app.py

import streamlit as st
import base64
from transformers import AutoModel, AutoTokenizer
from graphviz import Digraph
import json

def display_tree(output):
    size = str(int(len(output))) + ',5'
    dpi = '300'
    format = 'svg'
    print(size, dpi)
    
    # Initialize Digraph object
    dot = Digraph(engine='dot', format=format)
    dot.attr('graph', rankdir='LR', rank='same', size=size, dpi=dpi)
    
    # Add nodes and edges
    for i,word_info in enumerate(output):
        word = word_info['word']  # Prepare word for RTL display
        head_idx = word_info['dep_head_idx']
        dep_func = word_info['dep_func']
        
        dot.node(str(i), word)
        # Create an invisible edge from the previous word to this one to enforce order
        if i > 0:
            dot.edge(str(i), str(i - 1), style='invis')
        if head_idx != -1:
            dot.edge(str(i), str(head_idx), label=dep_func, constraint='False')


    # Render the Digraph object
    dot.render('syntax_tree', format=format, cleanup=True)
    # Display the image in a scrollable container
    st.markdown(
        f"""
            <div style="height:250px; width:75vw; overflow:auto; border:1px solid #ccc; margin-left:-15vw">
                <img src="data:image/svg+xml;base64,{base64.b64encode(dot.pipe(format='svg')).decode()}" 
                    style="display: block; margin: auto; max-height: 240px;">
            </div>
        """, unsafe_allow_html=True)
    
    #st.image('syntax_tree.' + format, use_column_width=True)
        
# Streamlit app title
st.title('BERT Syntax Dependency Tree Visualizer')

# Load Hugging Face token
hf_token = st.secrets["HF_TOKEN"]  # Assuming you've set up the token in Streamlit secrets

# Authenticate and load model
tokenizer = AutoTokenizer.from_pretrained('dicta-il/dictabert-joint', use_auth_token=hf_token)
model = AutoModel.from_pretrained('dicta-il/dictabert-joint', use_auth_token=hf_token, trust_remote_code=True)

model.eval()

# Checkbox for the compute_mst parameter
compute_mst = st.checkbox('Compute Maximum Spanning Tree', value=True)

output_style = st.selectbox(
    'Output Style: ',
    ('JSON', 'UD', 'IAHLT_UD'), index=1).lower()

# User input
sentence = st.text_input('Enter a sentence to analyze:')

if sentence:
    # Display the input sentence
    st.text(sentence)

    # Model prediction
    output = model.predict([sentence], tokenizer, compute_syntax_mst=compute_mst, output_style=output_style)[0]
    
    if output_style == 'ud' or output_style == 'iahlt_ud':
        ud_output = output
        # convert to tree format of [dict(word, dep_head_idx, dep_func)]
        tree = []
        for l in ud_output[2:]:
            parts = l.split('\t')
            if '-' in parts[0]: continue
            tree.append(dict(word=parts[1], dep_head_idx=int(parts[6]) - 1, dep_func=parts[7]))
        display_tree(tree)

        # Construct the table as a Markdown string
        table_md = "<div dir='rtl' style='text-align: right;'>\n\n"  # Start with RTL div
        
        # Add the UD header lines
        table_md += "##" + ud_output[0] + "\n"
        table_md += "##" + ud_output[1] + "\n"
        # Table header
        table_md += "| " + " | ".join(["ID", "FORM", "LEMMA", "UPOS", "XPOS", "FEATS", "HEAD", "DEPREL", "DEPS", "MISC"]) + " |\n"
        # Table alignment
        table_md += "| " + " | ".join(["---"]*10) + " |\n"
        for line in ud_output[2:]:
            # Each UD line as a table row
            cells = line.replace('_', '\\_').replace('|', '&#124;').split('\t')
            table_md += "| " + " | ".join(cells) + " |\n"
        table_md += "</div>"  # Close the RTL div
        
        # Display the table using a single markdown call
        st.markdown(table_md, unsafe_allow_html=True)

    else:
        # display the tree
        tree = [w['syntax'] for w in output['tokens']]
        display_tree(tree)
        
        # and the full json
        st.markdown("```json\n" + json.dumps(output, ensure_ascii=False, indent=2) + "\n```")