Preservation of Emotional Tone Patterns In Human and Machine Translations Between English and French

Problem and Method

The availability of two French dictionaries of connotative meaning made it possible to examine the preservation of emotional tone patterns in translations from English in children's stories (Anderson and McMaster, 1992; Anderson, McMaster and Mitchell, 1987). Good preservation was found for some of the stories but not others. It was not clear whether this was the result of poor translations or the indexing system was too sensitive to a few words in short and specialized text. In stylometric analysis Gurney and Gurney (1998) have shown that long texts cannot be assumed to be homogenous, so that reliance on sub-sets is dangerous. An examination of longer units of text would seem to be advisable before judgements of translation adequacy are based on emotional tone patterns.

The opportunity to consider larger units of French and English text arose as a consequence of an analysis of Saint Julian the Hospitaler (Anderson, 2000), which revealed that automated collection of the primary English emotional tone scores yielded values which agreed closely with those collected using operator disambiguation judgements. Since the two French dictionaries used in the LOGOS system (Anderson and McMaster, 1993) both collect the connotative meaning values by computer alone, processing of a large unit of text in both English and French guises could be undertaken.

It was proposed that the English text scores of evaluation, activity and potency generated by application of the 1000 words Heise dictionary (Heise, 1965) should be compared with the French text scores of emotionality and abstractness obtained with the 398 word Vikis-Freibergs dictionary (Vikis- Freibergs, 1976) and the scores of pleasantness obtained using the 904 nouns in the Messina dictionary (Messina, Moraid and Captaine, 1989). This would be done with word blocks of approximately 100 words, in corresponding units of Saint Julian the Hospitaler, using the French original of Flaubert (1890) and Grimaud's English translation (Berg, Moskos and Grimaud, 1982).

Grimaud's English human translation was assumed to be of the highest quality, so that correlations between corresponding scores would represent as much agreement as the system is able to capture. It was expected that the agreements between English and French patterns resulting from computer translation would be less, and that correlations between non-corresponding dimensions would be less than those between corresponding ones.

The testing of automated scoring in Saint Julian the Hospitaler that has been mentioned had been done with blocks of 100 total words. This division of the text was maintained in the present study with slightly different numbers of total words per block, so that in every case scores were available for 94 blocks. An English computer translation from the French original was made using SYSTRAN, (SYSTRAN, 1999) unloaded from the beyond.com web site. The correlations between these French and English texts were calculated for evaluation, activity and potency (English), emotionality and abstraction (French), and pleasantness (French).

Results and Conclusion

Within the French original the correlation of emotionality and pleasantness was 0.32, statistically significantly different from zero at the level p<.01, df=92, one-tailed. The correlations of these with the English human translation scores of evaluation were 0.20 and 0.18, respectively, p<.05 in both cases.

Other correlations across and between the scores were not statistically significant except that abstractness was negatively related to emotionality at the level -0.32, p<.01, one-tailed. It can be supposed that this is a consequence of some special quality of this fierce text. That may also be the explanation of the negative correlations with potency of all the evaluation corresponding scores. The correlations of potency with evaluation, emotionality and pleasantness were -0.47, -0.18 and -0.28 respectively, all significant at the 0.05 level or better. In this text lower levels of evaluation were accompanied by higher levels of potency, that is, more toughness. The agreements between the critical measures in the French original and the English human translation can be taken as supported by their common association with potency, and considered as revealing a considerable preservation of the emotional tone pattern in the English human translation.

The same correlations were calculated between the English computer translation and the French original. The intercorrelations within the French original were, of course, the same as had been observed before. The correlations of the French original text scores, emotionality and pleasantness, with the English translation scores of evaluation were 0.32 and 0.44 respectively, both significant at the level p<.01, df=92, one-tailed. These are larger values than the corresponding ones obtained before. The correlations of potency with evaluation, emotionality and pleasantness were -0.57, -0.11 and -0.08 respectively. In this case only that of dimensions within the English computer translation reached significance, the level being p<.01, df=92, one- tailed. So here potency correlations did not add evidence for the agreement of emotional tone patterns between the two texts, although the direct comparisons clearly did so. From the evaluation, emotionality and pleasantness agreements it appears that the computer translation had preserved the emotional tone pattern better than the human translation.

The differences in levels of emotional tone pattern agreement are not large, but they do fail to support the hypotheses with which the study was undertaken. The correlations between corresponding emotional tone dimensions was not greater than those with non-corresponding dimensions. The human translation correspondences were not higher than those achieved by computer translation.

If attention is focused on the corresponding dimensions in English and French it should be remembered that these scores were obtained entirely automatically and with separate dictionaries which were independently created using English subjects in the United States, and French subjects in Canada and Belgium. The two French measures correlate substantially with each other, yet they are independent enough that their relations with the English evaluation can both be taken as evidence of pattern similarity on this basic dimension of emotional tone.

These results suggest that a computer translation from French to English can retain the emotional tone of the original as well as a human translation, or better. This should be tested with another long text which has not the bloodthirsty quality of Saint Julian the Hospitaler, both to re-examine the comparative power of the computer translation to preserve emotional tone and to see if some of the large correlations with potency would disappear. The same research pattern should be extended to an examination of the preservation of emotional tone patterns in computer translation from English to French.

References

Anderson, C.W. (2000). Agreement between operator-assisted and fully automated scoring of emotional tone levels. Poster session submitted for the annual convention of the Canadian Psychological Association, Ottawa, ON, June.

Anderson, C.W. and McMaster, GE (1992). Connotation of emotional tone in French and English words and stories. Poster session presented at the annual convention of the Canadian Psychological Association, Quebec City, PQ, June.

Anderson, C.W. and McMaster, G. E. (1993). Pantextual indices of emotional tone. In Nardocchio, E. (Ed.), Reader Response to Literature. NY: Mouton de Gruyter, 35-56.

Anderson, C.W., McMaster, G. E, and Mitchell, T. A. (1987). Computer assisted enhancement of emotional tone accuracy in translated narrative. Literary and Linguistic Computing, 2, 1-6.

Berg, W.J., Moskos, G., and Grimaud, M. (1982). Saint/Oedipus, Psychocritical Approaches to Flaubert's Art. Ithica, NY: Cornell University Press.

Flaubert, G. (1890). "La L�gende de Saint Julien L'Hospitalier" from Trois Contes. Paris: G. Charpentier et Cie, Editeurs, Rue de Grenelle, 11, 89-164.

Gurney, P.J. and Gurney, L.W. (1998). Subsets and homogeneity: Authorship attribution in the Scriptores Historiae Augustae. Literary and Linguistic Computing, 13, No 3, 133-140.

Heise, P. R. (1965). Semantic differential profiles for 1000 most frequent English words. Psychological Monographs, 79, whole of 601.

Messina, D., Morais, J., and Contraine, F. (1989). Valeur affective de 904 mots de la langue Fran�aise. European Bulletin of Cognitive Psychology. 9, 2, 165-187.

SYSTRAN Translation Software (1999). AltaVista Connections Translation with SYSTRAN. Available at: http://bablefish.altavista.digital.com/cgi- bin/translate?

Vikis-Freibergs, V. (1976). Abstractness and emotionality values for 398 French words. Canadian Journal of Psychology, 30, 1, 22-30.