The Edit Blog

The Edit AI Christmas Carols

ARTICLE BY Hannah Newcombe
    READ TIME: 6 mins
    14th December 2018

    Written by AI. Crafted by Humans.

    This year, Edit used an AI to create a weird and wonderful series of Christmas carols, sharing all our favourite festive tunes with IBM Watson Tone Analyzer.

    Watson provides the name of the tone identified for each lyric analysed, plus a confidence level ranging from 0 to 1. With each sentence, Watson is looking to identify words or phrases that can be classified as one or more of the following tones: Anger, Analytical, Confident, Fear, Joy, Sadness and Tentative.

    We took the top-ranking lines from each emotion and crafted them into five new festive ditties with illustrations to match. Although the AI put together some creative combinations, all of the carols needed human skill to apply correct tenses, poetic rhythm, context and most importantly, cultural understanding.


    The Watson Tone Analyzer placed all of the selected lines in a spreadsheet in order of the highest confidence level for each of the emotions or combinations of emotions. This made picking the lines for each of the new carols the easy bit. The skilled component was then to draft the suggested text into verses that read well.  The boundaries between AI and skilled humans becomes very evident at this stage.  The real capability of AI is to process high volumes of data at speed and to provide a consistent, reliable output. However, from then on skilled humans are needed to take the raw output and craft it into a usable final product. For example, the mixing of tenses is not something that the AI picks up on. In ‘The Jolliest Carol’ AI wrote ‘We all know that Santa’s coming, He had a very shiny nose’, which would lead to questions like ‘What happened to Santa’s nose?’ and ‘How did he lose the shine!?’. In this instance human had to step in and change ‘had’ to ‘has’ to restore the order and keep Santa’s nose shining.

    Poetic Rhythm

    IBM Watson ranks all of the individual lines with a score from 0-1 based on a specific emotion, such as Joy. What it can’t do is bring linguistic rhythm or syllable satisfaction to traditional poetic conventions. No rhyming couplets, triplets or stanzas. Instead, it disrupts our secondary-school-knowledge from what we know about songwriting by offering up unconventional pairings like the abrupt ending to ‘The Darkest Carol’. The blunt 3 syllable last line ‘Heal the world’ stops the reader dead.

    The Edit humans have embraced the AI’s poetic rhythm however in ‘The Jolliest Carol’. ‘Fa la la la la la la la la’ occurred so frequently in the original carols we entered into IBM Watson, and scored so highly for Joy on the Watson rating (0.966751) that it was placed at the end of each verse.

    The AI is very good at taking the essence of a sentence and ranking it according to the individual words that can be read within it. What it’s not great at is applying the context when two words come together to create something unexpected. A brilliant example of this can be found in ‘The Happiest Carol’: the line ‘You have termites in your smile’ scored a very impressive 0.995826 for ‘Joy’, yet the reality of having those creepy crawly creatures scuttle around your teeth like  you’re on a Bushtucker Trial is nothing short of stomach-turning. However, it does open up  the human mind for creative interpretations as seen in the charming illustration of the Christmas tree who somehow looks delightful despite the infestation!

    Cultural understanding

    For AI to truly understand human sensitivities and political correctness it would need to digest thousands, if not millions, of years of information and human understanding. A colossal task.

    As humans, we are brought up by our own kind and have a shared historical knowledge, cultural understanding and social etiquette that’s key to our survival. Or simply put, we generally know how to successfully communicate with other humans. AI does not. It has no filter. A comedian leans into forbidden topics of conversation for humour. AI will share the dark content but doesn’t know how to deliver the punch line. An author may touch upon taboo subjects to explore its boundaries. AI will bring up controversy but can’t land a poignant message.

    During this AI Christmas experiment there was a religious point of tension as half of the lyrics given IBM Watson were from Christmas Carols. To demonstrate AI’s lack of cultural understanding for the Christmas story, here’s one of the dark, peculiar carols that never made it to the final five:


    The Scariest Carol.

    He’s a real victim of fear and of pride,

    The little lord Jesus asleep in the hay

    At the scene of the attack

    We’ll conspire

    They howl out for more.

    You’ll hear.

    You’ll hear.

    Unkind as any.

    You’ll hear.

    In summary, AI is a wonderfully, mischievous and bizarre toy for creatives to experiment with. It can be a great thought provoker; writers’ unblock, or simply the fresh eyes needed to put a twist on a campaign. Harmony and happy accidents between humans and machines. The caveat to this is that there will always be a bias with AI as it can only show you combinations of the information that humans inputted into it in the first place. For now, the world of AI is rich with opportunity – humans just need to play with the possibilities.


    The techie bits

    Edit has been working with IBM Watson Technology for many years. From our heritage in using IBM Campaign through WCA, Interact and other regular components we are also actively making use of the newer AI based capabilities. This ranges from an award-winning program making use of Visual Recognition through chatbots to Personality Insight analysis.  The Personality Insights analysis has been particularly helpful when working with understanding how items such as blog posts and product descriptions are being presented. The Tone Analyzer is a capability that has been used to good effect within chatbots, enabling the chatbot to gain an understanding of the emotion of the conversation and to adjust responses appropriately.

    Making use of the Tone Analyzer is another example of how we are using AI to bring together the bulk processing and analysis capability of the technology in support of the skilled humans who can then take the output and add the necessary human control to it.

    As with many of the various projects we deal with, the IBM technology is often the easier bit to put in place. In this case, we recycled some of the Python scripts that were used to support the Personality Insights projects.  Staff members were asked to provide a list of their favourite seasonal songs and carols to form the basis of the text that we could work from. The text was then processed by the Python scripts. Each document was read in and a call was made to the IBM Watson Tone Analyzer API for analysis. The Tone Analyzer is passed either sentences or paragraphs. It then processes each line of text and using linguistic analysis detects the emotional and language tones being used. In common with many of the Watson API’s a JSON response is returned.

    Some of the lines were deemed to have no recognised tones while others contained one and, in some cases, more than one tone. In order to help the humans select the lines for the final new carols, the output from the JSON object was processed to provide an easy to use spreadsheet. At this point, the AI process handed over to the skilled humans.

    Providing the output from the Watson Tone Analyzer in a spreadsheet allowed the selection of lines that exhibited the highest confidence level for each of the emotions or combinations of emotions. This made picking the lines for each of the new carols the easy bit.

    We hope you enjoy reading the weird and wonderful carols and we wish you a Merry Christmas from the team at Edit.