WannaDB allows users to explore unstructured text collections by automatically organizing the relevant information nuggets in a table. It supports ad-hoc SQL queries over text collections using a novel two-phased approach: First, a superset of information nuggets is extracted from the texts using existing extractors such as named entity recognizers. The extractions are then interactively matched to a structured table definition as requested by the user.
Watch our demo video or read our paper to learn more about the usage and underlying concepts.
Our tool is provided ready-to-use with a graphical user interface that allows to load own text collections and execute the table extraction on them.
Main Window, specify the query or attributes list, check current table filling state and inspect the result table.
Interactive matching, inspect the current list of feedback requests and confirm matches or request detailed inspection.
Inspect a single document, see the extractions in context and chose the correct one (including custom text spans) or state that the document does not contain a relevant value.
Run main.py to start the WannaDB GUI.
To run batch experiments instead, use experiments/experiment_runner.py.
To run the preprocessing stand-alone (e.g., on a server with GPUs) you can use scripts/preprocess.py.
See EXTRACTORS.md for more information on the interactive extraction component.
This project requires Python 3.10 or newer.
python -m venv venv
source venv/bin/activate
export PYTHONPATH="."
pip install --upgrade pip
pip install --use-pep517 -r requirements.txt
pip install --use-pep517 pytest
You may have to install torch by hand if you want to use CUDA:
https://pytorch.org/get-started/locally/
pytest
The code in this repository is the result of several scientific publications. If you build upon WannaDB, please cite:
@inproceedings{wannadb@BTW23,
author = {Hättasch, Benjamin AND Bodensohn, Jan-Micha AND Vogel, Liane AND Urban, Matthias AND Binnig, Carsten},
title = {WannaDB: Ad-hoc SQL Queries over Text Collections},
booktitle = {BTW 2023},
year = {2023},
editor = {König-Ries, Birgitta AND Scherzinger, Stefanie AND Lehner, Wolfgang AND Vossen, Gottfried} ,
doi = { 10.18420/BTW2023-08 },
publisher = {Gesellschaft für Informatik e.V.},
address = {}
}
If you want to reference specific features/parts, our further publications might be relevant:
@inproceedings{aset@SIGMOD22,
author = {H\"{a}ttasch, Benjamin and Bodensohn, Jan-Micha and Binnig, Carsten},
title = {Demonstrating ASET: Ad-Hoc Structured Exploration of Text Collections},
year = {2022},
isbn = {9781450392495},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3514221.3520174},
doi = {10.1145/3514221.3520174},
abstract = {In this demo, we present ASET, a novel tool to explore the contents of unstructured data (text) by automatically transforming relevant parts into tabular form. ASET works in an ad-hoc manner without the need to curate extraction pipelines for the (unseen) text collection or to annotate large amounts of training data. The main idea is to use a new two-phased approach that first extracts a superset of information nuggets from the texts using existing extractors such as named entity recognizers. In a second step, it leverages embeddings and a novel matching strategy to match the extractions to a structured table definition as requested by the user. This demo features the ASET system with a graphical user interface that allows people without machine learning or programming expertise to explore text collections efficiently. This can be done in a self-directed and flexible manner, and ASET provides an intuitive impression of the result quality.},
booktitle = {Proceedings of the 2022 International Conference on Management of Data},
pages = {2393–2396},
numpages = {4},
keywords = {matching embeddings, text to table, interactive text exploration},
location = {Philadelphia, PA, USA},
series = {SIGMOD '22}
}
@inproceedings{aset@AIDB21,
author = {H{\"a}ttasch, Benjamin and Bodensohn, Jan-Micha and Binnig, Carsten},
year = "2021",
title = "ASET: Ad-hoc Structured Exploration of Text Collections",
eventdate = "16.-20.08.2021",
language = "en",
booktitle = "3rd International Workshop on Applied AI for Database Systems and Applications (AIDB21). In conjunction with the 47th International Conference on Very Large Data Bases, Copenhagen, Denmark, August 16 - 20, 2021.",
location = "Copenhagen, Denmark"
}
@inproceedings{wannadb@DESIRES21,
author = {H{\"{a}}ttasch, Benjamin},
title = "WannaDB: Ad-hoc Structured Exploration of Text Collections Using Queries",
booktitle = "Proceedings of the Second International Conference on Design of Experimental Search Information REtrieval Systems, Padova, Italy, September 15-18, 2021",
series = "{CEUR} Workshop Proceedings",
volume = "2950",
pages = "179--180",
publisher = "CEUR-WS.org",
year = "2021",
url = "http://ceur-ws.org/Vol-2950/paper-23.pdf",
timestamp = "Mon, 25 Oct 2021 15:03:55 +0200",
biburl = "https://dblp.org/rec/conf/desires/Hattasch21.bib",
bibsource = "dblp computer science bibliography, https://dblp.org"
}
WannaDB is dually licensed under both AGPLv3 for the free usage by end users or the embedding in Open Source projects, and a commercial license for the integration in industrial projects and closed-source tool chains. More details can be found in our licence agreement.
We publish the source code four our system as discussed in the papers here. Additionally, we publish code to reproduce our experiments in a separate repository (coming soon).
Unfortunately, we cannot publish the datasets online due to copyright issues. We will send them via email on request to everyone interested and hope they can be of benefit for other research, too.
The core of WannaDB (extraction and matching) was previously developed by us under the name ASET (Ad-hoc Structured Exploration of Text Collections). To better reflect the whole application cycle vision we present with this paper, we switchted the name to WannaDB.
This repository is structured as follows:
wannadb,wannadb_parsql, andwannadb_uicontain the implementation of ASET and the GUI.scriptscontains helpers, like a stand-alone preprocessing script.testscontains pytest tests.
The core implementation of WannaDB is in the wannadb package and implemented as a library. The implementation allows you to construct pipelines of different data processors that work with the data model and may involve user feedback.
Data model
data contains WannaDB's data model. The entities are InformationNuggets, Attributes, Documents, and the DocumentBase.
A nugget is an information piece obtained from a document. An attribute is a table column that gets
populated with information from the documents. A document is a textual document, and the document base is a collection of documents and provides facilities for BSON serialization, consistency checks, and data access.
InformationNuggets, Attributes, and Documents can have BaseSignals, which provide a way to easily store additional information with them. Each signal is identified with a unique identifier and implements the serialization and deserialization. Furthermore, some signals may not be serialized. There are base implementations for different data types like floats or numpy arrays.
Configurations
configuration.py contains the abstract pipeline code. An Pipeline allows you to execute multiple
BasePipelineElements one after the other. These pipeline elements work on an DocumentBase and receive a
BaseInteractionCallback and BaseStatusCallback to facilitate user interactions and convey status updates.
Furthermore, they receive a Statistics object that allows them to record information during runtime.
Both BasePipelineElements and the Pipeline are BaseConfigurableElements. This means that they come with a unique identifier and provide methods to instantiate them from a given configuration dictionary and to serialize their configuration as a dictionary.
Each BasePipelineElement specifies which BaseSignals it requires and generates for the nuggets, attributes, and documents. This ensures the consistency of the pipeline. In other words, when a pipeline element is executed, all signals it requires must be set.
Callbacks
interaction.py and status.py contain BaseInteractionCallback and BaseStatusCallback, which allow the pipeline elements to request user interactions and convey status updates. They come with default implementations InteractionCallback and StatusCallback that receive a callback function when initialized, and EmptyInteractionCallback and EmptyStatusCallback that simply do nothing.
Resources
resources.py contains a resource manager that allows different parts of WannaDB to share resources like embeddings or transformer models. The module implements the singleton pattern, so there is always only one ResourceManager accessed via resources.MANAGER, which handles the loading, access, and unloading of BaseResources. You should use a context manager (with ResourceManager() as resource_manager:) to ensure that all resources are properly closed when the program stops/crashes.
Each BaseResource comes with a unique identifier and implements methods for loading, unloading, and access.
Statistics
The Statistics object allows you to easily record information during runtime. It is handed from the Pipeline to the BasePipelineElements, and from the BasePipelineElements to other components like distance functions.
The GUI implementation can be found in the wannadb_ui package. wannadb_api.py provides an asynchronous API for the wannadb library using PyQt's slots and signals mechanism. main_window.py, document_base.py, and interactive_window.py contain different parts of the user interface, and common.py contains base classes for some recurring user interface elements.