Sentence-Inference

For every given pair of sentences -- (sentence-1, sentence-2), we need to determine if sentence-2 can be logically inferred given sentence-1.

Dataset Description:

• Sentence1: String column of human entered text, Sentence 1
• Sentence2: String column of human entered text, Sentence 2
• gold_label: Categorical column inferring logical relation between sentence1 and sentence2

Implementation

• Length of document in sentence1:
• Length of document in sentence2:
• Heatmap of correlation between the features:
• Bidirectional LSTM Model performance(not good due to less data):
• Selected model's performance for predicting the testing gold_label.

Inference

• Since the dataset was very small, training a Neural network was not a good idea so I choose to move ahead with ML algorithms.
• So, working on a large dataset can improve the learning.
• Advanced NLP techniques can be implemented to find the semantic relationship between both the sentences to get a better result.
• Due to lack of time I decided to follow this approach but with various iterations during the development, model's performance can increase significantly.
• Data Cleaning was done signifantly well but can be done using other approaches.
• Feature engineering is one important part which require good knowledge of NLP which can be worked upon in future.
• Dimensionality reduction based on experimentation on using PCA or t-SNE can be perfromed to optimize model performance and remove useless features.
• Hypothesis testing can be done in making useful decissions about the feature, whether they contribute in predicting right gold_label or not.
• Word ebedding can be implemented to get a better semantic relationship between words.
• Working with more better Neural Networks will be a better choice for this kind of problem, although bidirectional LSTM should perform well with large dataset.
• Finally once we get a good model performance over the data, we can implement hyperparameter tuning to tune those small knobs in the bidirectional LSTM model to extract the best performance out of it.
• for any suggestions contact me at [email protected]