MarketingFramework / modelModelAccessible

Sentiment analysis

How can sentiment analysis improve people, teams, or organisational effectiveness?

AccessibleOperationalTeam2 min read

Contents

Sentiment analysis, also known as opinion mining, seeks to extract subjective opinion or sentiment from text ([Text Analytics](../text-analytics--b4af9c2e/index.md)), video ([Video Analytics](../video-analytics--3ed05a13/index.md)) or audio data ([Voice Analytics](../voice-analytics--c3044941/index.md)).

Sentiment analysis, also called opinion mining, classifies evaluative language in text and can be extended cautiously to audio or video. It is often used with Text Analytics, Video Analytics or Voice Analytics, but each medium presents different validity and ethical limits.

When to use it

Use sentiment analysis when a decision requires a consistent summary of expressed opinion across a volume of feedback.

It can help explore:

How customers describe a brand or experience.

Which product themes attract praise, criticism or uncertainty.

How perceptions differ from competitors under comparable data.

Which employee-experience topics require direct investigation.

Use it for triage and pattern detection, not as a covert measure of a person’s true emotional state, honesty, health or intent.

Origins

The field combines computational linguistics, information retrieval, text classification and earlier content analysis. “Opinion mining” and “sentiment analysis” became prominent as online reviews, forums and social media created large collections of evaluative text. Lexicon methods were followed by supervised machine learning and later contextual language models.

What it is

A basic model assigns text to positive, negative or neutral classes. More detailed systems identify target, aspect, intensity, stance, emotion or change over time. The unit matters: a review can praise delivery and criticise durability in the same sentence.

The claim that words account for only 7 per cent of communication is a misuse of Albert Mehrabian’s narrow experiments on inconsistent emotional messages. It is not a general law of comprehension and does not justify inferring truth from tone or body language.

Sentiment is expressed, contextual and culturally variable. Sarcasm, negation, dialect, mixed opinion, quoted speech and domain language create error. Aggregate results may also reflect who chose to speak, platform moderation and coordinated activity rather than the full stakeholder population.

How to use it

Define the decision, population, source, unit of analysis and label scheme. Sample the data, develop annotation guidance and use trained reviewers to establish a defensible reference set.

Select a lexicon, statistical or language-model method appropriate to the domain. Evaluate precision, recall, calibration and confusion by language, topic and relevant group. Keep an “uncertain” route and route high-impact cases to human review.

Report themes and sentiment together; a single score cannot tell management what to fix. Weighting, deduplication and bot or campaign detection may be needed, but methods should remain transparent.

Audio prosody can signal arousal under some conditions, while facial-expression systems are unreliable proxies for internal emotion across contexts. Neither voice nor face should be used as a lie detector. High-stakes employment, insurance, welfare or health use requires a much stronger legal, scientific and ethical basis.

Practical examples

Researchers at Microsoft Research examined public posts for language patterns associated with later reports of postpartum depression. Such work may generate hypotheses for voluntary clinical research, but a social-media classifier is not a diagnosis. Deployment would require prospective validation, consent, clinical oversight, crisis pathways, bias assessment and strong privacy protections.

Call centres may use aggregate acoustic and text signals to identify conversations needing support. Agents and callers should know the purpose, automated alerts should not determine adverse action, and direct statements and context should outweigh inferred emotion.

Top practical tip

Validate on recent, representative examples from the exact domain and publish the error profile. Use sentiment to locate themes for human investigation, then confirm them through direct research and operational evidence.

Top pitfall

Do not present model output as the “real truth” about a person, diagnose health, infer deception or take adverse action from voice, face or text sentiment. Expression is not an unmediated window into internal state.