Sentiment Analysis for Excel

Learn how to perform local sentiment analysis using Lexicon, ML, and BERT models. Includes function reference for BOARDFLARE.SENTIMENT and realistic examples.

Sentiment Analysis for Excel runs text-analysis inference in the browser. The text supplied for sentiment analysis is processed locally rather than sent to Boardflare or an external AI inference API; transformer model files are downloaded to the browser cache when first needed.

The sentiment calculation path is local after required model assets are available. This keeps the analyzed text out of a cloud inference service while still allowing the add-in to use the network for application or model assets.

Quick Start

1. Open the taskpane

After installation, click Sentiment Analysis in the Home ribbon. The taskpane provides a list of available functions and experimental features.

2. Insert the function

You can insert the sentiment analysis formula using the taskpane or by typing directly into any cell: =BOARDFLARE.SENTIMENT(A2)

3. Explore the results

The function returns a Data Type. Click the icon in the cell to view the full sentiment card with all available scores and labels.

Pattern Example

Function Reference

SENTIMENT

Analyzes the sentiment of a text string or range of cells.

Syntax

=BOARDFLARE.SENTIMENT(text, [sentiment_model])

Parameter	Required	Type	Default	Description
text	Yes	Range	-	A single cell or 2D range of text to analyze.
sentiment_model	No	String	"pattern"	The engine to use for analysis.

Models

The Sentiment add-in provides a variety of engines, ranging from fast lexicon-based analyzers to state-of-the-art transformer models. Each model returns a unique set of fields within the Data Type.

Pattern (Default)

The Pattern engine is a fast, lexicon-based analyzer. It is the best choice for general-purpose analysis where processing speed is a priority.

• When to use: Analyzing large datasets (thousands of rows) where deep contextual understanding is less critical than performance.
• Fields:
  • .Polarity (Primary Value): Sentiment score from -1.0 (negative) to 1.0 (positive).
  • .Subjectivity: Measures how much of the text is based on personal opinion versus factual information.
    • 0.0 (Objective): The text is purely factual (e.g., “The phone has a 6-inch screen.”)
    • 1.0 (Subjective): The text is highly opinionated (e.g., “The phone’s screen is absolutely beautiful.”)
    • Tip: This is particularly useful for filtering datasets. For example, you can use a formula like =FILTER(A2:A100, B2:B100.Subjectivity > 0.5) to focus your analysis only on rows where the user is expressing an opinion, ignoring purely factual statements.

Pattern Example

VADER

VADER (Valence Aware Dictionary and sEntiment Reasoner) is specifically tuned for social media sentiment. It is uniquely sensitive to both polarity (positive/negative) and intensity (strength) of emotion.

• When to use: Analyzing tweets, product comments, or short reviews that frequently use slang, emojis, or heavy punctuation (e.g., “The service was GREAT!!! :)”).
• Fields:
  • .Compound (Primary Value): A normalized, weighted composite score from -1.0 to 1.0. This is the primary metric for VADER.
  • .Positive: The proportion of the text that falls into the positive category.
  • .Neutral: The proportion of the text that falls into the neutral category.
  • .Negative: The proportion of the text that falls into the negative category.

VADER Example

Naive Bayes

A machine learning model trained on a large corpus of movie reviews. It uses statistical classification to determine sentiment.

• When to use: Categorizing longer, more structured text where simple keyword matching might fail, but you don’t want the resource overhead of a full transformer model.
• Fields:
  • .Classification (Primary Value): The predicted label (pos or neg).
  • .P-Positive: The calculated probability that the text is positive.
  • .P-Negative: The calculated probability that the text is negative.

Naive Bayes Example

BERT Models (Transformers)

State-of-the-art transformer models provide the highest level of accuracy by understanding the context and relationship between words in a sentence. These models run entirely in your browser using transformers.js.

• When to use: When accuracy is paramount and your dataset is small enough to accommodate the longer processing time. These models are far superior at detecting sarcasm, double negatives, and nuanced language.
• Available Variants:
  • Movie Reviews: High-accuracy general English sentiment. Returns uppercase labels like POSITIVE or NEGATIVE.
  • Product Reviews: Optimized for product feedback. Returns a star-rating sentiment format (e.g., 1 star to 5 stars).
  • Finance News: Specifically trained for financial news and earnings reports. Returns lowercase labels like positive, negative, or neutral.
  • Twitter Messages: Specifically tuned for social media language. Returns lowercase labels like positive, negative, or neutral.
• Fields:
  • .Label (Primary Value): The predicted sentiment label (e.g., POSITIVE, NEGATIVE, neutral, or 5 stars).
  • .Score: The confidence score (probability) of the predicted label, ranging from 0.0 to 1.0.
  

Aspect-Based Sentiment (Experimental)

Unlike standard sentiment analysis which gives a single score for a sentence, ABSA identifies specific entities (aspects) and the sentiment directed at each.

• When to use: When a single sentence contains multiple conflicting opinions. For example, in the sentence “The food was great but the service was slow”, ABSA will identify both “food” (positive) and “service” (negative).
• Fields:
  • .Aspect: The specific entity identified in the text.
  • .Sentiment: The sentiment associated with that specific aspect.
  • .Index: The original row index from your input data, allowing you to map multiple aspects back to a single source row.

Working with Data Types

The SENTIMENT function returns a rich Data Type instead of a single string. This allows you to store multiple scores in a single cell and extract them using the dot operator.

Extraction Examples

If your sentiment result is in cell B2, you can use these formulas to extract specific fields: - =B2.Polarity (For Pattern model) - =B2.Compound (For VADER model) - =B2.Label (For BERT models) - =B2.Score (Confidence score for BERT models)

Experimental: Aspect-Based Sentiment (ABSA)

Traditional sentiment analysis (like Pattern or BERT) calculates a single “net” score for an entire block of text. This often leads to inaccurate or “neutral” results when a user expresses multiple conflicting opinions in the same sentence.

Aspect-Based Sentiment Analysis (ABSA) solves this by identifying specific entities (aspects) and assigning an individual sentiment to each. This capability is based on the Instruct-ABSA model.

Why use ABSA?

Consider the review: “The food was great, but the service was lousy.” - Traditional Models: May return a “neutral” or “conflicted” result (e.g., 3 stars or a low positive score) because the positive and negative sentiments cancel each other out. - ABSA: Correctly disaggregates the feedback into two distinct results: food (positive) and service (negative).

ABSA Example

Key Differences

• Granularity: ABSA is the only model capable of pinpointing exactly what the user liked or disliked within a complex sentence.
• Taskpane-Only: Unlike the SENTIMENT function, ABSA is currently a taskpane-only feature. It processes your selected text and inserts the identified aspects as static values in your worksheet.
• Domain Specific: The current experimental model is optimized specifically for restaurant reviews. Non-restaurant related text (e.g., software feedback, financial news) will likely not extract aspects or assign sentiment correctly. If you would like to see other domains supported for ABSA, please contact us.

[!WARNING] ABSA is resource-intensive, requiring approximately 1GB of browser cache and 3GB of RAM. It is recommended for targeted analysis of key feedback rather than large-scale batch processing.

Examples & Scenarios

1. Customer Feedback Analysis (Pattern)

Quickly gauge the general sentiment of a large list of customer comments.

Cell	Comment	Polarity	Subjectivity
A2	I absolutely love the new design!	0.625	0.600
A3	The update is okay, but it’s a bit slow.	0.125	0.450
A4	This is the worst experience I’ve had.	-1.000	1.000

Formula

=BOARDFLARE.SENTIMENT(A2, "pattern")

2. Financial Sentiment (FinBERT)

Analyze earnings call snippets or financial news for market sentiment.

Cell	News Snippet	Label	Score
A2	Profits exceeded expectations by 20%.	positive	0.985
A3	Regulatory hurdles may delay the merger.	negative	0.892
A4	The company maintained its annual guidance.	neutral	0.998

Formula

=BOARDFLARE.SENTIMENT(A2, "bert_finance")

3. Social Media Monitoring (VADER)

Analyze tweets or short posts that include emojis and informal language.

Cell	Tweet	Compound Score
A2	The new movie was AMAZING!!! 🍿🔥	0.824
A3	I’m so disappointed with the service today 😠	-0.584
A4	Just another day at the office.	0.000

Formula

=BOARDFLARE.SENTIMENT(A2, "vader")

Tips & Best Practices

Privacy First

The sentiment models run locally in your browser, so the text supplied for inference is not sent to Boardflare or a third-party AI inference provider. Initial transformer-model downloads still use the network. Whether a dataset containing sensitive internal information or PII is appropriate for the add-in also depends on your organization’s browser, device, and data-handling policies.

Performance vs. Accuracy

• Use Pattern or VADER for large datasets where speed is critical.
• Use BERT models for high-accuracy requirements or specific domains (Finance, Twitter).
• Transformer models (BERT) are larger and will take longer to initialize on the first run as they download to your browser’s local cache.

Cleaning Data

For the best results, ensure your text is clean. Remove HTML tags or excessive boilerplate text before analysis.

Troubleshooting

The formula returns #VALUE! or #NAME?

Ensure the Sentiment taskpane is open. Try refreshing the page.

The first calculation is very slow

When using a BERT model for the first time, the add-in must download the model files (~100 MB). This only happens once; subsequent uses will load the model from your local cache.

Memory Issues

If you are using the BERT or ABSA models and your browser becomes unresponsive, you may be low on RAM. Try closing other tabs or switching to the lighter Lexicon models (Pattern, VADER).