Miniproject: Machine Learning

Supervised and Unsupervised Text Classification.

1. We created sections using ami3 which look like the following example:

https://github.com/petermr/openDiagram/blob/master/physchem/resources/oil26/PMC5454990/sections/2_back/0_ack.xml

<?xml version="1.0" encoding="UTF-8"?>
<ack>
 <title>Acknowledgments</title>
 <p>The authors are grateful to CNPq-Programa “Ciências sem fronteiras” (Grant No. 233761/2014-4) for financial support.</p>
</ack>

2. Further we flattened the xml into txt format for increased readability.

https://github.com/petermr/openDiagram/blob/master/physchem/resources/oil26/PMC5454990/sections/2_back/0_ack.txt

3. We have 2 major problem statements:

• Not all sections are labelled with universally accepted vocabulary.
• We want to improve our knowledge resource by clustering together similar articles on a paragraph or section basis. E.g. Using unsupervised learning we find out that gas chromatography is a frequently used phrase, we use it as a label to group together other articles that mention gas chromatography.

4. Goals

• We plan on extracting keywords and phrases using NLTK rake. We create a bag of words and tf-idf representation of the data. We manually agree on the labels we want to use for topic modelling. https://en.wikipedia.org/wiki/Tf%E2%80%93idf
• We want to work with different tools and libraries in python and discover the tools which serves our purpose best.
• Create a robust open access knowledge resource for invasive plant species and aromatic plants.

5. Tools:

• Scikit-learn clustering models https://scikit-learn.org/stable/modules/clustering.html
• gensim https://pypi.org/project/gensim/
• countvectorizer https://www.geeksforgeeks.org/using-countvectorizer-to-extracting-features-from-text/
• tf-idf https://en.wikipedia.org/wiki/Tf%E2%80%93idf
• LDA https://en.wikipedia.org/wiki/Latent_Dirichlet_allocation
• cosine similarity https://en.wikipedia.org/wiki/Cosine_similarity
• spacy https://spacy.io/models#conventions
• Jupyter Notebook for ease of collaboration, documentation and packaging.

Proposed method :

1. Use unsupervised K-means clustering to derive features names. This gives us a good feel for the data (aka data exploration) we are working with. We use these features to create an ami dictionary.
2. Implement key phrase extraction so as to make our list of feature names more comprehensive (clustering only extracts the most relevant features), this helps in accommodating different writing styles/expressions authors have from all over the world. We should use pke library for key phrase extraction purposes. Manually add all the key phrases which seem important to the given label. For e.g "authors thank" is a common theme across acknowledgment sections.
3. Perform supervised search across fulltext documentsso as to find and extract similar sentences (boiler plate similarity of sentences) which are hidden under wrongly labelled/unlabeled sections. Analyze the extracted sentence corpus to strike off false positives and noise.
4. Create a binary classifier using different classification models such as logistic regression, k nearest neighbors etc.
5. Label sentences and sections properly so as to make the knowledge resource more readable, useful and accurate.