Skip to main content

Writing a shorter account

Generate a summary, then check what it kept and changed

Learn how abstractive summarization differs from sentence selection and why shorter text still needs source review.

Rosa prepares the morning brief for a neighborhood archive. Two service notes have arrived, and each is longer than the one sentence her colleagues need. The shorter sentence still has to respect the source.

Abstractive summarization generates wording for a shorter account of a source. The output may reuse some source phrases, but it is not limited to selecting whole sentences. Lesson 54 used sentence selection. This lesson asks a local language model to write a new sentence, then checks the candidate against the original record.

TipWhat you will learn

This lesson shows how to:

• keep source records, prompts, and outputs tied to durable IDs;
• count the model’s input tokens before generation;
• compare generated summaries with a first-sentence baseline;
• separate compression from source support and useful coverage;
• inspect names, numbers, and explicit limits without calling them proof; and
• leave factual and editorial approval to a recorded human review.

Load the local model

The model files were downloaded before rendering at the revision recorded in data/nlg-models.csv. The code below uses huggingfaceR to load that local snapshot. Offline mode is set by R/use-nlg.R, so a missing model stops the lesson instead of starting a network download.

suppressPackageStartupMessages({
  library(dplyr)
  library(huggingfaceR)
  library(purrr)
  library(reticulate)
  library(stringr)
  library(tibble)
  library(tidyr)
})

source("R/use-nlg.R")

summary_model <- load_nlg_pipeline(
  "qwen_1_5b_instruct",
  "text-generation"
)

model_record <- summary_model$metadata |>
  select(model_id, revision, license, task_scope)

Pinned local model used for the generated summaries

Model	Revision	License	Use in this course
Qwen/Qwen2.5-1.5B-Instruct	989aa7980e4cf806f80c7fef2b1adb7bc71aa306	Apache-2.0	abstractive summarization;paraphrasing;long-text generation;question answering;chatbot dialogue

The model card documents an instruction-tuned causal language model. That description and license do not establish that its summaries are accurate.

Keep the source beside the brief

These records are fictional teaching examples. They were written for this lesson and are not a random sample, a held-out benchmark, or evidence about archive services elsewhere.

Each brief states information that matters to the reader. SUM-01 needs the service, Tuesday schedule, and boundary on volunteer help. SUM-02 needs the event, photo-identification rule, and restricted-storage limit.

summary_sources <- tribble(
  ~source_id, ~brief, ~source_text,
  "SUM-01",
  paste(
    "One sentence for staff; retain the job-search service, Tuesday help",
    "schedule, and boundary on volunteer help."
  ),
  paste(
    "The Riverton library opens a new job-search room on Monday.",
    "The room has 12 computers.",
    "Volunteers offer free help on Tuesdays, but they do not complete",
    "applications for visitors.",
    "The pilot runs for six weeks."
  ),
  "SUM-02",
  paste(
    "One sentence for visitors; retain the catalog workshop,",
    "photo-identification rule, and restricted-storage limit."
  ),
  paste(
    "The neighborhood archive will open Room 5 on 14 September for a",
    "catalog workshop led by Maya Chen.",
    "Registration is free, but visitors must bring photo identification.",
    "The session does not include access to restricted storage."
  )
)

Constructed source records and their summary briefs

Source ID	Brief	Source text
SUM-01	One sentence for staff; retain the job-search service, Tuesday help schedule, and boundary on volunteer help.	The Riverton library opens a new job-search room on Monday. The room has 12 computers. Volunteers offer free help on Tuesdays, but they do not complete applications for visitors. The pilot runs for six weeks.
SUM-02	One sentence for visitors; retain the catalog workshop, photo-identification rule, and restricted-storage limit.	The neighborhood archive will open Room 5 on 14 September for a catalog workshop led by Maya Chen. Registration is free, but visitors must bring photo identification. The session does not include access to restricted storage.

Declare required and optional facts

The checklist is fixed before generation. A missing required item sends a candidate back for revision. Optional clues remain useful during source review, but their omission alone does not violate the brief.

summary_checks <- tribble(
  ~source_id, ~check_id, ~needed_information, ~required, ~pattern,
  "SUM-01", "F1", "job-search service", TRUE,
  "\\bjob[ -]search\\b",
  "SUM-01", "F2", "Tuesday help", TRUE,
  "\\bTuesday\\b",
  "SUM-01", "F3", "volunteers do not complete applications", TRUE,
  "\\b(not|no|never|without|do not|does not)\\b",
  "SUM-01", "F4", "12 computers", FALSE,
  "\\b12\\b",
  "SUM-01", "F5", "six-week pilot", FALSE,
  "\\b(six|6)[ -]week",
  "SUM-02", "F6", "catalog workshop", TRUE,
  "\\bworkshop\\b",
  "SUM-02", "F7", "photo identification", TRUE,
  "\\b(photo|identification|ID)\\b",
  "SUM-02", "F8", "no restricted-storage access", TRUE,
  "\\b(not|no|without|does not)\\b",
  "SUM-02", "F9", "Room 5", FALSE,
  "\\bRoom 5\\b",
  "SUM-02", "F10", "14 September", FALSE,
  "\\b14\\b.*\\bSeptember\\b",
  "SUM-02", "F11", "Maya Chen", FALSE,
  "\\bMaya Chen\\b",
  "SUM-02", "F12", "free registration", FALSE,
  "\\bfree\\b"
)

Required facts and optional review clues declared before generation

Source ID	Check	Predeclared information	Required by brief
SUM-01	F1	job-search service	TRUE
SUM-01	F2	Tuesday help	TRUE
SUM-01	F3	volunteers do not complete applications	TRUE
SUM-01	F4	12 computers	FALSE
SUM-01	F5	six-week pilot	FALSE
SUM-02	F6	catalog workshop	TRUE
SUM-02	F7	photo identification	TRUE
SUM-02	F8	no restricted-storage access	TRUE
SUM-02	F9	Room 5	FALSE
SUM-02	F10	14 September	FALSE
SUM-02	F11	Maya Chen	FALSE
SUM-02	F12	free registration	FALSE

Check the input budget before generation

A model token is a tokenizer unit, not necessarily a word. The prompt includes the instruction, brief, and source text, so all three count against the input budget. This demonstration uses a 256-token ceiling and refuses to truncate.

summary_system <- paste(
  "Write one sentence for the stated brief.",
  "Use only information in the source.",
  "Do not add a cause, recommendation, or new event."
)

summary_inputs <- summary_sources |>
  mutate(
    user_prompt = paste0(
      "BRIEF: ", brief,
      "\nSOURCE ID: ", source_id,
      "\nSOURCE: ", source_text
    ),
    model_prompt = map2_chr(
      summary_system,
      user_prompt,
      \(system, user) {
        nlg_chat_prompt(
          summary_model$tokenizer,
          system,
          user
        )
      }
    ),
    input_tokens = map_int(
      model_prompt,
      \(prompt) nlg_token_count(summary_model$tokenizer, prompt)
    )
  )

Tokenizer counts before summarization

Source ID	Complete prompt tokens
SUM-01	120
SUM-02	117

No text is cut to fit. A longer document would need a stated policy such as sectioning or hierarchical summarization before the model call.

Generate the candidate summaries

Greedy decoding chooses the highest-scoring next token at each step. The output limit is 72 new tokens. The call returns the generated text and token-level stopping information from the local model.

summary_results <- summary_inputs |>
  mutate(
    generation = map(
      model_prompt,
      \(prompt) {
        nlg_generate(
          summary_model,
          prompt,
          max_new_tokens = 72L,
          repetition_penalty = 1.05
        )
      }
    ),
    generated_summary = map_chr(generation, "text"),
    output_tokens = map_int(generation, "output_tokens"),
    ended_by_eos = map_lgl(generation, "ended_by_eos"),
    hit_token_cap = map_lgl(generation, "hit_token_cap"),
    first_sentence_baseline = map_chr(
      str_split(source_text, "(?<=[.!?])\\s+"),
      \(sentences) sentences[[1]]
    )
  )

Generated summaries beside an extractive comparator

Source ID	Generated abstractive candidate	First-sentence teaching baseline
SUM-01	The Riverton library will continue to provide a job-search room with 12 computers, offer Tuesday volunteer help, and maintain a boundary on volunteer assistance during the pilot program’s six-week run.	The Riverton library opens a new job-search room on Monday.
SUM-02	Visitors are encouraged to attend the catalog workshop, participate in photo-identification, and adhere to the restricted-storage limit at the neighborhood archive’s opening on 14 September.	The neighborhood archive will open Room 5 on 14 September for a catalog workshop led by Maya Chen.

The first-sentence baseline is available before any model output is read. It is not a fair contest with a matched length budget; it shows the difference between selecting source wording and generating a new account.

Measure shortening without calling it truth

The next table counts whitespace-separated words. That rule is easy to inspect, but it is different from the model tokenizer used for the generation budget. A compression ratio below 1 means the candidate is shorter. It says nothing about whether the candidate kept the needed facts.

count_words <- function(text) {
  str_count(str_squish(text), "\\S+")
}

compression <- summary_results |>
  transmute(
    source_id,
    source_words = map_int(source_text, count_words),
    generated_words = map_int(generated_summary, count_words),
    first_sentence_words = map_int(
      first_sentence_baseline,
      count_words
    ),
    generated_to_source = generated_words / source_words
  )

Length diagnostics use whitespace-separated words

Source ID	Source words	Generated words	Baseline words	Generated/source ratio
SUM-01	35	30	10	0.86
SUM-02	36	25	18	0.69

The generated sentence can be short because it chose the right content, because it omitted necessary content, or both. Length cannot separate those cases.

Inspect facts chosen before generation

Rosa writes a checklist before reading the output. The regular expressions below flag visible strings related to that checklist. A missing flag points to an omission worth reading. A present flag cannot prove that the name, number, or negative statement has the right role or scope.

summary_flags <- summary_checks |>
  left_join(
    summary_results |>
      select(source_id, generated_summary),
    by = "source_id"
  ) |>
  mutate(
    text_flag = str_detect(
      generated_summary,
      regex(pattern, ignore_case = TRUE)
    )
  ) |>
  select(
    source_id,
    check_id,
    needed_information,
    required,
    text_flag
  )

summary_actions <- summary_flags |>
  group_by(source_id) |>
  summarise(
    decision = list(nlg_required_fact_action(required, text_flag)),
    optional_clues_missing = sum(!required & !text_flag),
    .groups = "drop"
  ) |>
  mutate(
    missing_required = map_int(decision, "missing_required"),
    action = if_else(
      map_lgl(decision, "revision_needed"),
      "revise",
      "send to human source review"
    )
  ) |>
  select(
    source_id,
    missing_required,
    optional_clues_missing,
    action
  )

Surface flags for required facts and optional review clues

Source ID	Check	Information chosen before generation	Required by brief	Surface flag present
SUM-01	F1	job-search service	TRUE	TRUE
SUM-01	F2	Tuesday help	TRUE	TRUE
SUM-01	F3	volunteers do not complete applications	TRUE	FALSE
SUM-01	F4	12 computers	FALSE	TRUE
SUM-01	F5	six-week pilot	FALSE	TRUE
SUM-02	F6	catalog workshop	TRUE	TRUE
SUM-02	F7	photo identification	TRUE	TRUE
SUM-02	F8	no restricted-storage access	TRUE	FALSE
SUM-02	F9	Room 5	FALSE	FALSE
SUM-02	F10	14 September	FALSE	TRUE
SUM-02	F11	Maya Chen	FALSE	FALSE
SUM-02	F12	free registration	FALSE	FALSE

Only missing required facts trigger automatic revision

Source ID	Missing required facts	Missing optional clues	Next action
SUM-01	1	0	revise
SUM-02	1	3	revise

This check is intentionally incomplete. It cannot detect a changed relation, a number attached to the wrong thing, an unsupported cause, or a fluent generalization. Those require reading the source and candidate together. In this render, the first candidate changes the boundary on volunteer help. The second uses negative wording while changing the relationship to restricted storage. The surface table catches part of the problem, and source comparison catches the rest. Likewise, n-gram overlap metrics such as ROUGE measure reference overlap under a specified normalization rule. They do not measure the share of claims supported by the source.

Record the review boundary

Execution can establish that a model ran, stayed inside the configured budgets, and returned text. It cannot grant factual or editorial approval.

summary_review <- tibble(
  review = c(
    "local model execution",
    "prompt input within 256 tokens",
    "output stopped at EOS or token cap",
    "required-fact surface screen",
    "human source-support review",
    "human usefulness and prose review"
  ),
  status = c(
    sprintf(
      "completed: %d candidates",
      sum(nzchar(summary_results$generated_summary))
    ),
    sprintf(
      "within budget: %d/%d candidates",
      sum(summary_inputs$input_tokens <= 256L),
      nrow(summary_inputs)
    ),
    sprintf(
      "recorded: %d/%d candidates",
      sum(summary_results$ended_by_eos | summary_results$hit_token_cap),
      nrow(summary_results)
    ),
    sprintf(
      "held: %d/%d candidates missing required facts",
      sum(summary_actions$missing_required > 0L),
      nrow(summary_actions)
    ),
    "pending",
    "pending"
  )
)

Automatic execution checks remain separate from human review

Review gate	Status
local model execution	completed: 2 candidates
prompt input within 256 tokens	within budget: 2/2 candidates
output stopped at EOS or token cap	recorded: 2/2 candidates
required-fact surface screen	held: 2/2 candidates missing required facts
human source-support review	pending
human usefulness and prose review	pending

A publishable summary needs a reader to trace each claim to the source, decide whether the brief’s essential information survived, and reject unsupported additions. A fluent one-sentence result is only a candidate.

What to remember

• Abstractive summarization generates wording; extractive summarization selects source text.
• Count the complete prompt with the model tokenizer and do not hide truncation.
• Compression describes length, not coverage or source support.
• Check omissions and additions against a brief written before generation.
• Names, numbers, and negation flags are prompts for review, not semantic proof.
• Keep model execution separate from human factual and editorial approval.

Sources

• Maynez et al. on faithfulness and factuality in abstractive summarization
• Lin on ROUGE and reference n-gram overlap
• Pagnoni et al. on classes of factual errors in summaries
• Qwen2.5-1.5B-Instruct model card
• Transformers text-generation controls