💬 NLP

Computational Analysis of Semantic Connections Between Herman Melville Reading and Writing

This study employs computational semantic similarity analysis using BERTScore to investigate the influence of Herman Melville's reading on his writing, demonstrating that the method effectively identifies known literary connections and highlights new passages for qualitative examination.

Original authors: Nudrat Habib, Elisa Barney Smith, Steven Olsen Smith

Published 2026-03-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nudrat Habib, Elisa Barney Smith, Steven Olsen Smith

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery: Did Herman Melville, the famous author of Moby Dick, secretly copy ideas from books he read, or did he just happen to think of the same things?

For a long time, literary detectives (scholars) had to read thousands of pages of Melville's work and hundreds of books from his personal library, manually hunting for clues. It was like trying to find a specific needle in a haystack by looking at every single piece of hay with a magnifying glass.

This paper introduces a new, high-tech detective tool to help solve the case faster. Here is the story of how they did it, explained simply.

The Big Idea: A "Semantic" Magnifying Glass

The researchers wanted to see if Melville's writing "smelled" like the books he owned. But they didn't just look for exact word-for-word copies (like a plagiarism detector). They wanted to find ideas that were similar, even if Melville changed the words.

Think of it like this: If you read a recipe for "Chocolate Cake" and then write a story about "a delicious, sweet, brown dessert," a simple word-search might miss the connection. But a smart detective knows that "Chocolate Cake" and "delicious, sweet, brown dessert" are essentially the same idea.

The tool they used is called BERTScore. Imagine it as a super-smart robot that reads two texts and asks, "Do these two sentences mean the same thing, even if they use different words?"

How They Set Up the Experiment

The Suspects: They picked four famous pairs of books. On one side, they had a book Melville wrote (like Moby Dick). On the other, they had a book Melville definitely read (like a book about whales).
The "Ground Truth": Before they turned on the robot, human experts had already pointed out specific spots where Melville definitely borrowed ideas. These were the "known clues."
The Two Ways to Look: To make sure they didn't miss anything, they looked at the text in two different ways:
- The "Phrase" Lens (5-grams): They broke the text into tiny chunks of five words. This is like looking for specific phrases or catchy sayings Melville might have stolen.
- The "Sentence" Lens: They looked at whole sentences. This is like looking at the bigger picture of an entire thought.

What They Found

The robot did a great job, but it also taught them some interesting lessons about how writing works.

1. The Robot Found the Known Clues
When the experts said, "Hey, Melville copied this part of Typee from this book about the South Seas," the robot agreed. It gave them high scores, confirming that the tool works. It could see that even though Melville changed some words, the meaning was still a perfect match.

2. The Robot Found New Clues (The "Eureka" Moments)
The robot didn't just stop at the known clues. It scanned the whole text and found new places where Melville's writing looked suspiciously like his reading. These were spots the human experts hadn't noticed yet. This is like the detective finding a hidden fingerprint on a window that no one else saw.

3. The "One-to-Many" Puzzle
Here is where it got tricky. Sometimes, Melville would read one long, complicated sentence in a source book and turn it into three shorter sentences in his own book.

The Problem: When the robot compared the source sentence to just one of Melville's new sentences, the score was low. It was like comparing a whole pizza to just one slice; they don't look the same size, even though the slice came from the pizza.
The Lesson: The robot is great at spotting direct matches, but it needs human help to realize that sometimes an author spreads one big idea across many small sentences.

4. The "Generic" Trap (False Alarms)
The robot sometimes got excited about boring sentences. For example, if a source book said, "The man walked into the room," and Melville wrote, "The knight visited his workers," the robot gave them a high score.

Why? Because "walking," "visiting," "men," and "workers" are common words that appear everywhere. The robot thought, "Hey, these are similar!"
The Reality: A human knows these are just generic phrases that don't prove Melville copied the specific story. The robot needs a human to say, "Okay, that's just a coincidence, not a clue."

The Takeaway

This study is like building a new kind of metal detector for literature.

Before: Scholars had to walk the beach with their eyes, hoping to spot a coin.
Now: They have a metal detector that beeps when it finds buried treasure (similar ideas).

The detector isn't perfect—it sometimes beeps at a soda can (generic words) or misses a coin if it's buried under a pile of sand (complex rewriting). But it is incredibly fast and helps scholars find the real treasures much quicker.

In short: By using AI to compare what Melville read with what he wrote, the researchers proved that computers can help us understand how authors borrow, twist, and improve upon the stories of others. It doesn't replace the human expert; it just gives them a powerful flashlight to see in the dark.

1. Problem Statement

The paper addresses the challenge of identifying literary influence in historical texts. Traditional scholarship relies on manual close reading to trace how authors borrow, paraphrase, or adapt source materials. However, for authors like Herman Melville, who owned and read hundreds of books, the sheer volume of potential source texts makes manual verification prohibitively difficult.

The core research question is: Can semantic similarity analysis effectively identify traces of literary influence in Melville's writings when compared against the books he owned or consulted?

The study aims to move beyond simple lexical overlap (word-for-word matching) to detect paraphrastic and conceptual similarities, acknowledging that authors often rewrite source material, altering vocabulary and phrasing while retaining underlying meaning.

2. Methodology

The authors propose a computational framework using BERTScore to evaluate semantic similarity between Melville's works and his documented reading list.

Data Sources

Target Texts: Selected passages from Herman Melville's novels (e.g., Typee, Moby Dick, Israel Potter, Mardi).
Reference Texts: Corresponding passages from books known to be in Melville's library or referenced in his journals/letters (e.g., A Visit to the South Seas by C.S. Stewart, Life and Remarkable Adventures of Israel R. Potter by Henry Trumbull).
Ground Truth: Four specific text pairs were selected based on existing scholarship (Mathiessen, Olsen-Smith, Bezanson) where expert analysis had already confirmed a source relationship.

Preprocessing and Segmentation

To test different granularities of influence, the text was segmented using two strategies:

Non-overlapping 5-grams: Captures fine-grained lexical overlap and recurring phrases.
Sentence-level: Captures broader syntactic and contextual units, though resulting in uneven chunk lengths.

Similarity Metric: BERTScore

Instead of traditional metrics like BLEU or ROUGE (which rely on exact n-gram overlap), the study utilizes BERTScore.

Mechanism: It uses contextualized token embeddings from a pre-trained transformer model (Microsoft/deberta-xlarge-mnli) to calculate semantic similarity.
Outputs: Precision ( $p$ ), Recall ( $r$ ), and F1-score (harmonic mean of $p$ and $r$ ).
Interpretation: The authors do not use fixed thresholds to declare "plagiarism." Instead, they interpret relatively higher scores as signals of potential semantic alignment warranting qualitative expert review.

Algorithm

The process involves:

Segmenting Melville's text ( $M$ ) and Source text ( $R$ ) into chunks ( $X$ and $Y$ ).
Creating a matrix of all pairwise comparisons between chunks of $M$ and $R$ .
Computing BERTScore ( $p, r, f$ ) for every pair.
Aggregating mean scores and visualizing results via heatmaps to identify high-similarity clusters.

3. Key Contributions

Application of BERTScore to Literary Studies: The paper positions BERTScore as a novel tool for investigating authorial influence, shifting the focus from plagiarism detection to the scholarly discovery of source appropriation.
Feasibility Study: It demonstrates that semantic similarity measures can successfully highlight known instances of influence and uncover previously unmarked parallels.
Granularity Analysis: The study provides insights into how segmentation levels (5-gram vs. sentence) affect the detection of different types of influence (local phrasing vs. broad thematic concepts).
Framework for Scholars: It offers a computational workflow that supports, rather than replaces, human literary analysis by flagging candidate passages for deeper investigation.

4. Experimental Results

The study evaluated the model against four expert-identified cases and analyzed the broader output for new discoveries.

Validation of Known Influences

High Correlation: The model successfully detected the expert-identified similarities. For example, in the comparison of Typee and A Visit to the South Seas, 5-gram analysis yielded an F1-score of 0.77 for specific overlapping phrases ("a plurality of husbands"), and sentence-level analysis showed an F1 of 0.63.
Complex Influence Patterns:
- One-to-One: High similarity scores were observed when a single sentence in Melville's work mirrored a single sentence in the source.
- One-to-Many: When Melville expanded a single source idea across multiple sentences (e.g., in Moby Dick vs. The Natural History of Sperm Whale), the sentence-level similarity scores dropped (diluted signal), but the precision remained high, indicating that Melville's tokens strongly matched specific parts of the source.

Model-Discovered Parallels

New Findings: In the analysis of Israel Potter vs. Trumbull's source text, the model identified high-similarity pairs (e.g., F1 = 0.56) in passages not previously highlighted by scholars, confirming the model's ability to find latent connections.
False Positives & Limitations: The study noted that short, generic sentences (e.g., "About noon the knight visited his workmen") produced artificially high similarity scores due to broad semantic matches. This highlights a limitation where generic text can trigger false positives.

Granularity Comparison

5-grams: More effective for detecting localized lexical overlaps and specific phrasing borrowings.
Sentences: Better for capturing broader thematic relationships but prone to variability due to sentence length differences and the "dilution" effect when one sentence draws from multiple source sentences.

5. Significance and Future Work

Significance: The study validates that computational semantic analysis is a viable method for source studies in literary scholarship. It allows researchers to systematically scan large corpora to find "vestigial source text" that might otherwise remain hidden due to Melville's habit of rewriting and improving source material.
Limitations:
- Non-overlapping 5-grams: May miss similarities that cross chunk boundaries.
- Generic Text: High scores for short, generic sentences require filtering.
- Structural Mismatch: Sentence-level comparisons struggle when the influence structure is non-linear (e.g., one paragraph in the source becoming one sentence in the target).
Future Directions:
- Implementing overlapping n-grams for better coverage (despite higher computational cost).
- Developing weighting schemes based on sentence specificity to reduce false positives from generic text.
- Exploring paragraph-level segmentation to better capture complex, multi-sentence influence patterns.

In conclusion, the paper establishes a robust computational framework that bridges the gap between digital humanities and traditional literary criticism, offering a scalable method to trace the "invisible" threads of influence in an author's work.