How to formulate Textract queries for documents in languages other than English?

0

I have been testing Textract on German insurance documents and text extraction seems to work pretty well. However, I'm not sure what is the best practice for extracting values using queries for best results?

  • Shall the queries be formulated in English? (e.g. "What is the customer's insurance policy number?")
  • Shall the queries be formulated in English with German field names? (e.g. "What is the customer's Versicherungsscheinnummer or VSNR?")
  • Shall the queries be completely formulated in German?

Thank you for your tips!

feita há um ano613 visualizações
3 Respostas
1

Hello I had the same problem as you but with French. In the beginning, my query was " What is the city or zip-code on this page? " but the query couldn't find the city on every page. So I try to be more explicit and finally, my new query is " This page is part of a French contract for selling a house. What is the city or zip-code of this house? " and it works all the time. You need to talk to Textract.

respondido há 9 meses
  • This is fascinating, we'll try to add some context to our queries, thanks! Never thought that would make a (positive) difference.

  • That is proper prompt-engineering. Thx @Andemion!

0
Resposta aceita

At the moment Textract Queries officially only supports English (see Textract Quotas). However, I have seen success in using other languages as well, but there is no official support and no official statement on what works best. At the moment testing and evaluating is your best option.

AWS
respondido há um ano
  • Thank you! Just to avoid opening yet another related question: the Quotas don't say anything about language limitations of receipt analysis API. Does it mean that receipts in German are officially supported?

0

The results of some limited testing on German documents:

  • Making queries in German seems to be possible, but generally leads to worse confidence
  • Interestingly, possessives improve confidence (company's name vs. company name)
  • Making queries in English with German nouns seems to give best confidence
  • Making English queries with translated nouns doesn't seem to work
  • Umlauts in queries are not supported

Not bad for an easy start!

respondido há um ano

Você não está conectado. Fazer login para postar uma resposta.

Uma boa resposta responde claramente à pergunta, dá feedback construtivo e incentiva o crescimento profissional de quem perguntou.

Diretrizes para responder a perguntas