Check that quoted evidence really is in the document

ans$evidence says what an answer rests on. For most readers those spans are verbatim chunk text and are true by construction. For skim they are what the model chose to write when asked to extract the relevant passages. They are presented as quotations, and this is what checks that they are.

Usage

gr_verify_evidence(answer, chunks = NULL)

Arguments

answer

A gr_answer.

chunks

The gr_chunks the answer was read from. Needed for readers whose evidence is verbatim, where the comparison is against the chunk the span claims to come from: its source_text where the segmenter recorded one (the document text behind a chunk whose text carries model-written context or propositions), and its text otherwise. skim answers already carry their sources, so they can be checked without it; an ensemble needs it for the rows its verbatim members contributed, even though its skim rows do not.

Pass the chunks the answer was actually read from. Chunk ids are positional, so a different chunk set of the same size will match on id and compare each span against unrelated text, reporting verified = FALSE for evidence that is perfectly sound. An id the chunk set does not contain reports NA, because there was nothing to compare against.

Value

A data frame with one row per evidence span: chunk_id, kind ("verbatim", "extracted" or "answer", per row; an ensemble mixes them in one table, and the same value appears as the kind column on ans$evidence itself), verified, match and span (the first 60 characters). verified is NA where the question does not apply: a map_reduce evidence row is a per-chunk answer, not a quotation, and asking whether it appears in the chunk is a category error.

For an extract answer, a quote has to carry the value it is cited for as well as appear in the chunk, and only the reader can check the first: it knows the value. A row the reader marked verified = FALSE for that reason stays FALSE here, with a match of 1 when the sentence itself is in the document. It is there, and it does not say this.

The quote_* columns come from a second matcher, recorded beside verified and match and deciding nothing yet; see "The edit-distance score" below.

Details

A fabricated citation is more convincing than a fabricated answer, because it looks like the thing that would let you check. Verification is a string operation on text you already have, so it costs nothing and there is no reason not to do it.

What the numbers mean

match is 1 for an exact quotation once whitespace, quote marks, dashes and case are folded away. These are the differences a faithful quotation introduces. Guillemets and the corner brackets of Chinese and Japanese count as quote marks, a full-width letter, digit or number sign as its ASCII form, the ideographic space as a space, and a letter written with combining accents (as macOS tools write text) as the accented letter, for Latin, Greek, Cyrillic and Korean letters. harnessr does no other Unicode normalisation, so a quotation that differs from its document only in how some other letter is encoded is reported unverified, never verified by mistake. It has to match whole words and whole numbers: "5%" is not found in "25%", nor "12%" in "-12%", nor "20" in "200", nor "200" in "1 200" written with a thin space, nor ".45" in "1.45", nor "65 years" in a range broken across two lines ("aged 18-" at the end of one). A hyphen between two letters that the quotation writes and the document does not, as where the cleaner joined a word broken at a line end ("investigator-blinded"), is not a difference in a quotation of one passage. Chinese, Japanese and Thai put no spaces between words, so a clause quoted from them may start and end anywhere; numbers in them are still whole, and a circled list number before one is not part of it. That includes ending inside a word: a quotation that stops at a verb's stem verifies where the document goes on to a negating ending written on to it, as an English quotation that stops before "not" does, with or without a full stop the model added. Markdown bold is not text, in the quotation or the document. A span made of several passages (in paragraphs, as a list, in separate quote marks, or joined by "..." or "[...]") is checked passage by passage and is verified when every passage is found, the parts either side of an elision in that order. What an elision leaves out may not be a negation ("the drug did ... reduce mortality" is not in "the drug did not reduce mortality"). That is checked in the major European languages, Russian, Greek, Turkish, Arabic, Hebrew, Hindi and Bengali by their negation words, and in Chinese, Japanese, Korean and Thai by the characters and endings that negate; what an elision leaves out in a script none of those cover is refused. Where an elision leaves out the end of a sentence the part after it has to start one ("Revenue ... rose 30%" is not in "Revenue fell 12%. Costs rose 30%."); a full stop inside a bracketed citation does not end one. A line break inside a sentence is a line wrap, not a new passage: the lines are checked as one, even where the next line opens with a capital, unless the document ends a sentence after the first line or starts one with the second. Passages the quotation separates (by a blank line, as a list, in quote marks) may come from anywhere in the document, but not from one sentence with words left out between them, which is an elision the quotation did not mark. None of this reads meaning: a quotation that stops before a negation later in its sentence still verifies, as it always has. Below 1 it is the fraction of the span's words carried by its longest consecutive run in the source, for the passage that matches worst; in a script written without spaces, the fraction of its characters.

Read that number with its shape in mind. Because it measures a run, where the change falls matters as much as how much changed: altering the last word of a ten-word span leaves a run of nine and scores 0.9, while altering a word in the middle splits the span and scores about 0.5. So a mid-sentence change (a swapped figure, the case this exists to catch) lands near 0.5, not near 0.9. Below roughly 0.3 there is no quotation left at all, only shared vocabulary. A run measure is still the right one: word overlap cannot tell a quotation from a paraphrase assembled out of the same words.

The edit-distance score

The run measure above cannot be given a threshold, because the same edit scores differently by where it falls. A second matcher is recorded next to it for every span, so the two can be compared on real runs before the default changes. It normalises both sides further than the check above: ligatures and the long s expanded, JSON escapes undone, soft hyphens and zero-width characters dropped, every full-width ASCII form folded, and a word broken across a line with a hyphen joined, as is any hyphen between two letters. It then finds the window of the source that the fewest character edits turn the quotation into.

quote_sim is 1 minus those edits over the longer of the quotation and the window, so it falls by the same step for each edit wherever it falls. quote_coverage is the share of the quotation's words found in order in the window, quote_edits the count, and quote_critical_ok whether its numbers (with sign, a comparison sign before them and unit, such as "-12%", "< 0.05" or "45.2 million"), numbers written in words ("twelve") and negations are the window's, exactly. quote_status is "verbatim" for no edit; "near_verbatim" for a similarity of at least 0.90 (0.95 under 40 characters) with the critical tokens identical and five words or more; "unverified" otherwise; and "not_checked" where there was nothing to compare, or where a passage was too long to score: the edit table has a cell for each pair of characters, and a passage whose length times the source's is over 5e7 is skipped. A quotation split by an elision or a blank line takes the worst of its passages. near_verbatim is not verified: a changed word that is not a number or a negation ("Sheffield" for "Leipzig") can stay above the threshold.

Citations

With cite = TRUE a reader asks the model to mark its sources as [chunk 3]. Every answer is checked for citations pointing at chunks that were never sent, whatever this function is called with; the result is ans$notes$cited_unknown, and an answer carrying one is partial. Lists and ranges are read ([chunks 1, 2, and 9], [chunks 1-9]), and a bracket that opens like a citation but cannot be read, such as [chunk nine], is listed in ans$notes$cited_unparsed and makes the answer partial too.

Examples

# A model that quotes faithfully.
honest <- gr_mock_client(function(messages, params) {
  txt <- messages[[length(messages)]]$content
  if (grepl("You extract evidence", messages[[1]]$content, fixed = TRUE)) {
    return("Revenue rose to 45.2 million dollars.")
  }
  "Revenue was 45.2 million dollars."
})

doc <- "Revenue rose to 45.2 million dollars.\n\nHeadcount grew to 1,204."
ch <- gr_segment(gr_ingest(doc), list(method = "paragraph", max_tokens = 40))
#> Ingested 2 block(s), ~26 tokens (0 chars removed by cleaning).
#> Segmenting with 'paragraph' (cap 40 tokens, overlap 0).
ans <- gr_read(ch, "What was revenue?", honest, "skim")
#> Reading with 'skim' (all|N+1|none) over 1 chunk(s).
gr_verify_evidence(ans)
#>   chunk_id      kind verified match quote_status quote_sim quote_coverage
#> 1        1 extracted     TRUE     1     verbatim         1              1
#>   quote_edits quote_critical_ok                                  span
#> 1           0              TRUE Revenue rose to 45.2 million dollars.

# A model that invents one. The span is fluent, plausible, and not in the
# document. That is exactly the case a reader cannot catch by eye.
liar <- gr_mock_client(function(messages, params) {
  if (grepl("You extract evidence", messages[[1]]$content, fixed = TRUE)) {
    return("Revenue rose to 88.9 billion dollars on record demand.")
  }
  "Revenue was 88.9 billion dollars."
})
bad <- gr_read(ch, "What was revenue?", liar, "skim")
#> Reading with 'skim' (all|N+1|none) over 1 chunk(s).
gr_verify_evidence(bad)
#>   chunk_id      kind verified match quote_status quote_sim quote_coverage
#> 1        1 extracted    FALSE 0.333   unverified    0.6792         0.4444
#>   quote_edits quote_critical_ok
#> 1          17             FALSE
#>                                                     span
#> 1 Revenue rose to 88.9 billion dollars on record demand.
bad$partial
#> [1] TRUE