPH
|
cadette
|
venir.PFV.INTR
|
QUOT
|
1SG
|
SBJV
|
|
ɲɔ̀
|
dí
|
à
|
mà
|
|
|
|
mil
|
donner
|
3SG
|
ADR
|
|
|
48‘Ma petite sœur est venue me demander du mil.’ = ‘My little sister came to ask me for millet.’ [“Juguya…”].
(3c)
|
Ála
|
y’
|
à
|
ládiya,
|
kà
|
nàfolo
|
cáman
|
|
Dieu
|
PFV.TR
|
3SG
|
récompenser
|
INF
|
biens
|
nombreux
|
|
dá
|
à
|
yé :
|
bà,
|
mìsi,
|
sàga,
|
fàli.
|
|
poser
|
3SG
|
PP
|
chèvre
|
bovidé
|
ovin
|
âne
|
49‘Dieu avait fait de lui un homme riche : il possédait en grand nombre des ânes, des vaches, des moutons et des chèvres.’ = ‘God had made him a rich man: he owned donkeys, cows, sheep and goats in large numbers.’ [“Juguya…”].
504.3. A descriptive synonymic translation can be used rather than a direct equivalent (4a–b).
(4a)
|
– Wò
|
bɛ́
|
dùnun
|
lá, ...
|
|
trou
|
être
|
tambour
|
à
|
51‘– Il y a une petite ouverture au bas du tam-tam.’ = ‘– There is a small opening at the bottom of the drum.’ [“Dununba…”]. Note that in example (2a) wònin ‘trou.DIM’ was utilized.
(4b)
|
… ò
|
kámalen
|
y’
|
í
|
kánto : …
|
|
ce
|
jeune.homme
|
PFV.TR
|
REFL
|
s’adresser
|
52‘… son petit ami lui déclare : …’ = ‘… her boyfriend tells her: …’ [“Juman…”]
53A few sentences further on, a diminutive is used instead:
(4c)
|
– Ɛ̀ɛ,
|
ń
|
térinin, …
|
|
pas.possible!
|
1SG
|
ami.DIM
|
54‘– Eh ! Mon petit ami, …’ = ‘– Hey ! My little friend, …’ [“Juman…”]
554.4. Free translations can be too loose or contain idioms. These range from a single extra word, like in the following example (5a) from [“Dununba…”], to more sophisticated approaches as represented by (5b).
(5a)
|
Bámànan1 kó2 : « Jànfajuru3 fyɛ́ku4 kójugu5 , à6 bɛ7 féreke8 í9 yɛ̀rɛ̂10 kán11 lá12 ».
|
|
‘{Les bambaras}1 {disent}2 : {A force de}5 {manier}4{ta petite}0 {corde de trahison}3, {tu}6 {finiras par}7 {la nouer}8 {autour de}12 {ton}9 {propre}10 {cou}11.’ = ‘'{The bamanas}1 {say}2: {By dint of}5 {wielding}4{your little}0 {cord of betrayal}3, {you}6 {will end up}7 {tying it}8 {around}12 {your}9 {own}10 {neck}11.’
|
56Here, jànfajuru ‘corde de trahison’ have neither an adjective nor the diminutive suffix in the Bamana text, while there is an extra description ‘ta petite’ in the French sentence.
(5b)
|
Hálì
|
bì
|
bámànan
|
tɛ́
|
bálimamuso
|
|
même
|
aujourd’hui
|
bambara
|
COP.NEG
|
sœur
|
|
kó
|
túlon
|
ná.
|
|
affaire
|
jeu
|
à
|
57‘Jusqu’à aujourd’hui, les bambaras ont une grande considération pour leurs sœurs.’ / ‘Until today, bamanas have had great regard for their sisters.’ [“Juguya…”].
584.5. Only in a few cases, size adjectives are used explicitly in Bamana. Two instances involving bèlebele ‘gros’ and fítinin ‘petit’ are shown in (6a,b):
(6a)
|
Kùnɲɔgɔn
|
kélen
|
bɛ
|
dáfa
|
tùma
|
mîn
|
ná,
|
|
semaine
|
un
|
IPFV.AFF
|
compléter
|
moment
|
REL
|
à
|
|
ò
|
y’
|
à
|
sɔ̀rɔ
|
ù
|
ye
|
fòrokɛnɛ
|
|
ce
|
PFV.TR
|
3SG
|
obtenir
|
3PL
|
PFV.TR
|
champ. clarté
|
|
bèlebele
|
yíriw
|
tìgɛ.
|
|
gros
|
arbre.PL
|
couper
|
59‘Une semaine après, ils avaient coupé les arbres sur une très grande surface.’ = ‘A week later, they had cut the trees over a very large area.’ [“Warabilen…”].
(6b)
|
À
|
tóra
|
ò
|
cógo
|
lá
|
fɔ́
|
dón
|
dɔ́,
|
|
3SG
|
rester. PFV.INTR
|
ce
|
manière
|
à
|
jusqu'à
|
jour
|
certain
|
|
Ncí
|
ye
|
búbaganton
|
fítinin
|
yé
|
tú
|
dɔ́
|
kɛ̀rɛfɛ̀
|
|
NOM.M
|
PFV.TR
|
termitière
|
petit
|
voir
|
touffe
|
certain
|
côté.par
|
60‘Tout resta comme ça jusqu'au jour où Nci vit une petite termitière près d'un bois.’ / ‘Everything remained like that until the day when Nci saw a small termite hill near a wood.’ [“Warabilen…”].
61Such examples include also the pleonastic use of fitinin ‘petit’, e. g.:
(6c)
|
Jálakɔrɔka
|
bɛ́ɛ,
|
hálì
|
dénmisɛnnin
|
fitininw, …
|
|
TOP.GENT
|
tout
|
même
|
petit.enfant.DIM
|
petit.PL
|
62‘Tout le monde à Dialakoro, même les petits enfants, …’ = ‘Everyone in Dialakoro, even little children, …’ [“Juman…”]. The word dénmisɛn is itself composed of dén ‘enfant’ and mìsɛn ‘petit’ and additionally gets here the diminutive suffix -nin.
63Studies of languages using approaches from the theory of complex networks date back to early 2000s (Dorogovtsev & Mendes 2001; Ferrer i Cancho & Solé 2001) and remain an active field of research (Holovatch & Palchykov 2016; Markovič et al. 2019).
64One of the approaches typically used to build a network from a text is as follows. Word types (in our case, autosemantic lemma types) are considered to be network vertices. Two vertices are connected by a link if the respective words are found in the same sentence. If there are several sentences where such two words occur, the links can be counted with multiplicity equal to the number of such sentences.
Figure 2. A sample network based on two sentences from the “Dununba…” tale.
Note a thicker line between n|só and n|dùgutigi: this pair of vertices occurs twice in the sample sentences, so the link has multiplicity two.
65The networks were built using own software (scripts in the Perl language). The visualization and calculation of the network parameters were made using the Pajek software (De Nooy, Mrvar & Batagelj 2011; Mrvar & Batagelj 1996–2018), which allows for the evaluation of many network properties, of which only those related to distances between vertices are analyzed below in detail.
66For illustration, consider a sample network in Figure 2, which is built using two consecutive sentences from the “Dununba…” tale. The sentences are (with autosemantic PoS given in boldface):
(7a)
|
Áw
|
yé
|
dùgutigi
|
ka
|
só
|
yɔ́rɔ
|
jìra
|
|
pers
|
pm
|
n
|
pp
|
n
|
n
|
v
|
|
2PL.EMPH
|
IMP
|
chef.de.village
|
POSS
|
maison
|
lieu
|
montrer
|
|
nê
|
lá
|
fɔ́lɔ.
|
|
pers
|
pp
|
adj
|
|
1SG.EMPH
|
à
|
premier
|
67‘Mais menez-moi d’abord chez le chef du village.’ = ‘But take me to the village chief first.’
(7b)
|
Mùso
|
kélen
|
bìlara
|
ɲɛ́najɛ
|
dúnan
|
ɲɛ́
|
kà
|
|
n
|
num
|
v
|
n
|
n
|
pp
|
pm
|
|
femme
|
un
|
mettre.PFV.INTR
|
réjouissance
|
étranger
|
devant
|
INF
|
|
tága
|
dùgutigi
|
ka
|
só.
|
|
v
|
n
|
pp
|
n
|
|
aller
|
chef.de.village
|
POSS
|
maison
|
68‘Une des femmes accepta de l’accompagner jusqu’à la maison du chef.’ = ‘One of the women agreed to accompany him to the chief’s house.’
69A vertex can be isolated, that is, not linked to any other. Usually, such situations correspond to very short sentences often found in the direct speech. For instance, the vertex corresponding to makɛ ‘maître’ is isolated in the “Ntalen” tale. It appeared once in an exclamation translated as ‘– Eh, chef !’:
(8)
|
<s>
|
– Eɛ !
|
</s>
|
<s>
|
Mákɛ!
|
</s>
|
|
|
intj
|
|
|
n
|
|
|
|
pas.possible!
|
|
|
maître
|
|
70The simplified tags for the beginning of sentence <s> and for the end of sentence </s> are shown explicitly.
71The distance between two non-isolated vertices is counted as the number of segments in the shortest path required to reach one vertex starting from the other. For instance, in Figure 2 the distance between num|kélen and n|mùso is d = 1, while the distance between num|kélen and v|jìra is d = 2. Usually, even in large text networks mean values of the distance remain within 2.2–2.5 (Cong & Liu 2014; Caldeira et al. 2006; Buk, Krynytskyi & Rovenchak 2019). The short texts of tales have the mean distance values shifted towards d = 2, as expected, see Table 8.
72The maximal distance between non-isolated vertices rarely exceeds 6, the language networks are thus regarded as “small worlds” (Ferrer i Cancho & Solé 2001) referring to the human society with “six handshakes rule” or “six degrees of separation” between people in the world (Watts 2004). Not surprisingly, in the analyzed case of short texts these values are smaller and most often equal to four, see Table 8.
Table 8. Some network properties of the Bamana and French versions of tales
|
Dununba…
|
Juguya…
|
Juman…
|
Ntalen
|
Sigidankelen…
|
Warabilenkɔrɔ…
|
|
bam
|
fra
|
bam
|
fra
|
bam
|
fra
|
bam
|
fra
|
bam
|
fra
|
bam
|
fra
|
Sentences
|
127
|
146
|
63
|
64
|
66
|
68
|
139
|
118
|
75
|
57
|
136
|
150
|
Sent len (all)
|
11,7
|
11,2
|
10,4
|
12,0
|
11,0
|
14,2
|
12,4
|
17,0
|
13,1
|
21,1
|
12,1
|
13,4
|
Sent len (aut)
|
8,3
|
7,8
|
7,8
|
9,1
|
7,9
|
9,8
|
9,1
|
12,6
|
9,3
|
15,1
|
8,7
|
9,7
|
Mean distance
|
2,19
|
2,30
|
2,13
|
2,10
|
2,33
|
2,13
|
2,14
|
2,05
|
2,17
|
2,01
|
2,14
|
2,17
|
Max distance
|
4
|
4
|
4
|
4
|
5
|
4
|
4
|
3
|
4
|
3
|
4
|
5
|
Vertices
|
324
|
426
|
191
|
233
|
224
|
268
|
377
|
437
|
291
|
351
|
362
|
448
|
Links
|
6780
|
7336
|
2612
|
3768
|
3606
|
5078
|
8582
|
11704
|
4930
|
9570
|
7504
|
9700
|
Links per vertex
|
20,9
|
17,2
|
13,7
|
16,2
|
16,1
|
18,9
|
22,8
|
26,8
|
16,9
|
27,3
|
20,7
|
21,7
|
Figure 3. Distribution of distances between vertices in the networks of the Bamana (green) and French (red) versions of the tales
The average percentage is shown on the vertical axis.
73It might be also interesting to look into the details of the path length distribution in the analyzed texts. A summary is shown in Figure 3. Most paths have length 2 (on average, over 70% in the Bamana texts and over 75% in the French texts). On the other hand, lengths 3 are slightly more frequent in the Bamana texts (22.3% versus 18.5%). The differences between the numbers, however, are not significant enough to draw any far-reaching conclusions.
74Interestingly, in Bamana the correlation between sentence length and mean distance is not very significant, while in French the inverse correlation in very well pronounced, i. e., shorter sentences yield larger mean distances. The correlation coefficient in French is –0.84 versus –0.38 in Bamana. The reason is that mean sentence lengths are more evenly distributed in the Bamana texts (7.8 to 9.3) than in the French ones (7.8 to 15.1).
75The number of vertices, as given in Table 8, is nothing but the number of autosemantic lemma types in Table 3. From the number of links per vertex one can conclude that, depending on the text, each lemma co-occurs in a sentence on average with 14–23 other lemmas in the Bamana texts and with 16–27 other lemmas in the French texts under study.
76The highest number of links ranges from 118 in “Juman…” to 295 in “Ntalen”. Almost always it is associated with the pronoun à ‘3SG’ and only in “Juman…” it corresponds to the pronoun ò ‘ce’, with à ‘3SG’ on the second place having 104 links. A similar behavior is found in French.
77My initial expectation was that mean distances in the networks for Bamana texts would be smaller compared to French ones. The reason is the smaller number of types covering a larger portion of texts in Bamana, see Table 7 and the frequency data by Rovenchak & Buk (2013). This was confirmed for the first analyzed text, “Dununba…”. However, an opposite relation was found for five other texts, see Table 8.
78The observed data suggest, in particular, that mean distances in a text network are mostly defined by mean sentence lengths rather than some deeper properties of languages. Mean sentence lengths, on the other hand, are believed to be good author style markers (Yule 1939; Sichel 1974; Pande & Dhami 2015).
79In the case of the texts under study it appears that differences in sentence lengths are often defined by the representation of the direct speech in the corpus. A proper treatment of the direct speech might require extending the end-of-sentence markers beyond the standard set of full-stop ‘.’, exclamation mark ‘!’, question mark ‘?’, and ellipsis ‘…’ (cf. Martin et al. 2003; Rovenchak & Buk 2013).
80The results presented in the present work allow for conclusions in several domains: lemmatization and tagging of French texts in the Bamana–French parallel corpus, which has not been implemented yet, parts of speech distributions in Bamana and French with a special focus on adjectives, and network properties of texts.
81From the preliminary preparation of the French texts for the analysis, namely, automated lemmatization and part-of-speech tagging, one can conclude that the TreeTagger software yields satisfactory results but requires additional manual tuning. The observations made in this work suggest how this tuning can be partly automated as well.
82As a by-product of the network analysis of texts, the need to unify approaches to the treatment of sentence breaks in the direct speech comes out. This applies to both Bamana and French texts and should be taken into considered in the subsequent development of the Bamana Reference Corpus. Another conclusion to be drawn from the network analysis has a negative hue: it seems that some properties of text networks are just defined by the mean sentence length and might be of little use for in-depth language studies, especially for relatively short texts. In prospect, approaches not relying on sentence boundaries can be used to build text networks and study their properties.
83The main body of the results concerns the distribution of autosemantic parts of speech in text and vocabulary. The analysis of the Bamana and French versions of the tales has revealed the similarities and differences between the languages. One of such peculiarities, the absence of adjectives among the most frequent words in Bamana, is discussed in detail through the analysis of size adjectives and several ways of their representation in Bamana compared to the French translations. Additional studies involving more texts are required to distinguish between language-related and genre-related features in detail.
84Further research would include analysis of other pairs of texts and text collections from the Bamana–French parallel corpus, especially of different genres, as well as eventual expansions to the Maninka–French parallel corpus and inclusion of other language pairs (cf. Vydrine, Togo & Bulman 2017).
85I am grateful to Valentin Vydrin for the discussions on some issues raised in the paper and for the help with the Bamana language and to Angela Kamianets for reading the manuscript.
861,2,3 — 1st, 2nd,3rd person
ADR — address postposition
AFF — affirmative
AUGM — augmentative
COP — copula
DEF — “new definite article”
DIM — diminutive
EMPH — emphatic
EQU — equative copula
GENT — “genitive” suffix
ID — identification copula
IMP — imperative
INF — infinitive
INTR — intransitive
IPFV — imperfective
NEG — negative
NOM.M — male name
OPT2 — optative
QUOT — quotation copula
PFV — perfective
PL — plural
PL2 — non—productive plural
POSS — possessive
PP — polysemic postposition
PROG — progressive
REFL — reflexive
REL — relativization
SBJV — subjunctive
SG — singular
TOP — toponym
TR — transitive
87adj — adjective
adv — adverb
conj — conjunctive
cop — copula
dtm — determinative
n — noun
n.prop — proper noun
num — numeral
pers — personal pronoun
pm — predicative marker
poss — possessive pronoun
pp — postposition
prn — pronoun
ptcp — participle
v — verb
vq — qualitative verb