Sentiment Analysis

The original promise of sentiment analysis — a single positive-or-negative score per message — has aged badly. A negative review that complains about delivery is operationally different from one that complains about quality, and both are different from a complaint about pricing. Lumped together as a polarity score, all three are invisible. Aspect-based sentiment dissolves the lump: each message is decomposed into the aspects discussed and the sentiment expressed about each, so that the resulting analytics align with the way the business is actually organised.

We treat aspect schemas as living artefacts. The categories that matter at launch — service, price, quality, delivery — quickly need to be supplemented with finer ones as the data reveals patterns the original schema cannot express. Our deployments include explicit processes for evolving the aspect taxonomy without invalidating historical reporting, so that the analytics remain comparable across time.

Alongside aspects we track intent (complaint, query, compliment, suggestion, churn signal), emotion where it is operationally useful, and explicit named entities (products, locations, competitors). The combination turns a stream of text into a structured event log that other systems can subscribe to.

• Aspect detection and aspect-level sentiment
• Intent classification (churn, complaint, query)
• Named entity recognition for products and brands
• Sarcasm-aware classification for review channels
• Multilingual scoring with language-aware models
• Explanation snippets for analyst trust

Brands operating across languages frequently end up with parallel sentiment systems for each market, scored on slightly different schemas, never quite comparable. We build single multilingual systems with consistent schemas and consistent scoring behaviour across languages, so that a regional dashboard means what a global one means and vice versa.

Our models combine multilingual transformer backbones with task-specific adapters fine-tuned on representative data from each language of interest. Where a particular language has thin training data, we use translation-augmented training, cross-lingual transfer and active labelling to bring quality up to acceptable parity. Crucially, evaluation is done per language, not on an average that can hide a mediocre performer behind a strong one.

Language-specific phenomena — Arabic dialectal variation, German compound nouns, Japanese politeness levels, English sarcasm — are addressed through targeted data collection and model adjustments rather than generic improvements. The result is multilingual analytics that hold up to scrutiny in each market.

“If your sentiment system gives the same answer in twelve languages, ask whether it is reading any of them.”

Sentiment analytics at rest are interesting; sentiment analytics in motion are valuable. We integrate classification output directly into the operational systems that act on it: support-ticket routing, CRM alerts, marketing-suppression lists for at-risk customers, product-roadmap dashboards for category managers, and PR-team notifications when emerging topics start to spike. The feedback loop closes when the people who act on the signal can also annotate it, feeding the labels that improve the next model.

Real-time alerting is treated with care. False alarms erode trust in alerting systems quickly, so we tune thresholds based on historical signal-to-noise, smooth them with appropriate windowing, and pair them with confidence intervals that distinguish a genuine spike from random variation. Where alerts depend on emerging topics — issues that did not exist last month — we use online clustering and topic-detection methods to surface them as soon as they reach an operationally meaningful volume.

Dashboards are designed around the questions that consumers of the data actually ask, not around what the data scientist found interesting. A product manager wants to see week-on-week movement in aspect sentiment for the products they own; a contact-centre lead wants to see the queue of newly negative tickets in real time; a brand director wants to see share-of-conversation against competitors. Each view is built from the same underlying classification stream.

Public sentiment benchmarks are an unreliable guide to performance on real business data, which is messier, more multilingual and more domain-specific. We measure quality on representative samples of your actual data, with labels collected from the people who would normally adjudicate ambiguous cases (your contact-centre team, your moderators, your category managers). The resulting baseline is the one against which all subsequent improvements are tested.

We pay particular attention to inter-annotator agreement. Sentiment is famously slippery; if your own experts disagree with each other on the labels, no model can be expected to produce consistent output beyond that bound. We measure agreement explicitly, refine guidelines where it is poor, and report model performance against the ceiling that the human label quality permits.

Bias and fairness are checked alongside accuracy. A sentiment model that is systematically harsher about messages in one language, or about content from one cohort of customers, will distort downstream analytics in ways that take a long time to detect. We test for these patterns explicitly and recalibrate where they appear.

A typical engagement starts with a discovery on a representative slice of your text data — perhaps three months of reviews, tickets and social mentions — to characterise volume, language mix, aspect distribution and the operational decisions the analytics will support. From there, we build an initial classifier, integrate it into one priority workflow, and grow the deployment as the appetite for signal grows. Most clients begin with a single use case (typically support ticket triage or product-issue detection) and reach the full multi-channel deployment over the following six to twelve months.