Mittetähistatud

Mittetähistatud is an Estonian term meaning "unlabeled" or "not marked." It is commonly used in data science, machine learning, and linguistics to describe data that lacks explicit annotations such as class labels, tags, or segmentation boundaries. The opposite term tähistatud denotes data that has been annotated or labeled. Mittetähistatud data can include text corpora, images, audio, or video and is often abundant compared with labeled data.

In natural language processing and computer vision, mittetähistatud data is used for self-supervised and unsupervised learning. Language models can be trained on unlabeled text using objectives like masked prediction or next-token prediction; image and audio models can employ contrastive learning, autoencoding, or predictive tasks that do not require labels. Such approaches enable the extraction of meaningful representations from large datasets without the cost of manual labeling.

Challenges and strategies: Because there is no ground-truth labeling, evaluating models trained on mittetähistatud data relies on performance on downstream supervised tasks or proxy metrics. To leverage unlabeled data effectively, researchers use semi-supervised learning, where a small labeled subset guides training; self-supervised learning, where signals are derived from the data itself; and techniques like pseudo-labeling, data augmentation, or representation learning to improve applicability to real tasks.

See also: Labeled data, Annotation, Semi-supervised learning, Unsupervised learning, Pseudo-labeling, Self-supervised learning.