Sentiment analysis scoring inconsistency between languages

Genesys Cloud WFM
1

We built a custom sentiment dashboard in React that visualizes the scoring inconsistency.

const SentimentChart = ({ data }) => (
  <div>
    {data.map(lang => (
      <div key={lang.code}>
        <span>{lang.name}: </span>
        <span style={{color: lang.avgScore > 0 ? 'green' : 'red'}}>
          {lang.avgScore.toFixed(2)}
        </span>
      </div>
    ))}
  </div>
);

English averages +0.15, Spanish +0.08, German -0.03. The German model is pessimistic.

2

The sentiment scoring inconsistency directly affects our outbound campaign targeting.

We use sentiment scores to prioritize re-engagement callbacks. German-speaking customers consistently score lower, so they are deprioritized by the algorithm. We are effectively discriminating against German customers due to a model calibration issue.

3

For callbacks, sentiment-based prioritization must account for the language bias.

We normalize sentiment scores per language before feeding them into the callback priority algorithm. A German score of -0.1 is equivalent to an English score of +0.05 after normalization. Without this adjustment, German callers wait longer for callbacks.

4

Our analytics dashboard shows the sentiment inconsistency clearly when filtered by language.

I built a query that segments sentiment by conversationLanguage:

{"filter":{"type":"and","predicates":[{"dimension":"conversationLanguage","value":"de"}]},
"metrics":["oSentimentScore"],"groupBy":["queueId"]}

The German queues consistently show 20% lower sentiment than English queues handling identical issue types.

5

Is there a financial impact to this sentiment scoring problem?

If our NPS surveys show German customers are equally satisfied as English customers, but the AI says they are unhappy, which data source do we trust? I need the analytics team to reconcile these numbers before the quarterly board review.

6

We benchmarked the sentiment models across languages using JMeter and a synthetic audio test set.

The English model has 87% accuracy. Spanish drops to 74%. Japanese is 61%. The models are trained on different corpus sizes, which explains the variance. English has the most training data by a significant margin.