How to make a dialect coach for a machine...and where we need it.

For a few years, I have wanted to start a guerrilla book club of sorts. We would send novels and poetry to software designers in the hope of introducing them to new cultures and languages (human languages).  This relies on them to extrapolate from the authors' artistic methods to articulate the human experience and use a sort of mimesis to create new, nuanced algorithms.

Current projects in ML/AI are trying to understand how and why humans are persuaded, why they trust, blame, coalesce and fragment around ideas or leaders, and then follow these patterns of behavior as though they were governed by rules we might one day anticipate.  We have a long way to go before any of this is possible.  In the meantime . . .

One of the most measurable things that writers have done that is transferable to current problems in NLP and ML involves using dialog with dialect.  When Mark Twain created characters that spoke distinctly American versions of English, how did readers recognize and place these local variations?  How do Japanese readers of Murakami recognize the Kansai dialect?  How do we instantly recognize that a character is from a certain region, city, class, or sub-culture?  The author purposefully and methodically transliterates the sound qualities of that language variant onto the page, and we are able to 'hear' it because we share the same concept, the same expectations for sounds, word choice, and phrasing to indicate a particular sub-group. If an author can plan and execute a dialect, and we can recognize it, then there must be rules. If there are rules, then we can teach a machine to find and follow them.  I am optimistic about this method for improving the acuity of NLP and sub-culture recognition by ingesting and parsing novels with strong dialects in their dialog.  This has been in the works while blog has been 'sleeping.' Updates soon on the research.

This approach to NLP is not as whimsical as it may seem.  A January 25th New York Times article illustrates how applied research in this area, prompted by observing and investigating issues of language and technology, could be beneficial.  In, Speaking Black Dialect in Courtrooms Can Have Striking Consequences, John Eligon wrote that:

"Researchers played audio recordings of a series of sentences spoken in African-American English and asked 27 stenographers who work in courthouses in Philadelphia to transcribe them. On average, the reporters made errors in two out of every five sentences, according to the study. 

The findings could have far-reaching consequences, as errors or misinterpretations in courtroom transcripts can influence the official court record in ways that are harmful to defendants, researchers and lawyers said."

To repeat the finding: 40% of the sentences have errors!?!

Sentencing and other judicial decisions will be based on the ingest and training of from court documents including transcripts.  We assume these documents are reasonably error free or at least not containing errors that reflect substantially different meanings than a witness or defendant intended.  If a dialect makes this much of a difference, improving how machines can be trained around dialect patterns will certainly be useful.  This is a great example of social science led ML design and evaluation for applied research.  (And maybe a little humanities if the book club takes off.)