Acture/MotifBench

EMNLP 2026

Motif recovery from published literary criticism

Dataset  ·  Usage  ·  Code  ·  Cite Us

Can LLMs connect scattered elements to grasp a story’s underlying ideas? A motif recurs across a work, even as its words, imagery and speakers change. Finding it means connecting distant passages and recognizing what they share.

MotifBench turns these recorded interpretations into a test of motif recovery. The task: given a line, rank the motifs it belongs to. Its labels come from published literary criticism.

What connects an ocean to a perfume bottle?

Act II · Macbeth 2.2.78	Act V · Lady Macbeth 5.1.54–55
“Will all great Neptune’s ocean wash this blood”	“the perfumes of Arabia will not sweeten this little hand”

Different elements → failed cleansing → guilt that cannot be erased.
An illustrative reading: neither excerpt names guilt.

Where LLMs fall short

LLMs struggle: the best tested configuration reaches 22.6% Hit@5. Adding three retrieved lines helps some models and hurts others.

How is Hit@5 measured?

Each example gives a model a line and 281 candidate motifs. The model ranks the motifs; a hit means the recorded target motif appears in its first five choices. The chart reports results over 283 query lines. The retrieval variant supplies three additional similar lines.

From criticism to a data record

Published criticism connects specific passages to recurring motifs. The dataset’s source bibliography lists 26 works of criticism. One example comes from Susan Sachon’s Shakespeare, Objects and Phenomenology (2020), p. 86:

The cited line is spoken by Macbeth:

“For the blood-boltered Banquo smiles upon me” — Macbeth 4.1.138

The released data stores the connection as ftln-1616 → motif_74, with the critic-derived descriptor:

the ghost’s appearance described as ‘blood‑baltered’ (4.1.138)

View the actual JSONL row

{
  "node_type": "line",
  "node_id": "ftln-1616",
  "act_id": "act_4",
  "scene_id": "act_4_scene_1",
  "section_id": "sp-1605",
  "speaker": "Macbeth_Mac",
  "text": "For the blood-boltered Banquo smiles upon me",
  "motif_id": "motif_74",
  "motif_descriptor": "the ghost’s appearance described as ‘blood‑baltered’ (4.1.138)"
}

Critics record interpretations rather than every possible reading: an unrecorded link is not necessarily wrong.

What's in the dataset?

The current Macbeth export contains 645 node–motif associations, spanning 314 distinct lines and 277 target motif IDs across all node types.

Node type	Association rows	Distinct source nodes
Line	585	314
Scene	44	14
Stage direction	7	5
Theme	9	9

One row is one association; a line can appear in several rows with different motifs.

Fields	What they contain
node_type, node_id	Narrative-node type and identifier
act_id, scene_id, section_id	Location and speaking-turn identifiers for line rows
speaker, text	Speaker identifier and Folger line text
motif_id, motif_descriptor	Target motif identifier and its descriptive text

All values are strings or null. Non-line rows have null text and structural-context fields. The companion source bibliography gives bibliographic metadata and available access links for the 26 criticism sources.

Use MotifBench

Each row associates a narrative node with a motif. Line rows include the play text, speaker and structural identifiers.

import json

from huggingface_hub import hf_hub_download

path: str = hf_hub_download(
    repo_id="Acture/MotifBench",
    filename="macbeth/macbeth.jsonl",
    repo_type="dataset",
)
with open(path, encoding="utf-8") as stream:
    rows: list[dict[str, str | None]] = [json.loads(line) for line in stream]

line_rows: list[dict[str, str | None]] = [
    row for row in rows if row["node_type"] == "line"
]
example: dict[str, str | None] = next(
    row for row in line_rows if row["node_id"] == "ftln-1616"
)
print(example)

Requires huggingface_hub. While this repository is private, authenticate with hf auth login using an account with access.

Files and use notes

File	Contents
macbeth/macbeth.jsonl	645 associations: 585 line, 44 scene, 7 stage-direction and 9 theme rows
macbeth/criticism_sources.csv	Bibliographic metadata and available access links for 26 criticism sources

Line identifiers use Folger Through-Line Numbers, such as ftln-1616. Text is supplied for the 314 line nodes represented in the export.

Version: the files are unchanged from revision fa060c5ad43ea3fdccb7cb6ad02a7bc85fa20c70. This Hub release is a flattened association table; the published LLM scores use the paper’s 283-line evaluation set.

For experiments on this export, group associations by node_id and keep each line’s associations together when defining your evaluation split. The Hub loader places all rows in a single train split. Pin revision to a Hub commit when recording experiments.

Sources and license

The dataset is released under CC BY-NC 4.0. Line text comes from the Folger Shakespeare Library’s digital edition of Macbeth, edited by Barbara A. Mowat and Paul Werstine, with encoding by Michael Poston and Rebecca Niles, licensed CC BY-NC 3.0. See the Folger edition.

The source bibliography provides metadata and available access links. Full criticism texts and PDFs are not part of the data export. The attributed Sachon excerpt is the same source excerpt used in the conference poster. The illustration, logos and vector chart are independent assets; explanatory text and data examples are native README content. See asset provenance.

Cite Us

MotifBench accompanies Critic’s Eye: Review-Grounded Graph Reasoning for Long-Form Motif Discovery (EMNLP 2026).

@inproceedings{sun2026criticseye,
  title     = {Critic's Eye: Review-Grounded Graph Reasoning for Long-Form Motif Discovery},
  author    = {Sun, Xinyu and Yin, Zhangyue and Wang, Xin},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026},
  year      = {2026}
}

Provisional entry; the final ACL Anthology record will supply publication details.

Contact: sunxy23@m.fudan.edu.cn.