Controlling Extra-Textual Attributes about Dialogue Participants: A Case Study of English-to-Polish Neural Machine Translation

Controlling Extra-Textual Attributes about Dialogue Participants: A Case Study of English-to-Polish Neural Machine Translation Sebastian T. Vincent, Loïc Barrault, Carolina Scarton Department of Computer Science, University of Sheffield Regent Court, 211 Portobello, Sheffield, S1 4DP, UK {stvincent1, l.barrault, c.scarton}@shef.ac.uk Abstract Unlike English, morphologically rich lan- guages can reveal characteristics of speak- ers or their conversational partners, such as gender and number, via pronouns, mor- phological endings of words and syntax. When translating from English to such languages, a machine translation model needs to opt for a certain interpretation of textual context, which may lead to seri- ous translation errors if extra-textual in- formation is unavailable. We investigate this challenge in the English-to-Polish lan- guage direction. We focus on the un- derresearched problem of utilising exter- nal metadata in automatic translation of TV dialogue, proposing a case study where a wide range of approaches for control- ling attributes in translation is employed in a multi-attribute scenario. The best model achieves an improvement of +5.81 chrF++/+6.03 BLEU, with other models achieving competitive performance. We additionally contribute a novel attribute- annotated dataset of Polish TV dialogue and a morphological analysis script used to evaluate attribute control in models. 1 Introduction In some languages, dialogue explicitly expresses certain information about the interlocutors: for example, while in English words describing the speaker “I” and the interlocutor “you” are ambigu- ous w.r.t. their gender, number and formality, lan- guages such as Polish, German or Spanish will © 2022 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. mark for one or more of these attributes. In in- dustrial settings such as dubbing and speech trans- lation, there is an abundance of available metadata about the interlocutors, such as their genders, that could be used to help resolve these ambiguities. Field Value source "Are you blind?" spoken by (=speaker) "Anne" speaker’s gender "feminine" spoken to (=interlocutor(s)) ["Mark", "Colin"] interlocutor(s)’ gender "masculine" formality "informal" Table 1: A TV segment along with available metadata. Table 1 shows an example of such a TV seg- ment: the English sentence ‘Are you blind?’, should translate to Polish as ‘Jeste´scie ´slepi?’ as the addressee is a group of men and the setting is informal; however, when spoken e.g. formally to a mixed-gender group of people, the correct transla- tion would read ‘S ˛a pa´nstwo ´slepi?’, using a differ- ent verb inflection and an honorific pa´nstwo. Since the contextual information required to resolve the ambiguity in this example does not belong to the text itself, traditional models do not use it. This yields hypotheses which introduce some assump- tions about that context, typically reflecting biases present in the (often unbalanced) training data. To avoid this, a better solution is to resolve such am- biguities by using both the available metadata and the source text as translation input. Alternatively, when such information is unavailable, all possible contextual variants could be provided as output, passing the choice from the model to the user (Ja- covi et …

Content truncated to 500 words