Model Card for Model Geo-BERT-multilingual

<!-- Provide a quick summary of what the model is/does. -->

This model predicts the geolocation of short texts (less than 500 words) in a form of two-dimensional distributions also referenced as the Gaussian Mixture Model (GMM).

Model Details

Number of predicted points: 5 Custom transformers pipeline and result visualization: https://github.com/K4TEL/geo-twitter/tree/predict

Model Description

<!-- Provide a longer summary of what this model is. -->

This project was aimed to solve the tweet/user geolocation prediction task and provide a flexible methodology for the geotagging of textual big data. The suggested approach implements BERT-based neural networks for NLP to estimate the location in a form of two-dimensional GMMs (longitude, latitude, weight, covariance). The base model has been finetuned on a Twitter dataset containing text content and metadata context of the tweets.

• Developed by: Kateryna Lutsai
• Model type: regression
• Language(s) (NLP): multilingual
• Finetuned from model: bert-base-multilingual-cased

Model Sources

<!-- Provide the basic links for the model. -->

• Repository: https://github.com/K4TEL/geo-twitter
• Paper: https://arxiv.org/pdf/2303.07865.pdf
• Demo: https://github.com/K4TEL/geo-twitter/blob/predict/prediction.ipynb

Uses

<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->

Geo-tagging of Big data

Direct Use

<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->

Per-tweet geolocation prediction

Out-of-Scope Use

<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->

Per-tweet geolocation prediction without "user" metadata is expected to show lower accuracy of predictions.

Bias, Risks, and Limitations

<!-- This section is meant to convey both technical and sociotechnical limitations. -->

Risk for unethical use on the basis of data that is not publicly available.

The limitation of text length is dictated by the BERT-based model's capacity of 500 tokens (words).

How to Get Started with the Model

Use the code below to get started with the model:

https://github.com/K4TEL/geo-twitter/tree/predict

A short startup guide is given in the repository branch description.

Training Details

Training Data

<!-- This should link to a Data Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->

The Twitter dataset contained tweets with their text content, metadata ("user" and "place") context, and geolocation coordinates.

Training Procedure

<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->

Information about the model training on the user-defined data could be found in the GitHub repository: https://github.com/K4TEL/geo-twitter

Training Hyperparameters

• Learning rate start: 1e-5
• Learning rate end: 1e-6
• Learning rate scheduler: cosine
• Number of epochs: 3
• Batch size: 10
• Optimizer: Adam
• Intra-feature loss: mean
• Inter-feature loss: mean
• Neg log-likelihood domain: positive
• Features: NON-GEO + GEO-ONLY

Evaluation

<!-- This section describes the evaluation protocols and provides the results. --> All performance metrics and results are demonstrated in the Results section of the article pre-print: https://arxiv.org/pdf/2303.07865.pdf

Testing Data, Factors & Metrics

Testing Data

<!-- This should link to a Data Card if possible. -->

Worldwide dataset of tweets with TEXT-ONLY and NON-GEO features

Metrics

<!-- These are the evaluation metrics being used, ideally with a description of why. -->

Spatial metrics: mean and median Simple Accuracy Error (SAE), Acc@161 Probabilistic metrics: mean and median Cumulative Accuracy Error (CAE), mean and median Prediction Area Region (PRA) for 95% density area, Coverage of PRA

Results

Tweet geolocation prediction task

• TEXT-ONLY: mean 1588 km and median 50 km, 61% of Acc@161
• NON-GEO: mean 800 km and median 25 km, 80% of Acc@161

User home geolocation prediction task

• TEXT-ONLY: mean 892 km and median 31 km, 74% of Acc@161
• NON-GEO: mean 567 km and median 26 km, 82% of Acc@161

Model Architecture and Objective

Implemented wrapper layer of liner regression with a custom number of output variables that operates with classification token generated by the base BERT model.

Hardware

NVIDIA GeForce GTX 1080 Ti

Software

Python IDE

Model Card Contact

lutsai.k@gmail.com