Text Tools·4 min read

Isolating Figures From Reports Without Retyping

Quarterly reports are written to be read, not analyzed. The figures you need for a model sit inside sentences, next to percent signs and currency symbols, and retyping them into a spreadsheet is where transcription errors are born. Extract Numbers from Text exists for exactly this job: paste the prose, get every value back as a one per line column, with the count, sum, average, minimum and maximum underneath. This piece shows a worked example, explains how the pattern reads decimals, negatives and thousands separators, and flags the places where prose and data genuinely disagree.

Open the free Extract Numbers from Text All insights

Why isolating figures from reports beats retyping them

Transcription is where data dies.

Manual retyping fails quietly. A transposed digit in one cell doesn't announce itself; it just skews an average three tabs later. Extraction removes the keyboard from the loop: the values in your column are character for character the values in the source, in the same order they appeared.

Order and repetition are the underrated part. By default the tool keeps duplicates and sequence, so a paragraph that mentions 12.5 twice gives you 12.5 twice. That preserves row alignment when you're matching extracted values back against the sentences they came from, and it lets you count occurrences. Tick Remove duplicates only when you want the set of distinct values.

Speed compounds the accuracy win. An earnings summary with forty scattered figures takes twenty minutes to transcribe carefully and two seconds to extract, and the two seconds version comes with zero fatigue errors at figure thirty seven. For anyone who does this weekly, the habit change pays for itself the first time a board deck depends on the column being right.

A worked number extraction on a report sentence

One sentence, four values.

Input: Revenue hit 4.2 million, up 8 percent, while margin slipped -1.5 points from 12.5. Output, one per line: 4.2, then 8, then -1.5, then 12.5. The decimals came through intact, the minus sign stayed attached to -1.5, and the sentence's final period didn't stick to 12.5 because a decimal point only counts when digits follow it. The summary under the list reads Count 4, Sum 23.2, Average 5.8, Min -1.5 and Max 12.5.

Notice what didn't survive: million, percent, and points. The extracted 4.2 is no longer 4.2 million, and the 8 is no longer 8 percent. Every unit and scale word stays behind in the prose, which is the single most important thing to remember about this workflow.

Decimals and negatives: what the number pattern accepts

Small rules, big consequences.

The pattern reads an optional minus sign, one or more digits, and an optional decimal tail. So 980000, 3.14159, and -12 all extract exactly as written. Decimals and negatives are first class citizens, which matters for financial deltas and temperature style data where the sign is the story. With Commas are thousands separators ticked, which is the default, 1,234,567 comes out as 1234567 and Indian style 1,23,456 as 123456. Tick Whole numbers only to skip anything with a decimal point.

The minus rules are careful. A minus only counts when digits follow it immediately, and a dash straight after a letter or digit is read as a hyphen, so COVID-19 gives 19, not -19, and a range like 2023-2024 gives 2023 and 2024. A decimal point only counts between digits, so trailing punctuation never contaminates a value. But the pattern still has no idea what the numbers mean, which is where the next section's traps come from.

Extraction traps: when prose numbers lie to the pattern

Four formats that need prep.

Every one of these is predictable once you know the pattern's rules, and every one has bitten someone's dataset.

Comma lists can merge. With the thousands option on, a row like 120,450,300 reads as one number, 120450300. Choose the 1234.56 number format when commas separate values, as in CSV rows.
Dates and phone numbers split. 2024-07-02 extracts as 2024, 07 and 02, and a phone number written 98765 43210 becomes two values. Remove dates and phone numbers first if they aren't the figures you want.
Version strings fragment. 2.5.1 becomes 2.5 and 1, since only one decimal point fits per value.
Scale words vanish. 4.2 million and 4.2 extract identically. If a report mixes raw and scaled figures, normalize the text first or tag the rows after.

Numbers with units: keeping the meaning attached

The column is clean. Is it still true?

Units never come along on extraction. The practical fix is scoping: instead of pasting a whole report, extract one table or one paragraph at a time, and label each resulting column with the unit that section used. Thirty seconds of scoping beats an hour of guessing which values were percentages.

A second habit worth keeping: when a report mixes scales, do a Find and Replace pass that rewrites 4.2 million as 4200000 before extracting. The pattern then delivers comparable magnitudes, and your spreadsheet formulas don't need a translation layer. And when a document quotes the same metric in multiple places, extract each mention rather than the first one you see; disagreements between them are exactly the inconsistencies a careful reader gets paid to find.

Sorting, exporting and the tools around extraction

From prose to a usable column.

The extractor does more of the follow-up work than it used to. Set Sort to low to high and the values are ordered numerically, not alphabetically, so 9 comes before 10 and outliers sit at either end. The Separator menu joins values with new lines, commas or semicolons, and the .txt and .csv buttons save the column for a spreadsheet. For a quick sanity check, the Sum, Average and Median in the summary often tell you whether a total in the report adds up.

Find and Replace is the preprocessing arm: it rewrites scale words like million into digits and removes dates or IDs you don't want counted. When the figures live inside exported markup instead of clean prose, Strip HTML Tags comes first, so attribute values like widths and IDs don't slip into the column.