635 lines
26 KiB
Python
635 lines
26 KiB
Python
import logging
|
|
import os
|
|
import re
|
|
import uuid
|
|
import random
|
|
from typing import Dict, List, Optional
|
|
|
|
from app.repositories.abc import IFileStorage, IDocumentStore
|
|
from app.services.abc import ISpeakingService, ILLMService, IVideoGeneratorService, ISpeechToTextService
|
|
from app.configs.constants import (
|
|
FieldsAndExercises, GPTModels, TemperatureSettings,
|
|
ELAIAvatars, FilePaths
|
|
)
|
|
from app.helpers import TextHelper
|
|
|
|
|
|
class SpeakingService(ISpeakingService):
|
|
|
|
def __init__(
|
|
self, llm: ILLMService, vid_gen: IVideoGeneratorService,
|
|
file_storage: IFileStorage, document_store: IDocumentStore,
|
|
stt: ISpeechToTextService
|
|
):
|
|
self._llm = llm
|
|
self._vid_gen = vid_gen
|
|
self._file_storage = file_storage
|
|
self._document_store = document_store
|
|
self._stt = stt
|
|
self._logger = logging.getLogger(__name__)
|
|
|
|
# TODO: Is the difficulty in the prompts supposed to be hardcoded? The response is set with
|
|
# either the difficulty in the request or a random one yet the prompt doesn't change
|
|
self._tasks = {
|
|
"task_1": {
|
|
"get": {
|
|
"json_template": {
|
|
"first_topic": "topic 1",
|
|
"second_topic": "topic 2",
|
|
"questions": [
|
|
(
|
|
"Introductory question about the first topic, starting the topic with "
|
|
"'Let's talk about x' and then the question."
|
|
),
|
|
"Follow up question about the first topic",
|
|
"Follow up question about the first topic",
|
|
"Question about second topic",
|
|
"Follow up question about the second topic",
|
|
]
|
|
},
|
|
"prompt": (
|
|
'Craft 5 simple and single questions of easy difficulty for IELTS Speaking Part 1 '
|
|
'that encourages candidates to delve deeply into personal experiences, preferences, or '
|
|
'insights on the topic of "{first_topic}" and the topic of "{second_topic}". '
|
|
'Make sure that the generated question does not contain forbidden subjects in '
|
|
'muslim countries.'
|
|
)
|
|
}
|
|
},
|
|
"task_2": {
|
|
"get": {
|
|
"json_template": {
|
|
"topic": "topic",
|
|
"question": "question",
|
|
"prompts": [
|
|
"prompt_1",
|
|
"prompt_2",
|
|
"prompt_3"
|
|
],
|
|
"suffix": "And explain why..."
|
|
},
|
|
"prompt": (
|
|
'Create a question of medium difficulty for IELTS Speaking Part 2 '
|
|
'that encourages candidates to narrate a personal experience or story related to the topic '
|
|
'of "{topic}". Include 3 prompts that guide the candidate to describe '
|
|
'specific aspects of the experience, such as details about the situation, '
|
|
'their actions, and the reasons it left a lasting impression. Make sure that the '
|
|
'generated question does not contain forbidden subjects in muslim countries.'
|
|
)
|
|
}
|
|
},
|
|
"task_3": {
|
|
"get": {
|
|
"json_template": {
|
|
"topic": "topic",
|
|
"questions": [
|
|
"Introductory question about the topic.",
|
|
"Follow up question about the topic",
|
|
"Follow up question about the topic",
|
|
"Follow up question about the topic",
|
|
"Follow up question about the topic"
|
|
]
|
|
},
|
|
"prompt": (
|
|
'Formulate a set of 5 single questions of hard difficulty for IELTS Speaking Part 3'
|
|
'that encourage candidates to engage in a meaningful discussion on the topic of "{topic}". '
|
|
'Provide inquiries, ensuring they explore various aspects, perspectives, and implications '
|
|
'related to the topic. Make sure that the generated question does not contain forbidden '
|
|
'subjects in muslim countries.'
|
|
)
|
|
}
|
|
},
|
|
}
|
|
|
|
async def get_speaking_part(
|
|
self, part: int, topic: str, difficulty: str, second_topic: Optional[str] = None
|
|
) -> Dict:
|
|
task_values = self._tasks[f'task_{part}']['get']
|
|
|
|
if part == 1:
|
|
task_prompt = task_values["prompt"].format(first_topic=topic, second_topic=second_topic)
|
|
else:
|
|
task_prompt = task_values["prompt"].format(topic=topic)
|
|
|
|
messages = [
|
|
{
|
|
"role": "system",
|
|
"content": (
|
|
'You are a helpful assistant designed to output JSON on this format: '
|
|
f'{task_values["json_template"]}'
|
|
)
|
|
},
|
|
{
|
|
"role": "user",
|
|
"content": task_prompt
|
|
}
|
|
]
|
|
|
|
part_specific = {
|
|
"1": 'The questions should lead to the usage of 4 verb tenses (present perfect, present, past and future).',
|
|
"2": (
|
|
'The prompts must not be questions. Also include a suffix like the ones in the IELTS exams '
|
|
'that start with "And explain why".'
|
|
)
|
|
}
|
|
|
|
if part in {1, 2}:
|
|
messages.append({
|
|
"role": "user",
|
|
"content": part_specific[str(part)]
|
|
})
|
|
|
|
if part in {1, 3}:
|
|
messages.append({
|
|
"role": "user",
|
|
"content": 'They must be 1 single question each and not be double-barreled questions.'
|
|
})
|
|
|
|
fields_to_check = ["first_topic"] if part == 1 else FieldsAndExercises.GEN_FIELDS
|
|
|
|
response = await self._llm.prediction(
|
|
GPTModels.GPT_4_O, messages, fields_to_check, TemperatureSettings.GEN_QUESTION_TEMPERATURE
|
|
)
|
|
|
|
if part == 3:
|
|
# Remove the numbers from the questions only if the string starts with a number
|
|
response["questions"] = [
|
|
re.sub(r"^\d+\.\s*", "", question)
|
|
if re.match(r"^\d+\.", question) else question
|
|
for question in response["questions"]
|
|
]
|
|
|
|
response["type"] = part
|
|
response["difficulty"] = difficulty
|
|
|
|
if part in {2, 3}:
|
|
response["topic"] = topic
|
|
|
|
return response
|
|
|
|
async def grade_speaking_task(self, task: int, answers: List[Dict]) -> Dict:
|
|
request_id = uuid.uuid4()
|
|
self._logger.info(
|
|
f'POST - speaking_task_{task} - Received request to grade speaking task {task}. '
|
|
f'Use this id to track the logs: {str(request_id)} - Request data: {str(answers)}'
|
|
)
|
|
|
|
text_answers = []
|
|
perfect_answers = []
|
|
|
|
if task != 2:
|
|
self._logger.info(
|
|
f'POST - speaking_task_{task} - {str(request_id)} - Received {str(len(answers))} total answers.'
|
|
)
|
|
|
|
for item in answers:
|
|
sound_file_name = FilePaths.AUDIO_FILES_PATH + str(uuid.uuid4())
|
|
|
|
self._logger.info(f'POST - speaking_task_{task} - {request_id} - Downloading file {item["answer"]}')
|
|
|
|
await self._file_storage.download_firebase_file(item["answer"], sound_file_name)
|
|
|
|
self._logger.info(
|
|
f'POST - speaking_task_{task} - {request_id} - '
|
|
f'Downloaded file {item["answer"]} to {sound_file_name}'
|
|
)
|
|
|
|
answer_text = await self._stt.speech_to_text(sound_file_name)
|
|
self._logger.info(f'POST - speaking_task_{task} - {request_id} - Transcripted answer: {answer_text}')
|
|
|
|
text_answers.append(answer_text)
|
|
item["answer"] = answer_text
|
|
os.remove(sound_file_name)
|
|
|
|
# TODO: This will end the grading of all answers if a single one does not have enough words
|
|
# don't know if this is intended
|
|
if not TextHelper.has_x_words(answer_text, 20):
|
|
self._logger.info(
|
|
f'POST - speaking_task_{task} - {request_id} - '
|
|
f'The answer had less words than threshold 20 to be graded. Answer: {answer_text}'
|
|
)
|
|
return self._zero_rating("The audio recorded does not contain enough english words to be graded.")
|
|
|
|
self._logger.info(
|
|
f'POST - speaking_task_{task} - {request_id} - '
|
|
f'Requesting perfect answer for question: {item["question"]}'
|
|
)
|
|
perfect_answers.append(await self._get_perfect_answer(task, item["question"]))
|
|
|
|
if task in {1, 3}:
|
|
self._logger.info(
|
|
f'POST - speaking_task_{task} - {request_id} - Formatting answers and questions for prompt.'
|
|
)
|
|
|
|
formatted_text = ""
|
|
for i, entry in enumerate(answers, start=1):
|
|
formatted_text += f"**Question {i}:**\n{entry['question']}\n\n"
|
|
formatted_text += f"**Answer {i}:**\n{entry['answer']}\n\n"
|
|
|
|
self._logger.info(
|
|
f'POST - speaking_task_{task} - {request_id} - '
|
|
f'Formatted answers and questions for prompt: {formatted_text}'
|
|
)
|
|
questions_and_answers = f'\n\n The questions and answers are: \n\n{formatted_text}'
|
|
else:
|
|
questions_and_answers = f'\n Question: "{answers[0]["question"]}" \n Answer: "{answers[0]["answer"]}"'
|
|
|
|
self._logger.info(f'POST - speaking_task_{task} - {request_id} - Requesting grading of the answer(s).')
|
|
response = await self._grade_task(task, questions_and_answers)
|
|
|
|
self._logger.info(f'POST - speaking_task_{task} - {request_id} - Answer(s) graded: {response}')
|
|
|
|
if task in {1, 3}:
|
|
self._logger.info(
|
|
f'POST - speaking_task_{task} - {request_id} - Adding perfect answer(s) to response.')
|
|
|
|
# TODO: check if it is answer["answer"] instead
|
|
for i, answer in enumerate(perfect_answers, start=1):
|
|
response['perfect_answer_' + str(i)] = answer
|
|
|
|
self._logger.info(
|
|
f'POST - speaking_task_{task} - {request_id} - Adding transcript and fixed texts to response.'
|
|
)
|
|
|
|
for i, answer in enumerate(text_answers, start=1):
|
|
response['transcript_' + str(i)] = answer
|
|
response['fixed_text_' + str(i)] = await self._get_speaking_corrections(answer)
|
|
else:
|
|
response['transcript'] = answers[0]["answer"]
|
|
|
|
self._logger.info(f'POST - speaking_task_{task} - {request_id} - Requesting fixed text.')
|
|
response['fixed_text'] = await self._get_speaking_corrections(answers[0]["answer"])
|
|
self._logger.info(f'POST - speaking_task_{task} - {request_id} - Fixed text: {response["fixed_text"]}')
|
|
|
|
response['perfect_answer'] = perfect_answers[0]["answer"]
|
|
|
|
response["overall"] = self._fix_speaking_overall(response["overall"], response["task_response"])
|
|
self._logger.info(f'POST - speaking_task_{task} - {request_id} - Final response: {response}')
|
|
return response
|
|
|
|
# ==================================================================================================================
|
|
# grade_speaking_task helpers
|
|
# ==================================================================================================================
|
|
|
|
async def _get_perfect_answer(self, task: int, question: str):
|
|
messages = [
|
|
{
|
|
"role": "system",
|
|
"content": (
|
|
'You are a helpful assistant designed to output JSON on this format: {"answer": "perfect answer"}'
|
|
)
|
|
},
|
|
{
|
|
"role": "user",
|
|
"content": (
|
|
'Provide a perfect answer according to ielts grading system to the following '
|
|
f'Speaking Part {task} question: "{question}"'
|
|
)
|
|
}
|
|
]
|
|
|
|
if task == 1:
|
|
messages.append({
|
|
"role": "user",
|
|
"content": 'The answer must be 2 or 3 sentences long.'
|
|
})
|
|
|
|
gpt_model = GPTModels.GPT_4_O if task == 1 else GPTModels.GPT_3_5_TURBO
|
|
|
|
return await self._llm.prediction(
|
|
gpt_model, messages, ["answer"], TemperatureSettings.GRADING_TEMPERATURE
|
|
)
|
|
|
|
async def _grade_task(self, task: int, questions_and_answers: str) -> Dict:
|
|
messages = [
|
|
{
|
|
"role": "system",
|
|
"content": (
|
|
f'You are a helpful assistant designed to output JSON on this format: {self._grade_template()}'
|
|
)
|
|
},
|
|
{
|
|
"role": "user",
|
|
"content": (
|
|
f'Evaluate the given Speaking Part {task} response based on the IELTS grading system, ensuring a '
|
|
'strict assessment that penalizes errors. Deduct points for deviations from the task, and '
|
|
'assign a score of 0 if the response fails to address the question. Additionally, provide '
|
|
'detailed commentary highlighting both strengths and weaknesses in the response.'
|
|
) + questions_and_answers
|
|
}
|
|
]
|
|
|
|
task_specific = {
|
|
"1": (
|
|
'Address the student as "you". If the answers are not 2 or 3 sentences long, warn the '
|
|
'student that they should be.'
|
|
),
|
|
"2": 'Address the student as "you"',
|
|
"3": 'Address the student as "you" and pay special attention to coherence between the answers.'
|
|
}
|
|
|
|
messages.append({
|
|
"role": "user",
|
|
"content": task_specific[str(task)]
|
|
})
|
|
|
|
if task in {1, 3}:
|
|
messages.extend([
|
|
{
|
|
"role": "user",
|
|
"content": (
|
|
'For pronunciations act as if you heard the answers and they were transcripted '
|
|
'as you heard them.'
|
|
)
|
|
},
|
|
{
|
|
"role": "user",
|
|
"content": 'The comments must be long, detailed, justify the grading and suggest improvements.'
|
|
}
|
|
])
|
|
|
|
return await self._llm.prediction(
|
|
GPTModels.GPT_4_O, messages, ["comment"], TemperatureSettings.GRADING_TEMPERATURE
|
|
)
|
|
|
|
@staticmethod
|
|
def _fix_speaking_overall(overall: float, task_response: dict):
|
|
grades = [category["grade"] for category in task_response.values()]
|
|
|
|
if overall > max(grades) or overall < min(grades):
|
|
total_sum = sum(grades)
|
|
average = total_sum / len(grades)
|
|
rounded_average = round(average, 0)
|
|
return rounded_average
|
|
|
|
return overall
|
|
|
|
@staticmethod
|
|
def _zero_rating(comment: str):
|
|
return {
|
|
"comment": comment,
|
|
"overall": 0,
|
|
"task_response": {
|
|
"Fluency and Coherence": {
|
|
"grade": 0.0,
|
|
"comment": ""
|
|
},
|
|
"Lexical Resource": {
|
|
"grade": 0.0,
|
|
"comment": ""
|
|
},
|
|
"Grammatical Range and Accuracy": {
|
|
"grade": 0.0,
|
|
"comment": ""
|
|
},
|
|
"Pronunciation": {
|
|
"grade": 0.0,
|
|
"comment": ""
|
|
}
|
|
}
|
|
}
|
|
|
|
async def _get_speaking_corrections(self, text):
|
|
messages = [
|
|
{
|
|
"role": "system",
|
|
"content": (
|
|
'You are a helpful assistant designed to output JSON on this format: '
|
|
'{"fixed_text": "fixed transcription with no misspelling errors"}'
|
|
)
|
|
},
|
|
{
|
|
"role": "user",
|
|
"content": (
|
|
'Fix the errors in the provided transcription and put it in a JSON. '
|
|
f'Do not complete the answer, only replace what is wrong. \n The text: "{text}"'
|
|
)
|
|
}
|
|
]
|
|
|
|
response = await self._llm.prediction(
|
|
GPTModels.GPT_3_5_TURBO,
|
|
messages,
|
|
["fixed_text"],
|
|
0.2,
|
|
False
|
|
)
|
|
return response["fixed_text"]
|
|
|
|
async def create_videos_and_save_to_db(self, exercises, template, req_id):
|
|
template = await self._create_video_per_part(exercises, template, 1)
|
|
template = await self._create_video_per_part(exercises, template, 2)
|
|
template = await self._create_video_per_part(exercises, template, 3)
|
|
|
|
await self._document_store.save_to_db_with_id("speaking", template, req_id)
|
|
self._logger.info(f'Saved speaking to DB with id {req_id} : {str(template)}')
|
|
|
|
async def _create_video_per_part(self, exercises: List[Dict], template: Dict, part: int):
|
|
avatar = (random.choice(list(ELAIAvatars))).name
|
|
template_index = part - 1
|
|
|
|
# Using list comprehension to find the element with the desired value in the 'type' field
|
|
found_exercises = [element for element in exercises if element.get('type') == part]
|
|
|
|
# Check if any elements were found
|
|
if found_exercises:
|
|
exercise = found_exercises[0]
|
|
self._logger.info(f'Creating video for speaking part {part}')
|
|
if part in {1, 3}:
|
|
questions = []
|
|
for question in exercise["questions"]:
|
|
result = await self._create_video(
|
|
question,
|
|
avatar,
|
|
f'Failed to create video for part {part} question: {str(exercise["question"])}'
|
|
)
|
|
if result is not None:
|
|
video = {
|
|
"text": question,
|
|
"video_path": result["video_path"],
|
|
"video_url": result["video_url"]
|
|
}
|
|
questions.append(video)
|
|
|
|
template["exercises"][template_index]["prompts"] = questions
|
|
if part == 1:
|
|
template["exercises"][template_index]["first_title"] = exercise["first_topic"]
|
|
template["exercises"][template_index]["second_title"] = exercise["second_topic"]
|
|
else:
|
|
template["exercises"][template_index]["title"] = exercise["topic"]
|
|
else:
|
|
result = await self._create_video(
|
|
exercise["question"],
|
|
avatar,
|
|
f'Failed to create video for part {part} question: {str(exercise["question"])}'
|
|
)
|
|
if result is not None:
|
|
template["exercises"][template_index]["prompts"] = exercise["prompts"]
|
|
template["exercises"][template_index]["text"] = exercise["question"]
|
|
template["exercises"][template_index]["title"] = exercise["topic"]
|
|
template["exercises"][template_index]["video_url"] = result["video_url"]
|
|
template["exercises"][template_index]["video_path"] = result["video_path"]
|
|
|
|
if not found_exercises:
|
|
template["exercises"].pop(template_index)
|
|
|
|
return template
|
|
|
|
async def generate_video(
|
|
self, part: int, avatar: str, topic: str, questions: list[str],
|
|
*,
|
|
second_topic: Optional[str] = None,
|
|
prompts: Optional[list[str]] = None,
|
|
suffix: Optional[str] = None,
|
|
):
|
|
params = locals()
|
|
params.pop('self')
|
|
|
|
request_id = str(uuid.uuid4())
|
|
self._logger.info(
|
|
f'POST - generate_video_{part} - Received request to generate video {part}. '
|
|
f'Use this id to track the logs: {request_id} - Request data: " + {params}'
|
|
)
|
|
|
|
part_questions = self._get_part_questions(part, questions, avatar)
|
|
videos = []
|
|
|
|
self._logger.info(f'POST - generate_video_{part} - {request_id} - Creating videos for speaking part {part}.')
|
|
for question in part_questions:
|
|
self._logger.info(f'POST - generate_video_{part} - {request_id} - Creating video for question: {question}')
|
|
result = await self._create_video(
|
|
question,
|
|
avatar,
|
|
'POST - generate_video_{p} - {r} - Failed to create video for part {p} question: {q}'.format(
|
|
p=part, r=request_id, q=question
|
|
)
|
|
)
|
|
if result is not None:
|
|
self._logger.info(f'POST - generate_video_{part} - {request_id} - Video created')
|
|
self._logger.info(
|
|
f'POST - generate_video_{part} - {request_id} - Uploaded video to firebase: {result["video_url"]}'
|
|
)
|
|
video = {
|
|
"text": question,
|
|
"video_path": result["video_path"],
|
|
"video_url": result["video_url"]
|
|
}
|
|
videos.append(video)
|
|
|
|
if part == 2 and len(videos) == 0:
|
|
raise Exception(f'Failed to create video for part 2 question: {questions[0]}')
|
|
|
|
return self._get_part_response(part, topic, videos, second_topic, prompts, suffix)
|
|
|
|
@staticmethod
|
|
def _get_part_questions(part: int, questions: list[str], avatar: str):
|
|
part_questions: list[str] = []
|
|
|
|
if part == 1:
|
|
id_to_name = {
|
|
"5912afa7c77c47d3883af3d874047aaf": "MATTHEW",
|
|
"9e58d96a383e4568a7f1e49df549e0e4": "VERA",
|
|
"d2cdd9c0379a4d06ae2afb6e5039bd0c": "EDWARD",
|
|
"045cb5dcd00042b3a1e4f3bc1c12176b": "TANYA",
|
|
"1ae1e5396cc444bfad332155fdb7a934": "KAYLA",
|
|
"0ee6aa7cc1084063a630ae514fccaa31": "JEROME",
|
|
"5772cff935844516ad7eeff21f839e43": "TYLER",
|
|
|
|
}
|
|
part_questions.extend(
|
|
[
|
|
"Hello my name is " + id_to_name.get(avatar) + ", what is yours?",
|
|
"Do you work or do you study?",
|
|
*questions
|
|
]
|
|
)
|
|
elif part == 2:
|
|
# Removed as the examiner should not say what is on the card.
|
|
# question = question + " In your answer you should consider: " + " ".join(prompts) + suffix
|
|
part_questions.append(f'{questions[0]}\nYou have 1 minute to take notes.')
|
|
elif part == 3:
|
|
part_questions = questions
|
|
|
|
return part_questions
|
|
|
|
@staticmethod
|
|
def _get_part_response(
|
|
part: int,
|
|
topic: str,
|
|
videos: list[dict],
|
|
second_topic: Optional[str],
|
|
prompts: Optional[list[str]],
|
|
suffix: Optional[str]
|
|
):
|
|
response = {}
|
|
if part == 1:
|
|
response = {
|
|
"prompts": videos,
|
|
"first_title": topic,
|
|
"second_title": second_topic,
|
|
"type": "interactiveSpeaking"
|
|
}
|
|
if part == 2:
|
|
response = {
|
|
"prompts": prompts,
|
|
"title": topic,
|
|
"suffix": suffix,
|
|
"type": "speaking",
|
|
# includes text, video_url and video_path
|
|
**videos[0]
|
|
}
|
|
if part == 3:
|
|
response = {
|
|
"prompts": videos,
|
|
"title": topic,
|
|
"type": "interactiveSpeaking",
|
|
}
|
|
|
|
response["id"] = str(uuid.uuid4())
|
|
return response
|
|
|
|
async def _create_video(self, question: str, avatar: str, error_message: str):
|
|
result = await self._vid_gen.create_video(question, avatar)
|
|
if result is not None:
|
|
sound_file_path = FilePaths.VIDEO_FILES_PATH + result
|
|
firebase_file_path = FilePaths.FIREBASE_SPEAKING_VIDEO_FILES_PATH + result
|
|
url = await self._file_storage.upload_file_firebase_get_url(firebase_file_path, sound_file_path)
|
|
return {
|
|
"video_path": firebase_file_path,
|
|
"video_url": url
|
|
}
|
|
self._logger.error(error_message)
|
|
return None
|
|
|
|
@staticmethod
|
|
def _grade_template():
|
|
return {
|
|
"comment": "extensive comment about answer quality",
|
|
"overall": 0.0,
|
|
"task_response": {
|
|
"Fluency and Coherence": {
|
|
"grade": 0.0,
|
|
"comment": (
|
|
"extensive comment about fluency and coherence, use examples to justify the grade awarded."
|
|
)
|
|
},
|
|
"Lexical Resource": {
|
|
"grade": 0.0,
|
|
"comment": "extensive comment about lexical resource, use examples to justify the grade awarded."
|
|
},
|
|
"Grammatical Range and Accuracy": {
|
|
"grade": 0.0,
|
|
"comment": (
|
|
"extensive comment about grammatical range and accuracy, use examples to justify the "
|
|
"grade awarded."
|
|
)
|
|
},
|
|
"Pronunciation": {
|
|
"grade": 0.0,
|
|
"comment": (
|
|
"extensive comment about pronunciation on the transcribed answer, use examples to justify the "
|
|
"grade awarded."
|
|
)
|
|
}
|
|
}
|
|
} |