Every number we publish starts as something someone said

Conversations containing keywords directly or indirectly related to the subject under analysis — across posts, videos, news, comments and shares.

All conversations about the subject under analysis, generated across news media, forums and social platforms.

A group of products, services, business types and organisations sharing a structure, operating toward the same commercial purpose or distributed product, and serving the same customer base.

Words or phrases used to assemble the dataset.

The sources that generate and hold the conversations to be collected and analysed.

Conversations arising from a company’s paid activity, from press coverage through to social platforms.

Conversations posted and managed by the administrators, plus member comments, on a company’s official website, fanpages and YouTube channels.

Discussions created organically by internet users that mention the brand.

Discussions tagged to a content group or related topic, capturing the character of the information or event.

The users who create posts, or who leave comments.

The relative weight of conversations about a brand, product or service across channels, measured against competitors in the market.

The ratio between positive and negative discussion. Neutral discussion does not enter the calculation.

User emotion on social platforms, in three classes:

• PositiveTích cựcLiking or loving the brand or campaign.
• NeutralTrung lậpMentions the brand without expressing emotion.
• NegativeTiêu cựcDisliking or rejecting the brand or campaign.

Used to compare a brand’s sentiment or attribute against the category as a whole. A value of 1 means level with the category; above 1 means ahead of it.

The emoji reactions users leave on social platforms: like, love, sad, angry, haha.

User behaviour and trends, drawn from the collected data.

The topics carrying the highest conversation volume.

The users generating the most conversation on social platforms.

Discussions or exchanges between two or more people and the interactions they leave: likes, comments, shares.

Conversations about online and offline promotional activity.

The value of the brand across social channels.

Brand health in media, determined by three metrics: total mentions, net sentiment rate (NSR), and unique voice.

The key factors securing the success of a brand’s communications activity.

The process a customer moves through while buying and using a product or service.

Responses and discussion arising directly from a campaign’s message.

Responses and discussion arising directly from the brand and from its campaign messages.

The cost of placing media articles and advertising slots on TV and radio.

The brand value earned from each communications activity across TV, radio slots and media articles.

Finding data by what a sentence means rather than by keyword match alone. A post that is plainly about the brand while naming none of its keywords is still found.

Happens at step 1 · Collect

Grouping discussions that carry the same content into one cluster by meaning, rather than by matching them character for character.

Happens at step 2 · Cluster

Two discussions that are not identical but carry the same content — an emoji added, a hashtag appended, a few words changed. Counted as one discussion.

Happens at step 2 · Cluster

One release or article republished across many outlets counts as one discussion with many sources, rather than as many separate discussions.

Happens at step 2 · Cluster

How strongly an item is believed to come from a real, independent user. Anything below threshold is excluded from headline figures and disclosed in the report.

Happens at step 3 · Authenticate

Several accounts publishing near-identical content inside a narrow window. Grouped, counted once, and reported separately from organic discussion.

Happens at step 3 · Authenticate

Sentiment assigned to each aspect within a discussion rather than to the post as a whole. One post can be positive about delivery and negative about price at once.

What the writer wanted: complaint, praise, purchase intent, comparison, question, or a churn signal. Two brands with identical NSR can differ completely here.

The habit of dropping English into Vietnamese sentences — “ship nhanh”, “sale off”, “oke em”. Code-switched sentences are segmented and labelled like any other.

The abbreviations and respellings common on Vietnamese social media — “ko”, “dc”, “vs”, “iu”. Normalised to their full forms before analysis.

A sentence that means the opposite of what it says. Machine and human readers can disagree here, so it is measured on its own rather than folded into a general error rate.

Happens at step 4 · Understand

Resolving a mention to the thing it refers to: a brand, a product, a competitor or a person. “Shop này”, “nó” and abbreviations all resolve to the right entity.

How certain the system is about a label it assigned, on a 0–100 scale. Carried by every automatic label.

The confidence level below which an item goes to a human reviewer instead of keeping its automatic label.

Happens at step 5 · Verify

A Vietnamese dataset labelled by hand, used as the standard against which system accuracy is measured. Model quality is measured on this set rather than described with adjectives.

Happens at step 5 · Verify

How often two independent human annotators give the same answer. Where people cannot agree with each other, a machine cannot be held to a higher standard.

Happens at step 5 · Verify

Every figure in a report resolves back to the exact set of discussions that produced it. Without this step a number can only be believed, never checked.

Happens at step 5 · Verify