This repository contains the code and dataset for the paper "Perceptions to Beliefs: Exploring Precursory Inferences for Theory of Mind in Large Language Models", presented at EMNLP 2024.

---

Table of Contents

• Requirements
• Perception-Augmented Theory of Mind Datasets
  • Percept-ToMi
  • Percept-FANToM
• Evaluating Language Models on Perception-Augmented ToM Datasets
  • Experiments on the Percept-ToMi
  • Experiments on the Percept-FANToM

---

Requirements

This project has been tested with Python 3.10.12. To install the required Python packages, run:

pip install -r requirements.txt

Perception-Augmented Theory of Mind Datasets

Percept-ToMi

File: dataset/Percept-ToMi.csv

Column	Description
index	Index of each story
qTypeRaw	Type of the question paired with the story (from original annotation of ToMi dataset)
qType	Type of the question paired with the story (with '0' and '1' removed from qTypeRaw)
story	Text of the story
question	Text of the question
char1	The character whose first-order belief the question is asking about
char2	The character whose second-order belief the question is asking about
perceivers	Perceivers of each scene in the story
story_with_perceivers	Scene descriptions in the story paired with corresponding perceivers. JSON-formatted to provide ground truth perception information to the language model.
persp_ctx	Perspective context corresponding to the story and the question
cands	Candidate answer choices
answer	Answer to the question

---

Percept-FANToM

1. File: dataset/Percept_FANToM/Percept-FANToM.csv

Column	Description
part_id	ID of each conversation part in FANToM
utterance	Single utterance in a conversation part
utterance_symbol_removed	Utterance with symbols removed
perceivers	Perceivers of the utterance

2. File: dataset/Percept_FANToM/Percept-FANToM-conv_with_perceivers.csv

Column	Description
part_id	ID of each conversation part in FANToM
conversation_with_perceivers	Utterances in the conversation part paired with corresponding perceivers. Formatted in JSON to provide ground truth perception information to the language model.

---

Evaluating Language Models on Perception-augmented ToM Datasets

The experimental implementations for Percept-ToMi and Percept-FANToM are based on the implementations of SymbolicToM and FANToM, respectively. We acknowledge and appreciate the work of the authors of these papers, which served as the foundation for our extensions.

The implemented methods include the following:

• vanilla: Vanilla language model
• cot: Chain of Thought
• s2a: System 2 Attention
• perc_inf: Perception inference
• perceptom: PercepToM
• ptob_inf: Perception-to-belief inference
• simtom: SimToM (implemented only for Percept-ToMi)

Experiments on the Percept-ToMi

We acknowledge that the evaluation results for Percept-ToMi exhibit variability across runs. Therefore, we recommend conducting multiple evaluations and using the average performance as a more reliable measure.

1. Run Evaluation Script

Execute the evaluation script with the desired model and method:

python eval_percept_tomi.py --model {model_name} --method {method_name}

Result files saved in results/Percept-ToMi/{model_name}-{method_name}/:

• acc.csv: Question answering accuracies for four qTypes — first order/second order, true belief (no ToM) / false belief (ToM)

• acc_raw.csv: Question answering accuracies for eight qTypeRaws

• result.csv: Evaluated results of each datapoint, including prompts, model responses, and correctness

Notes:

• For all methods except perception inference (perc_inf), running this script performs language model inference and evaluation, producing all evaluation results.

• The perception inference method (perc_inf) does not perform question answering. Running the script with this method only performs language model inference and produces result.csv.

2.Perception Inference Evaluation (Only for perc_inf)

To evaluate perception inference results, run the following script:

python eval_percept_tomi-perc_inf.py --model {model_name} --method {method_name}

Result files saved in results/Percept-ToMi/{model_name}-{method_name}/ (the same directory as step 1):

• scene_level_eval.csv: Model predictions and ground truth of scene-level perceivers

• ctx_level_eval.csv: Model predictions and ground truth of context (story)-level perceivers

• perc_inf_qType_acc.csv: Perception inference accuracies for four qTypes — first order/second order, true belief (no ToM) / false belief (ToM)

• perc_inf_scenario_acc.csv: Perception inference accuracies for two scenarios — true belief (no ToM) / false belief (ToM)

Experiments on the Percept-FANToM

1.Run Evaluation Script

Execute the evaluation script with the desired model and method:

python eval_percept_fantom.py --model {model_name} --method {method_name}

Result files saved in results/Percept-FANToM/{model_name}-{method_name}/:

• REPORT_short_input_{model_name}.json: Question answering accuracies for different question types of FANToM for the contexts of false belief scenarios

• control_task_report_short_input_{model_name}.json: Question answering accuracies for different question types of FANToM for the contexts of true belief scenarios

• evaluated_responses_short_input_{model_name}.json/csv: Evaluated results of each datapoint, including prompts, model responses, and correctness

Notes:

• For all methods except perception inference (perc_inf), running this script performs language model inference and evaluation, producing all evaluation results.

• The perception inference method (perc_inf) does not include the question answering step. Running the script with this method only performs language model inference and produces evaluated_responses_short_input_{model_name}.json/csv.

2.Perception Inference Evaluation (Only for perc_inf)

To evaluate perception inference results, run the following script:

python eval_percept_fantom-perc_inf.py --model {model_name} --method {method_name}

Result files saved in results/Percept-FANToM/{model_name}-{method_name}/:

• utter_level_eval.csv: Model predictions and ground truth of utterance-level perceivers

• ctx_level_eval.csv: Model predictions and ground truth of context (part of conversation)-level perceivers

• perc_inf_acc.csv: Perception inference accuracies for two scenarios — true belief (no ToM) / false belief (ToM)