AnAutomaticPencil's picture
data preprocessing update
a6326c7
"""
Several preprocessor classes.
Author: md
"""
from preprocessor.base import BasePreprocessorConfig, BasePreprocessor
from const import (
DIALOGUE_SUMMARY,
DIALOGUE_CONTEXT_TO_RESPONSE_GENERATION,
DIALOG,
KNOWLEDGE,
UTTERANCE,
ROLES,
EMOTION_RECOGNITION,
VALUE,
ABSA,
CHARACTER_IDENTIFICATION,
DIALOGUE_STATE_TRACKING,
DOCUMENT_GROUNDED_CONVERSATION,
TEXT2SQL,
SLOT_FILLING,
ROLE_RELATION_RECOGNITION,
QUESTION_IN_CONTEXT_REWRITING,
NATURAL_LANGUAGE_INFERENCE,
MACHINE_READING_COMPREHENSION,
MULTIPLE_CHOICE_QUESTION_ANSWERING,
INTENT_DETECTION,
DATA_TO_TEXT,
CHIT_CHAT,
TRAIN_SPLIT,
)
from typing import Dict, List, Callable
from copy import deepcopy
class SerialConfig(BasePreprocessorConfig):
def __init__(
self,
input_dir: str,
output_dir: str,
task: str,
task_bos_token: str = "<s>",
knowledge_bos_token: str = "[EK]",
prompt_bos_token: str = "[C]",
use_role: bool = True,
turn_sep: str = None,
roles_to_build_example: List = None,
dev_and_test_roles_to_build_example: List = None,
prompt_func: Callable = None,
knowledge_func: Callable = None,
label_func: Callable = None,
turn_knowledge_func: Callable = None,
roles_in_history: List[List] = None,
cur_turn_process_func: Callable = None,
all_turns_process_func: Callable = None,
multi_ref_sep: str = None,
*args,
**kwargs,
) -> None:
super().__init__(input_dir, output_dir, task, *args, **kwargs)
self.use_role = use_role
self.turn_sep = turn_sep
self.roles_to_build_example = roles_to_build_example
self.prompt_func = prompt_func
self.task_bos_token = task_bos_token
self.knowledge_bos_token = knowledge_bos_token
self.prompt_bos_token = prompt_bos_token
self.knowledge_func = knowledge_func
self.label_func = label_func
self.turn_knowledge_func = turn_knowledge_func
self.roles_in_history = roles_in_history
self.multi_ref_sep = multi_ref_sep
self.dev_and_test_roles_to_build_example = dev_and_test_roles_to_build_example
self.cur_turn_process_func = cur_turn_process_func
self.all_turns_process_func = all_turns_process_func
def concat_roles(roles):
return ", ".join(roles)
def concat_dial_history(config: SerialConfig, history: List[Dict]):
# utterance_list = [
# f"{concat_roles(turn[ROLES])}: {turn[UTTERANCE].strip()}"
# if config.use_role
# else turn[UTTERANCE].strip()
# for turn in history
# ]
utterance_list = []
for turn in history:
if (
config.roles_in_history is not None
and turn[ROLES] not in config.roles_in_history
):
continue
if config.use_role:
utterance_list.append(
f"{concat_roles(turn[ROLES])}: {turn[UTTERANCE].strip()}"
)
else:
utterance_list.append(turn[UTTERANCE].strip())
if not utterance_list:
return "None"
turn_sep = " "
if config.turn_sep is not None:
turn_sep = f" {config.turn_sep} "
return turn_sep.join(utterance_list)
def concat_history_knowledge_prompt(
config: SerialConfig, history: str, knowledge: str = "", prompt: str = ""
):
"""Concat `history`, `knowledge` and `prompt`.
NOTE: the order is fixed now.
"""
text = ""
if config.task_bos_token is not None:
text = f"{config.task_bos_token} "
text += history
if knowledge is not None:
text += f" {config.knowledge_bos_token} {knowledge}"
if prompt is not None:
text += f" {config.prompt_bos_token} {prompt}"
return text
def clean(text):
return text.replace("\r\n", " ").replace("\n", " ").replace("\r", " ")
def add_prefix_to_label(prefix, split, label):
tgt = f"{prefix} {label}" if split == "train" else label
return tgt
class SerialPreprocessor(BasePreprocessor):
def __init__(self, config: SerialConfig) -> None:
super().__init__(config)
def extract_knowledge(self, example: Dict):
if self.config.knowledge_func is None:
knowledge = None
elif (
KNOWLEDGE not in example
or not self.config.knowledge_func.__code__.co_argcount
):
knowledge = self.config.knowledge_func()
else:
knowledge = self.config.knowledge_func(example[KNOWLEDGE][VALUE])
return knowledge
def preprocess_for_dialogue_level(self, split: str, example: Dict, knowledge: str):
label = self.config.label_func(example)
tgt = add_prefix_to_label(self.config.task_bos_token, split, label)
history = concat_dial_history(self.config, example[DIALOG])
if self.config.prompt_func is None:
prompt = ""
elif not self.config.prompt_func.__code__.co_argcount:
prompt = self.config.prompt_func()
src = concat_history_knowledge_prompt(self.config, history, knowledge, prompt)
return [{"src": clean(src), "tgt": clean(tgt)}]
def preprocess_for_label_level(self, split: str, example: Dict, knowledge: str):
label_generator = self.config.label_func(example)
examples = []
for turn_id, label, extra_args in label_generator:
tgt = add_prefix_to_label(self.config.task_bos_token, split, label)
hist = deepcopy(example[DIALOG])
if self.config.all_turns_process_func is not None:
hist[turn_id] = self.config.all_turns_process_func(
hist[turn_id], *extra_args
)
history = concat_dial_history(self.config, hist)
if self.config.prompt_func is None:
prompt = ""
elif not self.config.prompt_func.__code__.co_argcount:
prompt = self.config.prompt_func()
src = concat_history_knowledge_prompt(
self.config, history, knowledge, prompt
)
examples.append({"src": clean(src), "tgt": clean(tgt)})
return examples
def get_label(
self, turn, include_current_turn, turn_idx, split, origin_knowledge=None
):
# skip the roles not requiring to build examples
if (
split != TRAIN_SPLIT
and self.config.dev_and_test_roles_to_build_example is not None
):
roles_to_build_example = self.config.dev_and_test_roles_to_build_example
else:
roles_to_build_example = self.config.roles_to_build_example
if (
roles_to_build_example is not None
and turn[ROLES] not in roles_to_build_example
):
return None
# skip the first turn if not including current turn
if not include_current_turn and turn_idx == 0:
return None
if self.config.task != DIALOGUE_STATE_TRACKING:
try:
label = self.config.label_func(turn, split=split)
except:
label = self.config.label_func(turn, origin_knowledge, split=split)
else:
label = self.config.label_func(
turn, self.ontologies[split], do_train=(split == TRAIN_SPLIT)
)
return label
def preprocess_for_turn_level(
self,
split: str,
example: Dict,
knowledge: str,
include_current_turn=False,
origin_knowledge=None,
):
examples = []
multiref = []
for turn_idx, turn in enumerate(example[DIALOG]):
label = self.get_label(
turn, include_current_turn, turn_idx, split, origin_knowledge
)
if label is None:
continue
multiref.append(label)
# requre to merge and arrive at the final consecutive label
if (
self.config.multi_ref_sep is not None
and split != "train"
and turn_idx < len(example[DIALOG]) - 1
and self.get_label(
example[DIALOG][turn_idx + 1],
include_current_turn,
turn_idx + 1,
split,
)
is not None
):
continue
if self.config.multi_ref_sep is not None and split != "train":
label = self.config.multi_ref_sep.join(multiref)
tgt = add_prefix_to_label(self.config.task_bos_token, split, label)
end = (turn_idx + 1) if include_current_turn else turn_idx
hist = deepcopy(example[DIALOG][:end])
if self.config.cur_turn_process_func is not None:
hist[-1] = self.config.cur_turn_process_func(hist[-1])
history = concat_dial_history(self.config, hist)
if self.config.prompt_func is None:
prompt = ""
elif not self.config.prompt_func.__code__.co_argcount:
prompt = self.config.prompt_func()
if self.config.turn_knowledge_func is not None:
knowledge_to_use = self.config.turn_knowledge_func(knowledge, turn)
else:
knowledge_to_use = knowledge
src = concat_history_knowledge_prompt(
self.config, history, knowledge_to_use, prompt
)
examples.append({"src": clean(src), "tgt": clean(tgt)})
multiref = []
return examples
def preprocess_line(self, split: str, example: Dict) -> List[Dict]:
knowledge = self.extract_knowledge(example)
# 1. Dialogue Summary
if self.config.task == DIALOGUE_SUMMARY:
return self.preprocess_for_dialogue_level(split, example, knowledge)
# 2. Emotion Recognition
if self.config.task == EMOTION_RECOGNITION:
return self.preprocess_for_turn_level(
split, example, knowledge, include_current_turn=True
)
# 3. Dialogue Context-to-Text Generation
if self.config.task == DIALOGUE_CONTEXT_TO_RESPONSE_GENERATION:
return self.preprocess_for_turn_level(
split, example, knowledge, include_current_turn=False
)
# 4. ABSA
if self.config.task.startswith(ABSA):
return self.preprocess_for_turn_level(
split, example, knowledge, include_current_turn=True
)
# 5. Character Identification
if self.config.task == CHARACTER_IDENTIFICATION:
# return self.preprocess_for_turn_level(
# split, example, knowledge, include_current_turn=True
# )
# return self.preprocess_for_dialogue_level(split, example, knowledge)
return self.preprocess_for_label_level(split, example, knowledge)
# 6. Dialogue State Tracking
if self.config.task == DIALOGUE_STATE_TRACKING:
return self.preprocess_for_turn_level(
split, example, knowledge, include_current_turn=True
)
# 7. Document Grounded Conversation
if self.config.task == DOCUMENT_GROUNDED_CONVERSATION:
return self.preprocess_for_turn_level(
split, example, knowledge, include_current_turn=False
)
# 8. Text2SQL
if self.config.task == TEXT2SQL:
seq_examples = self.preprocess_for_turn_level(
split, example, knowledge, include_current_turn=True
)
for idx in range(len(seq_examples)):
seq_examples[idx]["db_id"] = knowledge["db_id"]
return seq_examples
# 9. Slot Filling
if self.config.task == SLOT_FILLING:
return self.preprocess_for_turn_level(
split, example, knowledge, include_current_turn=True
)
# 10. Relation Recognition
if self.config.task == ROLE_RELATION_RECOGNITION:
return self.preprocess_for_dialogue_level(split, example, knowledge)
# 11. Question in Context Rewriting
if self.config.task == QUESTION_IN_CONTEXT_REWRITING:
return self.preprocess_for_turn_level(
split, example, knowledge, include_current_turn=True
)
# 12. Natural Language Inference
if self.config.task == NATURAL_LANGUAGE_INFERENCE:
return self.preprocess_for_turn_level(
split,
example,
knowledge,
include_current_turn=True,
origin_knowledge=example[KNOWLEDGE][VALUE],
)
# 13. Machine Reading Comprehension
if self.config.task == MACHINE_READING_COMPREHENSION:
return self.preprocess_for_turn_level(split, example, knowledge)
# 14. Multiple Choice Question Answering
if self.config.task == MULTIPLE_CHOICE_QUESTION_ANSWERING:
return self.preprocess_for_turn_level(
split,
example,
knowledge,
include_current_turn=True,
origin_knowledge=example[KNOWLEDGE][VALUE],
)
# 15. Intent Detection
if self.config.task == INTENT_DETECTION:
return self.preprocess_for_turn_level(
split, example, knowledge, include_current_turn=True
)
# 16. Data-to-Text
if self.config.task == DATA_TO_TEXT:
return self.preprocess_for_turn_level(
split, example, knowledge, include_current_turn=True
)
# 17. Chit-Chat
if self.config.task == CHIT_CHAT:
return self.preprocess_for_turn_level(
split, example, knowledge, include_current_turn=False
)
if self.config.task == "Semantic Parsing":
seq_examples = self.preprocess_for_turn_level(
split, example, knowledge, include_current_turn=True
)
return seq_examples