Svyat Vergun, Senior Data Scientist
February 6, 2019
Listening to long work calls is tedious. In finance, traders and sales people struggle remembering all the details of each lengthy call. They won’t listen to a full company call or earnings call, and likely won’t read a transcript if available. Faced with an unmanageable and intimidating wall of text or a long audio file, traders need a solution to quickly scan the call and identify key points.
Here’s how we’ve solved this problem. Highlights is an important natural language processing (NLP) feature in GK’s platform. Similar to using a highlighter to mark important sentences and phrases from a sales or earnings call, GK will deliver a manageable set of important sentences and phrases from a call, all through one of front end apps (like Focus).
Summarizing Earnings Calls
Below is an example of the Highlights from the first several minutes of an Amazon earnings call. The original text contains 19 sentences and 350 words. Highlights delivered it in 3 sentences and 49 words.
Original text (left) vs Highlights (right)
How it works
Highlights analyzes sentences individually and ‘inter-relatedly,’ and selects sentences that are objectively most important. The underlying AI model was trained and refined on financial domain data that revealed which sentences and phrases traders care about most.
Further, Highlights analyzes sentences based on sentiment, proper nouns, key terms (using a language graph ranking algorithm), digits, currencies, psychological factors and features of the audio itself (volume and stress, speed of speech, speaker change). Sentences are ranked, unimportant ones are removed and valuable ones are conserved to give a concise, manageable overview of the call or text record.
What’s next: Summarization
We’re working on a Summarization feature which provides an interpreted summary of transcripts, adding further abstractive interpretation value along with condensing a call.