What’s in a name?

Below is a snapshot of an email that I received recently,

I have experienced many written and verbal permutations of my name, but being addressed as {{first_name}} was a first. The email has the tracings of an algorithm mining the CEO’s LinkedIN contacts and broadcasting a marketing message to recent connects. The problem of identifying people’s names is a tricky problem to solve. In conversational AI names are the Achilles heel. And they are important to get right for a satisfying dialogue. In the context of the marketing email it was straightforward so, “Sorry {{Dear Sender}} better luck next time”.

Even as humans with evolved cognitive abilities names are difficult. My name is pronounced A-na-gha. A as in the word a, na as the sound of the alphabet n, gha is one sound like in gha-stly. It is a difficult name for some to pronounce. Many a meeting the host goes, “are there any questions Tim, Tom, …umm… Team?”. My sincerest apologies to Tim and Tom for having to field all the questions. Some meetings I am Anya, or Anag-a with the silent ‘h’, or Anag-ya. Then names can sometimes sound like other words. In college every time professors said the word “another”, a friend was startled and looked at me. He always heard it as ‘Anagha’. Since his name was Vinod, I returned the courtesy every time anyone said, “we know the”. And remember the early days of autocomplete in Outlook? To autocomplete I was ‘Anagram’. I have sent and received countless mails as Anagram.

If I had a universal digital assistant think about the gymnastics it would need to go through as it transcribed calls, attended lectures, or mined emails to accurately identify the different permutations.

The approaches to digitally identifying names in a dialogue is context specific and different techniques can be applied to increase the probability of getting names right. For a digital assistant deployed on the phone the contact list can be used as a reference to bias results for spoken names. Using dictionary matching algorithms like the Aho–Corasick that locate elements of a finite set of strings within an input text will give results with a high recall. But the approach also has low precision and will not work well for out-of-vocabulary words. When dealing with names, it is important for the model to be able to interpret such words as many names would rarely or never be seen in training.

Consider a registration system for new patients trying to decode the sentence “Schedule an appointment for May”. Depending on the context it could be the first name or the month. A neural network can be leveraged to memorize embeddings of frequent words, and construct representations of rare words keeping the context of the sentence in perspective to determine whether it is a true representation of a name. This method will give high precision and high recall results. Till our fertile imagination trips up the algorithm again.

So conversational AI continues to be one of the top focus areas for research. Enjoy this comprehensive blog that looks at the different sub-fields of NLP and lists the areas of research focus - https://medium.com/jasonwu0731/conversational-ai-research-roadmap-6920307cbbac

What do you think? Should a rose called by any other name, smell just as sweet?