ОБ ОДНОМ СПОСОБЕ ВЫЯВЛЕНИЯ ГЕНДЕРНО ЗНАЧИМОЙ ЛЕКСИКИ В АНГЛОЯЗЫЧНОМ ТЕКСТЕ

Научная статья
DOI:
https://doi.org/10.18454/RULB.3.05
Выпуск: № 3 (3), 2015
PDF

Аннотация

Гендерные особенности языка проявляются, прежде всего, на лексическом уровне. При этом основным исследовательским приемом является оценка употребительности: одни слова чаще встречаются в текстах, написанных женщинами, другие – текстах мужчин. Основанная на этом подходе процедура гендерной атрибуции текста была опробована на обширном англоязычном материале (22 автора) и оказалась достаточно эффективной. Расположение авторов на «шкале маскулинности» позволяет ставить и решать ряд нетривиальных задач. В частности, выдвинута гипотеза о том, что по сравнению с авторами XIX века степень маскулинности современных авторов-мужчин в целом значительно снизилась, а авторов-женщин – повысилась.

The issue of sex and gender has become very important nowadays and it is a matter of utmost interest not only scientifically but also in sociocultural sphere, sometimes even on the legislative level. Sex is regarded as a biological phenomenon; consequently men are opposed to women on the basis of purely biological characteristics, including behavioral peculiarities. As to gender, it touches upon the psychological features and in this respect contrasted notions will be masculinity and femininity (Kon 2004).

In most cases sex and gender coincide yet some examples of the opposite have been reported[1].

Since gender is a sharply defined framework of social and psychological settings, it is assumed to affect the linguistic behavior of an individual (Maslova 2004, p. 124).

Introducing the parameter of gender into the linguistic research has opened new prospects in the analysis of various aspects of language and speech. The term itself emerged in linguistics in the 1980s, i.e. a bit later than in other humanities – History, Psychology, and Sociology. The ideas of gender linguistics, put forward by various schools and movements, are being still moulded in a system (Mamaev 2011).

It must be stressed that the human being not only understands the meaning of a judgment uttered but realizes his or her involvement in it. Hence, the word becomes a cultural archetype and is regarded as an issue that both the outer world and the speaker have impact on (Lebedev 2008). It seems clear, then, that gender speech peculiarities are most likely to manifest themselves.

The study of gender aspects has become extremely popular among linguists lately. Up to the present efforts have been focused on specific features of men’s and women’s speech. For example, among feminine linguistic facts the researchers enumerate frequency of usage of euphemisms; adverbs helping to avoid the utterance categoricity (e.g. rather, quite); exclamatory sentences, tag questions, etc. Regarding masculine facts they reckon the pursuit for word creation; tendency to the usage of archaic, jargon and dialectal words; omissions of articles and auxiliary verbs; etc. (Antineskul 2001).

From numerous published results one may infer that gender peculiarities emerge, mostly, on the lexical level. Ordinarily, words or phrases are considered to be feminine if they appear substantially more frequently in the texts of female writers. Consequently, one should pay utmost attention to the usage frequency of language units.

The reverse approach may be of interest, too: basing on the lexical statistics it is possible to study the opposite task: is it possible to specify gender attribution of the text, i.e. to predict whether its author is a man or a woman. That would make the main task of our study: we undertake to find, in any random text, a group of lexical elements which would allow its gender attribution with a high degree of reliability.

Although it was ascertained that the usage of slang and Latin terms are characteristic of the male written speech, we should bear in mind that these phenomena are not frequent in any text. The same can be said about feminine texts which should presumably include many euphemisms and tag questions. The required solution would be turning to function words.

First of all, function words by their nature are very frequently used and are met in any type of texts. It is no less important that they are limited in number is limited and, thus, the process of analysis wouldn’t be too laborious.

It is worth noticing that gender attribution of function words hasn’t been practically studied. We can cite only one attempt to dwell with the problem. It was done on English material: (Argamon, Koppel 2006; Koppel, Argamon, Shimoni 2001).

The authors are trying to solve the problem of gender attribution by dividing function words into ‘mainly male’ or ‘mainly female’ words.[2] The algorithm was introduced which is to identify the gender of the text’s author with high probability. In has been tested on the large amount of text corpora and defined gender attribution with at least 80% success rate. The result seems good enough, however there still remains some opportunity to further perfect the procedure of the algorithm application.

According to the algorithm, the articles (a, the) as well as demonstrative pronouns (that, these) are considered masculine indicators, while a group of pronouns (I, you, she, her, their, myself, yourself, herself) indicate the belonging to the text written by a female author. The gist of algorithm is the evaluation of frequency of every word from the list. The frequency data are supplemented with the system of coefficients, which regulates the contribution of every word to the final result. For example, the coefficient, or the ‘weight’ of preposition with is 52, while pronoun who ‘weighs’ only 19, article the – 7, etc. Thus, if with is found in the text 4 times, then its total contribution will be 208 (4 х 52); if pronoun who isn’t present in the text at all, it will have zero contribution; and if article the appears 69 times, then its contribution will be rather substantial (69 х 7) – 483.

The total sum of weights for masculine words (who, the, as, etc.) and feminine ones (with, if, not, etc.) are counted and the results are summed up. If the total sum of feminine words turns out to be more than that of masculine ones then the text is attributed as feminine.

In spite of the authors’ claim that the algorithm functions rather efficiently, their suggested set of function words gives rise to some doubts. Particularly, it is not clear why the authors include forms of the verbs to be and to say in the list. The usage of to is also causes scrutiny: it is clear that there is serious difference if to functions as a particle (with the infinitive) or a preposition (the calculations would not take this difference into consideration at all).

With a glance to the above, we made up our minds to test the algorithm on the new selection of texts. There were selected 11 novels of British authors of the 19th–20th centuries (cf.: G.K. Сhesterton ‘Father Brown Stories’ / J. Austen ‘Pride and Prejudice’, etc.) and 11 modern novels of the 20th-21st centuries (cf.: P. Ness ‘A Monster Calls’ / M. Blackman ‘Noughts And Crosses’, etc.). Taking into consideration the character of the material (function words), the corpus seems quite representative. We only confined oneself to the analysis of three fragments from every novel – in the beginning, in the middle and at the end accordingly. The amount of every fragment was nearly the same – 1500 words approximately. The list of novels is given in the end of the article.

The processing of the texts was conducted with the help of the Gender Genie algorithm was applied to each text[3]. The algorithm allows every Internet user process a text, automatically highlighting the words relevant for the authors’ gender attribution.

The example of the initial matrix is presented in Table 1.

Table 1 – The initial matrix of the calculation results (fragment)

 

Feminine words

 

Masculine words

 

Word

Was used, times

 

Weight

Total

contri-

bution

 

Word

Was used, times

 

Weight

Total

contri-

bution

Charles Dickens

her

me

your

with

myself where

27

26

12

43

2

3

20

20

40

1

4

2

540

520

480

43

8

6

the

a

what

many

more

below

209

116

15

4

7

1

6

10

35

6

2

8

1254

1160

525

24

14

8

Total

4179

Total

5456

Ann Brontë

her

me

not

should

myself

with

where

61

65

65

8

17

51

3

20

18

8

50

4

1

2

1220

1170

520

400

68

51

6

the

as

a

what

more

these many

170

32

86

23

7

1

1

6

30

10

35

2

8

6

1020

960

860

805

14

8

6

Total

5400

Total

4939

According to Table 1, the sum of masculine words prevails in Ch. Dickens’ text (5456 > 4179), consequently, this text is written by a male author. On the contrary, in A. Brontë’s text feminine words prevail (5400 > 4939), so the text is obviously written by a female. As we see, the program has worked well in both cases.

However, having analyzed all the 22 texts, the results didn’t turn out that successful. Suffice it to say that the program has correctly attributed only 6 out of 11 texts by male-writers, i.e. made mistakes in nearly a half of the cases. Keeping to the hypothesis that basing on a limited number of function words we can make gender attribution of the text correctly enough, we have to concede that either some words from the given list aren’t appropriate, or the weight system needs correction, or both are true.

Our next step had two main aims:

1) making sure that the initial division of words into masculine and feminine was accurate and suggesting corrections, if necessary;

2) improving the weight system of the words used.

We summarized the initial matrix data of all the 22 authors. In case the algorithm attributes a word to the feminine group, than its total usage frequency in female-writers’ texts should substantially exceed that in male-writers’ texts, and vice versa.

The analysis shows that the confirmative data have been received for 10 words out of 15 (e.g. me was found 360 times in female-writers’ texts and only 180 times in male-writers’ texts; your 138 : 74 and so on). For 4 words no preferences have been spotted (e.g. with – 360 : 360, when – 160 : 150), and in one case there is even an opposite result (was – 860 : 1033).

Still more disappointing was the result for the masculine words. The confirmation was acquired for only 6 words (e.g. the – 3136 : 2330 or as – 486 : 293). In the rest 10 cases data were nearly the same for males and females (these – 36 : 35, below – 7 : 5, many – 24 : 20) or even with the prevalence for females (to – 1083 : 1240, are – 91 : 132).

In consideration of these data we have taken into account only the cases with the significant difference in the usage frequency of words in male- and female-writers’ texts (10 feminine and 7 male words). The rest of the words from the initial list were eliminated.

Our next task was to correct the weight system of the 17 words.

The way of reasoning in applying weights to words by the algorithm has been quite dim to us. Here are a few examples (Table 2).

Table 2 – The analysis of the lexical units’ weights suggested by the algorithm

 

Word

Found in male-authors’ texts

Found in female- authors’ texts

Word weight

 

1

should

32

48

50

around

43

49

10

myself

37

51

4

 

 

 

2

we

102

169

45

what

136

190

35

if

118

146

28

 

 

3

a

1347

1111

10

the

3136

2330

6

to

1083

1240

2

The groups (1–3) contain the words of nearly the same usage frequency. As one can see, the weights in every group differ significantly, so the necessity to somehow correct the weight system seems obvious.

In working out the new weight system we took two main factors into consideration. The first one is the range of frequency difference. For example, the words in Group 2 have more differentiating force than the words in Group 1, and consequently they should be given more weight.

On the other hand, some words (especially articles) have considerably higher frequency, and applying much weight to them will cause their taking the substantial part of the general result, minimizing all other words’ role. In setting a proper weight, we thought it reasonable to ensure that a word’s contribution to the whole sum shouldn’t exceed 25%. So, we applied the weights of 4 and 6 to the articles a and the respectively, while if got the weight of 28, and myself – 50.

Lastly, we have perfected the procedure in one more aspect. The initial algorithm only allows to attribute a text as ‘male’ or ‘female’. We’d like to go further and arrange the texts on a kind of “scale of masculinity”. This will allow not only distinguish a male text from a female one but also compare two texts of the same gender. As a result, we’ve introduced a masculinity index M, which may serve as an important feature of the author’s gender style. Index M is based directly on the masculine / feminine words’ indicator and is calculated according to the formula:

formula

Here comes an example of calculations for texts by G.S. Chesterton and A. Brontë. In Table 3 one can see fragments of initial data matrices.

Table 3 – Initial data matrices (fragments)

G.S. Chesterton

Feminine words

 

Masculine words

Word

Frequency

Weight

Total

Word

Frequency

Weight

Total

we

7

40

280

the

292

4

1168

be

22

10

220

a

163

6

978

myself

4

50

200

as

44

20

880

 

 

 

 

she

1

12

12

around

0

20

0

Total

1585

Total

4572

A. Brontë

me

65

20

1300

the

170

4

680

her

61

15

915

as

32

20

640

myself

17

50

850

a

86

6

516

 

 

 

 

be

17

10

170

around

2

20

40

Total

5515

Total

2684

Table 4 illustrates the stages of calculating the index M.

Table 4 – The calculations of parameters for the index M

Table 4 – The calculations of parameters for the index M

Table 5 – Index M of the whole group of authours

Table 5 – Index M of the whole group of authours

According to Table 5, index M attributes the texts rather accurately. Certainly, as it usually occurs, there are some borderline results when M approaches zero (in this category we find 3 male and 4 female authors; they are marked grey). Perhaps, further improvements of the procedure might reduce the number of such cases, however efforts in another direction seem no less promising. In fact, as we mentioned earlier, gender reflects an individual’s linguistic consciousness, and it shouldn’t inevitably coincide with his or her biological sex. Consequently, if of an author’s index M stands out from the rest, this should cast some doubt on his or her gender identity. To clarify the case, one may need additional facts from related sciences – Literary Criticism, Psychology, etc. From this point of view it would be interesting, for example, to study the results of Brontë sisters whose texts turned out to be in opposite ends of the “scale of masculinity”.

In conclusion let us raise one more point. As it is seen from the list of works, our selection of fiction texts is balanced not only according to gender (11 male and 11 female authors), but also according to the time of writing (11 texts of 19th–20th century and 11 texts of 20th–21st century). This enables us to juxtapose the index M values of the two epoques (Table 6).

Table 6 – The correlation between index M and the epoque of writing (average data)

Table 6 – The correlation between index M and the epoque of writing (average data)

One can definitely infer, from the data, that gender differentiation among the 19th century writers is more distinct (the range is 0,74). Among contemporary writers we see some tendency for convergence, the difference between male and female authors being only 0,45.

No doubt, such facts call for more research involving new literary material.

Certainly, the urgent problem might be to develop the procedure for defining the index M for the material in Russian. If it is successful, this may help discover new possibilities – for comparative analysis in the first place. In particular, a very interesting task would be to analyze the cases when, for example, the text written by a female-authour is translated by a male-translator and vice versa.


[1] It’s worth mentioning that English is among the few languages where “sex” and “gender” are differentiated on the lexical level.

 

 

 

[2] In this study the class of function words includes pronouns and modal verbs (30 words in total).

 

 

 

[3] The Gender Genie. URL: http://bookblog.net/gender/genie.php (date of reference: 19.02.2013).

 

 

 

Список литературы

  • Antineskul O.L. Gender as text-formation parameter: Special course teaching aid. – Perm, 2001.

  • Kon I.S. Sexology. Teaching aid. – M., 2004.

  • Lebedev S.A. Historical-philosophical methodological introduction in linguistic theory // Philosophy of social and studies and humanities. – M., 2008.

  • Mamaev M.M. Multiple-aspect character of studying gender factor in linguistics // Bulletin of Moscow State Regional University. Series “Linguistics”, №2. – M., 2011.

  • Maslova V.A. Lingvoculturology. – M., 2004.

  • Argamon S., Koppel M., Fine J., Shimoni A.R. Gender, Genre, and Writing Style in Formal Written Texts // Interdisciplinary Journal for the Study of Discourse. 2006, vol. 23, Issue 3, pp. 321–346.

  • M. Koppel, Argamon S., Shimoni A.R. Automatically determining the gender of a text’s author // Bar-Ilan University Technical Report BIU-TR-01-32. 2001.