Henry NevilleReading edition · 0.95.196
Book PDF
Chapter 14

The Pattern in Two-Word Sequences

Shakespeare’s plays occupy seven of the top ten places in the all-word bigram comparison with Neville’s writing, although they make up only 36 of the 236 plays tested. A bigram is a two-word sequence, such as “I will” or “write to”. Counting thousands of these combinations tests whether Neville’s habitual language resembles Shakespeare’s plays more closely than other contemporary drama. The three methods below compare all-word bigrams, pairs formed after filtering for frequent words, and samples of similar length.

What was compared

The dramatic texts come from Pervez Rizvi’s database of 527 early modern plays. The comparison here is limited to 236 dramatic texts dated 1590–1615: the 36 plays of the First Folio and 200 texts by other writers. Pericles, Edward III and The Two Noble Kinsmen are excluded from both groups. Collaborative First Folio plays, including Henry VIII, count once as complete plays in the group statistics. The Shakespeare section of Henry VIII is also shown separately in the tables, using the database’s division between Shakespeare and Fletcher.1

On Neville’s side, the study used 165 English letter records and three other documents, comprising 137,558 words after processing.1 The question is how each play resembles Neville’s collected writing, not which play resembles a single selected letter.

The comparison works with base forms of words, so that forms such as writes and writing can be counted under write. It therefore examines word use rather than the appearance of particular historical spellings. Contractions are expanded consistently before the same word mapping is applied to Neville and the plays: “on’t”, for example, becomes “on it”, whether an edition originally prints it as one token or two. A separate comparison without lemmatization also retains the overall pattern.1 A pair is never made by joining the last word of one Neville document to the first word of another.

What a two-word comparison measures

In “I will write to you,” the consecutive bigrams are “I will,” “will write,” “write to” and “to you.” A passage contributes many such overlapping pairs. Counting them records not just which words a writer uses, but which words repeatedly occur together.

Each text receives a frequency profile: some pairs occur often, others occasionally, and many not at all. The comparison asks how closely a play’s balance of those frequencies resembles Neville’s. Cosine similarity compares the relative pattern of pair frequencies, rather than raw totals. This removes the direct effect of multiplying every count by the same amount, but longer texts can still differ because they contain more varied and more stable samples of language. Its score runs from zero for no shared pairs towards one for increasingly similar profiles; it is not a percentage of identical wording.

This differs from selecting striking verbal parallels. A rare expression can be illuminating when read in context, but these tests also count ordinary combinations. A high score can arise from many small similarities in usage without any one phrase being remarkable. The three methods vary the vocabulary and the amount of text compared, allowing us to see which results depend on those choices.

How to read the percentages

Each play first receives a similarity score. The group comparison then pairs every Shakespeare play with every play by another writer and asks which has the higher score. With 36 Shakespeare plays and 200 others, there are 7,200 pairings. In the all-word test, the Shakespeare play scores higher in 5,756, or 79.9 per cent, of them.

Thus 79.9 per cent describes how often a Shakespeare play outranks another play in its similarity to Neville. It does not mean that 79.9 per cent of their words match, or that there is an 79.9 per cent probability of common authorship. A result of 50 per cent would mean that the two groups win equally often. The graph shows that Shakespeare’s advantage remains substantial under all three methods, although it becomes smaller when the comparison is restricted to samples of common words.

Shakespeare plays score higher in 79.9, 76.5 and 72.8 per cent of cross-group pairings. The dashed line marks equal ordering at 50 per cent. The first two comparisons contain 7,200 pairings each; the windowed comparison contains 6,660. These pairings reuse the same plays and are not independent observations. See note 1.
Shakespeare plays score higher in 79.9, 76.5 and 72.8 per cent of cross-group pairings. The dashed line marks equal ordering at 50 per cent. The first two comparisons contain 7,200 pairings each; the windowed comparison contains 6,660. These pairings reuse the same plays and are not independent observations. See note 1.

The tables rank twenty complete plays and insert Henry VIII (Shakespeare section), unnumbered, at its score position. Bold identifies First Folio plays and that section. All group statistics count complete plays only, so Henry VIII counts once.

1. All-word bigrams

The first method uses all eligible word pairs, without limiting the vocabulary to common words. It can therefore reflect both habitual phrasing and subject matter: correspondence about government, for example, may share vocabulary with a history play.

Henry VIII ranks first among complete plays, The Winter’s Tale second and Cymbeline third. Shakespeare plays occupy eleven of the top twenty complete-play places, although they make up only about 15.3 per cent of the 236-text comparison. The 79.9 per cent pairing result shows that the advantage extends beyond the plays at the very top.

Rank Play
1 Henry VIII (full play)
2 The Winter’s Tale
3 Cymbeline
Henry VIII (Shakespeare section)
4 All’s Well That Ends Well
5 Henry IV, Part 2
6 The Royal King and the Loyal Subject
7 The Tragedy of Charles Duke of Byron
8 Henry V
9 Cynthia’s Revels
10 Coriolanus
11 Every Man Out of His Humour
12 Bartholomew Fair
13 Measure for Measure
14 The Staple of News
15 Love’s Labor’s Lost
16 2 Edward the Fourth
17 The Widow’s Tears
18 Hamlet
19 King Lear
20 1 Edward the Fourth

2. Frequent-word bigrams

The second method concentrates on the 200 most frequent base words in the dramatic collection’s plays dated 1580–1620. The list is selected from the plays, rather than from expressions chosen because they favour Neville. It includes small grammatical words such as and, of and to, but also common words with substantial meaning, including king, love and honour.

Words outside that list are removed before the pairs are formed. This matters: two words counted as a pair may originally have had other words between them. The method tests sequences in the remaining common-word pattern, not necessarily phrases that a reader would find printed side by side. It reduces the contribution of unusual vocabulary while retaining much of the ordinary language through which writers construct their sentences.

For a simple illustration, take the invented sentence “I humbly ask you.” The study’s 200-word list includes I and you, but neither humbly nor ask:

Stage Words or pairs
Invented example I humbly ask you
Ordinary adjacent pairs I humbly; humbly ask; ask you
After the 200-word filter I you (humbly and ask removed)
Pair now counted I + you

The filtered pair “I you” records the sequence left after removal; it does not claim that those words stood together in the original sentence. The same operation is applied to Neville and the plays.

Henry VIII ranks first among complete plays, Chapman’s The Tragedy of Charles Duke of Byron second and Henry V third. Shakespeare plays occupy thirteen of the top twenty complete-play places and score higher in 76.5 per cent of the 7,200 pairings. Their advantage therefore survives when much of the less frequent vocabulary is removed.

Rank Play
1 Henry VIII (full play)
Henry VIII (Shakespeare section)
2 The Tragedy of Charles Duke of Byron
3 Henry V
4 The Winter’s Tale
5 Henry IV, Part 2
6 The Conspiracy of Charles Duke of Byron
7 Cymbeline
8 The Family of Love
9 Henry VI, Part 2
10 King John
11 Richard II
12 Sejanus His Fall
13 Cynthia’s Revels
14 Philotas
15 Henry VI, Part 1
16 Macbeth
17 Coriolanus
18 Richard III
19 The Widow’s Tears
20 Love’s Labor’s Lost

3. Comparing passages of similar length

The third method uses the same common-word list but compares passages rather than treating each full text as one sample. A long collection of letters offers a more extensive picture of usage than a short play. Sampling reduces that imbalance and asks whether the affinity remains when the amount of language being compared is more nearly alike.

The study takes fifty samples of approximately 10,000 retained words from Neville’s pooled writing and uses their average pattern as the comparison profile. For each sufficiently long play, it takes fifty passages of the same size and averages their similarity scores. Plays with between 5,000 and 10,000 retained words contribute their whole text once; shorter plays are excluded.

The sample size refers to words remaining after the vocabulary filter, not words on the original page. Neville’s samples contain slightly fewer than 10,000 actual words because document-boundary markers occupy some sampled positions. The passages can overlap; they are repeated samples of the same works, not new independent texts.

This leaves 221 plays: 36 First Folio plays and 185 others. Henry VIII ranks first among complete plays, The Tragedy of Charles Duke of Byron second and Henry V fifth. The separately scored Shakespeare section of Henry VIII comes above the full play in this method. Shakespeare takes nine of the top twenty complete-play places and scores higher in 72.8 per cent of the 6,660 cross-group pairings. Sampling reduces the advantage but does not remove it.

Rank Play
Henry VIII (Shakespeare section)
1 Henry VIII (full play)
2 The Tragedy of Charles Duke of Byron
3 The Conspiracy of Charles Duke of Byron
4 The Winter’s Tale
5 Henry V
6 Philotas
7 Macbeth
8 Richard II
9 The Family of Love
10 The Atheist’s Tragedy, or The Honest Man’s Revenge
11 King John
12 1 The Troublesome Reign of King John
13 Henry IV, Part 2
14 Henry VI, Part 1
15 The Hector of Germany, or The Palsgrave, Prime Elector
16 Locrine
17 Sejanus His Fall
18 Cymbeline
19 Summer’s Last Will and Testament
20 The Widow’s Tears

What the pattern shows

The second graph displays every eligible complete-text score. Filled dots represent the First Folio group and open dots the comparison group; a square identifies Henry VIII (full play) and a diamond its Shakespeare section. The section is supplementary and is excluded from the group means. Shakespeare scores lie further to the right on average in every panel, but the groups overlap. Chapman’s two Byron plays, for example, rank prominently in the common-word comparisons.

Every eligible complete-text record is shown: filled dots represent the 36 First Folio plays; open dots represent other writers; dark strokes mark group means. A square marks Henry VIII (full play), already included in the First Folio group, and a diamond marks Henry VIII (Shakespeare section), excluded from group means. Vertical spreading has no numerical meaning. Compare scores within each panel. See note 1.
Every eligible complete-text record is shown: filled dots represent the 36 First Folio plays; open dots represent other writers; dark strokes mark group means. A square marks Henry VIII (full play), already included in the First Folio group, and a diamond marks Henry VIII (Shakespeare section), excluded from group means. Vertical spreading has no numerical meaning. Compare scores within each panel. See note 1.

A further test asks how unusual the Shakespeare group’s high average position is. For each method, the computer selected 20,000 random groups of 36 plays from that comparison’s pool. None had an average rank as high as the actual Shakespeare group. Allowing for the finite number of random draws gives a p value of about 0.00005; an adjustment for conducting three tests gives about 0.00015, or 0.015 per cent.1

The rank concentration is statistically significant under this test, which compares Shakespeare’s group with randomly selected groups of plays. It does not control for date, length, genre or multiple plays by the same author—all possible influences on language. The adjustment for three tests addresses a separate issue. Chapter 12 compares Neville with other diplomatic writers using a different method. None of these probabilities is the probability that Neville wrote Shakespeare.

The three methods also share the same source material and much of the same vocabulary. Their agreement is useful evidence that the affinity survives different ways of making the comparison, but it is not three independent discoveries whose probabilities can be multiplied. The finding is a sustained resemblance in patterns of two-word usage. Establishing how distinctive that resemblance is to Neville requires comparisons with other writers’ prose as well as with other dramatists’ plays.

The Declaration, Case and Advice: a closer look at Hamlet

A separate comparison asks what happens when the focus narrows from Neville’s collected writing to particular documents. It tests the Declaration alone, then the Declaration together with the Case, and finally those two defences together with the later Advice. Here the comparison pool contains 331 plays dated 1590–1630. Its wider date range means its ranks should be read within that pool, rather than as positions in the 236-play tables above.2

Using the frequencies of all two-word sequences, Hamlet ranks first in all three comparisons. Adding the Case and Advice therefore preserves Hamlet’s leading position while broadening the Neville material under examination. Across these three document combinations, Shakespeare plays win 79.3, 79.9 and 80.2 per cent of pairings with other plays. With all three documents combined, eight Shakespeare plays appear in the leading twenty. Without lemmatization, Hamlet ranks second, third and second, while the group advantage remains.

This result depends on the kind of resemblance measured. Counting only distinct pairs found in no more than five plays puts Hamlet below first place under the shared lemma map. Without lemmatization, it leads the first two document comparisons. Its position therefore depends on which verbal features are counted. The specific passages examined in the Hamlet chapter provide the context that a numerical ranking alone cannot supply.

Comparisons with other letter-writers help define the limit of the result. In equal-sized correspondence samples processed consistently, John Chamberlain’s letters also show substantial affinity with Shakespeare: their median cross-group pairing result is 75.1 per cent, against 74.5 per cent for Neville’s letter samples. Without lemmatization, the figures are 75.2 and 75.6 per cent respectively.

Neville scores higher on a separate measure, average precision, which gives more credit when Shakespeare plays concentrate near the top of the ranking. Neither result makes general Shakespeare affinity unique to Neville; these letter samples also do not constitute a matched test of the Declaration, Case and Advice individually.2

Shakespeare plays are concentrated near the top when compared with Neville’s writing by the frequency of two-word sequences; several leading plays recur across the three methods, and Hamlet leads the focused document comparisons. These are suggestive patterns of linguistic affinity, to be considered alongside the documentary and verbal evidence, rather than definitive authorship identifications.

Notes and sources


  1. Controlled normalization rerun, 16 November 2026. The exact archived October input first reproduced all principal scores and sampled-score standard deviations; the corrected run uses the current canonical corpus: 165 English letter records and three other documents, totalling 168 analytical records and 137,558 normalized words. One duplicate witness of the 4 June 1606 letter was removed; enclosed and quoted material within English-labelled records remains. French records were excluded. The earlier descriptive census in Chapter 11 uses different counting rules. The principal comparison retains 36 First Folio plays and 200 other dramatic texts assigned dates of 1590–1615 in the database; Pericles, Edward III and The Two Noble Kinsmen are excluded from scoring and from the 1580–1620 vocabulary pool. Contractions are expanded before one surface-to-lemma mapping, fitted without author labels, is applied to both corpora. This ensures consistent processing, not perfect contextual interpretation. Ambiguous auxiliary contractions were tested under alternative is/would and has/had readings and by inserting boundaries. The principal branch gives 5,756/7,200, 5,509/7,200 and 4,850/6,660 winning pairs (79.9%, 76.5%, 72.8%); the alternative ambiguity treatments give 79.8–80.1%, 76.5–77.0% and 72.8–73.2%. With spelling normalized but no lemmatization, the results are 79.7%, 74.3% and 71.1%. These are sensitivity checks on the same texts, not independent discoveries. The top-200 lemma vocabulary is unchanged by the correction. The tables comprise twenty complete plays plus Henry VIII’s Shakespeare section; only complete plays enter group statistics. After correction, the assigned Shakespeare and Fletcher sections contain 10,076 and 14,867 words, or 6,917 and 10,552 after frequent-word filtering. Their assigned passages are concatenated, following the original implementation. Similarity is cosine similarity of normalized pair counts, with no pairs crossing Neville document boundaries. The windowed method averages fifty 10,000-position samples for longer texts; texts with 5,000–10,000 retained words are scored whole. Neville’s profile averages fifty samples, with boundary markers occupying some positions. Thus the section and full play are not strictly equal-length samples, and overlapping samples are not independent observations. Three additional sampling seeds, keyed separately by play, give 72.9–73.3% for the corrected common-word method and 71.4–71.8% without lemmatization. A dialogue-only sensitivity also retains the group advantage. Independent arithmetic checks reproduce the corrected score summaries and complete nonwindowed scores. Each corrected random-group test draws 20,000 groups of 36 complete texts: zero equalled or bettered the observed mean rank, giving add-one p = 1/20,001 and Holm-adjusted p = .000150 across three tests. The test assumes exchangeable labels and does not adjust for date, length, genre or shared authorship. Text sources and the database are documented by Pervez Rizvi, “Collocations and N-grams”. These analyses are this book’s research, not Rizvi’s results; study data and scripts have not yet been publicly deposited. Database dates and text records are used as supplied, without independently redating the plays. The collection includes masques, entertainments and additions as well as full-length plays.↩︎

  2. The separate focused comparison retains sentence- and document-bounded pairs in 331 plays assigned dates of 1590–1630, with 37 in its Shakespeare group, including Pericles. Its 16 November rerun applies the same spelling normalization, contraction expansion and optional lemma map to current Neville texts and archived comparison-letter surfaces. Full bigram-frequency cosine puts Hamlet first for Declaration, Declaration + Case, and Declaration + Case + Advice; their winning fractions are .792793, .798676 and .802078. The last combination places eight Shakespeare plays in the top twenty. Without lemmatization, the fractions are .768891, .778176 and .780382, and Hamlet ranks second, third and second. Counting distinct pairs occurring in at most five candidate plays gives Hamlet tied positions 9–18, 6–12 and 26–36 under shared lemmas, or positions 1, 1 and 5–8 without them. Equal-size correspondence controls use 100 samples per writer with the original Jonson-derived segment lengths and separate fixed random seeds. Median winning fractions for Neville and Chamberlain are .745 and .751; median average precision is .280 and .228. Without lemmatization these become .756 and .752, and .267 and .230. Cornwallis, Winwood, Edmondes, Wotton and Donne were also included; all results were retained. These overlapping samples are descriptive sensitivities, not independent replications or matched controls for each document combination. The play texts come from Rizvi’s collection, cited in note 1. The focused pool requires at least 4,000 dialogue words and retains its earlier membership rather than adopting the principal study’s First Folio-only selection. All 1,986 focused scores and 1,406 focused/sample winning fractions passed independent arithmetic checks. These normalization tests do not constitute fresh verification of the historical transcriptions.↩︎

Note

Search the book

Inspect the document