Large Language Models for Automated Open-domain Scientific Hypotheses Discovery

This repository is the official implementation of the paper <Large Language Models for Automated Open-domain Scientific Hypotheses Discovery>, which is accepted by ACL 2024 findings. It also receives the best poster award in ICML 2024 AI4Science workshop.

In general, with this repository, you can
(1) generate hypotheses with MOOSE framework,
(2) evaluate the generated hypotheses by GPT4,
(3) display results listed in the paper (Table 3-10) from existing checkpoints (where we store the generated hypotheses, and evaluation scores by GPT4 & human expert), and
(4) display hypotheses and corresponding intermediate generations from existing checkpoints (e.g., research background, research inspirations, future-feedback, hypothesis, and present-feedback).

Hypotheses Generation with MOOSE

MOOSE can be run with the python command in main.sh. Option parameters for the python command can be adjusted. Specifically, the function for the options are described below:
*--num_background_for_hypotheses: how many number of background to find for hypotheses generation. Each background will be used to generate a set of hypotheses.
*--num_CoLM_feedback_times: number of present-feedback iterations.
*--bkg_corpus_chunk_noter: start from which background corpus to find background.
*--if_indirect_feedback: if run past-feedback (0: no; 1: yes).
*--if_only_indirect_feedback: advanced options for past-feedback, by default is 0.
*--if_close_domain: if adopts groundtruth background and inspirations for hypotheses generation (0: not adopt; 1: adopt).
*--if_ban_selfeval: if ban future-feedback 1 (0: run future-feedback 1; 1: not run future-feedback 1).
*--if_baseline: baseline options, by default is 0.
*--if_novelty_module_have_access_to_surveys: 0: no access; 1: have access.
*--if_insp_pasg_for_bkg_and_bkg_pasg_included_in_insp: if randomized corpus (0: no; 1: yes), by default is 0.
*--if_hypothesis_suggstor: if run future-feedback 2 (0: not run future-feedback 2; 1: run future-feedback 2).
*--api_key: your openai api key to run gpt-3.5-turbo.

Hypotheses Evaluation with GPT4

Hypotheses can be evaluated by GPT4 with the python command in evaluation_main.sh. Specifically, the function for the options are described below:
*--if_groundtruth_hypotheses: if evaluate groundtruth hypotheses, by default is 0.
*--model_name: model used for evaluation, by default is gpt4.
*--num_CoLM_feedback_times: number of present-feedback iterations used for generating the hypotheses.
*--start_id: the background corpus id as start to generate hypotheses.
*--end_id: the background corpus id as end to generate hypotheses.
*--if_azure_api: whether the api is from azure; set to 0 if the api is from openai.
*--api_key: your openai api key to run gpt-4 for evaluation.

Display results in Table 3 and Table 4

python compare_score.py

Display results in Table 5 and Table 6

python read_expert_eval.py

Display results in Table 7 and Table 8

It can be done by adjusting the method_name1 varibale in compare_score.py.
Specifically, method_name1 can be set to "rand_background_baseline", "rand_background_rand_inspiration_baseline", "rand_background_BM25_inspiration_baseline", "gpt35_background_gpt35_inspiration", "MOOSE_wo_ff1", "MOOSE_wo_ff2", "MOOSE_wo_survey", and "MOOSE_w_random_corpus".

Display results in Table 9

python consistency_between_expert_gpt4.py
if_hard_consistency variable in main() can be adjusted (0 or 1) to check soft or hard consistency score.

Display results in Table 10

python consistency_between_experts.py
if_hard_consistency variable in main() can be adjusted (0 or 1) to check soft or hard consistency score.

Display hypotheses and corresponding intermediate generations (e.g., research background, research inspirations, future-feedback, hypothesis, and present-feedback)

python check_intermediate_hypothesis_and_feedback.py --research_background_id 5 --hypothesis_id 0 --hypothesis_refinement_round 0
*--research_background_id: id of research background. The range is [0, 49]
*--hypothesis_id: id of the hypotheses generated from a research background (and retrieved inspirations). The typical range is [0, 3 or 4]
*--hypothesis_refinement_round: id of hypothesis refinement round (referred as present-feedback). The range is [0, 3]

Functions of Other Python files

data_format_converter.py: used to transform expert annotated data file to usable data file (mostly format transformation).
expert_eval_random_order_to_normal_order.py: transform expert-evaluated file from random order (to minimize bias from expert) to normal order (to calculate consistency).
pick_hyp_for_expert_eval.py: used to randomly pick hypotheses for expert evaluation.
read_from_pdf.py: extract text contents from social science survey paper.