Summarising scientific publications from a tech-transfer perspective

This is a simplified example illustrating how great-ai is used in practice at ScoutinScience{ target=_blank }. The subpages show great-ai in action by going over the life-cycle of finetuning and deploying a BERT-based software service.

??? note "Propriety data" The purpose of this example is to show you different ways in which great-ai can assist you. The exact NLP task being solved is not central. Stemming from this, and from the difficult nature of obtaining appropriate training data, the propriety dataset used for the experiments is not shared.

Objectives

You will see how the great_ai.utilities can integrate into your Data Science workflow.
You will see how great_ai.large_file can be used to version and store your trained model.
You will see how [GreatAI][great_ai.GreatAI] should be used prepare your model for a robust and responsible deployment.
You will see multiple ways of customising your deployment.

Overview

One of the core features of the ScoutinScience platform is summarising research papers form a tech-transfer perspective. In short, extractive summarisation is preferred using a binary classifier trained on clients' judgement of sentence interestingness. Thus, documents are sentences and the expected output is a binary label showing whether a sentence is "worthy" of being in the tech-transfer summary. Providing an explanation for each decision is imperative since ScoutinScience embraces applying only explainable AI (XAI) methods wherever feasible.

!!! success You are ready to start the tutorial. Feel free to come back to the summary section once you're finished.

[:material-database: Examine data](data.ipynb){ .md-button .md-button--primary }

:fontawesome-solid-chart-simple: Train model{ .md-button .md-button--primary }

:material-cloud-tags: Deploy service{ .md-button .md-button--primary }

Summary

Data notebook

We load and analyse the data by calculating inter-rater reliability and checking the feasibility of using an AI-based approach by testing the accuracy of a trivial baseline method.

Training notebook

We simply fine-tune SciBERT.

After training and evaluating a model, it is exported using [great_ai.save_model][]. For more info, checkout the configuration how-to page.

Deployment notebook

We customise the GreatAI configuration, create custom cahcing for the model and implement an inference function that can be hardened by wrapping it in a [GreatAI][great_ai.GreatAI] instance. We also extract the attention weights as a quasi-explanation.

Finally, we test the model's inference function through the GreatAI dashboard. The only thing left is to deploy the hardened service properly.

Additional files

There are some other files required for deploying the notebook. For example, the config file for S3 and MongoDB or a Dockerfile for building a custom image. These are compiled and shown on a separate page.

3.3 KiB Raw Blame History