CodeFest on Green Plastics: Finalists announced

The
European Patent Office (EPO) has announced six finalists for its first ever public CodeFest. Towards the end of
2022, some 60 individuals and teams took up the challenge: to write an AI code
that improves access to patent information on green plastics, helping to rid
the planet of plastic waste.

Each
finalist will present their creative code at an online award ceremony on
Thursday 23 February 2023 at 14.00 hrs CET, where the winners will be revealed.
The ceremony will also highlight how the finalists‘ AI solutions could
contribute to the United Nations Sustainable Development Goals (SDGs),
particularly to responsible consumption and production (SDG 12). 

The
winner will receive a EUR 20 000 cash prize, with the second and third place
finalists each receiving EUR 15 000 and EUR 10 000 respectively. With internal
competitors from the EPO, external entrants from across Europe and several
mixed teams, which solution will take a coveted spot on the podium?

The solutions use various techniques, including neural networks,
gradient boosting machines, transformer models, fine-tuned language models, and
a sequence-to-sequence approach. Here are the six finalists; a full list of
team members and full summary of each solution is available on the official
event page, where you can also register to attend the award ceremony (see
below).

• AI4EPO, Greece and Netherlands
  This team’s model uses
  state-of-the-art AI pipelines and large language models from OpenAI for
  zero-shot, few-shot and other approaches to arrive at a custom MLP neural
  network for binary and multi-label classification.

• Green Hands, Netherlands
  As there is currently no
  classification scheme or labelled data available in this field, Green Hands
  proposed a new classification scheme, and developed a strategy to automatically
  assign labels to patents in order to create a labelled training dataset.

• Multimodal Patent Document Classification, Germany
  and Netherlands
  This team created a deep learning
  architecture to classify patent documents by fusing features from figures and
  text, thus exploiting the multimodal nature of patents.

• Nikolaos Gialitsis, Greece
  Nikolaos developed a machine
  learning model that incorporates both semantic and lexical features, and that was
  trained on a dataset of patents and scientific publications.

• Patent Variables, France,
  Germany and Switzerland
  The team converted the problem
  into a sequence-to-sequence challenge, asking the user to define green plastics
  and then testing any patent claim against that definition.

• Thomas Eißfeller, Germany
  Using a gradient boosting
  machine, Thomas focussed on high sample efficiency, unbiased validation metrics
  and maximising specificity.

An extremely high
level of submissions was received in response to the code challenge, which was:
To develop creative and reliable artificial intelligence (AI) models for
automating the identification of patents related to green plastics.

The competition, open to anyone
aged 18 or over and resident in one of the EPO member
states, included over 30 participating teams representing diverse
nationalities. The finalists were selected by a jury consisting of senior
specialists from across the EPO working in sustainability, IT, data science and
AI, as well as patent information and analysis.