Context-aware and gender-neutral Translation Memories

Context-aware and gender-neutral Translation Memories 1Marjolene Paulo, 1Vera Cabarrão, 2, 3Helena Moniz 2, 3Miguel Menezes, 1Rachel Grewcock, 1Eduardo Farah 1 Unbabel, Lisbon, Portugal 2 University of Lisbon, Portugal 3 INESC-ID, Lisbon Portugal {marjolene.paulo.int, vera.cabarrao, helena, luis.menezes.ext, rachel.grewcock, eduardo.farah}@unbabel.com Abstract This work proposes an approach to use Part-of-Speech (POS) information to auto- matically detect context-dependent Trans- lation Units (TUs) from a Translation Memory database pertaining to the cus- tomer support domain. In line with our goal to minimize context-dependency in TUs, we show how this mechanism can be deployed to create new gender-neutral and context-independent TUs. Our exper- iments, conducted across Portuguese (PT), Brazilian Portuguese (PT-BR), Spanish (ES), and Spanish-Latam (ES-LATAM), show that the occurrence of certain POS with specific words is accurate in iden- tifying context dependency. In a cross- client analysis, we found that 10% of the most frequent 13,200 TUs were context-dependent, with gender determin- ing context-dependency in 98% of all con- firmed cases. We used these findings to suggest gender-neutral equivalents for the most frequent TUs with gender con- straints. Our approach is in use in the Un- babel translation pipeline, and can be inte- grated into any other NMT pipeline. 1 Introduction Translation Memory (TM) servers are dynamic databases that store frequent and high-quality translated segments (Bowker and Fisher, 2010). In the context of Computer-aided translation (CAT) tools and in the Machine Translation (MT) indus- try, these segments are used as complementary © 2022 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. modules, to optimize the translation process, re- ducing costs and increasing the speed of transla- tions. A downside of these systems is that they are usually based on sentence-like units. Combining the sentences risks coherence problems at the doc- ument level. This is problematic as a high-quality text is a coherent and cohesive unit with ideas pre- sented in a logical way. Recent work in this field aims to achieve a context-aware MT system by incorporating more context than the current sentence and tackling dis- course phenomena across the document. Our use case is customer-support content: a con- tent type that differs significantly from other types of text since it represents an interaction between an agent and a customer, usually over email or chat messages. This implies that the text contains a lot of first and second person, and pronouns or anaphora. Even third person references can also be gendered in the target languages in question (e.g., "Está encantada?" "Está encantado?" / Are you thrilled?). Since the translation process is carried out sentence by sentence, gender information asso- ciated with that of the addressee is easily lost, re- sulting in agreement problems throughout the doc- ument. We hypothesize that, by using POS information, we can automatically identify context-dependent and context-independent segments. With the results of context identification, we can then transform segments with gender constraints into gender-neutral TUs, so that these can be re- used without damaging translation quality through agreement errors. As such, …

Content truncated to 500 words