NLG Reports in Hengshi BI: Turning Data Insights into Business Narratives

Article body

Full article

12 min read

Introduction

BI systems excel at turning data into bar charts, line charts, and heat maps. Decision-makers need concise written conclusions, such as: Q2 sales increased 15% year over year, with a new product line in East China as the main driver, while a two-percentage-point decline in gross margin needs attention.

Analysts have traditionally reviewed each chart and written these reports by hand. The process takes time, writing styles vary across analysts, and weekly or monthly reports can arrive late.

Natural Language Generation (NLG) enables a BI system to turn data insights into business-readable reports. Hengshi BI integrates NLG with metric management, anomaly detection, and Agentic BI to create an automated pipeline from data to narrative.

1. The Role of NLG in BI

1.1 The Bottleneck in Manual Reporting

A monthly operating report requires four steps:

An analyst exports charts from the BI system.
The analyst reviews each chart for metric changes, anomalies, and trends.
The analyst writes the findings and organizes the report.
The analyst formats, reviews, and submits the document.

This work can take one or two days and depends on the analyst’s availability and experience. The report is delayed when the analyst cannot complete it on time.

1.2 Three Problems NLG Addresses

Efficiency: Automated reporting reduces one or two days of work to minutes. The system can regenerate the report whenever the data changes.

Consistency: A shared set of generation rules keeps the style, structure, and metric definitions aligned across reporting periods.

Scale: An NLG system can produce reports for many business lines at the same time, reducing repetitive manual work.

1.3 How NLG Differs from ChatBI

ChatBI answers a user’s question during an interactive, one-off analysis.
NLG reporting produces structured documents for scheduled delivery and archiving.

Both can use large language models, but they support different workflows.

2. Hengshi BI’s NLG Architecture

2.1 A Four-Stage Generation Pipeline

Hengshi BI uses a structured pipeline instead of handing charts to a large language model and requesting open-ended prose.

Stage 1: Insight Extraction

Rules and statistical methods identify the findings worth including in a report:

Change detection compares each metric with the previous period or the same period last year.
Anomaly marking flags metrics outside their expected range.
Ranking changes identify substantial movements, such as a product line rising from fifth to second place.
Contribution analysis allocates an aggregate change across dimensions, such as East China contributing RMB 600,000 of a RMB 1 million increase.

This stage produces a structured list of insight objects. Each object records the metric, direction and size of change, and attribution result.

Stage 2: Narrative Planning

The system turns the insight objects into a report structure:

Prioritization places anomalies and large changes before small movements.
Logical grouping organizes findings into business areas such as sales and finance.
Length allocation gives more space to material findings and summarizes minor ones.

The output is a narrative outline.

Stage 3: Text Generation

The system generates each section from the outline. The prompt receives structured insight objects instead of a broad instruction to write a report.

Each insight category has a corresponding template strategy:

Change: Metric X reached Y, changing Z% from the previous period, with factor A as the main driver.
Anomaly: Metric X deviated from its baseline by Nσ; investigate cause B.
Ranking: Entity X rose K places within dimension D and now ranks M.

Stage 4: Review and Polish

The system checks the generated text before delivery:

Numerical consistency verifies that every number matches the data.
Terminology consistency keeps each metric’s name uniform throughout the report.
Tone adaptation keeps executive summaries concise while preserving detail in operating reports.

2.2 Combining Templates with Generated Text

Unconstrained generation can change the report structure, omit key facts, or introduce incorrect numbers. Hengshi combines templates with language-model generation.

Templates handle structured content: Fixed templates produce report introductions, standard metric-change language, and closing notices. This keeps them consistent and accurate.

The language model handles narrative content: The model writes explanations, qualitative trend descriptions, and recommendations, improving readability and flexibility.

Constraints limit the model’s source material: The model may use only the insight objects produced by the extraction stage. It cannot introduce numbers or conclusions outside that set.

2.3 Adding Domain Knowledge with RAG

Reports must reflect an enterprise’s business context. Hengshi uses Retrieval-Augmented Generation (RAG) to provide that context:

Industry knowledge supplies patterns such as retail seasonality.
Historical practices supply investigation paths for changes such as falling gross margin.
Metric lineage supplies relationships among gross margin, cost, and discount rates.

RAG helps the system interpret a change in business terms instead of restating the number.

3. Three Forms of NLG Reporting

3.1 Scheduled Reports

The system can generate and distribute daily, weekly, or monthly reports on a schedule:

A schedule triggers the workflow, such as 8:00 a.m. on the first day of each month.
The system loads the previous period’s data and extracts insights.
A template guides report generation.
Email, enterprise messaging, or BI notifications deliver the report to subscribers.

The report updates with the data. A retailer, for example, can generate daily reports for 200 stores at once, covering sales, traffic, average transaction value, period-over-period changes, and anomaly notices.

3.2 Anomaly-Driven Reports

An anomaly can trigger a root-cause report:

Anomaly detection identifies an unusual metric.
A Data Agent analyzes possible causes.
NLG organizes the analysis into a structured report.
The system sends the report to the responsible team.

These reports focus on the anomaly, its likely causes, and recommended action.

3.3 Interactive Reports

Readers can ask follow-up questions within a report. If a monthly report says that East China sales increased 15%, a reader can ask which products drove the increase.

The system interprets the question, queries the relevant data, and inserts a supplementary passage. Readers can continue their analysis from the report instead of starting a separate session.

4. Quality Controls for NLG

4.1 Numerical Accuracy

The primary risk in an NLG report is a mismatch between the narrative and the underlying data. Hengshi uses three controls.

Deterministic value insertion: The system extracts and inserts every number. The language model organizes the prose but does not calculate values. For example, the system fills the actual period-over-period change into a {change} placeholder.

Post-generation validation: The system extracts every number from the draft and compares it with the source data. Any mismatch blocks publication and raises an alert.

Confidence labels: A qualitative statement without sufficient data support is marked as an inference.

4.2 Consistent Style

Different readers need different versions of a report:

Executive reports summarize the main conclusions in a few paragraphs and include key charts.
Operations reports describe metric changes by business area and provide drill-down paths.
Technical reports retain metadata such as data sources, calculation definitions, and update times.

Hengshi combines a report template with the reader’s role. The same data can produce different versions for different audiences.

4.3 Explainability

Readers can inspect the data and calculation behind a conclusion. Selecting “East China contributed 60% of total growth,” for example, can reveal RMB 1 million in total growth and RMB 600,000 in East China growth.

5. Working with Agentic BI

5.1 From Reporting to Action

Combining NLG with Agentic BI connects report findings to follow-up work:

An Agent identifies action items, such as a decline in gross margin that needs investigation.
The Agent creates a task, notifies an owner, or tracks the item in the next report.
The system links detection, reporting, action, and follow-up.

5.2 Multi-Agent Report Production

Several Agents can divide the work for a complex report:

A Data Agent extracts insight objects.
An Analysis Agent investigates anomalies.
A Writing Agent turns the insights and analysis into prose.
A Review Agent checks numbers and metric terminology.

This division narrows each Agent’s responsibility and makes the process easier to control and audit.

6. Common Problems and Remedies

Problem 1: Free-Form Generation Produces Incorrect Numbers

Handing charts and data to a language model can produce numbers that do not exist in the source or turn 15% into 51%.

Remedy: Use deterministic value insertion. The model writes the narrative while the system supplies every number.

Problem 2: A Fixed Structure Hides Important Findings

Listing metrics in the same order each time can bury an important anomaly under routine observations.

Remedy: Rank findings by business importance. Put anomalies and material changes first, then shorten or omit routine movements.

Problem 3: Technical Terms Reduce Readability

Some readers may struggle with a report full of terms such as period-over-period change and contribution analysis.

Remedy: Adapt the language to the reader. Use business language in the executive version and move technical detail to an appendix.

7. Summary

NLG helps BI turn charts into business reports that readers can scan, archive, and question. Hengshi BI controls the quality of automated writing through three design choices:

A four-stage pipeline for insight extraction, narrative planning, text generation, and review.
Templates combined with generation, with the system inserting critical numbers and the language model organizing the prose.
Multiple report formats for scheduled, anomaly-driven, and interactive reporting.

The BI system can therefore support both data presentation and business communication.