|
--- |
|
library_name: transformers |
|
tags: |
|
- text summarization |
|
license: apache-2.0 |
|
language: |
|
- en |
|
metrics: |
|
- rouge |
|
pipeline_tag: text2text-generation |
|
--- |
|
|
|
# Model Card for Post-Disaster Digital Help Desk Summarization Model |
|
|
|
<!-- Provide a quick summary of what the model is/does. --> |
|
This model is designed to summarize digital help desk conversations in post-disaster scenarios, specifically tailored for non-profit organizations providing aid. It is based on the BART model, fine-tuned using parameter-efficient methods like LoRa adapters. |
|
|
|
## Model Details |
|
|
|
### Model Description |
|
This is a parameter efficient finetuned model based on the fine-tuning of the BART model. the methodology used is the LoRa adapter. this model focuses on automated text summarization of digital helpdesk conversations in post-disaster assistance scenarios in order to improve the efficiency and quality of the information gathered to provide timely and effective support to the affected people. |
|
## Uses |
|
|
|
<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. --> |
|
The model is designed to summarize digital help desk conversations for nonprofit organizations in post-disaster assistance scenarios, helping digital help desk staff to quickly extract key information and reduce the time it takes to manually write high-quality summaries. |
|
|
|
## Bias, Risks, and Limitations |
|
Generated summaries may contain certain errors, such as the inclusion of sensitive information, and require manual secondary correction to ensure accuracy and privacy protection. |
|
|
|
## How to Get Started with the Model |
|
|
|
Use the code below to get started with the model. |
|
|
|
```python |
|
# install package |
|
!pip install transformers[torch] -U |
|
!pip install -q -U peft |
|
|
|
import os |
|
import torch |
|
from google.colab import drive |
|
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer |
|
from huggingface_hub import notebook_login |
|
|
|
# login to hugging_face |
|
notebook_login() # use model on GPU |
|
device = "cuda" if torch.cuda.is_available() else "cpu" |
|
|
|
# load base model |
|
model_name = "knkarthick/MEETING_SUMMARY" |
|
model = AutoModelForSeq2SeqLM.from_pretrained(model_name).to(device) |
|
tokenizer = AutoTokenizer.from_pretrained(model_name) |
|
|
|
# load trained adapter |
|
adapter_id = "Joaaaane/510_ABW_LoRaAdapter_PostDisasterConv" |
|
model.load_adapter(adapter_id) # set the model to evaluation mode |
|
model.eval() |
|
input_text = """ |
|
PA: Hello, I need urgent housing help as a refugee from Ukraine. Can you assist? |
|
agent: Hello, thank you for reaching out to the Red Cross. We’re here to help with housing. |
|
agent: Have you registered with the local authorities yet? |
|
PA: Yes, but they mentioned delays, and we need something soon. It's urgent. |
|
agent: We have temporary shelters available. How many are with you, and are there any special needs? |
|
PA: It's just me and my elderly mother; we need accessible housing. |
|
agent: We can arrange for accessible temporary shelter. I’ll expedite your request and aim to place you within a few days. |
|
agent: I'll also connect you with a Ukrainian-speaking volunteer to help with your paperwork and make your mother more comfortable. |
|
PA: Thank you so much. This help means a lot to us right now. |
|
agent: You're welcome! Expect a call from our volunteer by tomorrow. We’ll make sure you both are settled quickly. |
|
PA: Thanks again. Looking forward to resolving this soon. |
|
""" |
|
|
|
# tokenized inputs |
|
inputs = tokenizer(input_text, return_tensors="pt", max_length=1024, truncation=True).to(device) |
|
# generate summary tokens |
|
outputs = model.generate(inputs['input_ids'], max_length=62, num_beams=5, early_stopping=True) |
|
# decode tokens |
|
summary = tokenizer.decode(outputs[0], skip_special_tokens=True) |
|
|
|
print("Generated Summary:", summary) |
|
|
|
|
|
## Training Details |
|
|
|
### Training Data |
|
Data provided by 510, an initiative of the Netherlands Red Cross (all confidential data has been masked). |
|
|
|
### Testing Data |
|
Data provided by 510, an initiative of the Netherlands Red Cross (all confidential data has been masked). |
|
|
|
### Metrics |
|
ROUGE Score |
|
|
|
### Results |
|
|
|
| Metric | Before LoRA | After LoRA | |
|
|--------------------|-------------|------------| |
|
| **ROUGE 1** | 22.50 | 28.30 | |
|
| **ROUGE 2** | 4.96 | 8.64 | |
|
| **ROUGE L** | 17.24 | 22.50 | |
|
|
|
|
|
## Technical Specifications |
|
|
|
### Model Architecture and Objective |
|
In post-disaster humanitarian assistance scenarios, the efficiency of digital help desks and the quality of the information they collect are crucial for providing effective and timely support to the people affected. This model leverages parameter-efficient fine-tuning techniques, including Low-Rank Adaptation (LoRA) and Prefix Tuning, to generate summaries and reduce the time spent on manually writing high-quality summaries. The results indicate that the adjusted LLMs not only improve the speed and quality of text summarization but also ensure adaptability to sensitive contexts. Potential challenges and recommendations for implementing the model in practice are also discussed. |
|
|
|
## Citation |
|
|
|
Base model: https://huggingface.co/knkarthick/MEETING_SUMMARY |
