Skip to content

Latest commit

 

History

11 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

READI: Read the Room, Read the Image

ACL Anthology arXiv

This is the official companion-code repository for READI, a bilingual multimodal benchmark for evaluating whether vision-language models can infer the intended function of indirect directive speech acts from an utterance and its visual sociopragmatic context.

Resource Status
Paper ACL Anthology · arXiv
Code Evaluation and release utilities available in this repository
Data Available for research use upon request
License Paper, code, and data follow their respective release terms

Benchmark

Each READI item contains:

  • an image encoding sociopragmatic context;
  • an indirect directive utterance;
  • a question about the utterance's pragmatic function;
  • four mutually exclusive intent choices; and
  • one gold answer.

The benchmark contains 102 multimodal items across two independently developed subsets:

Subset Language Items
KRISA Korean 57
ENGISA English 45

KRISA and ENGISA are not translation pairs. Each was constructed and validated through a language-specific process.

Indirectness levels

READI follows a CCSARP-based graded indirectness design.

Level Category Description
1 CID Conventionally indirect directive
2 NCID - strong hint Non-conventional directive with a lexical hint
3 NCID - mild/no hint Non-conventional directive requiring stronger contextual inference

Evaluation

READI formulates visual pragmatic understanding as a four-choice question-answering task. The primary metric is four-choice accuracy.

Predictions must be provided as JSONL, with one object per item:

{"item_id": "readi_ko_001", "predicted_choice": 2}

predicted_choice must be an integer from 1 to 4.

Run the evaluator with:

python scripts/evaluate.py \
  --gold data/release/readi.jsonl \
  --predictions predictions/model.jsonl

Missing, duplicate, unknown, and malformed predictions are reported and remain in the evaluation denominator. See the prediction format for details.

Release utilities

The repository includes deterministic utilities for converting authorized source workbooks, extracting embedded images, validating annotations and file hashes, and scoring model predictions.

python -m pip install -r requirements.txt
python -m unittest discover -s tests -v

Repository structure

├── data/                 # Data-access and format notes
├── docs/                 # Data card and prediction specification
├── scripts/              # Build, validation, and evaluation utilities
├── tests/                # Regression tests
└── requirements.txt

Usage terms

Paper, code, and data usage terms follow their respective official archival releases. This repository does not grant additional rights beyond those terms.

Data access

The READI dataset is available for research use upon request. To request access, email Park at cosmic4intelligence@gmail.com with your name, affiliation, and a brief description of the intended research use.

Source workbooks and manuscript files are not distributed through this repository. Access to the released dataset is subject to its accompanying usage terms.

The normalized public format is documented in the data card.

About

Official companion code for READI, a bilingual multimodal benchmark for visually grounded indirect speech act understanding (ACL 2026).

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages