Automated Text Summarisation with R

Introduction

This tutorial introduces automated text summarisation in R — a set of methods for condensing texts by extracting or generating their most important content. Text summarisation is increasingly important in corpus linguistics, digital humanities, and applied language research: the volume of textual data available far exceeds what any researcher can read manually, and automated methods provide scalable tools for navigating large collections.

The tutorial covers two broad families of summarisation: extractive summarisation, which selects a subset of the original sentences judged to be most representative or central, and abstractive summarisation, which generates new sentences that paraphrase the source text. We focus primarily on extractive methods — which are easier to implement, more transparent, and more widely used in linguistic research — and discuss abstractive approaches as an optional extension.

Within extractive summarisation we cover three approaches: the LexRank graph-based algorithm (via the lexRankr package), a term-frequency–inverse document frequency (TF-IDF) scoring approach, and a sentence-position heuristic. A worked corpus example applies summarisation to the chapters of a classic text to produce a structured summary.

Prerequisite Tutorials

Before working through this tutorial, we recommend familiarity with:

• Getting Started with R — R objects, basic syntax, RStudio orientation
• String Processing in R — regular expressions, stringr
• Loading and Saving Data in R — reading text files into R
• Web Scraping with R — extracting text from web pages

Learning Objectives

By the end of this tutorial you will be able to:

1. Explain the difference between extractive and abstractive summarisation and when each is appropriate
2. Apply LexRank summarisation to web-scraped and locally loaded texts using lexRankr
3. Implement TF-IDF sentence scoring as a transparent alternative to graph-based methods
4. Apply a sentence-position heuristic and understand its linguistic rationale
5. Compare summarisation outputs from different methods and evaluate their strengths
6. Apply summarisation across the chapters of a multi-section text using a loop
7. Understand the capabilities and limitations of large language model-based abstractive summarisation

Citation

Martin Schweinberger. 2026. Automated Text Summarisation with R. The Language Technology and Data Analysis Laboratory (LADAL), The University of Queensland, Australia. url: https://ladal.edu.au/tutorials/txtsum/txtsum.html (Version 2026.03.28), doi: .

---

What Is Text Summarisation?

Section Overview

What you will learn: The distinction between extractive and abstractive summarisation; the main algorithmic approaches to extractive summarisation (graph-based, TF-IDF, position-based); where automated summarisation is and is not appropriate for linguistic research; and how to evaluate a summary

Extractive vs. Abstractive Summarisation

Extractive summarisation selects a subset of sentences (or passages) from the source text and presents them as the summary. No new text is generated: every sentence in the output is a verbatim copy of a sentence from the input. Extractive methods are transparent, reproducible, and appropriate when fidelity to the original wording matters. They are the dominant approach in corpus linguistics and computational text analysis.

Abstractive summarisation generates new sentences that paraphrase or compress the source content. The output may contain wording that does not appear anywhere in the original text. Large language models (LLMs) like GPT-4 and Claude are capable of high-quality abstractive summarisation but introduce challenges: the output is not directly traceable to source sentences, models can introduce factual errors or “hallucinate” content, and the process is not fully reproducible. Abstractive summarisation is briefly discussed in the final section of this tutorial.

The key distinction is illustrated in the table below:

Property	Extractive	Abstractive
Source fidelity	Always verbatim	May paraphrase or fabricate
Transparency	High	Lower
Reproducibility	Deterministic	Stochastic (LLMs)
Language coverage	Language-agnostic	Model-dependent
Typical use in linguistics	Corpus navigation, content extraction	Paraphrase generation, annotation

Approaches to Extractive Summarisation

Three main approaches to extractive summarisation are covered in this tutorial:

Graph-based methods (LexRank, TextRank) model the text as a graph in which sentences are nodes and edges represent the lexical similarity between them. A sentence’s importance is measured by its centrality in the graph — how similar it is to many other sentences. The most central sentences are selected for the summary. This approach rewards sentences that are representative of the main topics discussed across the whole text, rather than locally prominent sentences.

TF-IDF scoring computes a relevance score for each sentence based on the term frequency–inverse document frequency weights of the words it contains. A sentence scores highly if it contains many words that are frequent in this text but rare in a reference corpus. This approach rewards sentences that contain the distinctive vocabulary of the document and is particularly transparent: the score of each sentence is a direct function of its word content.

Position heuristics exploit the well-established discourse organisation principle that important content tends to appear at the beginning and end of documents and at the beginning of paragraphs. A position score assigns higher weight to sentences appearing earlier in a text. This is the simplest approach and often performs surprisingly well for news articles and academic abstracts, which follow predictable organisational conventions.

When Is Automated Summarisation Appropriate?

Automated summarisation is appropriate when:

• the text is too long to read in full and an overview is needed
• a collection of documents needs to be navigated and key passages identified
• a researcher wants to identify which sentences are most representative of a topic
• summarisation is used as a preprocessing step for further analysis

It is not a substitute for careful reading when:

• the nuance of argument structure matters (summaries flatten hedging, counter-arguments, and qualifications)
• attribution and quotation accuracy are critical
• the text has non-standard organisation (e.g., poetry, transcribed conversation, highly technical tables)

Evaluating a Summary

The standard automatic metric for summarisation evaluation is ROUGE (Recall-Oriented Understudy for Gisting Evaluation), which measures the overlap between the candidate summary and a human-written reference summary. The most common variant, ROUGE-1, counts unigram overlaps; ROUGE-2 counts bigram overlaps. For linguistic research purposes, informal inspection is often more appropriate than ROUGE: does the summary cover the main claims? Does it avoid trivial or peripheral sentences?

Exercises: Concepts

Q1. A researcher uses extractive summarisation on a set of newspaper editorials and obtains a three-sentence summary for each article. A colleague argues that because the summaries contain the original sentences verbatim, they are “not really summaries — they are just quotations.” Is this a valid criticism?

Yes — only abstractive summaries count as true summarisation
No — extractive summarisation is always better than abstractive because it never fabricates content
Partially — extractive summaries do quote verbatim, but the selection process identifies the most representative sentences, which is the core task of summarisation
No — the criticism confuses summarisation with paraphrase; they are entirely different tasks

---

Q2. Which summarisation approach is most appropriate when you want to identify the sentences in a long academic article that are most representative of its overall content (not just its introduction)?

A position heuristic — the most important sentences are always in the introduction
A graph-based method like LexRank, which measures each sentence's similarity to all other sentences and selects the most central ones
Abstractive summarisation with an LLM — only language models can understand academic argument structure
TF-IDF scoring — sentences with the most distinctive words are always the most representative

---

Setup

Installing Packages

# Run once — comment out after installation
install.packages(c(
  "lexRankr",     # LexRank extractive summarisation
  "tidytext",     # TF-IDF scoring
  "dplyr",        # data manipulation
  "stringr",      # string processing
  "stringi",      # sentence boundary detection
  "tidyr",        # data reshaping
  "ggplot2",      # visualisation
  "flextable",    # formatted tables
  "xml2",         # HTML parsing (web scraping)
  "rvest",        # web scraping
  "here",         # file paths
  "purrr",        # functional programming
  "checkdown"     # interactive exercises
))

Loading Packages

library(lexRankr)
library(tidytext)
library(dplyr)
library(stringr)
library(stringi)
library(tidyr)
library(ggplot2)
library(flextable)
library(xml2)
library(rvest)
library(here)
library(purrr)
library(checkdown)

---

LexRank Summarisation

Section Overview

What you will learn: How the LexRank algorithm works; how to apply lexRankr::lexRank() to a web-scraped article and to a locally loaded text; how to control the number of summary sentences; and how to reorder extracted sentences into chronological order

How LexRank Works

LexRank (Erkan and Radev 2004) represents a document as a graph in which each sentence is a node. Edges between nodes are weighted by the cosine similarity of the sentences’ TF-IDF vectors — sentences that share many of the same important words have high-weight edges. The algorithm then applies a variant of Google’s PageRank to identify which nodes are most central in the graph. Sentences with high centrality scores are selected for the summary.

The key properties of LexRank are:

• It rewards globally representative sentences — those that are similar to many other sentences across the whole document
• It is unsupervised — no labelled training data or pre-trained model is needed
• It is language-agnostic — it works for any language with whitespace-delimited words
• It naturally avoids redundancy — the continuous variant (used via continuous = TRUE) down-weights sentences that are too similar to already-selected ones

Applying LexRank to a Web-Scraped Article

We download a Guardian article about the G20 summit and apply LexRank to identify the three most central sentences:

url <- "https://www.theguardian.com/world/2017/jun/26/angela-merkel-and-donald-trump-head-for-clash-at-g20-summit"

page <- xml2::read_html(url)

text <- page |>
  rvest::html_nodes("p") |>
  rvest::html_text() |>
  (\(x) x[nchar(x) > 0])()

# inspect the first few paragraphs
head(text, 4)

[1] "German chancellor plans to make climate change, free trade and mass migration key themes in Hamburg, putting her on collision course with US"                                                                                                       
[2] "A clash between Angela Merkel and Donald Trump appears unavoidable after Germany signalled that it will make climate change, free trade and the management of forced mass global migration the key themes of the G20 summit in Hamburg next week."  
[3] "The G20 summit brings together the world’s biggest economies, representing 85% of global gross domestic product (GDP), and Merkel’s chosen agenda looks likely to maximise American isolation while attempting to minimise disunity amongst others."
[4] "The meeting, which is set to be the scene of large-scale street protests, will also mark the first meeting between Trump and the Russian president, Vladimir Putin, as world leaders."                                                              

top3 <- lexRankr::lexRank(
  text,
  docId      = rep(1, length(text)),  # single document
  n          = 3,                      # number of sentences to extract
  continuous = TRUE                    # use continuous centrality scores
)

Parsing text into sentences and tokens...DONE
Calculating pairwise sentence similarities...DONE
Applying LexRank...DONE
Formatting Output...DONE

top3

  docId sentenceId
1     1        1_2
2     1        1_5
3     1       1_16
                                                                                                                                                                                                                                                       sentence
1             A clash between Angela Merkel and Donald Trump appears unavoidable after Germany signalled that it will make climate change, free trade and the management of forced mass global migration the key themes of the G20 summit in Hamburg next week.
2                     Trump has already rowed with Europe once over climate change and refugees at the G7 summit in Italy, and now looks set to repeat the experience in Hamburg but on a bigger stage, as India and China join in the criticism of Washington.
3 But the G7, and Trump’s subsequent decision to shun the Paris climate change treaty, clearly left a permanent mark on her, leading to her famous declaration of independence four days later at a Christian Social Union (CSU) rally in a Bavarian beer tent.
    value
1 0.06017
2 0.05656
3 0.04975

The sentenceId column encodes document and sentence positions. We extract and display the sentences in chronological order:

top3 |>
  dplyr::mutate(
    sentenceId = as.numeric(stringr::str_remove_all(sentenceId, ".*_"))
  ) |>
  dplyr::arrange(sentenceId) |>
  dplyr::pull(sentence)

[1] "A clash between Angela Merkel and Donald Trump appears unavoidable after Germany signalled that it will make climate change, free trade and the management of forced mass global migration the key themes of the G20 summit in Hamburg next week."            
[2] "Trump has already rowed with Europe once over climate change and refugees at the G7 summit in Italy, and now looks set to repeat the experience in Hamburg but on a bigger stage, as India and China join in the criticism of Washington."                    
[3] "But the G7, and Trump’s subsequent decision to shun the Paris climate change treaty, clearly left a permanent mark on her, leading to her famous declaration of independence four days later at a Christian Social Union (CSU) rally in a Bavarian beer tent."

Applying LexRank to a Locally Loaded Text

We now apply LexRank to Charles Darwin’s On the Origin of Species, stored as an .rda file:

darwin_raw <- base::readRDS("tutorials/txtsum/data/origindarwin.rda", "rb")

# Coerce to UTF-8, replacing any invalid bytes with a space
darwin <- stringi::stri_enc_toutf8(darwin_raw, is_unknown_8bit = TRUE,
                                    validate = TRUE)
darwin <- stringr::str_replace_all(darwin, "[^\u0001-\uFFEF]", " ")

# inspect structure
length(darwin)

[1] 20001

substr(paste(darwin, collapse = " "), 1, 500)

[1] "THE ORIGIN OF SPECIES  BY  CHARLES DARWIN  AN HISTORICAL SKETCH  OF THE PROGRESS OF OPINION ON  THE ORIGIN OF SPECIES  INTRODUCTION  When on board H.M.S. 'Beagle,' as naturalist, I was much struck  with certain facts in the distribution of the organic beings in-  habiting South America, and in the geological relations of the  present to the past inhabitants of that continent. These facts, as  will be seen in the latter chapters of this volume, seemed to throw  some light on the origin of species"

We split the text into individual sentences using stringi::stri_split_boundaries():

darwin_text <- paste(darwin, collapse = " ")

darwin_sents <- stringi::stri_split_boundaries(darwin_text,
                                                type = "sentence")[[1]] |>
  stringr::str_squish() |>
  (\(x) x[nchar(x) > 20])()

length(darwin_sents)

[1] 5217

darwin_top5 <- lexRankr::lexRank(
  darwin_sents[1:200],
  docId      = rep("1", 200),   # character, not integer
  n          = 5,
  continuous = TRUE
)

Parsing text into sentences and tokens...DONE
Calculating pairwise sentence similarities...DONE
Applying LexRank...DONE
Formatting Output...DONE

darwin_top5 |>
  dplyr::mutate(
    sentenceId = as.numeric(stringr::str_remove_all(sentenceId, ".*_"))
  ) |>
  dplyr::arrange(sentenceId) |>
  dplyr::pull(sentence)

[1] "ORIGIN OF SPECIES CHAPTER I Variation under Domestication Causes of variability Effects of habit and the use or disuse of parts- Correlated variation Inheritance Character of domestic varie- ties Difificulty of distinguishing between varieties and species Origin of domestic varieties from one or more species Domestic pigeons, their differences and origin Principles of selection, an- ciently followed, their efifects Methodical and unconscious selection Unknown origin of our domestic productions Circum- stances favourable to man's power of selection CAUSES OF VARIABILITY WHEN we compare the individuals of the same variety or sub-variety of our older cultivated plants and animals, one of the first points which strikes us is, that they generally differ more from each other than do the individuals of any one species or variety in a state of nature."
[2] "And if we reflect on the vast diversity of the plants and animals which have been cultivated, and which have varied during all ages under the most different climates and treatment, we are driven to conclude that this great varia- bility is due to our domestic productions having been raised under conditions of life not so uniform as, and somewhat different from, those to which the parent species had been exposed under nature."                                                                                                                                                                                                                                                                                                                                                                                                                                           
[3] "CHARACTER OF DOMESTIC VARIETIES; DIFFICULTY OF DISTINGUISHING BETWEEN VARIETIES AND SPECIES; ORIGIN OF DOMESTIC VARIETIES FROM ONE OR MORE SPECIES When we look to the hereditary varieties or races of our domestic animals and plants, and compare them with closely allied species, we generally perceive in each domestic race, as already remarked, less uniformity of character than in true species."                                                                                                                                                                                                                                                                                                                                                                                                                                                                            
[4] "I cannot doubt that if other animals and plants, equal in number to our domesticated productions, and belonging to equally diverse classes and countries, were taken from a state of nature, and could be made to breed for an equal number of genera- CHARACTER OF DOMESTIC VARIETIES 35 tions under domestication, they would on an average vary as largely as the parent species of our existing domesticated productions have varied."                                                                                                                                                                                                                                                                                                                                                                                                                                              
[5] "I have, after a laborious collection of all known facts, come to the conclusion that several wild species of Canidc'c have been tamed, and that their blood, in some cases mingled together, flows in the veins of our domestic breeds."                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                

Controlling the Number of Extracted Sentences

The n argument controls how many sentences are extracted. We compare one-sentence, three-sentence, and five-sentence summaries:

Computational Note: Sentence Subset

LexRank computes pairwise cosine similarities between all sentences — an O(n²) operation. Applied to the full Origin of Species (several thousand sentences) this takes hours on a standard laptop. Throughout this section we therefore work with the first 200 sentences. For your own texts, a subset of 200–500 sentences is usually sufficient to produce a representative summary; for longer texts consider splitting by chapter or section first (as demonstrated in the Worked Corpus Example section below).

for (n_sents in c(1, 3, 5)) {
  cat("\n--- Top", n_sents, "sentence(s) ---\n")
  res <- lexRankr::lexRank(
    darwin_sents[1:200],
    docId      = rep("1", 200),
    n          = n_sents,
    continuous = TRUE
  )
  res |>
    dplyr::mutate(
      sentenceId = as.numeric(stringr::str_remove_all(sentenceId, ".*_"))
    ) |>
    dplyr::arrange(sentenceId) |>
    dplyr::pull(sentence) |>
    cat(sep = "\n")
}

--- Top 1 sentence(s) ---
Parsing text into sentences and tokens...DONE
Calculating pairwise sentence similarities...DONE
Applying LexRank...DONE
Formatting Output...DONE
ORIGIN OF SPECIES CHAPTER I Variation under Domestication Causes of variability Effects of habit and the use or disuse of parts- Correlated variation Inheritance Character of domestic varie- ties Difificulty of distinguishing between varieties and species Origin of domestic varieties from one or more species Domestic pigeons, their differences and origin Principles of selection, an- ciently followed, their efifects Methodical and unconscious selection Unknown origin of our domestic productions Circum- stances favourable to man's power of selection CAUSES OF VARIABILITY WHEN we compare the individuals of the same variety or sub-variety of our older cultivated plants and animals, one of the first points which strikes us is, that they generally differ more from each other than do the individuals of any one species or variety in a state of nature.

--- Top 3 sentence(s) ---
Parsing text into sentences and tokens...DONE
Calculating pairwise sentence similarities...DONE
Applying LexRank...DONE
Formatting Output...DONE
ORIGIN OF SPECIES CHAPTER I Variation under Domestication Causes of variability Effects of habit and the use or disuse of parts- Correlated variation Inheritance Character of domestic varie- ties Difificulty of distinguishing between varieties and species Origin of domestic varieties from one or more species Domestic pigeons, their differences and origin Principles of selection, an- ciently followed, their efifects Methodical and unconscious selection Unknown origin of our domestic productions Circum- stances favourable to man's power of selection CAUSES OF VARIABILITY WHEN we compare the individuals of the same variety or sub-variety of our older cultivated plants and animals, one of the first points which strikes us is, that they generally differ more from each other than do the individuals of any one species or variety in a state of nature.
And if we reflect on the vast diversity of the plants and animals which have been cultivated, and which have varied during all ages under the most different climates and treatment, we are driven to conclude that this great varia- bility is due to our domestic productions having been raised under conditions of life not so uniform as, and somewhat different from, those to which the parent species had been exposed under nature.
I cannot doubt that if other animals and plants, equal in number to our domesticated productions, and belonging to equally diverse classes and countries, were taken from a state of nature, and could be made to breed for an equal number of genera- CHARACTER OF DOMESTIC VARIETIES 35 tions under domestication, they would on an average vary as largely as the parent species of our existing domesticated productions have varied.

--- Top 5 sentence(s) ---
Parsing text into sentences and tokens...DONE
Calculating pairwise sentence similarities...DONE
Applying LexRank...DONE
Formatting Output...DONE
ORIGIN OF SPECIES CHAPTER I Variation under Domestication Causes of variability Effects of habit and the use or disuse of parts- Correlated variation Inheritance Character of domestic varie- ties Difificulty of distinguishing between varieties and species Origin of domestic varieties from one or more species Domestic pigeons, their differences and origin Principles of selection, an- ciently followed, their efifects Methodical and unconscious selection Unknown origin of our domestic productions Circum- stances favourable to man's power of selection CAUSES OF VARIABILITY WHEN we compare the individuals of the same variety or sub-variety of our older cultivated plants and animals, one of the first points which strikes us is, that they generally differ more from each other than do the individuals of any one species or variety in a state of nature.
And if we reflect on the vast diversity of the plants and animals which have been cultivated, and which have varied during all ages under the most different climates and treatment, we are driven to conclude that this great varia- bility is due to our domestic productions having been raised under conditions of life not so uniform as, and somewhat different from, those to which the parent species had been exposed under nature.
CHARACTER OF DOMESTIC VARIETIES; DIFFICULTY OF DISTINGUISHING BETWEEN VARIETIES AND SPECIES; ORIGIN OF DOMESTIC VARIETIES FROM ONE OR MORE SPECIES When we look to the hereditary varieties or races of our domestic animals and plants, and compare them with closely allied species, we generally perceive in each domestic race, as already remarked, less uniformity of character than in true species.
I cannot doubt that if other animals and plants, equal in number to our domesticated productions, and belonging to equally diverse classes and countries, were taken from a state of nature, and could be made to breed for an equal number of genera- CHARACTER OF DOMESTIC VARIETIES 35 tions under domestication, they would on an average vary as largely as the parent species of our existing domesticated productions have varied.
I have, after a laborious collection of all known facts, come to the conclusion that several wild species of Canidc'c have been tamed, and that their blood, in some cases mingled together, flows in the veins of our domestic breeds.

Exercises: LexRank

Q3. LexRank selects sentences with high centrality in the sentence-similarity graph. What type of sentence tends to score highly, and what type tends to score low?

The first and last sentences of the document always score highest; middle sentences score low
Short sentences score highly because they are easier to compare; long sentences score low
Sentences that use vocabulary common across many other sentences in the document score highly; sentences with unique, specialised, or tangential content score low
Sentences with the most named entities (people, places) score highest; sentences without named entities score low

---

Q4. You run lexRankr::lexRank() with continuous = TRUE and continuous = FALSE. What is the main practical difference between these two options?

continuous = TRUE returns a continuous score column; continuous = FALSE returns only sentence text
continuous = TRUE is faster; continuous = FALSE is more accurate
continuous = TRUE uses a weighted (continuous) graph where edge weights reflect similarity degree; continuous = FALSE binarises edges using a threshold, treating sentences as either similar or not
There is no difference — the parameter name is misleading and both produce identical output

---

TF-IDF Sentence Scoring

Section Overview

What you will learn: How TF-IDF weights sentences by the distinctiveness of their vocabulary; how to compute sentence-level TF-IDF scores using tidytext; how to select the top-scoring sentences as an extractive summary; and how TF-IDF summarisation compares to LexRank

TF-IDF as a Sentence Scoring Method

Term Frequency–Inverse Document Frequency (TF-IDF) is a classic information retrieval measure that weights a word by how often it appears in a particular document (term frequency, TF) relative to how often it appears across a collection of documents (inverse document frequency, IDF). Words that are frequent in a document but rare across the collection are given high weight — they are the distinctive vocabulary of that document.

For summarisation, we score each sentence by summing the TF-IDF weights of its constituent words. A sentence that contains many high-TF-IDF words is considered more informative than one filled with common function words. This approach is:

• transparent: the score of each sentence is directly traceable to specific words
• interpretable: the top-scoring words reveal what the document is “about”
• flexible: it can be adapted for multi-document summarisation by computing IDF across the full collection

Computing TF-IDF Scores

We use the tidytext package to tokenise the text, compute TF-IDF weights, and score sentences:

# Create a sentence-level data frame
sents_df <- data.frame(
  sentence_id = seq_along(darwin_sents),
  text        = darwin_sents,
  stringsAsFactors = FALSE
)

# Tokenise to word level
words_df <- sents_df |>
  tidytext::unnest_tokens(word, text) |>
  dplyr::filter(!word %in% tidytext::stop_words$word) |>  # remove stop words
  dplyr::count(sentence_id, word, name = "n")

# Compute TF-IDF (treating each sentence as a "document")
tfidf_df <- words_df |>
  tidytext::bind_tf_idf(word, sentence_id, n)

head(tfidf_df |> dplyr::arrange(desc(tf_idf)), 15)

   sentence_id         word n tf   idf tf_idf
1         2347  paradoxical 1  1 7.864  7.864
2         4175        verse 1  1 7.864  7.864
3         5103        avail 1  1 7.864  7.864
4         2395         busk 1  1 7.458  7.458
5            8      induced 1  1 7.171  7.171
6          161 communicated 1  1 7.171  7.171
7         2591          cat 1  1 6.765  6.765
8         1657      noticed 1  1 6.477  6.477
9         2323       fairly 1  1 6.159  6.159
10        5028      endless 1  1 6.072  6.072
11         613   expression 1  1 5.992  5.992
12        4428        false 1  1 5.992  5.992
13        1551  undoubtedly 1  1 5.918  5.918
14        2193       return 1  1 5.849  5.849
15        3439        agree 1  1 5.849  5.849

Scoring and Ranking Sentences

We aggregate word-level TF-IDF scores to the sentence level by summing:

sent_scores <- tfidf_df |>
  dplyr::group_by(sentence_id) |>
  dplyr::summarise(tfidf_score = sum(tf_idf), .groups = "drop") |>
  dplyr::arrange(desc(tfidf_score))

head(sent_scores, 10)

# A tibble: 10 × 2
   sentence_id tfidf_score
         <int>       <dbl>
 1        1665        8.21
 2        1880        8.21
 3         391        8.11
 4        2096        8.08
 5        1775        8.01
 6        2819        7.96
 7        3599        7.86
 8        2347        7.86
 9        4175        7.86
10        5103        7.86

# Extract the top 5 sentences and return them in document order
top5_tfidf <- sent_scores |>
  dplyr::slice_max(tfidf_score, n = 5) |>
  dplyr::left_join(sents_df, by = "sentence_id") |>
  dplyr::arrange(sentence_id)

top5_tfidf |>
  dplyr::select(sentence_id, tfidf_score, text) |>
  flextable::flextable() |>
  flextable::set_table_properties(width = .95, layout = "autofit") |>
  flextable::theme_zebra() |>
  flextable::fontsize(size = 10) |>
  flextable::set_caption(caption = "Top 5 sentences by TF-IDF score (Darwin, On the Origin of Species).") |>
  flextable::border_outer()

sentence_id	tfidf_score	text
391	8.109	Isi- dore Geoffroy St.
1,665	8.210	And it would appear from infor- mation given me by Mr.
1,775	8.007	But may not this inference be presumptuous?
1,880	8.210	Criigcr in the Coryanthes.
2,096	8.080	In Helian- themum the capsule has been described as unilocular or 3-locular; and in H. mutabile, "Une lame, plus on mains large, s'etend entre le pericarpe et le placenta."

Visualising the Top Words

We visualise the highest-TF-IDF words to understand what vocabulary is driving the sentence scores:

tfidf_df |>
  dplyr::group_by(word) |>
  dplyr::summarise(mean_tfidf = mean(tf_idf), .groups = "drop") |>
  dplyr::slice_max(mean_tfidf, n = 20) |>
  ggplot(aes(x = reorder(word, mean_tfidf), y = mean_tfidf)) +
  geom_col(fill = "#1f77b4") +
  coord_flip() +
  labs(title = "Top 20 Words by Mean TF-IDF",
       x = NULL, y = "Mean TF-IDF score") +
  theme_bw()

Top 20 words by TF-IDF score across the full text. These are the most distinctive non-stop words in the document.

Exercises: TF-IDF Summarisation

Q5. A sentence contains only the words the, a, is, and it. What TF-IDF score would it receive, and why?

A high score — short sentences with common words are easy to process and therefore central
Close to zero — these are common stop words with very high document frequency, so their IDF weight is near zero
Exactly zero — stop words are removed before scoring, so the sentence contributes nothing
An undefined score — sentences with no content words cannot be scored by TF-IDF

---

Q6. Compare LexRank and TF-IDF summarisation. For a very short document (5 sentences), which would you expect to work better, and why?

LexRank always outperforms TF-IDF regardless of document length
TF-IDF is more reliable for very short texts — LexRank needs a sufficiently large sentence graph to produce meaningful centrality scores
Both methods will fail on documents shorter than 20 sentences
Position heuristics are always better than both for short documents

---

Position-Based Summarisation

Section Overview

What you will learn: The discourse-level rationale for position-based summarisation; how to implement a simple lead-sentence heuristic; how to combine position and TF-IDF scores into a hybrid ranking; and when position-based methods are and are not appropriate

The Linguistic Rationale for Position

Many text genres follow predictable organisational conventions that make sentence position an informative cue for importance. In news articles, the inverted-pyramid structure places the most newsworthy information in the first paragraph; the classic “Five Ws” (who, what, when, where, why) are typically answered in the lead sentence. In academic abstracts, the structure follows Introduction–Methods–Results–Conclusion, with the main claim often in the first or last sentence. In structured reports, topic sentences at the beginning of paragraphs signal the main point of each section.

The lead-sentence heuristic — simply selecting the first n sentences — is a strong baseline that is difficult to beat for news text. For general documents, a more flexible approach assigns a position score that decays with distance from the beginning:

\[\text{position\_score}(i) = \frac{1}{\log(i + 1)}\]

This gives the first sentence a high score, the second a somewhat lower score, and so on, with diminishing returns as the sentence index grows.

Implementing a Position Score

n_sents <- length(darwin_sents)

position_df <- data.frame(
  sentence_id    = seq_along(darwin_sents),
  text           = darwin_sents,
  position_score = 1 / log(seq_along(darwin_sents) + 1),
  stringsAsFactors = FALSE
)

head(position_df |> dplyr::select(sentence_id, position_score), 8)

  sentence_id position_score
1           1         1.4427
2           2         0.9102
3           3         0.7213
4           4         0.6213
5           5         0.5581
6           6         0.5139
7           7         0.4809
8           8         0.4551

# Top 5 sentences by position (these will simply be the first 5)
position_df |>
  dplyr::slice_max(position_score, n = 5) |>
  dplyr::pull(text)

[1] "THE ORIGIN OF SPECIES BY CHARLES DARWIN AN HISTORICAL SKETCH OF THE PROGRESS OF OPINION ON THE ORIGIN OF SPECIES INTRODUCTION When on board H.M.S."                                                                                                                            
[2] "'Beagle,' as naturalist, I was much struck with certain facts in the distribution of the organic beings in- habiting South America, and in the geological relations of the present to the past inhabitants of that continent."                                                 
[3] "These facts, as will be seen in the latter chapters of this volume, seemed to throw some light on the origin of species that mystery of mysteries, as it has been called by one of our greatest philosophers."                                                                 
[4] "On my return home, it occurred to me, in 1837, that something might perhaps be made out on this question by patiently accumulating and reflecting on all sorts of facts which could possibly have any bearing on it."                                                          
[5] "After five years' work I allowed myself to specu- late on the subject, and drew up some short notes; these I enlarged in 1844 into a sketch of the conclusions, which then seemed to me probable; from that period to the present day I have steadily pursued the same object."

Hybrid Position + TF-IDF Scoring

A more powerful approach combines position and TF-IDF scores. We normalise both to the [0, 1] range and average them:

# Normalise TF-IDF scores to [0, 1]
tfidf_norm <- sent_scores |>
  dplyr::mutate(
    tfidf_norm = (tfidf_score - min(tfidf_score)) /
                 (max(tfidf_score) - min(tfidf_score))
  )

# Join with position scores
hybrid_df <- position_df |>
  dplyr::select(sentence_id, text, position_score) |>
  dplyr::mutate(
    pos_norm = (position_score - min(position_score)) /
               (max(position_score) - min(position_score))
  ) |>
  dplyr::left_join(tfidf_norm |> dplyr::select(sentence_id, tfidf_norm),
                   by = "sentence_id") |>
  dplyr::mutate(hybrid_score = (pos_norm + tfidf_norm) / 2) |>
  dplyr::arrange(desc(hybrid_score))

# Top 5 hybrid-scored sentences in document order
hybrid_df |>
  dplyr::slice_max(hybrid_score, n = 5) |>
  dplyr::arrange(sentence_id) |>
  dplyr::select(sentence_id, hybrid_score, text) |>
  flextable::flextable() |>
  flextable::set_table_properties(width = .95, layout = "autofit") |>
  flextable::theme_zebra() |>
  flextable::fontsize(size = 10) |>
  flextable::set_caption(caption = "Top 5 sentences by hybrid position + TF-IDF score.") |>
  flextable::border_outer()

sentence_id	hybrid_score	text
1	0.8025	THE ORIGIN OF SPECIES BY CHARLES DARWIN AN HISTORICAL SKETCH OF THE PROGRESS OF OPINION ON THE ORIGIN OF SPECIES INTRODUCTION When on board H.M.S.
2	0.5339	'Beagle,' as naturalist, I was much struck with certain facts in the distribution of the organic beings in- habiting South America, and in the geological relations of the present to the past inhabitants of that continent.
6	0.5662	I hope that I may be excused for entering on these personal details, as I give them to show that I have not been hasty in coming to a decision.
8	0.5405	I have more especially been induced to do this, as Mr.
12	0.5402	Wallace's excellent memoir, some brief extracts from my manuscripts.

Exercises: Position-Based Summarisation

Q7. You apply the lead-sentence heuristic to a mystery novel and find the summaries are poor. You apply it to a collection of newspaper articles and find the summaries are excellent. What explains this difference?

Newspaper articles are shorter, so position-based methods work better on short texts
Mystery novels use unusual vocabulary that confuses position-based methods
Newspaper articles follow the inverted-pyramid convention, placing key information first; novels deliberately withhold key information and place it later for narrative effect
The heuristic works for newspapers because they have paragraph breaks; novels do not

---

Q8. In the hybrid score formula (pos_norm + tfidf_norm) / 2, both components are given equal weight. For a scientific research article, would you adjust this weighting, and how?

Increase the position weight — the abstract always summarises the whole paper
Increase the TF-IDF weight — scientific articles do not front-load findings; key results and conclusions may appear anywhere in the text
Keep equal weighting — scientific articles are long enough that both signals are equally reliable
Remove TF-IDF entirely — scientific vocabulary is too specialised for TF-IDF to work

---

Comparing Summarisation Methods

Section Overview

What you will learn: How to compare the outputs of LexRank, TF-IDF, and position-based summarisation on the same text; how to visualise sentence score distributions; and how to assess agreement between methods

Side-by-Side Comparison

We extract the top 5 sentence IDs under each method and compare which sentences each approach selects:

# LexRank top 5 sentence IDs
lexrank_ids <- lexRankr::lexRank(
  darwin_sents[1:200],
  docId      = rep("1", 200),
  n          = 5,
  continuous = TRUE
) |>
  dplyr::mutate(
    sent_num = as.integer(stringr::str_remove_all(sentenceId, ".*_"))
  ) |>
  dplyr::pull(sent_num)

Parsing text into sentences and tokens...DONE
Calculating pairwise sentence similarities...DONE
Applying LexRank...DONE
Formatting Output...DONE

# TF-IDF top 5 sentence IDs
tfidf_ids <- sent_scores |>
  dplyr::slice_max(tfidf_score, n = 5) |>
  dplyr::pull(sentence_id)

# Position top 5 sentence IDs (simply the first 5)
position_ids <- 1:5

# Build comparison table
all_ids <- sort(unique(c(lexrank_ids, tfidf_ids, position_ids)))

comparison <- data.frame(
  sentence_id = all_ids,
  LexRank     = all_ids %in% lexrank_ids,
  TF_IDF      = all_ids %in% tfidf_ids,
  Position    = all_ids %in% position_ids,
  stringsAsFactors = FALSE
) |>
  dplyr::mutate(
    n_methods = LexRank + TF_IDF + Position,
    text = darwin_sents[all_ids]
  )

comparison |>
  dplyr::select(sentence_id, LexRank, TF_IDF, Position, n_methods) |>
  flextable::flextable() |>
  flextable::set_table_properties(width = .7, layout = "autofit") |>
  flextable::theme_zebra() |>
  flextable::fontsize(size = 11) |>
  flextable::set_caption(caption = "Sentences selected by each method (top 5 per method). n_methods = number of methods that selected each sentence.") |>
  flextable::border_outer()

sentence_id	LexRank	TF_IDF	Position	n_methods
1	false	false	true	1
2	false	false	true	1
3	false	false	true	1
4	false	false	true	1
5	false	false	true	1
54	true	false	false	1
55	true	false	false	1
136	true	false	false	1
151	true	false	false	1
161	true	false	false	1
391	false	true	false	1
1,665	false	true	false	1
1,775	false	true	false	1
1,880	false	true	false	1
2,096	false	true	false	1

Sentences selected by all three methods (if any) are the most robust candidates for the summary. Sentences selected by only one method are more method-specific.

Visualising Score Distributions

sent_scores |>
  ggplot(aes(x = tfidf_score)) +
  geom_histogram(bins = 40, fill = "#1f77b4", colour = "white") +
  geom_vline(
    xintercept = sent_scores |> dplyr::slice_max(tfidf_score, n = 5) |>
                   dplyr::pull(tfidf_score) |> min(),
    linetype = "dashed", colour = "red", linewidth = 0.8
  ) +
  labs(title = "Distribution of TF-IDF Sentence Scores",
       subtitle = "Red dashed line = threshold for top 5 sentences",
       x = "TF-IDF score", y = "Count") +
  theme_bw()

Distribution of TF-IDF scores across all sentences. The long right tail shows a small number of highly distinctive sentences.

---

Worked Corpus Example: Chapter-by-Chapter Summarisation

Section Overview

What you will learn: How to apply text summarisation to a multi-chapter text using a loop; how to extract and display the top sentences from each chapter; and how to combine the per-chapter summaries into a structured document overview

Splitting the Text into Chapters

Darwin’s On the Origin of Species is divided into chapters. We split the text on the word “CHAPTER” and apply LexRank to each chapter independently:

# Split by chapter marker
chapters_raw <- darwin_text |>
  stringr::str_split("CHAPTER") |>
  unlist() |>
  (\(x) x[nchar(x) > 200])()  # remove short fragments

length(chapters_raw)

[1] 20

# Function: summarise one chapter
summarise_chapter <- function(chapter_text, n = 3) {
  sents <- stringi::stri_split_boundaries(chapter_text,
                                           type = "sentence")[[1]] |>
    stringr::str_squish() |>
    (\(x) x[nchar(x) > 40])()

  if (length(sents) < 5) return(tibble::tibble(sentence = sents))

  lexRankr::lexRank(
    sents,
    docId      = rep(1, length(sents)),
    n          = n,
    continuous = TRUE
  ) |>
    dplyr::mutate(
      sent_num = as.numeric(stringr::str_remove_all(sentenceId, ".*_"))
    ) |>
    dplyr::arrange(sent_num) |>
    dplyr::select(sentence)
}

# Apply to all chapters
chapter_summaries <- purrr::map(chapters_raw, summarise_chapter, n = 3)

Parsing text into sentences and tokens...DONE
Calculating pairwise sentence similarities...DONE
Applying LexRank...DONE
Formatting Output...DONE
Parsing text into sentences and tokens...DONE
Calculating pairwise sentence similarities...DONE
Applying LexRank...DONE
Formatting Output...DONE
Parsing text into sentences and tokens...DONE
Calculating pairwise sentence similarities...DONE
Applying LexRank...DONE
Formatting Output...DONE
Parsing text into sentences and tokens...DONE
Calculating pairwise sentence similarities...DONE
Applying LexRank...DONE
Formatting Output...DONE
Parsing text into sentences and tokens...DONE
Calculating pairwise sentence similarities...DONE
Applying LexRank...DONE
Formatting Output...DONE
Parsing text into sentences and tokens...DONE
Calculating pairwise sentence similarities...DONE
Applying LexRank...DONE
Formatting Output...DONE
Parsing text into sentences and tokens...DONE
Calculating pairwise sentence similarities...DONE
Applying LexRank...DONE
Formatting Output...DONE
Parsing text into sentences and tokens...DONE
Calculating pairwise sentence similarities...DONE
Applying LexRank...DONE
Formatting Output...DONE
Parsing text into sentences and tokens...DONE
Calculating pairwise sentence similarities...DONE
Applying LexRank...DONE
Formatting Output...DONE
Parsing text into sentences and tokens...DONE
Calculating pairwise sentence similarities...DONE
Applying LexRank...DONE
Formatting Output...DONE
Parsing text into sentences and tokens...DONE
Calculating pairwise sentence similarities...DONE
Applying LexRank...DONE
Formatting Output...DONE
Parsing text into sentences and tokens...DONE
Calculating pairwise sentence similarities...DONE
Applying LexRank...DONE
Formatting Output...DONE
Parsing text into sentences and tokens...DONE
Calculating pairwise sentence similarities...DONE
Applying LexRank...DONE
Formatting Output...DONE
Parsing text into sentences and tokens...DONE
Calculating pairwise sentence similarities...DONE
Applying LexRank...DONE
Formatting Output...DONE
Parsing text into sentences and tokens...DONE
Calculating pairwise sentence similarities...DONE
Applying LexRank...DONE
Formatting Output...DONE
Parsing text into sentences and tokens...DONE
Calculating pairwise sentence similarities...DONE
Applying LexRank...DONE
Formatting Output...DONE
Parsing text into sentences and tokens...DONE
Calculating pairwise sentence similarities...DONE
Applying LexRank...DONE
Formatting Output...DONE
Parsing text into sentences and tokens...DONE
Calculating pairwise sentence similarities...DONE
Applying LexRank...DONE
Formatting Output...DONE
Parsing text into sentences and tokens...DONE
Calculating pairwise sentence similarities...DONE
Applying LexRank...DONE
Formatting Output...DONE
Parsing text into sentences and tokens...DONE
Calculating pairwise sentence similarities...DONE
Applying LexRank...DONE
Formatting Output...DONE

names(chapter_summaries) <- paste0("Chapter_", seq_along(chapter_summaries))

Displaying Chapter Summaries

# Display summaries for the first 4 chapters
for (i in 1:min(4, length(chapter_summaries))) {
  cat("\n=========================================\n")
  cat("CHAPTER", i, "\n")
  cat("=========================================\n")
  cat(chapter_summaries[[i]]$sentence, sep = "\n\n")
}

=========================================
CHAPTER 1 
=========================================
These facts, as will be seen in the latter chapters of this volume, seemed to throw some light on the origin of species that mystery of mysteries, as it has been called by one of our greatest philosophers.

In considering the Origin of Species, it is quite conceivable that a naturalist, reflecting on the mutual affinities of organic beings, on their embryological relations, their geographical distribution, geological succession, and other such facts, might come to the con- clusion that species have not been independently created, but had descended, like varieties, from other species.

I will then pass on the variability of species in a state of nature; but I shall, unfortunately, be compelled to treat this subject far too briefly, as it can be treated properly only by giving long catalogues of facts.

=========================================
CHAPTER 2 
=========================================
I cannot doubt that if other animals and plants, equal in number to our domesticated productions, and belonging to equally diverse classes and countries, were taken from a state of nature, and could be made to breed for an equal number of genera- CHARACTER OF DOMESTIC VARIETIES 35 tions under domestication, they would on an average vary as largely as the parent species of our existing domesticated productions have varied.

One circumstance has struck me much ; namely, that nearly all the breeders of the various domestic animals and the cultivators of plants, with whom I have conversed, or whose treatises I have read, are firmly convinced that the several breeds to which each has attended, are descended from so many aboriginally distinct species.

Although I do not doubt that some domestic animals vary less than others, yet the rarity or absence of distinct breeds of the cat, the donkey, peacock, goose, &c., may be attributed in main part to selection not having been brought into play : in cats, from the difficulty in pairing them ; in donkeys, from only a few being kept by poor people, and little attention paid to their breeding; for recently in certain parts of Spain and of the United States this animal has been surprisingly modified and improved by careful selection ; in peacocks, from not being very easily reared and a large stock not kept; in geese, from being valuable only for two purposes, food and feathers, and more especially from no pleasure having been felt in the dis- play of distinct breeds ; but the goose, under the conditions to which it is exposed when domesticated, seems to have a sin- gularly inflexible organisation, though it has varied to a slight extent, as I have elsewhere described.

=========================================
CHAPTER 3 
=========================================
Some few naturalists maintain that animals never present varieties; but then these same naturalists rank the slightest difference as of specific value ; and when the same identical form is met with in two distinct countries, or in two geologi- cal formations, they believe that two distinct species are hid- den under the same dress.

Certainly no clear line of demarcation has as yet been drawn between species and sub-species that is, the forms which in the opinion of some naturalists come very near to, but do not quite arrive at, the rank of species: or, again, between sub-species and well-marked varieties, or between lesser varieties and individual differences.

From these remarks it will be seen that I look at the term species as one arbitrarily given, for the sake of convenience, to a set of individuals closely resembling each other, and that it does not essentially differ from the term variety, which is given to less distinct and more fluctuating forms.

=========================================
CHAPTER 4 
=========================================
III Struggle for Existence Its bearing on natural selection The term used in a wide sense Geometrical ratio of increase Rapid increase of naturalized animals and plants Nature of the checks to increase Competi- tion universal Effects of climate Protection from the number of individuals Complex relations of all animals and plants throughout nature Struggle for life most severe between indi- viduals and varieties of the same species : often severe between species of the same genus The relation of organism to organism the most important of all relations.

The amount of food for each species of course gives the extreme limit to which each can increase; but very fre- NATURE OF THE CHECKS TO INCREASE 83 quently it is not the obtaining food, but the serving as prey to other animals, which determines the average numbers of a species.

That climate acts in main part indirectly by favouring other species, we clearly see in the prodigious number of plants which in our gardens can perfectly well endure our climate, but which never became naturalised, for they can- not compete with our native plants nor resist destruction by our native animals.

Building a Structured Summary Table

We combine all chapter summaries into a single data frame for easy export:

summary_table <- purrr::imap_dfr(chapter_summaries, function(df, nm) {
  dplyr::mutate(df, chapter = nm, sentence_rank = dplyr::row_number())
}) |>
  dplyr::select(chapter, sentence_rank, sentence)

head(summary_table, 9) |>
  flextable::flextable() |>
  flextable::set_table_properties(width = .95, layout = "autofit") |>
  flextable::theme_zebra() |>
  flextable::fontsize(size = 10) |>
  flextable::set_caption(caption = "Structured chapter-by-chapter summary (top 3 sentences per chapter, first 3 chapters shown).") |>
  flextable::border_outer()

chapter	sentence_rank	sentence
Chapter_1	1	These facts, as will be seen in the latter chapters of this volume, seemed to throw some light on the origin of species that mystery of mysteries, as it has been called by one of our greatest philosophers.
Chapter_1	2	In considering the Origin of Species, it is quite conceivable that a naturalist, reflecting on the mutual affinities of organic beings, on their embryological relations, their geographical distribution, geological succession, and other such facts, might come to the con- clusion that species have not been independently created, but had descended, like varieties, from other species.
Chapter_1	3	I will then pass on the variability of species in a state of nature; but I shall, unfortunately, be compelled to treat this subject far too briefly, as it can be treated properly only by giving long catalogues of facts.
Chapter_2	1	I cannot doubt that if other animals and plants, equal in number to our domesticated productions, and belonging to equally diverse classes and countries, were taken from a state of nature, and could be made to breed for an equal number of genera- CHARACTER OF DOMESTIC VARIETIES 35 tions under domestication, they would on an average vary as largely as the parent species of our existing domesticated productions have varied.
Chapter_2	2	One circumstance has struck me much ; namely, that nearly all the breeders of the various domestic animals and the cultivators of plants, with whom I have conversed, or whose treatises I have read, are firmly convinced that the several breeds to which each has attended, are descended from so many aboriginally distinct species.
Chapter_2	3	Although I do not doubt that some domestic animals vary less than others, yet the rarity or absence of distinct breeds of the cat, the donkey, peacock, goose, &c., may be attributed in main part to selection not having been brought into play : in cats, from the difficulty in pairing them ; in donkeys, from only a few being kept by poor people, and little attention paid to their breeding; for recently in certain parts of Spain and of the United States this animal has been surprisingly modified and improved by careful selection ; in peacocks, from not being very easily reared and a large stock not kept; in geese, from being valuable only for two purposes, food and feathers, and more especially from no pleasure having been felt in the dis- play of distinct breeds ; but the goose, under the conditions to which it is exposed when domesticated, seems to have a sin- gularly inflexible organisation, though it has varied to a slight extent, as I have elsewhere described.
Chapter_3	1	Some few naturalists maintain that animals never present varieties; but then these same naturalists rank the slightest difference as of specific value ; and when the same identical form is met with in two distinct countries, or in two geologi- cal formations, they believe that two distinct species are hid- den under the same dress.
Chapter_3	2	Certainly no clear line of demarcation has as yet been drawn between species and sub-species that is, the forms which in the opinion of some naturalists come very near to, but do not quite arrive at, the rank of species: or, again, between sub-species and well-marked varieties, or between lesser varieties and individual differences.
Chapter_3	3	From these remarks it will be seen that I look at the term species as one arbitrarily given, for the sake of convenience, to a set of individuals closely resembling each other, and that it does not essentially differ from the term variety, which is given to less distinct and more fluctuating forms.

How Many Sentences per Chapter?

We visualise how many sentences each chapter contributes (as an indicator of chapter length and information density):

chapter_lengths <- purrr::map_int(chapters_raw, function(ch) {
  stringi::stri_split_boundaries(ch, type = "sentence")[[1]] |>
    stringr::str_squish() |>
    (\(x) x[nchar(x) > 40])() |>
    length()
})

data.frame(
  chapter = paste0("Ch.", seq_along(chapter_lengths)),
  n_sents = chapter_lengths
) |>
  ggplot(aes(x = chapter, y = n_sents)) +
  geom_col(fill = "#1f77b4") +
  labs(title = "Sentences per Chapter",
       x = "Chapter", y = "Number of sentences") +
  theme_bw() +
  theme(axis.text.x = element_text(angle = 45, hjust = 1))

Number of extractable sentences per chapter after sentence boundary detection and filtering.

Exercises: Worked Corpus Example

Q9. You apply lexRankr::lexRank() to a very short chapter (8 sentences) and get unexpected results — the function selects sentences that seem peripheral. What is likely causing this, and how would you address it?

The chapter file is probably corrupted; re-download the text
LexRank requires stop word removal before it can work correctly on short texts
LexRank's graph is too sparse with only 8 sentences to produce reliable centrality scores; use TF-IDF scoring or simply return all sentences for very short chapters
The issue is sentence boundary detection; short chapters always have more sentence fragments than long ones

---

Q10. You want to summarise a corpus of 500 Wikipedia articles, each 15–30 paragraphs long. You need to process all 500 in a reasonable time. What practical consideration favours TF-IDF over LexRank for this task?

TF-IDF produces more accurate summaries for encyclopaedic text than graph-based methods
TF-IDF scoring is O(n) in the number of sentences; LexRank requires computing pairwise sentence similarities, which is O(n²) and much slower for long documents
LexRank cannot handle multiple documents; TF-IDF is designed for multi-document settings
TF-IDF is always faster because it uses simpler arithmetic than LexRank's PageRank algorithm

---

Abstractive Summarisation

Section Overview

What you will learn: What abstractive summarisation is; how large language models generate abstractive summaries via the Anthropic API; key limitations and risks (hallucination, non-reproducibility, attribution loss); and when abstractive summarisation is appropriate for linguistic research

Abstractive Summarisation with Large Language Models

Large language models (LLMs) such as Claude, GPT-4, or Llama can generate abstractive summaries by reading a text and producing a new, condensed paraphrase. Unlike the extractive methods above, LLM summaries:

• may use wording that does not appear in the source text
• can compress multiple sentences into one or split a single complex sentence into simpler ones
• can synthesise information from non-adjacent sections of the text
• adapt the summary style to the specified audience and purpose

The following example calls the Anthropic API with a passage from Darwin to generate a three-sentence abstractive summary:

# This chunk requires an Anthropic API key — set eval=TRUE once configured
library(httr2)

darwin_passage <- paste(darwin_sents[1:30], collapse = " ")

response <- httr2::request("https://api.anthropic.com/v1/messages") |>
  httr2::req_headers(
    "x-api-key"         = Sys.getenv("ANTHROPIC_API_KEY"),
    "anthropic-version" = "2023-06-01",
    "content-type"      = "application/json"
  ) |>
  httr2::req_body_json(list(
    model      = "claude-sonnet-4-20250514",
    max_tokens = 256,
    messages   = list(
      list(
        role    = "user",
        content = paste0(
          "Please provide a three-sentence summary of the following text. ",
          "Focus on the main argument and key evidence.\n\n",
          darwin_passage
        )
      )
    )
  )) |>
  httr2::req_perform() |>
  httr2::resp_body_json()

cat(response$content[[1]]$text)

Limitations of Abstractive Summarisation for Research

Abstractive summarisation with LLMs introduces several concerns that researchers should weigh carefully:

Hallucination: LLMs can generate plausible-sounding but factually incorrect statements, particularly for specialised or technical content. Always verify abstractive summaries against the source text before citing them.

Non-reproducibility: LLM outputs are stochastic — the same input will produce different outputs on different runs (and over time as models are updated). Extractive summaries are fully deterministic.

Attribution loss: An abstractive summary is not quotable as a source; it is a paraphrase generated by a model. In research contexts requiring traceable evidence, extractive methods are preferable.

Language and domain bias: LLMs perform better on languages and genres well-represented in their training data (primarily English, primarily recent web text). Performance on historical texts, minority languages, or highly technical domains may be lower.

Exercises: Abstractive Summarisation

Q11. A researcher uses an LLM to abstractively summarise 200 historical letters for a corpus study. The summaries will be used as metadata annotations. What is the most serious methodological concern?

Speed — LLMs are too slow to process 200 letters in a reasonable time
Hallucination — the LLM may generate content that does not appear in the source letters, introducing false claims into the metadata
The summaries will be too long to serve as metadata annotations
LLMs cannot process historical text because the vocabulary has changed too much

---

Q12. A colleague argues that abstractive summarisation is always inferior to extractive summarisation because “it makes things up.” Is this characterisation fair?

Yes — extractive summarisation is always preferable because it never generates new text
No — abstractive summarisation can be superior when compression, synthesis across sections, or plain-language paraphrase is needed; the risk of hallucination is real but manageable with verification
Yes — LLMs are not reliable enough to be used in academic research at all
No — abstractive summarisation is always superior because it produces more natural-sounding text

---

Summary and Further Reading

This tutorial introduced automated text summarisation in R, covering both the conceptual framework and practical implementation of three extractive methods and an overview of abstractive approaches.

We began with the conceptual distinction between extractive summarisation (selecting verbatim sentences) and abstractive summarisation (generating new paraphrases), and discussed the three main extractive approaches: graph-based (LexRank), TF-IDF scoring, and position heuristics.

LexRank was demonstrated using both web-scraped and locally loaded texts. The algorithm identifies sentences that are most central in a sentence-similarity graph — those that share key vocabulary with many other sentences throughout the document. The continuous = TRUE option uses a weighted graph and is recommended for most applications.

TF-IDF sentence scoring computes a relevance score for each sentence based on the distinctiveness of its vocabulary. It is more transparent than LexRank — the contribution of each word is traceable — and more robust for very short texts or collections where a reference IDF can be computed across multiple documents.

Position-based summarisation exploits genre-specific information-structural conventions. The lead-sentence heuristic is a strong baseline for news text; a hybrid position + TF-IDF score combines the two signals and is more appropriate for genres where key content does not consistently front-load.

The worked corpus example applied LexRank chapter-by-chapter to Darwin’s On the Origin of Species, producing a structured per-chapter summary table. A utility function handled the common edge case of very short chapters.

Abstractive summarisation via LLMs was discussed as an extension, with emphasis on the key limitations for research use: hallucination risk, non-reproducibility, and attribution loss. Extractive methods remain the default recommendation for corpus linguistics and text analysis applications where fidelity to source text is important.

Further reading: Erkan and Radev (2004) is the original LexRank paper. Mihalcea and Tarau (2004) describes the closely related TextRank algorithm. Jones et al. (1999) provides a comprehensive early overview of automatic summarisation. For TF-IDF and information retrieval fundamentals, see Manning (2008). For the use of LLMs in linguistic research, see Bender et al. (2021) for a critical perspective on large language model capabilities and limitations.

---

Citation & Session Info

Citation

Martin Schweinberger. 2026. Automated Text Summarisation with R. The Language Technology and Data Analysis Laboratory (LADAL), The University of Queensland, Australia. url: https://ladal.edu.au/tutorials/txtsum/txtsum.html (Version 2026.03.28), doi: .

@manual{martinschweinberger2026automated,
  author       = {Martin Schweinberger},
  title        = {Automated Text Summarisation with R},
  year         = {2026},
  note         = {https://ladal.edu.au/tutorials/txtsum/txtsum.html},
  organization = {The Language Technology and Data Analysis Laboratory (LADAL), The University of Queensland, Australia},
  edition      = {2026.03.28}
  doi      = {}
}

sessionInfo()

R version 4.4.2 (2024-10-31 ucrt)
Platform: x86_64-w64-mingw32/x64
Running under: Windows 11 x64 (build 26100)

Matrix products: default


locale:
[1] LC_COLLATE=English_United States.utf8 
[2] LC_CTYPE=English_United States.utf8   
[3] LC_MONETARY=English_United States.utf8
[4] LC_NUMERIC=C                          
[5] LC_TIME=English_United States.utf8    

time zone: Australia/Brisbane
tzcode source: internal

attached base packages:
[1] stats     graphics  grDevices datasets  utils     methods   base     

other attached packages:
 [1] here_1.0.1      igraph_2.1.4    quanteda_4.2.0  lubridate_1.9.4
 [5] forcats_1.0.0   stringr_1.5.1   dplyr_1.1.4     purrr_1.0.4    
 [9] readr_2.1.5     tidyr_1.3.1     tibble_3.2.1    ggplot2_3.5.1  
[13] tidyverse_2.0.0 textmineR_3.0.5 Matrix_1.7-2    lexRankr_0.5.2 
[17] rvest_1.0.4     xml2_1.3.6     

loaded via a namespace (and not attached):
 [1] generics_0.1.3     renv_1.1.1         stringi_1.8.4      lattice_0.22-6    
 [5] hms_1.1.3          digest_0.6.37      magrittr_2.0.3     evaluate_1.0.3    
 [9] grid_4.4.2         timechange_0.3.0   fastmap_1.2.0      rprojroot_2.0.4   
[13] jsonlite_1.9.0     RcppProgress_0.4.2 httr_1.4.7         stopwords_2.3     
[17] selectr_0.4-2      scales_1.3.0       codetools_0.2-20   klippy_0.0.0.9500 
[21] cli_3.6.4          rlang_1.1.5        munsell_0.5.1      withr_3.0.2       
[25] yaml_2.3.10        tools_4.4.2        tzdb_0.4.0         colorspace_2.1-1  
[29] fastmatch_1.1-6    curl_6.2.1         assertthat_0.2.1   vctrs_0.6.5       
[33] R6_2.6.1           lifecycle_1.0.4    htmlwidgets_1.6.4  pkgconfig_2.0.3   
[37] pillar_1.10.1      gtable_0.3.6       glue_1.8.0         Rcpp_1.0.14       
[41] xfun_0.51          tidyselect_1.2.1   rstudioapi_0.17.1  knitr_1.49        
[45] SnowballC_0.7.1    htmltools_0.5.8.1  rmarkdown_2.29     compiler_4.4.2    

AI Transparency Statement

This tutorial was revised and substantially expanded with the assistance of Claude (claude.ai), a large language model created by Anthropic. Claude was used to restructure the document into Quarto format, write the new “Creating Surveys in R” section covering shiny and surveydown, expand the design principles and scale types sections, add the mosaic plot and cumulative density plot examples, expand the ordinal regression section with the Brant test and improved visualisation, write the checkdown quiz questions and feedback strings, and revise all callout boxes, section overviews, and the summary section. All content was reviewed, edited, and approved by the author (Martin Schweinberger), who takes full responsibility for the accuracy and pedagogical appropriateness of the material.

Back to top

Back to LADAL home

---

References

Bender, Emily M, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021. “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?🦜.” In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 610–23.

Erkan, Günes, and Dragomir R Radev. 2004. “Lexrank: Graph-Based Lexical Centrality as Salience in Text Summarization.” Journal of Artificial Intelligence Research 22: 457–79.

Jones, K Sparck et al. 1999. “Automatic Summarizing: Factors and Directions.” In Advances in Automatic Text Summarization, 1–12. 1. MIT press Cambridge, MA.

Manning, Christopher D. 2008. Introduction to Information Retrieval. Syngress Publishing,.

Mihalcea, Rada, and Paul Tarau. 2004. “Textrank: Bringing Order into Text.” In Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing, 404–11.