Text annotation for NLP and document processing: a complete guide

What is Text Annotation in Machine Learning?

Simply put, text annotation in machine learning (ML) is the process of assigning labels to a digital file or document and its content. Labeling a text can consist in assigning tags to text attributes such as keywords, sentences, and paragraphs or simply classifying the text based on its content (i.e. text classification).

These include various NLP technologies like neural machine translation (NMT) programs, auto Q&A (question and answer) platforms, smart chatbots, sentiment analysis, text-to-speech synthesizers, and auto speech recognition (ASR) tools, among other related projects. These technologies can streamline the activities and transactions of many organizations across different industries.

If you want to learn more about the history of NLP in computer science, you can read our article here.

What are the different types of text annotations?

Text / Document classification

Text and document classification consists in attributing one or multiple attributes to a single text or full document.

Examples:

1. 1. Classifying emails as spam or regular emails

2. 2. Doing sentiment analysis on tweets

3. 3. Labeling legal documents based on their content (legal notices, agreements, bonds, …)

Named entity recognition

At a high level, named entity recognition is the action of identifying named entities within a text and assigning it a predefined category. Common categories that are used for this type of text annotation include names of organizations, locations, persons, numerical values, month or time and day of the week, etc. But depending on the type of NER performed, categories such as paragraph, title, and content can also be used.

Entity linking

This is the process for linking entities to better understand the structure of the text and the relationships between entities. Here you can see a table being extracted from a scanned invoice and item lines being linked together to preserve its structure.

Layout analysis

Layout analysis consists of labeling document structures to transform them into another format (ex: JSON).

How to label texts, PDFs, and Images?

In real life, textual data exists under a wide range of different formats txt, pdf or even text in images or scanned documents. In this part are going to dive deep into the specificities of those data formats and what features are mandatory to label efficiently.

Labeling text data

When labeling text data in simple txt format, the following features are important:

Multilingual Support

When labeling textual data, language and tokenization are key to consider. The labeling tool used must allow the labeling of all languages. Using Kili, you can natively label any language using our interfaces.

Learn more!

Discover how training data can make or break your AI projects, and how to implement the Data Centric AI philosophy in your ML projects.

Download White Paper

Propagate regex from UI

When labeling long documents, a time saver is to tag all equal tokens in one click automatically. We call this feature regex propagation. In the example below, you will see that the name of the company RentalCo is identified once and propagated by the labeler. A good way of accelerating the NER process is to generate pre-annotations.

Formatting text for ease of labeling

Labeling PDFs

PDF documents can be a lot more complex than simple text documents. Often, PDFs are made from a couple to hundreds of pages and they contain other data types such as images, tables, and graphs. When choosing a tool to label PDFs you should have in mind the following features to ensure success:

Native PDFs vs. Scanned PDFs

A native PDF is a PDF of a document that was “born digital” because the PDF was created from an electronic version of a document, rather than from print. A scanned PDF, by contrast, is a PDF of a print document, such as when you scan in pages from a print journal and then save this file as a PDF.

When labeling a Native PDF, the content of the PDF can be directly extracted by using the underlining text layer (i.e text can be selected from the PDF). On a scanned PDF this layer might not be present. In Kili, you can add OCRs to the scanned PDF to label it just as a native one. Now, you just need one interface to label all of your PDF documents!

Page-level classification vs. Document-level classification

Because PDFs can contain multiple pages, it’s important you can make classifications at the page level and document level.

For example, let's say you want to label the type of document (claim, invoice, quote, …) and for each document, you also want to classify each page to differentiate the main document from the attachments. Tooling needs to be adapted for the metadata to be extracted.

Extracting mixed content (images, tables, graphs, stamps, signatures, …)

PDF documents also contain elements other than texts, such as images, tables, or signatures. In various use cases, such as automatic document parsing and understanding, it is mandatory to be able to extract such elements during the annotation process.

Therefore, you need to have on your PDF interface the possibility to use both text annotation tools like classification NER and transcription and also image annotation tools like bounding boxes.

Labeling images with OCR

Lastly, you might also need to perform text annotation to extract texts from images, like extracting license plates video streams, or reading traffic signs.

If this is the case being able to use an OCR model with the labeling interface will be mandatory to speed up drastically labeling times. With Kili, you can use any OCR model that you want like Google Vision API, Tesseract OCR, Azure OCR, Amazon Textract, or your own custom model!

Going further

Text annotation is the backbone of AI applications handling text under all its aspects. To scale annotation and produce reliable AI models in production data needs to be of the highest quality and relevance. Our mission at Kili is to help organizations tackle the bottleneck of training data quality. To do so, we are building a state-of-the-art data labeling platform with high labeling interfaces supported by data quality monitoring and automation.

If you want to learn more about data quality and labeling automation, go check out our documentation:

• https://docs.kili-technology.com/docs/introduction-to-kili-technology

• https://docs.kili-technology.com/docs/model-based-pre-annotation