Verbum

62,209 verses, hand-labelled

Most Bible-analytics projects that publish "emotional landscape" charts use TextBlob or VADER on the English text. That pipeline is fast, free, and reproducible. It is also wrong in roughly the way a sentiment classifier trained on movie reviews would be wrong on Hebrew prophecy.

This post is about why Verbum did the labour anyway, and what we found when we did. The dataset described here ships with the app and is queried live on the emotional landscape page.

The problem with off-the-shelf sentiment

TextBlob's polarity score for Lamentations 3:1, in the BBE translation, comes out as +0.16. The verse:

"I am the man who has seen trouble by the rod of his wrath."

A model trained on Yelp reviews sees "I am the man who has seen" — neutral — followed by "trouble" — slightly negative — and "rod of his wrath" — which contains "rod", which on retail data correlates with positive sentiment (fishing rod, lightning rod, "spare the rod", "rod of God"). The negative "wrath" gets averaged with positives. The result is mildly positive.

The verse is not mildly positive. It is the opening sentence of one of the most pure expressions of grief in the Hebrew Bible.

This pattern repeats through every poetic and prophetic book. Job, Ecclesiastes, Lamentations, Jeremiah, Ezekiel, the imprecatory Psalms, the woes of Matthew 23, the apocalyptic chapters of Revelation — none of them survive lexical sentiment analysis. The vocabulary doesn't match. The genre doesn't match. The biblical register of "grief" uses words like abandon, forsake, cup, cut off, hand of the LORD, none of which carry the right polarity in a model that learned English from product reviews.

The choice

Three options:

  1. Ship TextBlob with a disclaimer.
  2. Train a custom model on biblical text. Requires labelled data, which we didn't have.
  3. Hand-label every verse, in two languages where the app's primary audience reads.

We picked option 3. Specifically: 31,107 verses in Portuguese (Nova Versão Internacional as the canonical text) and 31,102 verses in Spanish (Reina-Valera 1960). 62,209 total annotations. The labels are in verses_sentiment_multilang in the DuckDB ship; the API serves them through LEFT JOIN ... COALESCE so the TextBlob fallback covers anything missing.

The rubric

A polarity float in [-1, +1] plus a discrete label drawn from:

  • joy, peace, gratitude, hope, love — strongly positive
  • praise, awe, confidence — moderately positive
  • instruction, wisdom, narrative — neutral
  • lament, confession, warning — moderately negative
  • grief, judgment, wrath, despair, imprecation — strongly negative

Anchors fixed at the start of labelling:

  • John 3:16 → +0.85, hope
  • Psalm 23:1 → +0.80, peace
  • Lamentations 3:1 → −0.85, grief
  • Romans 3:23 → −0.20, confession
  • Genesis 1:1 → +0.10, narrative
  • Matthew 5:4 → +0.30, hope (the comfort, not the mourning, is the verb)

Difficult cases were resolved by re-anchoring to those reference verses. Genre overrides individual word choice — a wisdom proverb that contains the word "death" stays neutral if its rhetorical purpose is instruction.

What 62,209 labels looks like

Aggregated by book:

  • Highest mean polarity in Portuguese: 1 John (+0.42), Ephesians (+0.37), Philippians (+0.35).
  • Lowest mean polarity in Portuguese: Lamentations (−0.61), Habakkuk (−0.47), Joel (−0.39).
  • Most volatile (largest standard deviation): Job (0.51 σ), Psalms (0.48 σ), Jeremiah (0.46 σ).
  • Most uniform: Philemon (0.07 σ), Jude (0.09 σ), Ephesians (0.13 σ).

Hebrews — a book often imagined as gloomy because of its themes of perseverance under persecution — runs +0.18 mean polarity in PT. The encouragement machinery dominates the warnings. The dataset surfaces that immediately.

What we got wrong

A second-pass audit on Lamentations and Job found ~40 verses where the first pass had been mechanical (matching narrative phrasing rather than reading the parallelism in poetry). They were re-labelled. Similar audits are pending on Jeremiah, Ezekiel, and the longer Pauline epistles in Spanish.

What we couldn't do

English. The English BBE and KJV verses still use TextBlob, which means the emotional landscape view defaults to a fallback when the active translation is English. This is documented in the UI as a small ⓘ disclaimer; it is not hidden.

That asymmetry will not be fixed by re-running TextBlob with better tuning. It will be fixed by labelling another 31,000 verses, which is a finite amount of work, and which is on the roadmap.

What this is good for

Pastoral planning ("which Psalms set the tone for grief care?"). Lectionary work. Comparative studies between books. Teaching: the texture of Scripture becomes visible when you can plot it.

It is not good for proving doctrinal claims. A book's mean polarity doesn't tell you whether the book is "happy" or "sad" — Lamentations is −0.61 but is also one of the most spiritually generative books of the Hebrew Bible. Sentiment is a lens. It is not a verdict.

Open the emotional landscape and see for yourself.

The labelling work was carried out by Claude (Anthropic) under a strict rubric, anchored to the verses listed above, and validated by spot-checks across genres. The full dataset, the loader script, and the JSONL audit trail live in the public repo at github.com/DavidKGBR/verbum.