Kelvin

Develop a KPI dashboard to analyze the sentiment of agile teams based on aspects.

• Business problem: How do we measure how teams / organizations feel across the multiple stages of a certain project?
• Technical question: How do we perform aspect-based sentiment analysis over conversational data?

Requirements

Basic functional requirements:

• Visualization of Sentiment Analysis Dashboard
• Fine-tuning and deployment of ML Model as microservice
• Integrating the data of communication tools

Basic non-functional requirements:

• Easy to use user interface
• Appropriate deployment strategy

Aspect-based Sentiment Analysis

ABSA is text analysis technique that categorizes data by aspect and identifies the sentiment attributed to each one of them.

• aspect: category, feature, or topic that is being talked about (can be any Part-of-Speech) sentiment
• score**: a value between 0 and 1 which resembles the confidence of a certain text corpus expressing neutral, negative, or positive "opinions" about a particular aspect

NLP - "behind the scenes": Aspect Extraction

• In order for a sequence s of n words to bed into the BERT architecture, they are represented as [CLS], w1, w2, …, wn, [SEP] where:
  • [CLS] is the token indicating of the beginning of the sequence
  • [SEP] is the token to separate a sequence from the subsequent one
  • wi are the words of the sequence
• Goal: Each word of the input is labeled as one of the three letters:
  • B: beginning of the aspect terms,
  • I: inside or the aspect terms' continuation
  • O: outside or the non-aspect terms
• Input => BERT model => vector representation of the size 768 for each sequence => word-to-label mapping

Visualizations

Messages Sentiment Classification

• Bucketing all filtered messages in 3 ways (+,-,neu).
• In case that 1 message has more than 1 aspect that message is duplicated and processed twice.
• The leading sentiment (the sentiment with the highest %) iis used to place the message in the right bucket.
• Pie slice % = bucket.length / (3 buckets).length

Daily Sentiment Overview / Calendar

• We group all filtered messages by day and calculate the total +,-,neu sentiment from each message and aspect of that day.
• We then calculate the avg (+,-,neu) = total (+,-,neu) / counter
• We use these individual +,-,neu value to build the daily sentiment graph.
• We only use the leading sentiment (highest avg score) to build the sentiment calendar.

Aspect Based Heat Map

• Construct a flat array of all aspects found in filtered messages.
• Use this flat array to group messages by the specific aspects.
• For each aspect group we calculate: avg (+,-,neu) per aspect = total (+,-,neu) / counter
• We use the acquired values to construct the heat map.
• Strongest Color = highest score
• Weakest Color = lowest score

Total Count by Word / Word Cloud

• We create a map of all aspects included in the filtered messages.
• We then analyze the filtered messages and calculate the total count for each aspect.
• This mapping of (aspect : total_count) is used in both graphs:
  • Total Count by Word (Aspect)
  • Word (Aspect) Cloud

Total Score by Word

• The same map of ( aspect : total_count ) is used in combination with weights.
• These weights are proposed from the user.
• A higher weight dictates a greater interest on the aspect.

Tech Stack