Yelena Dobric
Yelena Dobric
Language analyzer, NLP (Natural Language Processing), for an asset management company
Situation:
Sapient had pitched this idea to the client: Model analyzes psychological features of companies, by quantifying language usage in earnings call transcripts. One of the stakeholders bought into it, and collaborated with Sapient Data Scientists, to build the model in Excel.
Earnings call is a conference call between the management of a public company, analysts, investors and the media to discuss the company's financial results during a given reporting period, such as a quarter or a fiscal year.
Task:
Transfer the model from Excel, to UI.Team:
Data Scientist with a PhD in Linguistics, Risk Analyst, Product Manager and Front End developer.
Action:
As the only UX resource on this project, my role was:
Digital Transformation: Initial iteration of the model was in Excel. I envisioned the structure for the digital product, interactions and data visualization.Explainable AI: I created visuals to explain to client stakeholders how the model works.
NLP (Natural Language Processing): translate scientific terms into common words.
Personas:
Creators: Sapient's Data Scientist/PhD Linguist (language expert) pitched the idea of analyzing language, to help inform investment strategy, to the client. Client embraced it. Risk Analyst from the client side, became the project's sponsor.
Users: Analysts & Traders.Result:
● By the time we completed work on this model, it was able to process 100,000 calls at once.
● There was clear correlation between certain detected Emotions, and drop in Stock Price, for example.
● Model was successful in providing signals, and additional insight, to inform investment research and strategy.Initial product in Excel is below:
Yellow line represents Fear, detected in Corporate Presentation portion, of the General Electric Earnings Call Transcript, measured over time, between 2003-2018. GE is used as a sample; the model can show one company, or aggregate. Such as Sector.
Blue line is Fear, detected in Analyst Questions-Corporate Responses portion of the Call.
Thick black line is the Stock Price.
Pink bar in the background, is where we want viewer to focus. This area particularly highlights how, where Fear peaks, stock price drops most drastically.
Heinz: as Fear, Anger, and Disgust are detected over time, stock price begins to drop.
Behind the scenes: grid view. Shows how words are identified, and coded, to detect a particular Emotion, or Perspective.
It is interesting to note that there is no machine learning in this model. This model does not "learn and evolve". It groups the words, counts them, and applies labels. That's it. It's not machine learning, but it is NLP (Natural Language Processing).
Explainable AI: I used a visual metaphor, to explain to client stakeholders how this model works: if a Lexicon is an apartment building, Lexemes are families who occupy the apartments.
The initial analog, Excel version of the product.
My first step was to translate the expressions our PhD Linguist used, into common words. I experienced some push-back from the Data Scientist.
Next step was to identify patterns and group similar words into categories.
My UI: I outlined what it is the user is looking at, for first time users, by adding a clear chart title. Filters are shown in the next slide; they expand/collapse. Word buttons are persistent, although the display can be customized.
MVP: User can focus on a particular segment of the Earnings Call Transcript. Corporate Presentation section of the call is prepared, and often rehearsed. While "Analyst Questions" section is more probing, and answers more immediate.
Disabled filters will be available in Day #2. We kept them, to show our future vision.
Word buttons are placed in touch screens' hot spots.
Model in Production. Video is 1:04 minutes long. No sound.
Future feature: adding Stock Price.
