Textual entailment task
Budget: $60 – $100 USD
Textual entailment task
· Given two sentences classify if first sentence follows from second sentence or if they are unrelated
· We will use two datasets for this task Fever 2.0 dataset https://s3-eu-west-1.amazonaws.com/fever.public/fever2-fixers-dev.jsonl (Links to an external site.)and https://www.kaggle.com/c/fake-news-pair-classification-challenge
Project details.
Check if know the task clearly,
Is it supervised? or unsupervised task?
Pattern detection or anomoly detection?
Write down the task description, input, expected output, potential challenges you expect
if you have the appropriate data for the task
Look for multiple datasets preferably
if it is supervised task do you have labels?
If you want to use deep learning do you have enough data?
Decide your roles in the project
Divide tasks team work is important
Roughly each member should contribute for 1/3rd of the load (or whatever your team size is)
Cleanup, Explore the data
Remove noise if any
Check for outliers
Normalize, feature scaling, binarize see what is necessary
apply SVD or PCA or TSNE to visuzalize
Decide which data mining algorithm is suitable?
Possibly several
Try a few algorithms you are free to use sklearn or whatever library you prefer
Possibly ensemble techniques like random forest or adaboost work better
Possibly a new method, or novel set of features for the same methods
Experiments
paramter tuning
Learning rate, regurlarization parameter, dropout etc
Depth and width of neural network
Model specific parameters for CNNs and RNNS or any other type of neural network you are training
For unsupervised techniques cluster size etc
compare all the approaches
document the limitations of all the approaches you try
Choose the best performing method and do more detailed analysis
evaluation measures
Precision, Recall, F1, ROC curve
Statistical significance test (see the examples here (Links to an external site.))
Report
At least 4 pages (max 10 pages) in 2 column ACM conference style in latex (Here is the link to the template on overleaf (Links to an external site.))
Writing should be high quality technical writing not copy pasting think of this as a writing practice for your thesis (Refer to this book (Links to an external site.) for tips on good writing )
Describe the task, dataset, methods you use and discuss their limitations you saw in the experiments
Don't forget to include the github link to your code in the report along with a README.md in github repository which mentions how to reproduce your results in the report.
Very important: Document who did what and justify you did equal share of the tasks
Presentation
5-10 minutes presentation per team
Discuss the problem, data, interesting findings etc
In person project presentations are preferred. But you can present over zoom as well.
· Given two sentences classify if first sentence follows from second sentence or if they are unrelated
· We will use two datasets for this task Fever 2.0 dataset https://s3-eu-west-1.amazonaws.com/fever.public/fever2-fixers-dev.jsonl (Links to an external site.)and https://www.kaggle.com/c/fake-news-pair-classification-challenge
Project details.
Check if know the task clearly,
Is it supervised? or unsupervised task?
Pattern detection or anomoly detection?
Write down the task description, input, expected output, potential challenges you expect
if you have the appropriate data for the task
Look for multiple datasets preferably
if it is supervised task do you have labels?
If you want to use deep learning do you have enough data?
Decide your roles in the project
Divide tasks team work is important
Roughly each member should contribute for 1/3rd of the load (or whatever your team size is)
Cleanup, Explore the data
Remove noise if any
Check for outliers
Normalize, feature scaling, binarize see what is necessary
apply SVD or PCA or TSNE to visuzalize
Decide which data mining algorithm is suitable?
Possibly several
Try a few algorithms you are free to use sklearn or whatever library you prefer
Possibly ensemble techniques like random forest or adaboost work better
Possibly a new method, or novel set of features for the same methods
Experiments
paramter tuning
Learning rate, regurlarization parameter, dropout etc
Depth and width of neural network
Model specific parameters for CNNs and RNNS or any other type of neural network you are training
For unsupervised techniques cluster size etc
compare all the approaches
document the limitations of all the approaches you try
Choose the best performing method and do more detailed analysis
evaluation measures
Precision, Recall, F1, ROC curve
Statistical significance test (see the examples here (Links to an external site.))
Report
At least 4 pages (max 10 pages) in 2 column ACM conference style in latex (Here is the link to the template on overleaf (Links to an external site.))
Writing should be high quality technical writing not copy pasting think of this as a writing practice for your thesis (Refer to this book (Links to an external site.) for tips on good writing )
Describe the task, dataset, methods you use and discuss their limitations you saw in the experiments
Don't forget to include the github link to your code in the report along with a README.md in github repository which mentions how to reproduce your results in the report.
Very important: Document who did what and justify you did equal share of the tasks
Presentation
5-10 minutes presentation per team
Discuss the problem, data, interesting findings etc
In person project presentations are preferred. But you can present over zoom as well.