Word resume template — name, phone, and skills mis-parsed (completeness 12/30)

[sriyau64]

Word resume template — name, phone, and skills mis-parsed (completeness 12/30)

Using a resume sample from a Microsoft Word "Office Manager" template, the phone number was visible in the extracted plain-text panel but did not appear in the score card. Deeper investigation shows the phone is one of three parsing failures on this document; overall completeness is 12/30.

Chanchal sharma sample.pdf (synthetic persona — PII-verified safe: body text, Info dict, XMP metadata, and fonts/ICC layers all checked clean)

Screenshot 1 — phone present in plain-text panel
Screenshot 2 — phone absent from score card

Summary

Running the full cascade on the attached PDF yields overall 12, completeness 12/30, with name, phone, and skills all reported as missing. The text layer is clean (ASCII + one en-dash; verified at codepoint level — no ToUnicode/CMap mojibake). The failures are entirely in the Tier-1 heuristic parser, which makes three layout assumptions that Word table/text-box templates break simultaneously.

Reproduction

1. Open the attached PDF in ResumeLint.
2. Observe completeness = 12/30.
3. Score card shows name, phone, and skills missing (among others — see "Out of scope" below for the non-bugs).

Expected: name, phone, and skills all parsed and scored.

Root cause

Word's resume templates lay out the header name and skills section in tables/text-boxes. This produces three independent assumption-breaks in the single-pass heuristic parser:

1. Name (src/lib/heuristics/extract-fields.ts:173) — The real name renders as two separate single-word lines (Chanchal / Sharma). extractName rejects any candidate with words.length < 2, so both name lines are individually disqualified. The two-word tagline "Office Manager" wins the slot at confidence 0.25 (below the anonymous contact floor of 0.5), and name is reported missing.

2. Phone (src/lib/heuristics/phone.ts:187-200) — The phone number is (718) 555–0100 where the separator is an en-dash U+2013, not an ASCII hyphen. mightHavePhone() pre-gates on PHONE_RE whose separator class [\s.-]? excludes Unicode dashes, so PHONE_RE.test() returns false and findPhoneNumbersInText (PR feat(phone): locale-aware region derivation + fix two-column fixture phone #100, libphonenumber-js) is never invoked. findPhoneNumbersInText handles en-dashes natively — the fix is to make the pre-filter gate wide enough to let the call reach it, not to widen PHONE_RE itself.

3. Skills (src/lib/heuristics/extract-fields.ts:469) — The skills are in a 3-column table. SKILL_SPLIT_RE = /[,;·•|/]+|\s{2,}/ splits correctly when pdfjs assembles a column row with 2+ spaces between entries. When a row is assembled with only a single space between columns, the entire row collapses to one token ("Organization Problem-solving Management"), which counts as 1 skill — below the minimum of 5 (extract-fields.ts:512) — so skills is reported missing.

Note: the text layer is not the problem. Fonts are subsetted Georgia + Calibri with a valid ToUnicode map; the raw text the cascade sees is byte-for-byte the visible text. ToUnicode/CMap mojibake is not the cause here.

Scope — what to fix

• Name: reconstruct two-word name from adjacent single-word lines. When consecutive short lines at the top of the document together form a plausible full name (two words, title-cased), treat them as a single candidate before applying the words.length < 2 guard. (src/lib/heuristics/extract-fields.ts:165-205)

• Phone: expand mightHavePhone's pre-filter to recognize Unicode dashes. In phone.ts:187-200, widen the pre-gate so that en-dash (U+2013), em-dash (U+2014), and figure dash (U+2012) allow the string through to findPhoneNumbersInText. PR feat(phone): locale-aware region derivation + fix two-column fixture phone #100 already wires findPhoneNumbersInText (which handles these chars natively) — the only missing piece is that mightHavePhone gates it out before it runs. Do not widen PHONE_RE itself — the fix is the pre-filter gate, not the regex matcher.

• Skills: context-detect a single-space multi-token column run and split accordingly. When inside a recognized skills section, detect a run of tokens that are separated by single spaces and together form a plausible skills list, then split it. Do not widen SKILL_SPLIT_RE to \s+ unconditionally — that causes false-positive splits in ordinary prose. (src/lib/heuristics/extract-fields.ts:469)

Explicitly out of scope

• Role dates showing as missing — The template uses placeholder text January 20XX / Sep 20XX. YEAR_RE = /\b(19|20)\d{2}\b/ correctly rejects 20XX. This is a fixture artifact (template placeholder), not a product bug. Do not modify YEAR_RE to accept 20XX as a valid year.

• Summary missing — The template genuinely has no Summary, Objective, or Profile section. Not a detection failure.

• LinkedIn missing — No LinkedIn URL appears in the document. Not a detection failure.

Acceptance criteria

• parsed.name is the real name (not "Office Manager") for the Chanchal Word PDF.
• parsed.phone is (718) 555-0100 (or equivalent normalized form) for the Chanchal Word PDF.
• parsed.skills contains at least the skills visible in the template columns for the Chanchal Word PDF.
• Completeness score rises above the 12/30 baseline for this document (role-dates, summary, and LinkedIn remain correctly absent — no strict numeric floor beyond the current baseline).
• The committed Word fixture's expected snapshot is updated to reflect the corrected parsed output.

Test fixture

Commit the attached PDF to tests/fixtures/pdfs/word/chanchal-sharma-sample.pdf. The file has been verified PII-safe across all four layers (body text via pdftotext, Info dict via pdfinfo, XMP metadata, and fonts/ICC). After implementing fixes:

npm run bake-fixtures

Review and commit the updated expected snapshot. This fixture regression-guards all three fixes simultaneously.

[s-annam]

Clarification (rewrite) — Body rewritten from phone-only scope to the validated Word-template multi-field root cause: name mis-picked due to words.length < 2 guard rejecting split single-word lines (extract-fields.ts:172), phone blocked by en-dash U+2013 not in PHONE_RE separator class (regex.ts:20), and skills column collapsing to one token under single-space join (extract-fields.ts:460); codebase reproduction confirmed completeness 12/30 exactly; role-dates (20XX placeholder), summary, and LinkedIn ruled out as non-bugs.

[s-annam]

Clarification (rewrite) — Issue remains valid; all 3 bugs reproduce. Phone-fix scope updated for PR #100: the fix is expanding mightHavePhone's pre-filter gate to pass Unicode dashes through to findPhoneNumbersInText (which handles them natively), replacing the old "normalize at text-assembly" guidance that predated PR #100. Skills approach narrowed to context-detect a single-space column run in the skills section rather than widening SKILL_SPLIT_RE unconditionally. AC floor set to "rises above 12/30 baseline." Fixture marked PII-verified-safe (all four layers checked).