Invoice Mining Assignment
Approach
The problem revolves around extracting the data corresponding to the year 2019 and all the line items given in the borderless table.
This is not a conventional NLP problem as the spatial information regarding the alignment of data is quite important in this scenario,
which cannot be obtained from conventional NLP techniques.
So my approach is a combination of computer vision and NLP ( gathering spatial information using tesseract and classification using naïve Bayes).
Steps undertaken are as follows:
1) Convert the given text to an image using ascii art algorithm. ( txttoimg.py )
2) To Identify the line Items or table rows from the rest of the component,
I had to train a Naive Bayes model to classify the rows as relevant or not.
For this purpose I collected all the lines of every textfile and manually annotated 1000 lines ( dataCollectionNLP.py )
3) Using the data in step 2 I trained a Model for predicting the line tems. ( createModelNLP.py, predict.py )
4) Iterate through each image from step 1 in a row-wise(line by line) manner while Obtaining spatial information of each sparse line.
Using the model in step 3 identify the table records, now identify the corresponding sparse record (line/word) which is occurring
under the year 2019 by comparing the x co-ordinates of the words.
Format output as per requirement and write to csv. ( invoiceReader_Main_Program.py )