This repository is based on this Competition.

The task is trying to classify each sentences in the Abstract of the theses with the topic of Computer Science sourced from axXiv. There are 6 categories: Background, Objectives, Methods, Results, Conclusions and Others.

• only task1_trainset.csv contains labels
• still provide testset but might be useless

Almost all the .py files are for preprocessing and extract the features. The model training and evaluation are in Notebook.

• Preprocessing

  • GetAbstract.py: Get the structured data from the provided files.
  • CountPosition.py: Check the position of the sentences inside its source Abstract.
  • CheckNameNumber.py: Check if the sentence contains name entities or number.
  • Preprocessing.py: Text preprocessing. 3 cleaning punctuation functions are included. All of them using regular expression to remove some unwanted words or punctuation
  • Utility:
    • Combine_pre&post-sentence.py: Combine previous/next sentences of Target sentence together
    • Combine_2pre&post-sentence.py: Combine first/last two sentences of Target sentence together
    • Statistics.py: For data exploration
  • Useless files:
    • Extract_Date.py: Extract the data of each thesis in the provided files as feature. (Not useful after my experience)
    • AdjustThreshold.py/ Submit.py: For competition submission
    • WordEmbedding.py: Embed each words in the sentences(after tokenize) using different pre-trained Word2Vec. Take lots of time and enormous resulting file. We should do wordembedding online during training. There's one more detailed guided instruction in Notebook

• Some Model that can be trained locally

  • TFIDF.py: Classifying the text using TFIDF as feature vectors and xgboost as classifier. There's one more detailed guided instruction in Notebook
  • CRF.py: Using Conditional Random Field as classifier. The sentence from same Abstract need to be combine together. The feature of each sentence is also TFIDF.

Some models (especially deep learning models) require lots of computing power, we can take the advantage of Google Colab as our computing resource. Notebooks also provide some detailed instruction.

• Tokenize.ipynb: How to preprocessing ans tokenize
• TFIDF.ipynb: How to use TFIDF as feature vectors (same as TFIDF.py)
• Spacy_embed.ipynb: How to conduct word embedding using Spacy (same as WordEmbedding.py)
• Deep_Averaging_Network(DAN).ipynb: Using DAN which is provided by TF-hub as features of the sentences. Then, use a naive Fully-Connected Neural Network as classifier
• BidirectionalRNN.ipynb: Build a Bidirectional GRU as classifier and word embedding of each words as representation of the sentences
• ELMO.ipynb: Using ELMO as embedded layer and Bi-GRU as classifier
• ELMO_with_pre_post_sentences.ipynb: Best model. Using ELMO as embedded layer and Bi-GRU as classifier. The input is:

  • previous/next sentences of target sentences      or

  • first/last two sentences of target sentence

  • Model architecture: (Best performance is the ensemble of these two models)

    • MODEL_1

    

    • MODEL_2

    

• Position of each label in its source Abstract

  • BACKGROUND

  

  • OBJECTIVES

  

  • METHODS

  

  • RESULTS

  

  • CONCLUSIONS

  

  • OTHERS

  

• Whether sentences with different labels contain Name Entity or Number

  • Name Entity

  

  • Number