U.S. Patent Phrase to Phrase Matching

Introduction

This repository contains the code that achieved 31st place in U.S. Patent Phrase to Phrase Matching. You can see the detailed explanation of the solution here.

You can also view all the experiments logs on Weights & Biases dashboard here.

This competition was organized by USPTO and Kaggle. The main aim of this competition was to extract relevant information by matching key phrases in patent documents. Determining the semantic similarity between phrases is critically important during the patent search and examination process to determine if an invention has been described before. You can read more about the problem statement on the overview page of the competition.

Requirements

You can simply install all the requirements and setup your machine completely to run the code by install_deps.sh script in bash folder. Before running this script make sure that you include your kaggle key into the script in the respective field.

$ bash bash/install_deps.sh

Preparing the data

You can download the data by the following command. Make sure to download the data into a folder named as input

$ kaggle datasets download -d atharvaingle/uspppm-data

Setting up the configuration

You can specify your configuration in config.yaml (present in config/ folder) whether to log artifacts to GCP/W&B or not log them at all by boolean flags in the file.

You can run an experiment by writing a bash file as following:

#!/bin/bash

# sample run file
cd /home/US-Patent-Matching-Kaggle/src

for fold in 0 1 2 3 4
do
    python3 train.py \
    paths="jarvislabs" \
    trainer="deberta_v3_base" \
    run.debug=False \
    run.fold=$fold \
    run.exp_num="53" \
    trainer.dataloader_num_workers=6 \
    data.use_custom_seperator=True \
    model.model_name="microsoft/deberta-v3-base" \
    model.class_name="DebertaV2ForSequenceClassificationGeneral" \
    model.loss_type="mse" \
    model.multi_sample_dropout=True \
    model.attention_pool=False \
    run.name="mse-stable-drop-msd" \
    run.comment="mse loss + multi sample dropout with deberta StableDropout"
done

# use following line only while training on jarvislabs.ai
# pause instance programatically after running a series of experiments
python3 -c "from jarviscloud import jarviscloud; jarviscloud.pause()"

• You can override any config field from bash file
• You can pick any model class imported in modeling/__init__.py. You just have the specify the name in model.class_name field of Hydra configuration.
• Loss choices: mse, bce and pearson
• You can choose any trainer configuration you want from various *.yaml files in config/trainer folder.

Inference

You can check out the final ensemble inference code submission for this competition in Kaggle notebook here