saminens / hmm_tagger

Comparison of various tagging mehods. I also deep dive into HMM tagger.

Introduction

In this notebook, used the Pomegranate library to build a hidden Markov model for part of speech tagging with a universal tagset. Hidden Markov models have been able to achieve >96% tag accuracy with larger tagsets on realistic text corpora. Hidden Markov models have also been used for speech recognition and speech generation, machine translation, gene recognition for bioinformatics, and human gesture recognition for computer vision, and more.

Installation

1.Open a terminal and clone the project repository:

$ git clone https://github.com/saminens/HMM-tagger

Train your own model

Place datasets in the data folder and run HMM tagger.py

Example : Decoding Sequences with the HMM Tagger

for key in data.testing_set.keys[:3]:
    print("Sentence Key: {}\n".format(key))
    print("Predicted labels:\n-----------------")
    print(simplify_decoding(data.sentences[key].words, basic_model))
    print()
    print("Actual labels:\n--------------")
    print(data.sentences[key].tags)
    print("\n")

Sentence Key: b100-28144

Predicted labels:
-----------------
['CONJ', 'NOUN', 'NUM', '.', 'NOUN', 'NUM', '.', 'NOUN', 'NUM', '.', 'CONJ', 'NOUN', 'NUM', '.', '.', 'NOUN', '.', '.']

Actual labels:
--------------
('CONJ', 'NOUN', 'NUM', '.', 'NOUN', 'NUM', '.', 'NOUN', 'NUM', '.', 'CONJ', 'NOUN', 'NUM', '.', '.', 'NOUN', '.', '.')


Sentence Key: b100-23146

Predicted labels:
-----------------
['PRON', 'VERB', 'DET', 'NOUN', 'ADP', 'ADJ', 'ADJ', 'NOUN', 'VERB', 'VERB', '.', 'ADP', 'VERB', 'DET', 'NOUN', 'ADP', 'NOUN', 'ADP', 'DET', 'NOUN', '.']

Actual labels:
--------------
('PRON', 'VERB', 'DET', 'NOUN', 'ADP', 'ADJ', 'ADJ', 'NOUN', 'VERB', 'VERB', '.', 'ADP', 'VERB', 'DET', 'NOUN', 'ADP', 'NOUN', 'ADP', 'DET', 'NOUN', '.')


Sentence Key: b100-35462

Predicted labels:
-----------------
['DET', 'ADJ', 'NOUN', 'VERB', 'VERB', 'VERB', 'ADP', 'DET', 'ADJ', 'ADJ', 'NOUN', 'ADP', 'DET', 'ADJ', 'NOUN', '.', 'ADP', 'ADJ', 'NOUN', '.', 'CONJ', 'ADP', 'DET', 'NOUN', 'ADP', 'ADJ', 'ADJ', '.', 'ADJ', '.', 'CONJ', 'ADJ', 'NOUN', 'ADP', 'ADJ', 'NOUN', '.']

Actual labels:
--------------
('DET', 'ADJ', 'NOUN', 'VERB', 'VERB', 'VERB', 'ADP', 'DET', 'ADJ', 'ADJ', 'NOUN', 'ADP', 'DET', 'ADJ', 'NOUN', '.', 'ADP', 'ADJ', 'NOUN', '.', 'CONJ', 'ADP', 'DET', 'NOUN', 'ADP', 'ADJ', 'ADJ', '.', 'ADJ', '.', 'CONJ', 'ADJ', 'NOUN', 'ADP', 'ADJ', 'NOUN', '.')

Improve Model Performance

The data sparsity problem arises because the same amount of data split over more tags means there will be fewer samples in each tag, and there will be more missing data tags that have zero occurrences in the data.

• Laplace Smoothing (pseudocounts) Laplace smoothing is a technique where you add a small, non-zero value to all observed counts to offset for unobserved values.

• Backoff Smoothing Another smoothing technique is to interpolate between n-grams for missing data. This method is more effective than Laplace smoothing at combatting the data sparsity problem.

• Extending to Trigrams HMM taggers have achieved better than 96% accuracy on this dataset with the full Penn treebank tagset using an architecture described in this paper. Altering your HMM to achieve the same performance would require implementing deleted interpolation (described in the paper), incorporating trigram probabilities in your frequency tables, and re-implementing the Viterbi algorithm to consider three consecutive states instead of two.