In order to process bank statement narratives using optical character recognition, we will be using a multi-stage system which we call OCR/narrative prediction. The stages are:
(1) Scan the bank statements using OCR. After narratives are scanned, we process them to remove clutter from the leading and trailing parts. For example, if the narrative says “1rect deb1t payment to David Porthouse and Co on 18 Sep” then the initial stage will be to recognise that this should be amended to read “Direct debit payment …”. However, the leading bit of verbosity can be removed altogether and the trailing date is also redundant, so our software will amend this to just “David Porthouse and Co”.
(2) Compare this to last year’s narratives, where we might find “David Porthouse”. The narrative will then be amended to say just “David Porthouse”. A variety of narratives such as “David Porthouse & Co” or “David Porthouse Ltd” will all be cut down to read just “David Porthouse”.
(3) If a narrative is not in last year’s lexicon, then it needs to be perfect to survive, such as letters only. Otherwise it is deleted.
(4) Deleted narratives are then predicted based upon numeric values. For example, if we had a large number of payments of 41.33 to David Porthouse, then we can predict that the next payment of 41.33 will also be to David Porthouse even if we cannot scan the narrative because it falls on a paper crease or has been scribbled over.
(5) Reprogram the function keys F1 to F10 with the commonest narratives encountered so far. Then the clerk needs to inspect each narrative and overtype it with the correct narrative if required. Overtyping can often use a function key rather than having to type it all in. If we need to type it, autocomplete often helps.
When we reprogram the function keys, we try to ensure that F7 is some type of telephone expense, and key F9 is some type of motor expense. This makes the system more useable.
This system has plenty of graceful degradation and backup systems. OCR and NP are both capable of getting it right first time, but in cases where they don’t, we have other systems which take a little longer, but never disproportionately longer. We can use this system on handwritten records which resemble bank statements. OCR will be useless, but the rest of the system is still there. We can use it where we need to type in invoices one at a time, because while OCR and NP are both useless as they stand, NP has a secondary action of reprogramming the function keys in bulk which can be helpful. It can be arranged that if the client drives a Diesel, then key F9 will generate “Diesel fuel”, while if the client drives a petrol-engined vehicle then F9 will generate “Petrol”.
All the keys on our keyboard are covered in stickers which are normally white lettering on black. Keys 3, R and F are reversed-art black on yellow and in the dates column they generate the dates 3, 13 and 23 respectively with month and year being copied down automatically. Keys 6, U and J are reversed-art black on white and generate 6, 16 and 26 in the dates column. We can whiz through a pile of receipts for petrol with this system. Key F9 is reversed-art black on yellow so it looks the colour of petrol.