Unstructured and Semistructured Data

Imagine a phone that has collected ten thousand photos and several hundred short videos over the years. Each file has a timestamp, but none of them say what is in the picture. Scroll through the library and you will find beaches, receipts, pets, and street signs, yet nothing in the files themselves tells a search tool which is which.

The same problem grows as the amount of data grows. The source points out that the volume of data that people and organizations produce keeps rising, and that most of it does not arrive in neat rows and columns. Multiply the phone's library by every company, every customer record, every system log, and every signed contract, and the challenge becomes organizing a very large body of material that a search tool cannot read in its raw form. This article follows how AI helps turn that raw material into something that can be found later.

The source describes a general direction for the process. Content that has no fixed structure goes in. AI produces either a description or a set of extracted fields. The result is stored so that a later search can reach it. The diagram below traces that path for a single file, from upload to the moment a query can find it.

Read the diagram from left to right. The raw file on the left has no fixed fields, so a search tool has nothing to match against. The AI step in the middle is where the file gains searchable form. Its output, whether a text description or structured fields, is what gets stored. Only that stored output is available when a query arrives at the right-hand end.

The two words describe different degrees of organization. Unstructured content, such as a photo or a video, has no fixed set of fields. The meaning lives inside the material itself, and a search tool cannot read that meaning without help. Semistructured content carries some organization, such as labeled parts or tagged fields, but the full meaning still is not presented as clean rows and columns. The source groups logs and PDF documents together as examples of unstructured or semistructured data, so the boundary between the two terms is not sharp for every file type.

To make the difference concrete, consider a generated illustration that is not taken from the source. A photo of a beach has no labeled parts at all. A log entry that records a login event might contain a tagged timestamp, a user field, and a message field. A search tool can look at the labeled parts of the log entry directly, but it cannot look inside the beach photo the same way.

The source describes two related jobs that AI performs for organizing data. The first is generating text descriptions for images and videos automatically. Instead of a person typing a caption for every file, a model produces a written account of what the file contains. Because that description is text, it can be stored and searched with the same tools used for words.

The second job is matching text queries with visuals. A person types a phrase, such as a description of a scene or an object, and the system returns images that fit that phrase. Together these two jobs turn visual content, which is hard to search directly, into something that behaves like a searchable index.

The diagram shows the core idea. The image never gets compared with the query as raw pixels. Instead, the image has been turned into text beforehand, and the query is matched against that text. The source does not describe the matching method itself, so the learner's takeaway is the flow: the query meets stored text, and the image is returned because of the description that was generated for it.

The source gives two real-world illustrations. Google Photos uses AI to surface images that match a search query, so a user who types a phrase finds pictures without having labeled them by hand. Google Image Search goes a step further. If there is no existing image that matches what a user needs, it can generate one.

The following example is an original illustration rather than a case from the source. A company receives scanned receipts as images. Each receipt is unstructured: the store name, date, items, and total are all present, but they sit in a free layout with no fixed fields. Extraction aims to read those four kinds of information off each image and place them into labeled fields.

The source groups extraction tasks by difficulty. Simple use cases include automatically extracting information from credit cards, driver's licenses, receipts, tickets, and contact information found in email footers. These documents tend to have recognizable patterns, which makes the extraction task comparatively contained.

More complex use cases include extracting data from contracts, reports, and charts. These sources carry more varied layouts and richer meaning, so the extraction task is harder. The source uses the term Intelligent Data Processing, abbreviated IDP, for the industry built around this kind of AI-driven extraction.

Organizing data and analyzing it are closely connected. The source notes that AI is very good with data analysis and that it can write programs that generate data visualizations, identify outliers, and make predictions such as revenue forecasts. Each of these tasks takes data that has already been gathered and asks a question of it.

An outlier is a value that sits far from the rest of a dataset. Spotting outliers helps a team notice errors or unusual events in their records. A revenue forecast uses past figures to estimate future ones. The source names these as examples of what AI can do, and it does not describe the specific methods used, so this article keeps to the tasks themselves.

Extraction, description, and search therefore form one chain. Extraction and description make raw material organized, search lets people reach that organized material, and analysis asks questions of the data once it has been gathered. Organizing comes first in the chain because none of the later questions can be answered until the material is in a usable form.