如何在nlp中进行特征提取

Question

I am building a multi class text classifier , that has data set of a job portal. The data set consists of names of organisations mapped to actual name (see below). I want to make a ml model which can predict actual organisation name.
My data set looks like this:

Flipkart.com flipkart

FlipKart pvt ltd flipkart

flipkart.com flipkart

My question is this:

A.) What kind of features can I extract?
B.) Should my feature extractor use the labels of training set too?
C.) What should my features look like, since they are supposed to be dict for nbclassifier. What key to what value?

I'm new to NLP, any help would be appreciated. Source code on github

Answer 1

I would leave machine learning out of the equation. What you're trying to do is fuzzymatching, with potentially some synonym deprecation.

An expensive technique is the levenshtein distance formula, a cheaper, although just as effective in some cases, technique is token/ngram chunking and indexing.

Make a dictionary of n-grams where n is the length of the gram. n = 3, then grams for Flipkart.com are 'Fli', 'lip', 'ipk', 'pka', etc..., with the key being the ngram and the value being a list of matches that contain that ngram. For each of your n-grams in the input string, look it up in the dict (achieved at O(log(n)m) where n is number of total indexed n-grams and m is number of n-grams in input string), and tally the results until you have a 'score' for each match according to how many n-grams it shares with the input string.

The 'chunking' I mentioned is indexing 'chunks' or sets of n-grams and performing the same task. Aka ['Fli', 'lip', 'ipk'] would be what is indexed and used to tally results.

These techniques can be peformed using 'tokens' as well, rather than or in addition to n-grams, to capture entire words that match.

None of this requires statistics and, instead, leverages an understanding of language.

Or, you can try to derive a meaningful set of features from a list of short strings to map to an extremely large set of classes. This will be an extremely difficult task, and so I recommend the fuzzymatching approach.

Answer 2

First of all you have to convert all the text data into machine readable form as machine learning algorithms only understand vectors.

1) Find the vocabulary of the dataset

2) Use CountVectorizer() or tfidfVectorizer() to convert the text into vectors

3) Now train a naive bayes classifier on the pre-processed dataset

for more detail check this out https://machinelearningmastery.com/prepare-text-data-machine-learning-scikit-learn/

Answer 3

Since you are trying to work with the text I think you should train your model using the GloVe model. This model is a word to vector model Which has a Large Dataset with the vectors for all the words in the dataset. GloVe Model: https://nlp.stanford.edu/projects/glove/

The advance version is the Sentence Encoder https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/46808.pdf

I believe you must study the word embeddings to get a brief idea to proceed with text processing. You can see the details to proceed with Text Processing here: https://www.analyticsvidhya.com/blog/2018/02/the-different-methods-deal-text-data-predictive-python/

I hope this helps. All the best.