Datasets:

Acture
/

MotifBench

@@ -1,98 +1,124 @@
 ---
 license: cc-by-nc-4.0
 language:
-- en
 tags:
-- long-context
-- llm-evaluation
-- graph-neural-networks
-- reasoning
 pretty_name: MotifBench
 size_categories:
-- 1K<n<10K
 configs:
-- config_name: macbeth
-  data_files: macbeth/macbeth.jsonl
-  default: true
 ---
-# MotifBench
-[![Critic’s Eye — EMNLP 2026 poster](assets/EMNLP%202026_Find-1836.png)](https://huggingface.co/datasets/Acture/MotifBench/resolve/main/assets/EMNLP%202026_Find-1836.pdf)
-[Open the full-resolution poster (PDF)](https://huggingface.co/datasets/Acture/MotifBench/resolve/main/assets/EMNLP%202026_Find-1836.pdf)
-**A benchmark for whether models can assemble meaning that no single passage
-states — not just retrieve it.**
-[![Code](https://img.shields.io/badge/code-critics--eye-blue)](https://github.com/Acture/critics-eye)
-[![License](https://img.shields.io/badge/license-CC%20BY--NC%204.0-lightgrey)](LICENSE)
-## Why this benchmark
-Long-context benchmarks test **locating**: the answer sits somewhere in the
-input, and the task is to find it. Needle-in-a-haystack, multi-hop QA,
-book-length summarization — in all of them, retrieval eventually wins.
-MotifBench tests something retrieval cannot do. A **motif** is a recurring
-salient thematic element a work returns to — blood in *Macbeth*, surfacing
-in lines that share no vocabulary:
-> "the multitudinous seas incarnadine" (2.2) · "Out, damned spot!" (5.1) ·
-> "all the perfumes of Arabia" (5.1)
-None of these lines contains the word *blood*. No line states the motif —
-it exists only across them. Recovering it means noticing that scattered,
-dissimilar passages belong together: **assembly, not search**.
-The labels are not ours to invent. An interpretive target has no annotator
-we could trust more than the alternative, so MotifBench draws its ground
-truth from published literary criticism (1765–2020) — critics who named
-these motifs, in print, long before any benchmark existed. A language model
-is used only to *locate* where a critic asserts a link and resolve the
-citation; nothing enters the benchmark on the model's word.
-Paper: *Critic's Eye: Review-Grounded Graph Reasoning for Long-Form Motif
-Discovery* (EMNLP 2026). Code: <https://github.com/Acture/critics-eye>.
-## Dataset structure (`macbeth/`)
-| file | contents |
-|---|---|
-| `macbeth.jsonl` | One row per (narrative node, motif) association: `node_type`, `node_id`, `act_id`, `scene_id`, `section_id`, `speaker`, `text`, `motif_id`, `motif_descriptor`. `node_type` is `line`, `scene`, or `stage_direction`; `node_id` for a line is the Folger Through-Line-Number (FTLN), e.g. `ftln-0001`. `text` holds the Folger line text for `line` rows and is `null` for `scene`/`stage_direction` rows. |
-| `criticism_sources.csv` | The 26 criticism works this benchmark draws on: author, year, title, venue, public-domain status, access URL/DOI. |
-### Example row
-```json
-{"node_type": "line", "node_id": "ftln-0004", "act_id": "act_1", "scene_id": "act_1_scene_1", "section_id": "sp-0003", "speaker": "WITCHES.2_Mac", "text": "When the battle’s lost and won.", "motif_id": "motif_23", "motif_descriptor": "lady macbeth's line \"wouldst not play false, / and yet wouldst wrongly win\""}
-```
-## Usage
 ```python
-import pandas as pd
 from huggingface_hub import hf_hub_download
-path = hf_hub_download(repo_id="Acture/MotifBench", filename="macbeth/macbeth.jsonl", repo_type="dataset")
-df = pd.read_json(path, lines=True)
-print(df[["node_id", "text", "motif_id", "motif_descriptor"]].head())
 ```
-## Text and copyright
-`text` on `line` rows is from the Folger Shakespeare Library's digital
-edition of *Macbeth* (Barbara A. Mowat and Paul Werstine, eds.; encoded by
-Michael Poston and Rebecca Niles; <https://www.folgerdigitaltexts.org>),
-licensed CC BY-NC 3.0.
-Of the 26 criticism works, public-domain ones (pre-1922) are pointed to in
-full (mostly via Project Gutenberg); copyrighted ones are referenced by
-metadata only — see `criticism_sources.csv`.
 ## Citation
-Provisional BibTeX; the final ACL Anthology entry is pending.
 ```bibtex
 @inproceedings{sun2026criticseye,
@@ -103,6 +129,4 @@ Provisional BibTeX; the final ACL Anthology entry is pending.
 }
 ```
-## License
-CC BY-NC 4.0 — see `LICENSE`.

 ---
 license: cc-by-nc-4.0
 language:
+  - en
 tags:
+  - long-context
+  - llm-evaluation
+  - graph-neural-networks
+  - reasoning
 pretty_name: MotifBench
 size_categories:
+  - n<1K
 configs:
+  - config_name: macbeth
+    data_files: macbeth/macbeth.jsonl
+    default: true
 ---
+<h1 align="center">Critic’s Eye</h1>
+<h3 align="center">Review-Grounded Graph Reasoning for Long-Form Motif Discovery</h3>
+<p align="center">Xinyu Sun · Zhangyue Yin · Xin Wang<br/>Fudan University · EMNLP 2026</p>
+<p align="center">
+  <a href="https://github.com/Acture/critics-eye">Code</a> ·
+  <a href="#use-motifbench">Dataset</a> ·
+  <a href="https://huggingface.co/datasets/Acture/MotifBench/resolve/main/assets/EMNLP%202026_Find-1836.pdf">Poster</a> ·
+  <a href="#citation">Citation</a>
+</p>
+## Can LLMs connect distant descriptions and make sense of them?
+A **motif** is an element that recurs with variation across scenes. In *Macbeth*, an ocean and a perfume bottle evoke the same problem: guilt that cannot be erased. Recognizing the connection means bringing distant passages together despite their different imagery.
+[![The poster’s ocean-and-perfume example: different acts, speakers and imagery return to failed cleansing.](assets/motif-example.png)](https://huggingface.co/datasets/Acture/MotifBench/resolve/main/assets/motif-example.png)
+*An illustrative reading, not a scored model response.*
+**MotifBench** turns this challenge into a task: given a line, rank a fixed inventory of motifs. **Critic’s Eye** learns from the play’s narrative structure to recover associations recorded in literary criticism.
+## Let critics supply the connections
+The paper draws on **26 published works of criticism**. Their cited passages yield **351 line–motif membership links** over **283 query lines**, with **281 candidate motifs**. For a query line, Hit@5 asks whether a motif identified by critics appears among the first five choices.
+Here is one recorded connection: Susan Sachon’s discussion points to *Macbeth* 4.1.138, associated with **Blood-Stained Ghost**. Macbeth speaks the line; Banquo is the person it describes.
+[![Susan Sachon’s published criticism points to “For the blood-boltered Banquo smiles upon me”, associated with Blood-Stained Ghost.](assets/critic-evidence.png)](https://huggingface.co/datasets/Acture/MotifBench/resolve/main/assets/critic-evidence.png)
+Critics record interpretations, not every possible reading. **Missing does not mean wrong.**
+## Connect lines through the play’s structure
+Lines belong to speaking turns, follow other lines, and are spoken by characters. These relations connect the text across the play. The diagram keeps the words and speakers visible so that the structure can be read directly.
+[![Three scene groups from the poster: Macbeth’s bloodied hands, Banquo’s ghost, and Lady Macbeth’s sleepwalking. Lines are grouped by speaking turn and linked to speakers.](assets/narrative-structure.png)](https://huggingface.co/datasets/Acture/MotifBench/resolve/main/assets/narrative-structure.png)
+Critic’s Eye first learns structural relations, then learns recorded line–motif links, and finally scores the candidate motifs for a query line.
+[![The poster’s three learning stages: learn narrative structure, learn recorded motif links, then rank motifs.](assets/learning-motif-links.png)](https://huggingface.co/datasets/Acture/MotifBench/resolve/main/assets/learning-motif-links.png)
+## What did we find?
+**LLMs struggle to recover recorded motifs.** The best tested LLM configuration reaches **22.6% Hit@5**; adding three retrieved lines helps some models and hurts others. In a separate motif-supervised setting, the graph configurations show that narrative structure carries useful signal.
+[![Poster results. Left: LLM motif recovery with and without retrieval. Right: separate supervised graph configurations using speaker, speaking-turn and next-line relations.](assets/results.png)](https://huggingface.co/datasets/Acture/MotifBench/resolve/main/assets/results.png)
+The two panels use different training and scoring settings. The LLM evaluation supplies a target line and the motif inventory, optionally with three retrieved lines, and scores one recorded motif per query. The graph panel reports transductive, frozen-encoder configurations trained with critic labels and credits any recorded motif. These are separate experimental findings, not a matched comparison of adding a graph to an LLM.
+## Use MotifBench
+The `macbeth` configuration contains a flattened export of narrative-node–motif associations, with Folger line text and structural identifiers. Use an account with access while the repository is private.
 ```python
+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"
+]
+print(line_rows[0])
 ```
+Install `huggingface_hub` and authenticate with `hf auth login` if needed. Pin `revision` to a Hub commit when recording experiments.
+<details>
+<summary><b>Files, fields and data version</b></summary>
+| File | Contents |
+| --- | --- |
+| [macbeth/macbeth.jsonl](macbeth/macbeth.jsonl) | 645 associations: 585 line, 44 scene, 7 stage-direction and 9 theme rows |
+| [macbeth/criticism_sources.csv](macbeth/criticism_sources.csv) | Bibliographic metadata and available access links for 26 criticism sources |
+The JSONL contains 314 distinct line nodes and 277 distinct target motif IDs. Its fields are `node_type`, `node_id`, `act_id`, `scene_id`, `section_id`, `speaker`, `text`, `motif_id` and `motif_descriptor`. All values are strings or null. Line identifiers use Folger Through-Line Numbers, such as `ftln-1616`; non-line rows have null text and structural-context fields.
+The data files are unchanged from revision `fa060c5ad43ea3fdccb7cb6ad02a7bc85fa20c70`. Their population differs from the frozen paper evaluation described above. This flattened export does not include relation types, the complete candidate inventory, narrative graph, evaluation folds or per-association source citations. It should not be treated as the complete reproduction package for the reported scores.
+Multiple associations can share a query line. Group by `node_id` before defining queries, and keep a line’s associations together when constructing a split. The single Hub `train` split is a loading convention, not the paper’s training partition.
+</details>
+<details>
+<summary><b>Text, illustration sources and license</b></summary>
+The dataset is released under [CC BY-NC 4.0](LICENSE). 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](https://www.folger.edu/explore/shakespeares-works/macbeth/read/).
+The source bibliography provides metadata and available access links. Full criticism texts and PDFs are not part of the data export. The criticism figure reproduces the attributed excerpt already shown in the conference poster.
+Figures on this page are individually rendered from the authors’ final poster: the explanatory illustrations, source example, narrative diagram, learning stages and results retain their original roles. Illustrative passages and ranking placeholders are not model predictions. See [figure provenance](assets/README.md).
+</details>
 ## Citation
+Provisional BibTeX; replace with the final ACL Anthology entry when available.
 ```bibtex
 @inproceedings{sun2026criticseye,
 }
 ```
+Contact: [sunxy23@m.fudan.edu.cn](mailto:sunxy23@m.fudan.edu.cn).

assets/README.md ADDED Viewed

	@@ -0,0 +1,17 @@

+# Figure provenance
+These panels are rendered individually from the final V22 HTML poster in the authors’ [paper repository](https://github.com/Acture/Critic-s-Eye-Paper/tree/d2250009573ba563c43a7170f423746537ef59eb/acl/poster), using the existing fonts, illustrations, diagrams and charts. The README assembles the separate panels in reading order; it does not embed the full poster as its main content.
+| File | Poster elements |
+| --- | --- |
+| `motif-example.png` | `.story-gallery`: ocean/perfume example and reasoning hurdles |
+| `critic-evidence.png` | `.record-gallery`: attributed Sachon source example, reflowed horizontally |
+| `narrative-structure.png` | `.structure-figure`: scene groups, speaking turns, speakers and line relations |
+| `learning-motif-links.png` | `.learning-flow`: structural learning, recorded motif links and ranking |
+| `results.png` | `.llm-result` and `.graph-result`: original charts and protocol captions, placed side by side |
+Only presentation geometry and figure-level spacing are adapted for the README. All results and chart marks come from the poster. The LLM and supervised graph panels describe separate experimental settings. The ocean/perfume reading, character illustrations and ranking placeholders are illustrative, not scored predictions.
+The fictional character and object artwork was created with OpenAI image generation for the poster. Native SVG diagrams and charts are source-authored. The source page and excerpt in the criticism panel come from Susan Sachon (2020), *Shakespeare, Objects and Phenomenology*, printed page 86, and retain the poster’s attribution. They are source evidence, not generated text.
+The existing full poster PDF remains available through the README’s Poster link.