LLM-TLS

Repo for paper "From Moments to Milestones: Incremental Timeline Summarization Leveraging Large Language Models".

Paper link: https://aclanthology.org/2024.acl-long.390.pdf

Event-TLS

Dataset

The tweet ID and summary data we used in ./data is from CrisisLTLSum. Due to tweet privacy policy, we cannot release the content of the tweet. We recommend to download the tweet content through https://www.tweepy.org/ using the tweet ID.

Workflow

1. To prepare data for tweet timeline membership classification training, please refer to prepare_cls_data.py

python prepare_cls_data.py --input "./data" --output "./data/cls"

2. To prepare data for timeline summarization training, please refer to preprocess_timelines.py and prepare_tls_data.py

python preprocess_timelines.py
python prepare_tls_data.py --input "./data" --output "./data/ft"

In our implementation, we host the classification model through vLLM using the below command in a separate process:

python -m vllm.entrypoints.openai.api_server --model "CLS MODEL PATH" --port 8000

3. Perform incremental clustering process by generate_clusters.py

python generate_clusters.py --input "./data" --output "./cluster_output" --dataset_type "test" --cls_model_name "[CLS MODEL PATH]" --is_retrieval

4. Perform timeline summarization and evaluation by cluster_tls_eval.py

python cluster_tls_eval.py --input "./data" --output "./cluster_output" --dataset_type "test" --sum_model_base "SUMMARIZATION MODEL PATH" --sum_model_lora_checkpoint "LORA CHECKPOINT PATH" --is_retrieval

---

Topic-TLS

Download Dataset

To download datasets(T17, Crisis, Entities), please refer to complementizer/news-tls .

Workflow

1. To preprocess dataset articles to certain format, please refer to preprocess_articles.py

python preprocess_articles.py --ds_path "./datasets" --dataset "entities" --save_path "./corpus"

In our implementation, we host the model through vLLM using the below command in a separate process:

python -m vllm.entrypoints.openai.api_server --model "meta-llama/Llama-2-13b-hf" --port 8000

2. Perform LLM for event generation by generate_events.py.

python generate_events.py --dataset entities --model "meta-llama/Llama-2-13b-hf" --extraction_path "./event_outputs"

3. Perform incremental clustering process by generate_clusters.py

python generate_clusters.py \
    --model meta-llama/Llama-2-13b-hf \
    --output ./timelines_output \
    --input ./event_outputs/entities/Llama-2-13b-hf \
    --top_n 20 \
    --dataset entities \
    --incremental \
    --keyword all

4. Perform timeline summarization and evaluation by cluster_tls_eval.py

python cluster_tls_eval.py \
    --timelines_path ./timelines_output/entities/Llama-2-13b-hf \
    --events_path ./event_outputs/entities/Llama-2-13b-hf \
    --output ./result/entities \
    --dataset entities \
    --text_rank

If you encounter any problem with the code running, please contact [email protected] .

---

Citation

If you make use of our released source code, please cite the following paper:

Qisheng Hu, Geonsik Moon, Hwee Tou Ng (2024). From Moments to Milestones: Incremental Timeline Summarization Leveraging Large Language Models. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024). (pp. 7232 – 7246). Bangkok, Thailand.

---

License

The source code in this repository is licensed under the GNU General Public License Version 3. For commercial use of this code, separate commercial licensing is also available. Please contact:

• Qisheng Hu ([email protected])
• Hwee Tou Ng ([email protected])