jingcshi / bert_subsumption

Subsumption prediction on e-commerce taxonomies

bert_subsumption

Subsumption prediction on e-commerce taxonomies

Abstract

This work studies the characteristics of e-commerce taxonomies, and propose a new subsumption prediction method based on the pre-trained language model BERT that is well adapted to the e-commerce setting. The proposed model utilises textual and structural semantics in a taxonomy, as well as the rich and noisy instance (item) information. Experiments have been conducted on two large-scale e-commerce taxonomies from eBay (proprietary) and AliOpenKG (public), that our method offers substantial improvement over strong baselines. Click here for the extracted item-augmented taxonomy of AliOpenKG, the subset of AliOpenKG that this work is concerned with.

Dependencies

• torch
• transformers
• sklearn
• tqdm
• box-embeddings

Experiments

Setup

Download the AliOpenKG taxonomy to the repository folder and set data[tbox_path] and data[abox_path] to the corresponding paths for TBox and ABox data. A pretrained BERT model is required to run some experiments. Please set pretrained[pretrained_path] and pretrained[tokenizer_path] to your BERT model/ tokenizer.

Running

The experiments with linear classifier, box embedding classifier and k-NN classifier are implemented in /experiments/run/linear.ipynb, /experiments/run/box.ipynb and /experiments/run/knn.ipynb respectively.