Retain unique PDF formatting to view translated paperwork with Amazon Textract, Amazon Translate, and PDFBox

Corporations throughout varied industries create, scan, and retailer massive volumes of PDF paperwork. In lots of circumstances, the content material is text-heavy and sometimes written in a unique language and requires translation. To handle this, you want an automatic resolution to extract the contents inside these PDFs and translate them shortly and cost-efficiently.

Many companies have various world customers and have to translate textual content to allow cross-lingual communication between them. It is a guide, sluggish, and costly human effort. There’s a have to discover a scalable, dependable, and cost-effective resolution to translate paperwork whereas retaining the unique doc formatting.

For verticals similar to healthcare, as a consequence of regulatory necessities, the translated paperwork require an extra human within the loop to confirm the validity of the machine-translated doc.

If the translated doc doesn’t retain the unique formatting and construction, it loses its context. This will make it troublesome for a human reviewer to validate and make corrections.

On this submit, we reveal easy methods to create a brand new translated PDF from a scanned PDF whereas retaining the unique doc construction and formatting utilizing a geometry-based method with Amazon Textract, Amazon Translate, and Apache PDFBox.

Resolution overview

The answer offered on this submit makes use of the next elements:

Amazon Textract – A totally managed machine studying (ML) service that robotically extracts printed textual content, handwriting, and different information from scanned paperwork that goes past easy optical character recognition (OCR) to determine, perceive, and extract information from kinds and tables. Amazon Textract can detect textual content in quite a lot of paperwork, together with monetary reviews, medical information, and tax kinds.
Amazon Translate – A neural machine translation service that delivers quick, high-quality, and inexpensive language translation. Amazon Translate supplies high-quality on-demand and batch translation capabilities throughout greater than 2,970 language pairs, whereas reducing your translation prices.
PDF Translate – An open-source library written in Java and printed on AWS Samples in GitHub. This library incorporates logic to generate translated PDF paperwork in your required language with Amazon Textract and Amazon Translate. It additionally makes use of the open-source Java library Apache PDFBox to create PDF paperwork. There are related PDF processing libraries obtainable in different programming languages, for instance Node PDFBox.

Whereas performing machine translations, you’ll have conditions the place you want to protect particular sections of textual content from being translated, similar to names or distinctive identifiers. Amazon Translate permits tag modifications, which lets you specify what textual content shouldn’t be translated. Amazon Translate additionally helps formality customization, which lets you customise the extent of ritual in your translation output.

For particulars on Amazon Textract limits, confer with Quotas in Amazon Textract.

The answer is restricted to the languages that may be extracted by Amazon Textract, which at the moment helps English, Spanish, Italian, Portuguese, French, and German. These languages are additionally supported by Amazon Translate. For the complete record of languages supported by Amazon Translate, confer with Supported languages and language codes.

We use the next PDF to reveal translating the textual content from English to Spanish. The answer additionally helps producing the translated doc with none formatting. The place of the translated textual content is maintained. The supply and translated PDF paperwork will also be discovered within the AWS Samples GitHub repo.

Within the following sections, we reveal easy methods to run the interpretation code on a neighborhood machine and have a look at the interpretation code in additional element.

Stipulations

Earlier than you get began, arrange your AWS account and the AWS Command Line Interface (AWS CLI). For entry to any AWS Providers similar to Textract and Translate, applicable IAM permissions are wanted. We suggest using least privilege permissions. To be taught extra about IAM permissions see Policies and permissions in IAM in addition to How Amazon Textract works with IAM and How Amazon Translate works with IAM.

Run the interpretation code on a neighborhood machine

This resolution focuses on the standalone Java code to extract and translate a PDF doc. That is for simpler testing and customizations to get the best-rendered translated PDF doc. The code can then be built-in into an automatic resolution to deploy and run in AWS. See Translating PDF documents using Amazon Translate and Amazon Textract for a pattern structure that makes use of Amazon Simple Storage Service (Amazon S3) to retailer the paperwork and AWS Lambda to run the code.

To run the code on a neighborhood machine, full the next steps. The code examples can be found on the GitHub repo.

Clone the GitHub repo:

git clone https://github.com/aws-samples/amazon-translate-pdf

Run the next command:

Run the next command to translate from English to Spanish:

java -jar goal/translate-pdf-1.0.jar --source en --translated es

Two translated PDF paperwork are created within the paperwork folder, with and with out the unique formatting (SampleOutput-es.pdf and SampleOutput-min-es.pdf).

Code to generate the translated PDF

The next code snippets present easy methods to take a PDF doc and generate a corresponding translated PDF doc. It extracts the textual content utilizing Amazon Textract and creates the translated PDF by including the translated textual content as a layer to the picture. It builds on the answer proven within the submit Generating searchable PDFs from scanned documents automatically with Amazon Textract.

The code first will get every line of textual content with Amazon Textract. Amazon Translate is used to get translated textual content and save the geometry of the translated textual content.

Area area = Area.US_EAST_1;
TextractClient textractClient = TextractClient.builder()
        .area(area)
        .construct();

// Get the enter Doc object as bytes
Doc pdfDoc = Doc.builder()
        .bytes(SdkBytes.fromByteBuffer(imageBytes))
        .construct();

TranslateClient translateClient = TranslateClient.builder()
        .area(area)
        .construct();

DetectDocumentTextRequest detectDocumentTextRequest = DetectDocumentTextRequest.builder()
        .doc(pdfDoc)
        .construct();

// Invoke the Detect operation
DetectDocumentTextResponse textResponse = textractClient.detectDocumentText(detectDocumentTextRequest);

Checklist<Block> blocks = textResponse.blocks();
Checklist<TextLine> strains = new ArrayList<>();
BoundingBox boundingBox;

for (Block block : blocks) {
    if ((block.blockType()).equals(BlockType.LINE)) {
        String supply = block.textual content();

        TranslateTextRequest requestTranslate = TranslateTextRequest.builder()
                .sourceLanguageCode(sourceLanguage)
                .targetLanguageCode(destinationLanguage)
                .textual content(supply)
                .construct();

        TranslateTextResponse resultTranslate = translateClient.translateText(requestTranslate);

        boundingBox = block.geometry().boundingBox();
        strains.add(new TextLine(boundingBox.left(),
                boundingBox.high(),
                boundingBox.width(),
                boundingBox.peak(),
                resultTranslate.translatedText(),
                supply));
    }
}
return strains;

The font dimension is calculated as follows and may simply be configured:

int fontSize = 20;
float textWidth = font.getStringWidth(textual content) / 1000 * fontSize;
float textHeight = font.getFontDescriptor().getFontBoundingBox().getHeight() / 1000 * fontSize;
 
if (textWidth > bbWidth) {
    whereas (textWidth > bbWidth) {
        fontSize -= 1;
        textWidth = font.getStringWidth(textual content) / 1000 * fontSize;
        textHeight = font.getFontDescriptor().getFontBoundingBox().getHeight() / 1000 * fontSize;
     }
} else if (textWidth < bbWidth) {
     whereas (textWidth < bbWidth) {
         fontSize += 1;
         textWidth = font.getStringWidth(textual content) / 1000 * fontSize;
         textHeight = font.getFontDescriptor().getFontBoundingBox().getHeight() / 1000 * fontSize;
      }
}

The translated PDF is created from the saved geometry and translated textual content. Modifications to the colour of the translated textual content can simply be configured.

float width = picture.getWidth();
float peak = picture.getHeight();
 
PDRectangle field = new PDRectangle(width, peak);
PDPage web page = new PDPage(field);
web page.setMediaBox(field);
this.doc.addPage(web page); //org.apache.pdfbox.pdmodel.PDDocument
 
PDImageXObject pdImage;
 
if(imageType == ImageType.JPEG){
    pdImage = JPEGFactory.createFromImage(this.doc, picture);
} else {
    pdImage = LosslessFactory.createFromImage(this.doc, picture);
}
 
PDPageContentStream contentStream = new PDPageContentStream(doc, web page, PDPageContentStream.AppendMode.OVERWRITE, false);
 
contentStream.drawImage(pdImage, 0, 0);
contentStream.setRenderingMode(RenderingMode.FILL);
 
for (TextLine cline : strains){
    String clinetext = cline.textual content;
    String clinetextOriginal = cline.originalText;
                       
    FontInfo fontInfo = calculateFontSize(clinetextOriginal, (float) cline.width * width, (float) cline.peak * peak, font);
    //config to incorporate unique doc construction - overlay with unique
    contentStream.setNonStrokingColor(Coloration.WHITE);
    contentStream.addRect((float) cline.left * width, (float) (peak - peak * cline.high - fontInfo.textHeight), (float) cline.width * width, (float) cline.peak * peak);
    contentStream.fill();
 
    fontInfo = calculateFontSize(clinetext, (float) cline.width * width, (float) cline.peak * peak, font);
    //config to incorporate unique doc construction - overlay with translated
    contentStream.setNonStrokingColor(Coloration.WHITE);
    contentStream.addRect((float) cline.left * width, (float) (peak - peak * cline.high - fontInfo.textHeight), (float) cline.width * width, (float) cline.peak * peak);
    contentStream.fill();
    //change the output textual content shade right here
    fontInfo = calculateFontSize(clinetext.size() <= clinetextOriginal.size() ? clinetextOriginal : clinetext, (float) cline.width * width, (float) cline.peak * peak, font);
    contentStream.setNonStrokingColor(Coloration.BLACK);
    contentStream.beginText();
    contentStream.setFont(font, fontInfo.fontSize);
    contentStream.newLineAtOffset((float) cline.left * width, (float) (peak - peak * cline.high - fontInfo.textHeight));
    contentStream.showText(clinetext);
    contentStream.endText();
}
contentStream.shut()

The next picture exhibits the doc translated into Spanish with the unique formatting (SampleOutput-es.pdf).

The next picture exhibits the translated PDF in Spanish with none formatting (SampleOutput-min-es.pdf).

Processing time

The employment software pdf took about 10 seconds to extract, course of and render the translated pdf. The processing time for textual content heavy doc such because the Declaration of Independence PDF took lower than a minute.

Value

With Amazon Textract, you pay as you go primarily based on the variety of pages and pictures processed. With Amazon Translate, you pay as you go primarily based on the variety of textual content characters which are processed. Consult with Amazon Textract pricing and Amazon Translate pricing for precise prices.

Conclusion

This submit confirmed easy methods to use Amazon Textract and Amazon Translate to generate translated PDF paperwork whereas retaining the unique doc construction. You’ll be able to optionally postprocess Amazon Textract outcomes to enhance the standard of the interpretation, for instance extracted phrases may be handed by ML-based spellchecks similar to SymSpell for information validation, or clustering algorithms can be utilized to protect studying order. You may as well use Amazon Augmented AI (Amazon A2I) to construct human overview workflows the place you should utilize your individual non-public workforce to overview the unique and translated PDF paperwork to supply extra accuracy and context. See Designing human review workflows with Amazon Translate and Amazon Augmented AI and Building a multi-lingual document translation workflow with domain-specific and language-specific customization to get began.

Concerning the Authors

Anubha Singhal is a Senior Cloud Architect at Amazon Internet Providers within the AWS Skilled Providers group.

Sean Lawrence was previously a Entrance Finish Engineer at AWS. He specialised in entrance finish growth within the AWS Skilled Providers group and the Amazon Privateness group.