Here are described the main concepts related to TEXTA Toolkit (TTK).


Anonymizer is a tool for anonymizing names in a text-based on a predefined list of names. Each name detected from the text will be substituted with randomly generated pair of initials. You can read more about it here.

Bert Tagger

Bert Tagger uses Google’s state-of-the-art Bidirectional Encoder Representations from Transformer for word embeddings. Read more about its Toolkit implementation here.

Dataset Importer

Dataset Importer is a tool for conveniently uploading files into Elasticsearch to make them accessible for the Toolkit. Read more about the usage here.

Document Importer

Document importer provides API endpoints for adding, deleting and replacing documents one by one in or into a certain index in an easy way. Read more about the usage here.


Elasticsearch is a distributed, open-source search and analytics engine and database for all types of data. You can read more about it here.


Embedding is a statistical model describing the distributional properties of words and phrases, which enables the computation of similarity between words and phrases. In TTK, embeddings are used for finding contextually similar keywords to extend search results (by building lexicons). Furthermore, vectors from word embeddings can also be used in neural classification models (Taggers, Tagger Groups, :ref:` Torch Tagger <torch_tagger_concept>`, Taggers) in Texta Toolkit. TTK will also employ cross-lingual embeddings developed by EMBEDDIA to support multilingual text classification.

ES Analyzer

ES Analyzer is a tool for applying various Elasticsearch analyzers to indices. Currently supported analyzers are:

  • Stemmer

  • Tokenizer


Evaluator is a tool for evaluating labels predicted with classification or entity extraction models.


In TTK text annotations are regarded as facts:

    "doc_path": "content.text",
    "fact": "PER",
    "spans": "[[12, 24]]"
    "str_val": "Donald Trump",
    "author": "admin",
    "id": "f1f34f62-33e8-471d-bb88-78a96f17cca5",
    "source": "annotator"

Facts are objects with following properties:

  • fact - a name for the annotation (e.g. PER, LOC, or any arbitrary string)

  • str_val - string value for the annotation (e.g. “Donald Trump” for PER). Values don’t need to match annotated text segments.

  • doc_path - path to the field containing the annotated string (e.g. “text”, “text.author”, “comment.text.author”, etc.)

  • spans - JSON string containing the start and end positions in the annotated string

  • author - User that has generated or modified the fact

  • id - Auto generated UUID used to identify the fact

  • source - The source which has generated or modified the fact

The Toolkit comes with several predefined fact names that are used by our Multilingual Processor:






Name of a person.



Name of an organisation, gotten statistically.



Name of a location, gotten statistically.



Names of companies registered in Estonia, gotten from Estonian Open Data.



Estonian address, gotten from Estonian Open Data.



Name of a medicine.



Ingredient of a drug.



E-mail address.



Phone number.


Own tag

Tags we have trained in the Taggers under Models

Table 1. Predefined Fact Names used by Multilingual Processor

It is important to notice that COMPANY and ADDR identify only companies and addresses registered in Estonian Open Data. It won’t tag any foreign company nor address while ORG and LOC identify all of them.


Each document is a collection of fields. Field is a separated piece of information in that document. For example, we can have an article with fields containing the text, the date of publishing, headline, etc.


Server health is a comprehensive overview of the status and performance of individual servers. It includes hardware parameters and, in our case, the TTK version numbers and the state of needed services. See here how to check TTK’s health.


Elasticsearch’s index is a collection of documents in Elasticsearch. In Search output the table consists of all the suitable documents filtered out from that index and one row indicates a document.

Language Detector

Language Detector is a tool for detecting languages of the chosen indices documents. It uses :ref:`langdetect python module <https://pypi.org/project/langdetect/>`_. This is useful for getting a quick overview of the languages presented in your dataset and parse out documents in a certain language for future work. Read more about its usage here.

Lexicon miner

Lexicon Miner is a tool for creating topic-related lexicon. It uses :ref:’embeddings <embedding_concept>’ for finding words used in similar context. Read more about its usage here.

Multilingual Preprocessor (MLP)

MLP offers different analyzing options that give more features (information) for machine learning applications in TTK and helps to parse the information out of textual data user is interested in. Read more about the usage here.


Project is the main unit of access and management for datasets and resources (embeddings, text classifiers, etc.). A project is defined by its description, list of Elasticsearch indices related to the project (this is where the data is!), and a list of users who can access the project and its resources. All resources in TTK belong to some project and by adding or removing users, one can manage their access to the project.

Additionally, creators of a project can assign other people with the Project Administrator permission to share the responsibility of who should have access to the resources and who shouldn’t

Project is active when it is chosen in the upper panel on the right. Read more about projects here.

Project Administrator

Project Administrators are users who have been added into the Project to manage it, which can be done either by superusers or the author of the Project itself (by default, the Projects author is also its Project Admin). Within the permission of a Project Administrator lies: Adding and removing access to users into said Project, removing indices from the Project, editing the title of the Project.

Project Administrators who aren’t superusers though, lack the option to add existing indices into the Project and to search for usernames by a list (users need to be added by username manually).


Elasticsearch’s query is a little JSON command that retrieves objects with given attributes from Elasticsearch. It can be thought of as subset of the dataset the query is performed on. Queries can be used in various Toolkit’s tools and it can be saved under the Searcher as saved search for using it later or in other tools.

Rakun Keyword Extractor

Rakun Keyword Extractor is a tool for extracting keywords from texts. The tool is based on an unsupervised graph-based method RaKUn.

Regex Tagger

Regex Tagger is a tool for tagging texts based on pattern matches. The user can define patterns (words/phrases/regexes) to search from the text and fine-tune a variety of parameters to extract specific information and tag the text based on the information retrieved. Read more about its usage here.

Regex Tagger Group

Regex Tagger Group allows combining different Regex Taggers. Read more about its usage here.


Reindexer is a tool for making copies or subsets of Elasticsearch indices. Reindexer can be used for various data management tasks:

  • Making a copy of an index;

  • Merging multiple indices into one;

  • Making a random subset of an index;

  • Making a subset based on a predefined query;

  • Renaming fields;

  • Updating Elasticsearch field types.

Read more about it here.

Search Query Tagger

Search Query Tagger is a tool for tagging documents based on predefined queries.


Search Query Tagger is applicable only for tagging documents indexed in Elasticsearch and can be only used on indices that contain the same fields as the defined search.

Search Fields Tagger

Search Fields Tagger is a tool for converting tags/keywords in a separate field into Texta Facts as most applications in Texta Toolkit assume that the labels are stored as Texta Facts.


Taggers are monolingual binary text classifiers used to predict tags for documents, e.g. whether a tweet is toxic or talking about Brexit. Taggers are trained using a subset of documents defined by a TTK search or a raw Elasticsearch query (big ugly JSON object). The subset of documents defines the “positive” set of examples (the documents about the topic being tagged), whilst “negative” examples will be selected automatically (to be unlike the positive ones). TTK taggers are trained using scikit-learn pipelines and include models like logistic regression and SVM. TTK automatically splits the training data into training and testing data (by default 80-20) and applies grid search combined with k-fold cross-validation to identify best hyperparameters. An SVM model is also used for feature selection to remove unimportant features from the model making it smaller. For features, both word-based and character-based n-grams are used. Read more about it here.

Tagger Group

Tagger Group is an extension to TTK’s binary taggers to support monolingual multi-label classification. As its name suggests, Tagger Groups incorporate multiple (binary) taggers, which are executed in parallel to produce a list of tags to the user. Tagger Group has been successfully tested with over 6000 binary models and prediction times are usually less than 1 second. To achieve this, TTK has employed a hybrid approach for multi-label tagging, which uses unsupervised machine learning (document vectors) to limit the number of binary models used for prediction. In such a scenario, the input document is compared to training data to determine the most probable models to produce valid tags. Read more about it here.


TEXTA Toolkit allows building several different statistical models to process the text. The training process is initiated via TTK API or GUI, which results in creating the object in TTK data model and starting the asynchronous training task. Tasks are data objects for keeping track of the training progress of trainable statistical models.

Topic Analyzer

Topic Analyzer is a tool that detects groups of similar documents in the data. It can be used for example to explore the structure of the unlabeled data to get an understanding of what it contains. However, the main purpose of the tool is to take it one step further and make actual use of one’s exploration by transforming their discoveries into the labelling. It can then be used to build supervised machine learning models.

Topic Analyzer puts to use the best parts of unsupervised clustering and manual labelling. While clustering is a fast and efficient method to create groups of rather similar texts, it still lacks preciseness required to use these clusters as labelling of the data. For example, while some clusters may indeed contain similar documents, some other will completely fail in capturing the similarity and most of them will probably be so-and-so.

Topic Analyzer allows user to look inside each cluster and make the decision about the quality manually. Users can also perform various actions on the cluster: remove documents, add more similar documents, and move documents to another more suitable cluster. Finally, the user can choose to label the documents inside the cluster if it has reached a sufficient quality.

Read more about its usage here.

Torch Tagger

While Taggers and Tagger Groups use classical machine learning to produce binary classification models, Texta Toolkit also incorporates deep neural models for binary and multi-class text classification. As the models are all programmed using PyTorch, the TTK’s component is called Torch Tagger.

It allows for the user to use several state-of-art text classification models, including fastText, TextRNN using bi-direction LSTM networks, RCNN using recurrent convolutional neural nets. Since all models have been developed using PyTorch, introducing new models is fairly straightforward. TorchTagger models also include the possibility to use pre-trained word vectors (e.g. Word2Vec trained in TTK) in the embedding layer of the models. To create data processing pipelines, Torch Tagger uses torchtext package. Torch Tagger has been validated on monolingual toxic comment detection, reaching accuracy and F1-score of 96%. Read more about its usage here.

UUA server

User Account and Authentication is an identity management service for making sure that only selected users have access to certain datasets. Read more about it here.