app.py

import os

os.system("pip install Cython")
os.system("pip install benepar")

import networkx as nx
import matplotlib.pyplot as plt
import jraph
import jax.numpy as jnp
from datasets import load_dataset
import spacy
import gradio as gr
import en_core_web_trf
import numpy as np
import benepar
import re


dataset = load_dataset("gigant/tib_transcripts")

nlp = en_core_web_trf.load()

benepar.download('benepar_en3')
nlp.add_pipe('benepar', config={'model': 'benepar_en3'})

def parse_tree(sentence):
    stack = []  # or a `collections.deque()` object, which is a little faster
    top = items = []
    for token in filter(None, re.compile(r'(?:([()])|\s+)').split(sentence)):
        if token == '(':
            stack.append(items)
            items.append([])
            items = items[-1]
        elif token == ')':
            if not stack:
                raise ValueError("Unbalanced parentheses")
            items = stack.pop()  
        else:
            items.append(token)
    if stack:
        raise ValueError("Unbalanced parentheses")    
    return top

class Tree():
  def __init__(self, name, children):
    self.children = children
    self.name = name
    self.id = None
  def set_id_rec(self, id=0):
    self.id = id
    last_id=id
    for child in self.children:
      last_id = child.set_id_rec(id=last_id+1)
    return last_id
  def set_all_ids(self):
    self.set_id_rec(0)
  def print_tree(self, level=0):
    to_print = f'|{"-" * level} {self.name} ({self.id})'
    for child in self.children:
      to_print += f"\n{child.print_tree(level + 1)}"
    return to_print
  def __str__(self):
    return self.print_tree(0)
  def get_list_nodes(self):
    return [self.name] + [_ for child in self.children for _ in child.get_list_nodes()]

def rec_const_parsing(list_nodes):
  if isinstance(list_nodes, list):
    name, children = list_nodes[0], list_nodes[1:]
  else:
    name, children = list_nodes, []
  return Tree(name, [rec_const_parsing(child) for i, child in enumerate(children)])

def tree_to_graph(t):
  senders = []
  receivers = []
  for child in t.children:
    senders.append(t.id)
    receivers.append(child.id)
    s_rec, r_rec = tree_to_graph(child)
    senders.extend(s_rec)
    receivers.extend(r_rec)
  return senders, receivers

def construct_constituency_graph(docs):
  doc = docs[0]
  sent = list(doc.sents)[0]
  print(sent._.parse_string)
  t = rec_const_parsing(parse_tree(sent._.parse_string)[0])
  t.set_all_ids()
  senders, receivers = tree_to_graph(t)
  nodes = t.get_list_nodes()
  graphs = [{"nodes": nodes, "senders": senders, "receivers": receivers, "edge_labels": {}}]
  return graphs

def half_circle_layout(n_nodes, sentence_node=True):
  pos = {}
  for i_node in range(n_nodes - 1):
    pos[i_node] = ((- np.cos(i_node * np.pi/(n_nodes - 1))), 0.5 * (-np.sin(i_node * np.pi/(n_nodes - 1))))
  pos[n_nodes - 1] = (0, -0.25)
  return pos

def get_adjacency_matrix(jraph_graph: jraph.GraphsTuple):
  nodes, edges, receivers, senders, _, _, _ = jraph_graph
  adj_mat = jnp.zeros((len(nodes), len(nodes)))
  for i in range(len(receivers)):
    adj_mat = adj_mat.at[senders[i], receivers[i]].set(1)
  return adj_mat

def dependency_parser(sentences):
  return [nlp(sentence) for sentence in sentences]

def construct_dependency_graph(docs):
  """
  docs is a list of outputs of the SpaCy dependency parser
  """
  graphs = []
  for doc in docs:
    nodes = [token.text for token in doc]
    senders = []
    receivers = []
    edge_labels = {}
    for token in doc:
        for child in token.children:
            senders.append(child.i)
            receivers.append(token.i)
            edge_labels[(token.i, child.i)] = child.dep_
    graphs.append({"nodes": nodes, "senders": senders, "receivers": receivers, "edge_labels": edge_labels})
  return graphs

def construct_both_graph(docs):
  """
  docs is a list of outputs of the SpaCy dependency parser
  """
  graphs = []
  for doc in docs:
    nodes = [token.text for token in doc]
    nodes.append("Sentence")
    senders = [token.i for token in doc][:-1]
    senders.extend([token.i for token in doc][1:])
    receivers = [token.i for token in doc][1:]
    receivers.extend([token.i for token in doc][:-1])
    edge_labels = {(token.i, token.i + 1): "next" for token in doc[:-1]}
    for token in doc[:-1]:
        edge_labels[(token.i + 1, token.i)] = "previous"
    for node in range(len(nodes) - 1):
      senders.append(node)
      receivers.append(len(nodes) - 1)
      edge_labels[(node, len(nodes) - 1)] = "in"
    for token in doc:
        for child in token.children:
            senders.append(child.i)
            receivers.append(token.i)
            edge_labels[(token.i, child.i)] = child.dep_
    graphs.append({"nodes": nodes, "senders": senders, "receivers": receivers, "edge_labels": edge_labels})
  return graphs

def construct_structural_graph(docs):
  graphs = []
  for doc in docs:
    nodes = [token.text for token in doc]
    nodes.append("Sentence")
    senders = [token.i for token in doc][:-1]
    senders.extend([token.i for token in doc][1:])
    receivers = [token.i for token in doc][1:]
    receivers.extend([token.i for token in doc][:-1])
    edge_labels = {(token.i, token.i + 1): "next" for token in doc[:-1]}
    for token in doc[:-1]:
        edge_labels[(token.i + 1, token.i)] = "previous"
    for node in range(len(nodes) - 1):
      senders.append(node)
      receivers.append(len(nodes) - 1)
      edge_labels[(node, len(nodes) - 1)] = "in"
    graphs.append({"nodes": nodes, "senders": senders, "receivers": receivers, "edge_labels": edge_labels})
  return graphs

def to_jraph(graph):
  nodes = graph["nodes"]
  s = graph["senders"]
  r = graph["receivers"]

  # Define a three node graph, each node has an integer as its feature.
  node_features = jnp.array([0]*len(nodes))

  # We will construct a graph for which there is a directed edge between each node
  # and its successor. We define this with `senders` (source nodes) and `receivers`
  # (destination nodes).
  senders = jnp.array(s)
  receivers = jnp.array(r)

  # We then save the number of nodes and the number of edges.
  # This information is used to make running GNNs over multiple graphs
  # in a GraphsTuple possible.
  n_node = jnp.array([len(nodes)])
  n_edge = jnp.array([len(s)])


  return jraph.GraphsTuple(nodes=node_features, senders=senders, receivers=receivers,
  edges=None, n_node=n_node, n_edge=n_edge, globals=None)

def convert_jraph_to_networkx_graph(jraph_graph: jraph.GraphsTuple) -> nx.Graph:
  nodes, edges, receivers, senders, _, _, _ = jraph_graph
  nx_graph = nx.DiGraph()
  if nodes is None:
    for n in range(jraph_graph.n_node[0]):
      nx_graph.add_node(n)
  else:
    for n in range(jraph_graph.n_node[0]):
      nx_graph.add_node(n, node_feature=nodes[n])
  if edges is None:
    for e in range(jraph_graph.n_edge[0]):
      nx_graph.add_edge(int(senders[e]), int(receivers[e]))
  else:
    for e in range(jraph_graph.n_edge[0]):
      nx_graph.add_edge(
          int(senders[e]), int(receivers[e]), edge_feature=edges[e])
  return nx_graph

def plot_graph_sentence(sentence, graph_type="constituency"):
  # sentences = dataset["train"][0]["abstract"].split(".")
  docs = dependency_parser([sentence])
  if graph_type == "dependency":
    graphs = construct_dependency_graph(docs)
  elif graph_type == "structural":
    graphs = construct_structural_graph(docs)
  elif graph_type == "structural+dependency":
    graphs = construct_both_graph(docs)
  elif graph_type == "constituency":
    graphs = construct_constituency_graph(docs)
  g = to_jraph(graphs[0])
  adj_mat = get_adjacency_matrix(g)
  nx_graph = convert_jraph_to_networkx_graph(g)
  pos = half_circle_layout(len(graphs[0]["nodes"]))
  if graph_type == "constituency":
    pos = nx.planar_layout(nx_graph)
  plot = plt.figure(figsize=(12, 6))
  nx.draw(nx_graph, pos=pos,
          labels={i: e for i,e in enumerate(graphs[0]["nodes"])},
          with_labels = True, edge_color="blue",
          # connectionstyle="arc3,rad=0.1",
          node_size=1000, font_color='black', node_color="yellow")
  nx.draw_networkx_edge_labels(
    nx_graph, pos=pos,
    edge_labels=graphs[0]["edge_labels"],
    font_color='red'
  )
  adj_mat_plot, ax = plt.subplots(figsize=(6, 6))
  ax.matshow(adj_mat)
  return [gr.update(value=plot), gr.update(value=adj_mat_plot)]
    
def get_list_sentences(id):
  id =  int(min(id, len(dataset["train"]) - 1))
  return gr.update(choices = dataset["train"][id]["transcript"].split("."))

with gr.Blocks() as demo:
  with gr.Row():
    graph_type = gr.Dropdown(label="Graph type", choices=["structural", "dependency", "structural+dependency", "constituency"], value="structural+dependency", interactive = True)
  with gr.Tab("From transcript"):
    with gr.Row():
      with gr.Column():
        id = gr.Number(label="Transcript")
      with gr.Column(scale=3):
        sentence_transcript = gr.Dropdown(label="Sentence", choices = dataset["train"][0]["transcript"].split(".")[1:], interactive = True)
  with gr.Tab("Type sentence"):
    with gr.Row():
      sentence_typed = gr.Textbox(label="Sentence", interactive = True)
  with gr.Row():
    with gr.Column(scale=2):
      plot_graph = gr.Plot(label="Word graph")
    with gr.Column():
      plot_adj = gr.Plot(label="Word graph adjacency matrix")

  id.change(get_list_sentences, id, sentence_transcript)  
  sentence_transcript.change(plot_graph_sentence, [sentence_transcript, graph_type], [plot_graph, plot_adj])
  sentence_typed.change(plot_graph_sentence, [sentence_typed, graph_type], [plot_graph, plot_adj])

demo.launch()