Update app.py

app.py

@@ -1,597 +1,8 @@
-import spacy
-import re
-import nltk
-from nltk.corpus import wordnet
-import numpy as np
-
-from sklearn.metrics.pairwise import cosine_similarity
-
-nltk.download('wordnet')
-nltk.download('omw-1.4')
-nltk.download('punkt')
-
-# load the spacy model
-spacy.cli.download("en_core_web_sm")
-
-# use spacy small because in that way we are closer to a BOW model which is the one we care in our case since we just compare words
-nlp = spacy.load('en_core_web_sm', disable=["parser", "ner"])
-
-
-def find_comptives_symbols(sentence):
-    """
-    Capture unique cases of symbols like <, >, =
-    If more than one symbols exist, return []
-    """
-
-    # symbols regex pattern
-    pattern = r"(?<![<=>])[%s](?![<=>])" % (re.escape("<=>"))
-    matches = re.findall(pattern, sentence)
-
-    found_symbols = []
-    for matching in matches:
-      # add symbol to list for each occurrence found
-      found_symbols.append({'comparative': ['symbol', matching]})   
-
-    # return the found symbols
-    return found_symbols
-
-
-def find_comptives_straight_patterns(sentence):
-    """
-    Function to identivy mentions of comparatives. The form is "comparative adverbs/adjectives followed by than", "words like more/less followed by than", "equal to" 
-    """
-
-    doc = nlp(sentence)
-    comparatives = []
-  
-    for token in doc:
-
-        # find mentions of "equal" followed by "to"
-        if token.text.lower() == "equal":
-            next_token = token.nbor()
-  
-            if next_token.text.lower() == "to":
-                prev_token = token.nbor(-1)
-  
-                if prev_token.pos_ == "NOUN":
-                    comparatives.append({'comparative': ["equal to", "="]})
-
-
-        # find mentions of "more"/"less" followed by "than"
-        elif token.text.lower() in ["more", "less"]:
-          
-            next_token = token.nbor()
-  
-            if next_token.text.lower() == "than":
-                prev_token = token.nbor(-1)
-  
-                # this part is to check what will be before more/less. We can add a NOUN as mandatory (e.g magnitude) or even specifically the word magnitude
-                # for the moment we have disable it
-                # if prev_token.pos_ == "NOUN":
-
-                if token.text.lower() == 'more': 
-                  comparatives.append({'comparative': [token.text+" "+next_token.text, '>']})
-                elif token.text.lower() == 'less':
-                  comparatives.append({'comparative': [token.text+" "+next_token.text, '<']})
-
-
-        # find mentions of comparative adjectives or comparative adverbs followed by "than"
-        elif token.tag_ == "JJR" or token.tag_ == "RBR":
-            next_token = token.nbor()
-  
-            if next_token.text.lower() == "than" and next_token.nbor().pos_ != "NOUN":
-
-                # check if the token is a synonym of "bigger"
-
-                # retrieve a set of synonyms for the concepts of 'big' and 'bigger'
-                big_synonyms = set(wordnet.synsets('big') + wordnet.synsets('large') + wordnet.synsets('great') + wordnet.synsets('huge') + wordnet.synsets('enormous') + wordnet.synsets('heavy') + wordnet.synsets('strong') + wordnet.synsets('enormous') + wordnet.synsets('massive') + wordnet.synsets('immense') + wordnet.synsets('substantial'))
-                bigger_synonyms = set(wordnet.synsets('bigger') + wordnet.synsets('larger') + wordnet.synsets('greater') + wordnet.synsets('higher') + wordnet.synsets('taller') + wordnet.synsets('heavier') + wordnet.synsets('stronger'))
-
-                bigger_related_words = big_synonyms.union(bigger_synonyms)
-                bigger_rel_words = [word.name().split('.')[0] for word in bigger_related_words]
-
-                flag_bigger = 0
-
-                if token.text.lower() in bigger_rel_words:
-                  flag_bigger = 1
-                  comparatives.append({'comparative': [token.text+" "+next_token.text, '>']})
-
-                # if no synonym of bigger was found, check for smaller synsets
-                if not flag_bigger:
-
-                  # retrieve a set of synonyms for the concepts of 'small' and 'smaller'
-                  small_synonyms = set(wordnet.synsets('small') + wordnet.synsets('little') + wordnet.synsets('tiny') + wordnet.synsets('petite') + wordnet.synsets('miniature') + wordnet.synsets('slight') + wordnet.synsets('meager') + wordnet.synsets('inconsequential') + wordnet.synsets('minor'))
-                  smaller_synonyms = set(wordnet.synsets('smaller') + wordnet.synsets('lesser') + wordnet.synsets('lower') + wordnet.synsets('shorter') + wordnet.synsets('lighter') + wordnet.synsets('weaker'))
-
-                  smaller_related_words = small_synonyms.union(smaller_synonyms)
-                  smaller_rel_words = [word.name().split('.')[0] for word in smaller_related_words]
-
-                  if token.text.lower() in smaller_rel_words:
-                      flag_bigger = 0
-                      comparatives.append({'comparative': [token.text+" "+next_token.text, '<']})
-
-    return comparatives
-
-
-# helper functions for 'identify_pattern_bigger_smaller'
-
-def identify_comparison(sentence):
-  """
-  Capture patterns of 'word-er' followed by 'than' (e.g. 'better than', 'lesser than', etc)
-  """
-
-  pattern = r'\b(\w+er than)\b'
-  matches = re.findall(pattern, sentence)
-
-  if matches:
-    return matches
-  else:
-    return 0
-
-
-def find_more_than_reference(sentence):
-  """
-  Capture patterns of 'more' followed by 'word' followed by 'than' (e.g. more advanced than)
-  """
-
-  pattern = r"(more) (\w+) than"
-  matches = re.findall(pattern, sentence)
-
-  if matches:
-      return [' '.join(match) for match in matches]
-  else:
-      return 0
-
-
-def find_less_than_reference(sentence):
-  """
-  Capture patterns of 'less' followed by 'word' followed by 'than' (e.g. less advanced than)
-  """
-
-  pattern = r"(less) (\w+) than"
-  matches = re.findall(pattern, sentence)
-  
-  if matches:
-      return [' '.join(match) for match in matches]
-  else:
-      return 0      
-
-
-def is_related_to(word, target_word):
-    """
-    Returns True if the input 'word' is semantically related to the 'target_word', otherwise False.
-    """
-
-    target_synsets = set(wordnet.synsets(target_word))
-    word_synsets = set(wordnet.synsets(word))
-
-    if word_synsets.intersection(target_synsets):
-        return True
-    else:
-        return False
-
-
-def is_related_to_bigger(word):
-    """
-    Returns True if the input 'word' is semantically related to the concept 'bigger', otherwise False.
-    """
-
-    if word.lower() == "more" or word.lower().startswith("more "):
-        return True
-    
-    # retrieve a set of synonyms for the concepts of 'big' and 'bigger'
-    big_synonyms = set(wordnet.synsets('big') + wordnet.synsets('large') + wordnet.synsets('great') + wordnet.synsets('huge') + wordnet.synsets('enormous') + wordnet.synsets('heavy') + wordnet.synsets('strong') + wordnet.synsets('enormous') + wordnet.synsets('massive') + wordnet.synsets('immense') + wordnet.synsets('substantial'))
-    bigger_synonyms = set(wordnet.synsets('bigger') + wordnet.synsets('larger') + wordnet.synsets('greater') + wordnet.synsets('higher') + wordnet.synsets('taller') + wordnet.synsets('heavier') + wordnet.synsets('stronger'))
-
-    related_words = big_synonyms.union(bigger_synonyms)
-    
-    # Check if the input word is semantically related to any of those 'big'/'bigger' synonyms
-    for related_word in related_words:
-        if is_related_to(word, related_word.name().split('.')[0]):
-            return True
-    return False
-
-
-def is_related_to_smaller(word):
-    """
-    Returns True if the input word is semantically related to the concept of 'smaller', otherwise False.
-    """
-    if word.lower() == "less" or word.lower().startswith("less "):
-        return True
-    
-    # retrieve a set of synonyms for the concepts of 'small' and 'smaller'
-    small_synonyms = set(wordnet.synsets('small') + wordnet.synsets('little') + wordnet.synsets('tiny') + wordnet.synsets('petite') + wordnet.synsets('miniature') + wordnet.synsets('slight') + wordnet.synsets('meager') + wordnet.synsets('inconsequential') + wordnet.synsets('minor'))
-    smaller_synonyms = set(wordnet.synsets('smaller') + wordnet.synsets('lesser') + wordnet.synsets('lower') + wordnet.synsets('shorter') + wordnet.synsets('lighter') + wordnet.synsets('weaker'))
-
-    related_words = small_synonyms.union(smaller_synonyms)
-    
-    # Check if the input word is semantically related to any of those 'small'/'smaller' synonyms
-    for related_word in related_words:
-        if is_related_to(word, related_word.name().split('.')[0]):
-            return True
-    return False
-
-
-def identify_bigger_smaller_advanced(sentence):
-  """
-  This is a complementary function to capture cases of 'words ending with -er' followed by 'than' and cases of 'more'/'less' followed 'word' followed by 'than'
-  """
-
-  # pattern 'words ending with -er' followed by 'than' (pattern1)
-  word_er_than = identify_comparison(sentence)
-  
-  # pattern 'more' followed 'word' followed by 'than' (pattern2)
-  more_word_than = find_more_than_reference(sentence)
-
-  # pattern 'less' followed 'word' followed by 'than' (pattern3)
-  less_word_than = find_less_than_reference(sentence)
-
-  bigger_list = []
-  smaller_list = []
-
-  # in case any pattern is captured
-  if word_er_than or more_word_than or less_word_than:
-
-    # in case of pattern1
-    if word_er_than:
-      for word in word_er_than:
-
-        # perform relevant substitutions
-        target_word = word.replace("than", "").strip()
-
-        # examine if it is a bigger-related or smaller-related word
-        bigger_word = is_related_to_bigger(target_word)
-        smaller_word = is_related_to_smaller(target_word)
-        
-        # case of bigger word
-        if bigger_word and not smaller_word:          
-          bigger_list.append({"comparative":[word, ">"]})
-
-        # case of smaller word
-        elif smaller_word and not bigger_word:
-          smaller_list.append({"comparative":[word, "<"]})
-
-    # in case of pattern2
-    if more_word_than:
-      for word in more_word_than:
-      
-        # perform relevant substitutions
-        target_word = word.replace("than", "").replace("more", "").strip()
-
-        # in this case it must be a bigger-related word
-        bigger_word = is_related_to_bigger(target_word)
-
-        # case of bigger word
-        if bigger_word:
-          bigger_list.append({"comparative":[word, ">"]})
-
-
-    # in case of pattern3
-    if less_word_than:
-      for word in less_word_than:
-
-        # perform relevant substitutions
-        target_word = word.replace("than", "").replace("less", "").strip()
-
-        # in this case it must be a lesser-related word
-        lesser_word = is_related_to_smaller(target_word)
-
-        # case of bigger word
-        if lesser_word:
-          smaller_list.append({"comparative":[word, "<"]})
-  
-  # return the combined list
-  return bigger_list + smaller_list
-
-
-def find_equal_to_comptives_ngrams(sentence):
-    """
-    This function takes a sentence as input and returns a reference phrase based on semantic similarity using n-grams.
-    The possible reference phrases are provided as a list. 
-    """
-
-    # This is a reference list for the concept of 'equal to'. It has many references to perform on them the semantic similarity examination 
-    possible_references = ["equal to", "same as", "similar to", "identical to", "equivalent to", "tantamount to", "corresponding to", "comparable to", "akin to", "commensurate with", "in line with", "on a par with" , "indistinguishable from" , "corresponding with", "congruent with"]
-    
-    # that thershold is enough empirically
-    max_similarity = 0.85
-
-    possible_reference_list = []
-
-    # parse with the spacy model (embeddings each of the references)
-    embedding_references = []
-    for reference in possible_references:
-        reference_doc = nlp(reference)
-        embedding_references.append(reference_doc)
-
-    # Check 2-grams, 3-grams, and 4-grams
-    for n in range(2, 5): 
-
-        # get n-grams
-        sentence_ngrams = list(nltk.ngrams(sentence.split(), n))
-
-        for sent_ngram in sentence_ngrams:
-            sentence_ngram_str = ' '.join(sent_ngram)
-            sentence_ngram_doc = nlp(sentence_ngram_str)
-
-            for emb_ref in embedding_references:
-              similarity = sentence_ngram_doc.similarity(emb_ref)
-
-              if similarity >= max_similarity:
-                  possible_reference_list.append({'comparative': [sentence_ngram_str, "="]})
-                  break
-
-    # if we have found a possible refernce that is similar enough with an n-gram of the input sentence, return the comparative '=', otherwise return 0
-    if possible_reference_list:
-      return possible_reference_list
-    else:
-      return []
-
-
-
-def single_verb_comptives(sentence):
-    """
-    This function takes a sentence and identifies any mention of bigger than, smaller than, equal to, expressed
-    as single-word verb. It uses wordnet synsets to examine for synonyms and antonyms
-    """
-
-    # base references
-    bigger_references_sg = ["surpass", "exceed", "outstrip", "outdo", "outmatch", "outclass", "eclipse", "overshadow", "outrank", "overtake", "top", "beat", "transcend", "dominate", "prevail", "trump", "vanquish", "outperform", "outgun", "outdistance", "outshine"]
-    lesser_references_sg = ["lag", "trail", "lose", "underperform", "yield", "surrender", "submit", "succumb", "straggle", "dawdle", "lollygag", "loiter", "delay", "defer", "postpone", "procrastinate", "linger", "hesitate", "prolong", "drag"]
-    equal_references_sg = ["match", "equal", "tie", "correspond", "conform", "agree", "harmonize", "coordinate", "comply", "fit", "parallel", "resemble", "mirror", "emulate", "equilibrate", "balance", "counterbalance", "offset", "compensate"]
-
-    doc = nlp(sentence)
-    
-    bigger_list = []
-    smaller_list = []
-    equal_list = []
-
-    # search for all verbs and examine their lemma with all the synonyms of each of the previous references. Assign a label accordingly
-    for token in doc:
-        if token.pos_ == "VERB":
-    
-            for lemma in token.lemma_.split('|'):
-                synsets = wordnet.synsets(lemma, pos='v')
-                
-                for syn in synsets:
-                    if any(lemma in bigger_references_sg for lemma in syn.lemma_names()):
-                        bigger_list.append({'comparative': [token.text, ">"]})
-                        break
-
-                    elif any(lemma in lesser_references_sg for lemma in syn.lemma_names()):
-                        smaller_list.append({'comparative': [token.text, "<"]})
-                        break
-
-                    elif any(lemma in equal_references_sg for lemma in syn.lemma_names()):
-                        equal_list.append({'comparative': [token.text, "="]})
-                        break
-
-                # for syn in synsets:
-                #     antonyms = syn.lemmas()[0].antonyms()
-
-                #     if antonyms and any(lemma in bigger_references_sg for lemma in antonyms[0].name()):
-                #       return 0
-
-                #     elif antonyms and any(lemma in lesser_references_sg for lemma in antonyms[0].name()):
-                #       return 0
-
-
-                #     elif antonyms and any(lemma in equal_references_sg for lemma in antonyms[0].name()):
-                #       return 0
-
-    final_list = bigger_list + smaller_list + equal_list
-
-    if final_list:
-      return final_list
-    else:
-      return []
-
-
-# helper functions for 'identify_multi_word_verbs'
-
-# Define multi-word verb lists
-bigger_list = ["is a cut above", "is ahead of", "is superior to", "is greater than", "raise the bar", "climb the ladder", "set the standard", "set the pace", "break the mold", "push the envelope", "raise the game", "is a class apart"]
-smaller_list = ["fall behind", "is inferior to", "is smaller than", "lag behind", "trail behind", "is second to", "bring up the rear", "lose ground", "bring up the tail end", "fall short", "fall beneath", "fail to measure up", "put off"]
-equal_list = ["is in line with", "is equal to", "is on a par with", "is on par with", "is the same as", "is comparable to", "is in sync with", "is in harmony with", "is in step with", "is in tune with", "is in accord with", "is consistent with", "is consonant with", "keep pace with", "keep up with", "is equivalent to", "balance out", "even out"]
-
-# Calculate embeddings of multi-word verbs
-bigger_embeddings = [np.mean([token.vector for token in nlp(verb)], axis=0) for verb in bigger_list]
-smaller_embeddings = [np.mean([token.vector for token in nlp(verb)], axis=0) for verb in smaller_list]
-equal_embeddings = [np.mean([token.vector for token in nlp(verb)], axis=0) for verb in equal_list]
-
-
-# Define function to check if n-gram is in multi-word verb list
-def check_list(ngram, verb_list):
-    """
-    This is a function to check if n-gram is in multi-word verb list
-    """
-
-    if ngram in verb_list:
-        return True
-    else:
-        return False
-
-
-def cosine_sim(a, b):
-    """
-    This is a function to calculate cosine similarity
-    """
-
-    return cosine_similarity(a.reshape(1,-1), b.reshape(1,-1))[0][0]
-
-
-# we examine the n-grams reversely and any time we find a match, we "delete" that match, so that lesser ngrams will not be matched \
-# (e.g. is on a par with, would also match afterwords on a par with, par with, etc)
-
-def multiword_verb_comptives(sentence):
-    """
-    This function takes a sentence and identifies any mention of bigger than, smaller than, equal to, expressed
-    as multi-word verbs. Based on three refernces lists it performs initially a simple string comparison with each
-    of their elements and the ngrams of the input sentence. If there is no match there, it performs the same procedure
-    with cosine similarity to identify any similar ngrams.
-    """
-
-    # Split sentence into tokens
-    tokens = sentence.split()
-
-    # Initialize variables to store label and max similarity
-    label = None
-    max_sim = 0
-
-    # these lists are used to capture any possible reference
-    bigger_l = []
-    smaller_l = []
-    equal_l = []
-
-    # Define set to keep track of matched ngrams
-    matched_ngrams = set()
-
-    # Iterate through n-grams of sentence, starting with the largest n-grams
-    for n in range(5, 1, -1):
-        for i in range(len(tokens)-n+1):
-            ngram = ' '.join(tokens[i:i+n])
-
-            # Skip ngrams that have already been matched
-            if ngram in matched_ngrams:
-                continue
-
-            # Check if n-gram is in bigger_list
-            if check_list(ngram, bigger_list):
-                matched_ngrams.update(set(ngram.split()))
-                bigger_l.append({"comparative": [ngram, '>']})
-
-            # Check if n-gram is in smaller_list
-            elif check_list(ngram, smaller_list):
-                matched_ngrams.update(set(ngram.split()))
-                smaller_l.append({"comparative":[ngram, '<']})
-
-            # Check if n-gram is in equal_list
-            elif check_list(ngram, equal_list):
-                matched_ngrams.update(set(ngram.split()))
-                equal_l.append({"comparative":[ngram, '=']})
-
-            # Check if n-gram is similar to any verb in bigger_list using pre-calculated embeddings
-            else:
-                ngram_emb = np.mean([token.vector for token in nlp(ngram)], axis=0)
-                similarities_bigger = [cosine_sim(ngram_emb, verb_emb) for verb_emb in bigger_embeddings]
-                max_sim_bigger = max(similarities_bigger)
-
-                # Check if n-gram is similar to any verb in smaller_list using pre-calculated embeddings
-                similarities_smaller = [cosine_sim(ngram_emb, verb_emb) for verb_emb in smaller_embeddings]
-                max_sim_smaller = max(similarities_smaller)
-
-                # Check if n-gram is similar to any verb in equal_list using pre-calculated embeddings
-                similarities_equal = [cosine_sim(ngram_emb, verb_emb) for verb_emb in equal_embeddings]
-                max_sim_equal = max(similarities_equal)
-
-                # Determine the maximum similarity value among the three lists
-                if max_sim_bigger > max_sim_smaller and max_sim_bigger > max_sim_equal and max_sim_bigger > max_sim:
-                    max_sim = max_sim_bigger
-                    if max_sim > 0.9:
-                        matched_ngrams.update(set(ngram.split()))
-                        bigger_l.append({"comparative":[ngram, '>']})
-                    else:
-                        matched_ngrams.update(set(ngram.split()))
-
-
-                elif max_sim_smaller > max_sim_bigger and max_sim_smaller > max_sim_equal and max_sim_smaller > max_sim:
-                    max_sim = max_sim_smaller
-                    if max_sim > 0.9:
-                        matched_ngrams.update(set(ngram.split()))
-                        smaller_l.append({"comparative":[ngram, '<']})
-                    else:
-                        matched_ngrams.update(set(ngram.split()))
-
-
-                elif max_sim_equal > max_sim_bigger and max_sim_equal > max_sim_smaller and max_sim_equal > max_sim:
-                    max_sim = max_sim_smaller
-                    if max_sim > 0.9:
-                        matched_ngrams.update(set(ngram.split()))
-                        equal_l.append({"comparative":[ngram, '=']})
-                    else:
-                        matched_ngrams.update(set(ngram.split()))
-
-
-    return bigger_l + smaller_l + equal_l
-
-
-
-def identify_comparatives(sentence):
-    """
-    This function combines the results of all the aforementioned techniques (simple and advance) to identify bigger than, smaller than, equal to patterns
-    """
-
-    # Identify straightforward patterns
-    straight_comptives = find_comptives_straight_patterns(sentence)
-
-    # Identify advanced bigger/smaller comparatives
-    bigger_smaller_comparatives = identify_bigger_smaller_advanced(sentence)
-
-    # Identify advanced equal-to comparatives
-    equal_to_comparatives = find_equal_to_comptives_ngrams(sentence)
-
-    single_verb = single_verb_comptives(sentence)
-    multi_verb = multiword_verb_comptives(sentence)
-
-    # return all the patterns that were captured
-    comparatives = straight_comptives + bigger_smaller_comparatives + equal_to_comparatives + single_verb + multi_verb
-
-    # since those different techniques might capture similar patterns, we keep only unique references. More precisely 
-    unique_comparatives = {}
-
-    for item in comparatives:
-        if item['comparative'][0] not in unique_comparatives:
-            unique_comparatives[item['comparative'][0]] = item
-
-    unique_output = list(unique_comparatives.values())
-
-    return unique_output
-
-
-
-def comparatives_binding(sentence):
-
-  try:    
-    comparative_symbols = find_comptives_symbols(sentence)
-    comparative_mentions = identify_comparatives(sentence)
-
-    # starting with the symbols, if one was captured 
-    if len(comparative_symbols) == 1:
-
-      # if the rest of the functions are empty (meaning that there are no other references)
-      if len(comparative_mentions) == 0:
-        return comparative_symbols
-
-      else:
-        return (0, "COMPARATIVES", "more_comparatives_mentions")
-
-    # in case that there is no symbol
-    elif len(comparative_symbols) == 0:
-
-      # we need only one mention of comparatives
-      if len(comparative_mentions) == 1:
-        return comparative_mentions
-
-      # case of no comparative mentions
-      elif len(comparative_mentions) == 0:
-        return (0, "COMPARATIVES", "no_comparatives")
-
-      # case of no more than one comparative mentions
-      else:
-        return (0, "COMPARATIVES", "more_comparatives_mentions")
-
-    # case of multiple symbol references
-    else:
-      return (0, "COMPARATIVES", "more_symbol_comparatives")
-
-  except:
-    return (0, "COMPARATIVES", "unknown_error")
-
-
-
 from transformers import pipeline
 import gradio as gr
 
+from helper import comparatives_binding
+
 title = "Comparatives Demo"
 description = "This is a simple demo just for demonstration purposes for Serco team, to validate the results of the Natural Language module concerning comparatives identification, while in progress"