How biba10/generative-cross-lingual-absa works
This repository implements multilingual aspect-based sentiment analysis (ABSA): reviews are loaded and formatted for a selected task, then analyzed by a generative sequence-to-sequence model or by prompted language models; predictions are evaluated with task-specific F1 metrics. Its principal workflow is training and testing the generative model, with separate open-source LLM and ChatGPT prompting workflows and an instruction-tuning entry point. Relationships for unsampled modules are included only where the README or SOURCE INDEX supports them; the source excerpts do not establish every runtime call.
Main entry points: main.py Train and test · data_loader.py Seq2seq data loader · llm_api_classifier.py ChatGPT classifier · main_llm.py Open LLM run
Components 26
Application workflows
- Train and test
main.py - Open LLM run
main_llm.py - ChatGPT run
main_prompting.py - Instruction tuning
instruction_tuning.py
Data preparation
- Seq2seq data loader
src/data_utils/data_loader.py - Seq2seq dataset
src/data_utils/dataset_seq2seq.py - Data preprocessing
src/data_utils/data_utils.py - LLM dataset
src/data_utils/llm_dataset.py - API dataset
src/data_utils/llm_api_dataset.py
Models and prompting
- Model and tokenizer loader
src/model/model_tokenizer_loader.py - Generative ABSA model
src/model/model_generative.py - Open LLM classifier
src/llm_prompting/llm_classifier.py - ChatGPT classifier
src/llm_prompting/llm_api_classifier.py - LLM API access
src/llm_prompting/llm_access.py - Prompt templates
src/llm_prompting/templates/prompt_templates.py - Language few-shot prompts
src/llm_prompting/templates
Evaluation and metrics
- Seq2seq evaluation
src/evaluation/evaluation.py - Seq2seq F1
src/evaluation/f1_score_seq2seq.py - Combined task F1
src/evaluation/total_f1_score.py - Prompted-model evaluation
src/evaluation/llm_evaluation.py - General F1
src/evaluation/f1_score.py
Configuration and logging
- CLI arguments
src/utils/args_utils.py - Task and language config
src/utils/config.py - Experiment logging
src/utils/logger_utils.py
Other components
- Experimenter
- ChatGPT service
Connections 34
- Experimenter → Train and test: runs training
README.md - Experimenter → Open LLM run: runs prompting
README.md - Experimenter → ChatGPT run: runs prompting
README.md - Experimenter → Instruction tuning: runs tuning
README.md - Train and test → CLI arguments: parses arguments
main.py - Train and test → Model and tokenizer loader: loads model
main.py - Train and test → Seq2seq data loader: creates loader
main.py - Train and test → Generative ABSA model: fits and tests
main.py - Train and test → Experiment logging: configures logging
main.py - Open LLM run → Seq2seq data loader: builds data loader
main_llm.py - Open LLM run → Open LLM classifier: runs classifier
main_llm.py - Open LLM run → Model and tokenizer loader: loads tokenizer
main_llm.py - ChatGPT run → ChatGPT classifier: runs classifier
main_prompting.py - ChatGPT run → Seq2seq data loader: builds data loader
main_prompting.py - Instruction tuning → LLM dataset: uses training data
instruction_tuning.py - Instruction tuning → Open LLM classifier: uses classifier
instruction_tuning.py - Instruction tuning → CLI arguments: parses arguments
instruction_tuning.py - Seq2seq data loader → Seq2seq dataset: creates datasets
src/data_utils/data_loader.py - Seq2seq data loader → LLM dataset: creates datasets
src/data_utils/data_loader.py - Seq2seq data loader → API dataset: creates datasets
src/data_utils/data_loader.py - LLM dataset → Data preprocessing: encodes prompts
src/data_utils/llm_dataset.py - LLM dataset → Prompt templates: gets instructions
src/data_utils/llm_dataset.py - Prompt templates → Language few-shot prompts: selects templates
src/llm_prompting/templates/prompt_templates.py - Model and tokenizer loader → Generative ABSA model: constructs model
src/model/model_tokenizer_loader.py - Generative ABSA model → Seq2seq evaluation: evaluates predictions
src/model/model_generative.py - Seq2seq evaluation → Seq2seq F1: updates metrics
src/evaluation/evaluation.py - Seq2seq evaluation → Combined task F1: updates aggregate
src/evaluation/evaluation.py - Seq2seq F1 → General F1: extends metric
src/evaluation/f1_score_seq2seq.py - Combined task F1 → General F1: uses metric
src/evaluation/total_f1_score.py - Open LLM classifier → Seq2seq F1: scores predictions
src/llm_prompting/llm_classifier.py - ChatGPT classifier → LLM API access: requests completions
src/llm_prompting/llm_api_classifier.py - ChatGPT classifier → Prompt templates: builds prompts
src/llm_prompting/llm_api_classifier.py - ChatGPT classifier → Prompted-model evaluation: evaluates predictions
src/llm_prompting/llm_api_classifier.py - ChatGPT classifier → ChatGPT service: sends requests