This work describes and answers a problem in Mathematical Named-Entity Recognition: Given a mathematical object and the context in which it, can we extract its definition?
This repo contains the Mathematical Token Definition Extraction (MTDE) dataset as well as the implemetations of five different neural defintion extraction models. [1] is a peer-reviewed full analysis on the MTDE problem and how the models perform on the MTDE dataset.
The MTDE dataset contains around 10,000 entries of variable names, the contexts in which they are defined, their ‘short’ definitions and their ’long’ definition. Here, a short defintion is a 1-word-long definition and a long definition is a one-or-more-word-long definition. The data was collected from a random sampling of mathematical and scientific arXiv preprint manuscripts. The manuscripts cover a wide range of mathematic and scientific disciplines including Physics, Computer Science, and Biology. Candidate data was generated via a corpus crawler and then pruned and cleaned manually.
In this repo, the following models are implemented in jypter notebook tutorials:
- Vanilla Seq2Seq
- Transformer Seq2Seq
- Pointer Network
- Match-LSTM
- BERT (Huggingface's BertForQuestionAnswering)
These tutorials are aimed to throughly explain the mechanisms behind each mathematical defintion extraction model examined in [1] as well as serve as a blueprint for future experiements on this problem.
Email: [email protected]
Please message me with any feedback or errors you may find! Any help is appriciated :)
There is a small error in right-most subfigure of figure 1 in [1]. The correct figure should be:
[1] Hamel, E., Zheng, H., & Kani, N. (2022). An Evaluation of NLP Methods to Extract Mathematical Token Descriptors. In International Conference on Intelligent Computer Mathematics (pp. 329-343). Springer, Cham.