- latexml converts the source into an HTML5 document
- llamapun tokenizes the first paragraph into a plain-text representation with sub-formula lexemes
- tensorflow executes a pre-trained BiLSTM model with 13 classification targets
- served as a rocket web service
For the scientific work behind this showcase, please read our paper
The current deployed model is a Keras BiLSTM(128)→BiLSTM(64)→LSTM(64), with a Dense(13) softmax output.
The model file 13_class_statement_classification_bilstm.pb can be downloaded from this repository via git-lfs. It is compatible with the rust wrapper for tensorflow and compiled to use a CPU implementation of LSTM, as our demo server has no dedicated GPU.
The input layer is embedded via the arxmliv 08.2018 GloVe embeddings, as well as padded/truncated to a maximum length of 480 words.
A paragraph is hence a fixed (480,300) matrix, as passed into the bilstm layer.
The specific model in this demo was trained on 8.3 million paragraphs from the arxmliv 08.2018 dataset, and tested on 2.1 million paragraphs respectively, obtaining a 0.91 F1 score on a target of 13 classes. The base rate baseline was 0.38, the frequency of the "proposition" class.
For more experimental details, please see the main experiment repository.
For practical evaluation, a likelihood threshold could be used, where entries with smaller likelihoods (e.g. <0.3) can be considered as an "other" label.