Implemented a Part Of Speech Tagger using Support Vector Machine(SVM), Hidden Markov Model(HMM) and Bi-directional Long-Short Term Memory(Bi-LSTM)

All the results, detailed error analysis, strengths and weakness of the models and references are also present in the Report.pdf

• Instructions for running the code

  SVM_PoS_tagger.ipynb contains the implementation of SVM

  Run the code from https://colab.research.google.com/drive/178a6M4J3lt-twzGr1Nv-Y2EliW77NEt9?usp=sharing

  It can also be run as a python script from svm.py in the Support Vector Machine directory.

  No need to download any dependencies if running from the colab file.

• Results

  • Per-POS accuracy vs Relative Frequency

    

  • Accuracy

    Model	Test Accuracy(%)	Train Accuracy(%)
    HMM	83.25	83.36

  • Feature Engineering

    Features Selected	Accuracy(%)
    Word length, capitalisation, upper-case, lower-case, isNumeric	40
    Prefix and suffix for nouns, verbs, adjectives and adverbs	55
    Word stems using PorterStemmer	60
    Tag of Previous Word	65
    Pre-trained Glove word-embeddings (1 lakh 50-dimensional vectors)	83
    Pre-trained word2vec embeddings (10 crore 300-dimensional vectors)	90
    Including the features of previous 3 words and following 3 words	95

• Instructions for running the code

  The file main.ipynb contains the implementation of HMM. Use jupyter notebook to access it.

  Following utilities are to be installed: nltk, pandas, seaborn, matplotlib, sklearn, scikit-learn, tqdm

  Also make sure to input the following commands in jupyter notebook:

  nltk.download('brown')
  nltk.download('universal_tagset')
  

  Hidden Markov Model directory has a copy of all the images, plots and data produced from main.ipynb

• Results

  • Per-POS accuracy vs Relative Frequency

    

  • Accuracy

    Model	Test Accuracy(%)	Train Accuracy(%)
    HMM	96.01	97.35

• Baseline Model Architecture

  Layer (type)	Output Shape	# Param
  embedding (Embedding)	(None, 180, 300)	14944800
  bidirectional	(None, 180, 64)	85248
  time_distributed	(None, 180, 13)	845

  Total params: 15,030,893
  Trainable params: 86,093
  Non-trainable params: 14,944,800

• Instructions for running the code

  BiLSTMBaseline.ipynb contains the implementation of baseline code for HMM

  Run the baseline code from https://colab.research.google.com/drive/1lhBd-gxsXNVeQJtBXoLZ7HABI5yYttrw?usp=sharing

  The CNN based code is available in the following two formats:

  1. CNNBiLSTMCRF.ipynb for jupyter notebooks and
  2. main.py can be run as a python script

  The following utilities are to be installed for the CNN based code: pytorch, nltk, torchvision, numpy, seaborn, pandas, matplotlib, tqdm, scikit-learn, sklearn

  Also download the glove embedding from http://nlp.stanford.edu/data/wordvecs/glove.6B.zip and extract it in the source directory:

  Execute the following commands for the CNN based code

  nltk.download('brown')
  nltk.download('universal_tagset')
  

  The results of running the CNN based code are stored in the Bi-LSTM folder.

• Results

  • Per-POS accuracy vs Relative Frequency

    

  • Accuracy

    Model	Test Accuracy(%)	Train Accuracy(%)
    Bi-LSTM Baseline	79.38	78.39
    Bi-LSTM-CNN	87.15	87.19

1. Leon Bottou “UNE APPROCHE THEORIQUE DE L’APPRENTISSAGE CONNEXIONNISTE ET AP-PLICATIONS A LA RECONNAISSANCE DE LA PAROLE” PhD thesis (1991)
2. Xuezhe Ma and Eduard Hovy “End-to-end sequence labeling via bi-directional lstm-cnns-crf” (2016)
3. Jesus Gimenez and Llus Marquez "Fast and Accurate Part-of-Speech Tagging: The SVM Approach Revisited" (2003)
4. Mathieu Blondel, Akinori Fujino, Naonori Ueda Large-scale Multiclass Support Vector Machine Training via Euclidean Projection onto the Simplex (2014)