Class: Informers::TokenClassificationPipeline

Inherits:
Pipeline
  • Object
show all
Defined in:
lib/informers/pipelines.rb

Instance Method Summary collapse

Methods inherited from Pipeline

#initialize

Constructor Details

This class inherits a constructor from Informers::Pipeline

Instance Method Details

#call(texts, ignore_labels: ["O"], aggregation_strategy: "simple") ⇒ Object 



55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
 
# File 'lib/informers/pipelines.rb', line 55

def call(
  texts,
  ignore_labels: ["O"],
  aggregation_strategy: "simple"
)
  is_batched = texts.is_a?(Array)

  # Run tokenization
  model_inputs = @tokenizer.(is_batched ? texts : [texts],
    padding: true,
    truncation: true,
    return_offsets: true
  )

  # Run model
  outputs = @model.(model_inputs)

  logits = outputs.logits
  id2label = @model.config.id2label

  to_return = []
  logits.length.times do |i|
    ids = model_inputs[:input_ids][i]
    batch = logits[i]
    offsets = model_inputs[:offsets][i]

    # List of tokens that aren't ignored
    tokens = []
    batch.length.times do |j|
      token_data = batch[j]
      top_score_index = Utils.max(token_data)[1]

      entity = id2label ? id2label[top_score_index.to_s] : "LABEL_#{top_score_index}"
      if ignore_labels.include?(entity)
        # We predicted a token that should be ignored. So, we skip it.
        next
      end

      # TODO add option to keep special tokens?
      word = @tokenizer.decode([ids[j]], skip_special_tokens: true)
      if word == ""
        # Was a special token. So, we skip it.
        next
      end

      scores = Utils.softmax(token_data)

      tokens << {
        entity: entity,
        score: scores[top_score_index],
        index: j,
        word: word,
        start: offsets[j][0],
        end: offsets[j][1]
      }
    end

    case aggregation_strategy
    when "simple"
      tokens = group_entities(tokens)
    when "none"
      # do nothing
    else
      raise ArgumentError, "Invalid aggregation_strategy"
    end

    to_return << tokens
  end
  is_batched ? to_return : to_return[0]
end

#get_tag(entity_name) ⇒ Object 



142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
 
# File 'lib/informers/pipelines.rb', line 142

def get_tag(entity_name)
  if entity_name.start_with?("B-")
    bi = "B"
    tag = entity_name[2..]
  elsif entity_name.start_with?("I-")
    bi = "I"
    tag = entity_name[2..]
  else
    # It's not in B-, I- format
    # Default to I- for continuation.
    bi = "I"
    tag = entity_name
  end
  [bi, tag]
end

#group_entities(entities) ⇒ Object 



158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
 
# File 'lib/informers/pipelines.rb', line 158

def group_entities(entities)
  entity_groups = []
  entity_group_disagg = []

  entities.each do |entity|
    if entity_group_disagg.empty?
      entity_group_disagg << entity
      next
    end

    # If the current entity is similar and adjacent to the previous entity,
    # append it to the disaggregated entity group
    # The split is meant to account for the "B" and "I" prefixes
    # Shouldn't merge if both entities are B-type
    bi, tag = get_tag(entity[:entity])
    _last_bi, last_tag = get_tag(entity_group_disagg[-1][:entity])

    if tag == last_tag && bi != "B"
      # Modify subword type to be previous_type
      entity_group_disagg << entity
    else
      # If the current entity is different from the previous entity
      # aggregate the disaggregated entity group
      entity_groups << group_sub_entities(entity_group_disagg)
      entity_group_disagg = [entity]
    end
  end
  if entity_group_disagg.any?
    # it's the last entity, add it to the entity groups
    entity_groups << group_sub_entities(entity_group_disagg)
  end

  entity_groups
end

#group_sub_entities(entities) ⇒ Object 



126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
 
# File 'lib/informers/pipelines.rb', line 126

def group_sub_entities(entities)
  # Get the first entity in the entity group
  entity = entities[0][:entity].split("-", 2)[-1]
  scores = entities.map { |entity| entity[:score] }
  tokens = entities.map { |entity| entity[:word] }

  entity_group = {
    entity_group: entity,
    score: scores.sum / scores.count.to_f,
    word: @tokenizer.convert_tokens_to_string(tokens),
    start: entities[0][:start],
    end: entities[-1][:end]
  }
  entity_group
end