使用Amazon Textract解析多页PDF，如何获取LAYOUT类BLOCK内容？

我尝试用Amazon Textract抓取多页PDF，需要将内容按章节、子章节、表格格式化为JSON。通过Textract UI演示（开启LAYOUT和Table功能）能正常显示布局标题、章节、文本等信息，下载的layout.csv和analyzeDocResponse.json也包含LAYOUT相关数据。

编写代码1打印Block完整JSON结构时能看到LAYOUT类块，但代码2打印Block类型对应文本时，仅LINES、WORDS类型有内容，LAYOUT_TITLE、LAYOUT_SECTION_HEADER等LAYOUT类块的Text字段为空。我是Textract新手，查过文档和相关问题没找到解法，寻求帮助。测试PDF为药品SmPC文件。

代码1 #

def start_textract_job(bucket, document):
    response = textract.start_document_analysis(
        DocumentLocation={
            'S3Object': {
                'Bucket': bucket,
                'Name': document
            }
        },
        FeatureTypes=["LAYOUT"]  # 可根据需求调整FeatureTypes
    )
    return response['JobId']


def print_blocks(job_id):
    next_token = None
    while True:
        if next_token:
            response = textract.get_document_analysis(JobId=job_id, NextToken=next_token)
        else:
            response = textract.get_document_analysis(JobId=job_id)

        for block in response.get('Blocks', []):
            print(json.dumps(block, indent=4))

        next_token = response.get('NextToken', None)
        if not next_token:
            break

代码2 #

# start_textract_job函数与代码1相同，启用LAYOUT

def print_blocks(job_id):
    next_token = None
    while True:
        if next_token:
            response = textract.get_document_analysis(JobId=job_id, NextToken=next_token)
        else:
            response = textract.get_document_analysis(JobId=job_id)

        for block in response.get('Blocks', []):
            print(f"{block['BlockType']}: {block.get('Text', '')}")

        next_token = response.get('NextToken', None)
        if not next_token:
            break

LAYOUT类型的Block（如LAYOUT_TITLE、LAYOUT_SECTION_HEADER）本身不存储文本内容，它们是布局容器块，仅用于标记文档的结构区域。实际文本内容存储在它们关联的子Block（如WORDS、LINES类型）中，通过Block的Relationships字段建立关联。

要获取LAYOUT类块对应的文本，需按以下步骤处理：

1. 先遍历所有Block，建立以BlockId为键的字典，方便快速查找关联子块
2. 对每个LAYOUT类块，通过其Relationships中的CHILD类型提取关联的子Block ID
3. 从字典中取出这些子Block，拼接它们的Text内容

修改后的print_blocks函数示例：

def print_blocks(job_id):
    next_token = None
    block_dict = {}
    
    # 先收集所有Block到字典
    while True:
        if next_token:
            response = textract.get_document_analysis(JobId=job_id, NextToken=next_token)
        else:
            response = textract.get_document_analysis(JobId=job_id)
        
        for block in response.get('Blocks', []):
            block_dict[block['Id']] = block
        
        next_token = response.get('NextToken', None)
        if not next_token:
            break
    
    # 遍历所有Block，处理LAYOUT类块
    for block_id, block in block_dict.items():
        block_type = block['BlockType']
        if block_type.startswith('LAYOUT_'):
            # 获取关联的子Block ID
            child_ids = []
            if 'Relationships' in block:
                for rel in block['Relationships']:
                    if rel['Type'] == 'CHILD':
                        child_ids.extend(rel['Ids'])
            
            # 拼接子Block的文本内容
            text_parts = []
            for child_id in child_ids:
                child_block = block_dict.get(child_id)
                if child_block and 'Text' in child_block:
                    text_parts.append(child_block['Text'])
            
            full_text = ' '.join(text_parts)
            print(f"{block_type}: {full_text}")
        else:
            # 保留其他类型Block的原有打印逻辑
            print(f"{block_type}: {block.get('Text', '')}")

通过这种方式就能正确获取LAYOUT类块对应的文本，后续可基于这些布局块的层级关系，组织出章节、子章节结构，最终生成符合要求的JSON格式文档。

内容的提问来源于stack exchange，提问作者Santhosh