Publicação
Analysis on the impact of the source text quality: Building a data-driven typology
| Resumo: | In this study we propose a typology which concerns source errors and linguistic structures that might have an impact on Machine Translation (MT). Although most typologies are built on a bilingual level, the source text (ST) also presents issues that cannot be expected to be resolved by MT. In this study, we were able to test whether or not the quality of the ST has an impact on the target text (TT) quality. For that purpose, source data was annotated. The data analyzed was both inbound (user-generated content) and outbound (agent) in the context of chat. Through this analysis, it was possible to build a data driven typology. To aid the construction of a new typology, there was also a comparison between multiple typologies, whether they have a bilingual or a monolingual focus. This allowed us to see what could be applied to a monolingual typology and what was missing. With the annotation results, it was possible to build a new typology — Source Typology. To assist future annotators, we provided annotation guidelines with a listing of all the issue types, an explanation of the different span types, the severities to be used and the tricky cases that might occur during the annotation process. In order to test the reliability of the typology, three different case studies of an internal pilot were conducted. Each case study had a different goal and took into account different language pairs. By testing the Source Typology, we could see its effectiveness and reliability and what should be improved. In the end, we demonstrated that the quality of the ST can actually have an impact on the TT quality, where, at times, minor errors on the source would become or originate critical errors on the target. The typology is now being applied at Unbabel. |
|---|---|
| Autores principais: | Gonçalves, Madalena Sofia Nunes |
| Assunto: | Tradução automática Erros de linguagem Tradução Teses de mestrado - 2021 |
| Ano: | 2021 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório da Universidade de Lisboa |
| Resumo: | In this study we propose a typology which concerns source errors and linguistic structures that might have an impact on Machine Translation (MT). Although most typologies are built on a bilingual level, the source text (ST) also presents issues that cannot be expected to be resolved by MT. In this study, we were able to test whether or not the quality of the ST has an impact on the target text (TT) quality. For that purpose, source data was annotated. The data analyzed was both inbound (user-generated content) and outbound (agent) in the context of chat. Through this analysis, it was possible to build a data driven typology. To aid the construction of a new typology, there was also a comparison between multiple typologies, whether they have a bilingual or a monolingual focus. This allowed us to see what could be applied to a monolingual typology and what was missing. With the annotation results, it was possible to build a new typology — Source Typology. To assist future annotators, we provided annotation guidelines with a listing of all the issue types, an explanation of the different span types, the severities to be used and the tricky cases that might occur during the annotation process. In order to test the reliability of the typology, three different case studies of an internal pilot were conducted. Each case study had a different goal and took into account different language pairs. By testing the Source Typology, we could see its effectiveness and reliability and what should be improved. In the end, we demonstrated that the quality of the ST can actually have an impact on the TT quality, where, at times, minor errors on the source would become or originate critical errors on the target. The typology is now being applied at Unbabel. |
|---|