Address parsing for a hyper-connected world - a neural network approach

It’s no surprise that we’re seeing more parcels shipped than ever before. With the relentless growth of ecommerce sales, new digitally comfortable consumers come online every day.

Wed Feb 07 10:57:00 EST 2018
,

It’s no surprise that we’re seeing more parcels shipped than ever before. With the relentless growth of ecommerce sales, new digitally comfortable consumers come online every day. For retailers with an online presence, that’s mostly good news.

However, this growth introduces challenges in ensuring accurate data is collected at the point of on boarding new customers to ensure merchandise makes it to the right location without the cost of failed deliveries.

When a consumer enters an address, its’ important to separate out the various data elements – to parse it, in the right way to ensure the location is verified as deliverable.

Traditionally, delivery address verification involves manual effort, or at the best, is dependent on some hardcoded rule-based processing.

These hard-coded rules bring in a number of problems such as understanding country specific conventions and variations, effort in developing new rules and maintaining existing ones as address conventions change. This requires a lot of effort, time and money, and is a mammoth task a data manager or data team. Even after putting in the effort, the quality of the solution is not full proof. All of these limitations lead to production issues, costly delivery failures and delays for the customer.

Due to the volumes involved, this is now becoming a big data problem and can be addresses with machine learning.

Pitney Bowes has developed an approach using supervised machine learning neural network-based techniques to understand the structure and variations of different type of addresses. A single model engine learns and interprets the thousands of rules required to parse an address in an automated manner.

The Global Address Parser, has been introduced to the Spectrum Technology Platform and ­­uses this state of the art approach for accurate, reliable and scalable address parsing.

Its inherent nature of understanding convention and variation in addresses, just ‘seeing the data’ enables Global Address Parser to handle cases of fuzziness such as spelling errors, ambiguities and abbreviations where traditional systems are at risk of failure.

A machine learning approach, using neural networks, outperforms the earlier approaches both in terms of quality and performance.

Going forward it will be possible to expand the Global Address Parser for new countries in a generic manner with a quick turn-around time and without the requirement of much manual intervention.

This approach was presented at the 12th IEEE International Conference on Semantic Computing, Los Angeles, California in January 2018.

The paper, “Automated Parsing of Geographical Addresses: A Multilayer Feedforward Neural Network based approach” was authored by Shikhar Sharma, Ritesh Ratti, Himanshu Kapoor, Ishaan Arora and Gaurav Bhatt from Pitney Bowes.

To find out more about the conference, go to: http://www.ieee-icsc.org/

To learn more about how entity resolution solutions provide valuable insights, download the IDC Whitepaper.