Quinex: An Automatic Quantitative Information Extraction Tool from Text Based on Lightweight Open-Source Language Models

Quinex, developed by the Forschungszentrum Jülich (Germany), is a Python library specifically for extracting quantitative information from text. It uses a domain-specific lightweight architecture, significantly reducing computational resource requirements while maintaining high accuracy, addressing the pain points of traditional manual extraction (time-consuming and labor-intensive) and general-purpose large language models (redundant parameters and high costs).

In fields like scientific research, business analysis, and policy-making, quantitative information is scattered across massive amounts of text. Traditional manual extraction methods are inefficient, and general-purpose large language models have issues with redundant parameters and high costs. Quinex (Quantitative Information Extraction) was created to address these pain points—it's a Python library specifically designed for extracting and analyzing quantitative information, using a lightweight architecture that balances accuracy and resource requirements.

Quinex is based on the Transformer architecture and optimized for quantitative extraction tasks. Its core functions are divided into two main modules:

Quantitative Span Recognition: Similar to NER tasks, it locates quantitative expressions in text through sequence labeling, capable of identifying precise numerical values and vague quantifications (e.g., "several" in "several turbines").

Measurement Context Extraction: Extracts relevant entities, attributes, and context for each quantitative value. For example, from "The output power of the Reykjanesvirkjun geothermal power plant is 130 MW", it identifies the entity "Reykjanesvirkjun", attribute "power output", and spatial scope "Iceland".

Quinex has strong data standardization capabilities: extracted quantitative values are converted into standard numerical forms, and units are linked to the QUDT ontology library, facilitating comparison and integration of data from different sources. Additionally, the system can identify implicit attributes—for example, inferring that the 5 MW in "this 5 MW power plant" refers to power generation capacity, enhancing the completeness of information extraction.

Quinex supports large-scale scientific literature screening and quantitative search, helping to build thematic databases. The project provides an experimental web service, including a visualization dashboard (world map for spatial distribution, timeline for temporal changes, citation network for source tracing) and text annotation functions. It also experimentally supports classification of quantitative statement types, distinguishing between normative descriptions, goal setting, observation records, etc., to help understand the context of data.

Advantages: The model has far fewer parameters than general-purpose models, leading to low deployment and operation costs; results are anchored to the original text with no hallucinations, transparent and verifiable; open-source self-deployment ensures data privacy.

Limitations: Primarily trained for English, with limited support for other languages; excels in genres like scientific literature but performs poorly on first-person narratives, lyrics, etc.; cannot extract quantitative information from tables/charts or implicit quantities (e.g., the 1 in "an elephant").

Quinex is released as open-source, with resources distributed across multiple repositories: the main repository provides the core library and models, the dataset repository provides training data, and the tool library includes quantity parsers, etc. The model is hosted on the Hugging Face platform, with easy installation (via pip). It can run with the spaCy English model and a few dependencies, and the project provides detailed guides and example code, making it easy for beginners to get started.

Quinex is a typical case of specialization and lightweight development in the NLP field, proving that domain specialization can achieve efficiency and interpretability with small models for specific tasks. It provides a powerful and cost-effective solution for personnel dealing with large amounts of quantitative information in text. Future expectations include improving multilingual support and table extraction functions.