• Combination of Edit Distance and Compged: We have done experiments of layering one technique over another technique. Here the edit distance technique is used on the standardized dataset with a certain score cut off. Then the compged function is layered over the edit distance output where only the vowels are tampered with & the desired operations are specified with low associated costs using the compcost function.
• Soundex (Original): As explained earlier, it is a phonetic technique which generates a sound code for each name and assigns same code to the similar sounding names.
• Soundex (Modified): We have modified the Soundex code to handle names of Indian as well as US origins. We have generated phonetic codes in Roman script for Indian script’s consonants and vowels. The end result is a modified Soundex code of length 14 characters (instead of original length of 4 characters) while keeping vowels for matching (whereas Soundex drops vowels in original form).

Our experimental results reveal that a combination of Edit Distance with Compged and modified Soundex yield the best results in minimizing both “Wrong Captures” as well as “Non Captures” in a database having names from Indian as well as US origin as shown in the graph below (Non Capture error occurs when a name is not matched to any other name and Wrong Capture occurs when a name is matched to a wrong name).

* Shingles, Compged, Spedis and Edit Distance are String Comparison Techniques while Soundex is a Phonetic Matching Technique

We have further segregated these numbers for names from India and the US and have found that modified Soundex results in high accuracy across names from both origins:

Inductis experience has shown that Name Matching can be significantly enhanced by either modifying an existing technique or creating hybrid techniques through the combination of multiple techniques to take into account the complexities of multi-cultural name databases. In addition, if an input name is compared to what is known about name strings from each culture, and the probability of the name belonging to each of the cultures is computed, it can add further accuracy in Name Matching.

Inductis has developed and implemented a customized algorithm for matching online queries to an online database of tens of millions of names, resulting in an improvement of 5-8% over the previous algorithm and a business impact of $60 million per year. In addition, we are continuing our research in name matching methodology in order to ensure maximum benefit for our clients from this evolving technology.

References:

Acknowledgements:
We are grateful to the following people for their continued support and insights during the experiments and in writing the whitepaper.