File size: 1,775 Bytes
e9e9f7f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
# Guideline to run the DemoApp using Streamlit

Using Anaconda or create an environment to run streamlit  
* Create env:  
    ```python3 -m venv env```   
    ```source ./env/bin/activate```  
* Using Anaconda:  
    ```conda create -n maple python=3.11.5```  
    ```conda activate maple```

In the file "app.py" in "demoapp" folder:
* ```pip install streamlit```
* Install all imported libraries: ```pip install pandas langchain openai chromadb tiktoken``` or you can refer to the requirement.txt
* ```streamlit run demoapp/app2.py```  
  (The "app2.py" is our work on the most popular 12 bills. It is our latest code with RAG, vectara.)

# Additional Pointers (Source:Research Paper)
In the demo app itself we have included evaluation metrics that help gauge the quality of the generated summaries in our use-case
* The metrics we have used are : ROUGE-L, ROUGE-1, ROUGE-2, Cosine Similarity, and Factual Consistency Score.
* ROUGE-1 is the the overlap of unigrams (each word) between the original bill and generated summaries.
* ROUGE-2 is the overlap of bigrams between the original bill and generated summaries.
* ROUGE-L tell us about the Longest common subsequence, taking into account sentence-level structure similarity naturally and helps identify longest co-occurring in sequence n-grams.
* Cosine Similarity in this case tells us about the text similarity of two documents.
* Factual Consistency Score: We used Vectara that trained transformer model to output probability from 0 to 1 by comparing the source and summary. 0 being hallucination, 1 being factually consistent.

# Understand this folder
extracted_mgl.txt is the relevant mgl content for the 12 bills that MAPLE team wanted. Extracted from the column using the csv files <b> 12billswithmgl.csv </b>