import json import logging import os.path import tempfile from functools import partial from typing import List, Optional, Tuple import gradio as gr import pandas as pd from backend import DocumentStore, create_and_annotate_document, get_annotation_from_document from pie_modules.taskmodules import PointerNetworkTaskModuleForEnd2EndRE from pytorch_ie import Pipeline from pytorch_ie.auto import AutoPipeline from pytorch_ie.documents import TextDocumentWithLabeledSpansBinaryRelationsAndLabeledPartitions from rendering_utils import render_displacy, render_pretty_table from transformers import AutoModel, AutoTokenizer, PreTrainedModel, PreTrainedTokenizer logger = logging.getLogger(__name__) RENDER_WITH_DISPLACY = "displaCy + highlighted arguments" RENDER_WITH_PRETTY_TABLE = "Pretty Table" DEFAULT_MODEL_NAME = "ArneBinder/sam-pointer-bart-base-v0.3" DEFAULT_MODEL_REVISION = "76300f8e534e2fcf695f00cb49bba166739b8d8a" # local path # DEFAULT_MODEL_NAME = "models/dataset-sciarg/task-ner_re/v0.3/2024-05-28_23-33-46" # DEFAULT_MODEL_REVISION = None DEFAULT_EMBEDDING_MODEL_NAME = "allenai/scibert_scivocab_uncased" def render_annotated_document( document: TextDocumentWithLabeledSpansBinaryRelationsAndLabeledPartitions, render_with: str, render_kwargs_json: str, ) -> str: render_kwargs = json.loads(render_kwargs_json) if render_with == RENDER_WITH_PRETTY_TABLE: html = render_pretty_table(document, **render_kwargs) elif render_with == RENDER_WITH_DISPLACY: html = render_displacy(document, **render_kwargs) else: raise ValueError(f"Unknown render_with value: {render_with}") return html def wrapped_process_text( text: str, doc_id: str, models: Tuple[Pipeline, Optional[PreTrainedModel], Optional[PreTrainedTokenizer]], document_store: DocumentStore, ) -> Tuple[dict, TextDocumentWithLabeledSpansBinaryRelationsAndLabeledPartitions]: document = create_and_annotate_document( text=text, doc_id=doc_id, models=models, ) document_store.add_document(document) # Return as dict and document to avoid serialization issues return document.asdict(), document def process_uploaded_files( file_names: List[str], models: Tuple[Pipeline, Optional[PreTrainedModel], Optional[PreTrainedTokenizer]], document_store: DocumentStore, ) -> pd.DataFrame: try: for file_name in file_names: if file_name.lower().endswith(".txt"): # read the file content with open(file_name, "r", encoding="utf-8") as f: text = f.read() base_file_name = os.path.basename(file_name) gr.Info(f"Processing file '{base_file_name}' ...") document = create_and_annotate_document(text, base_file_name, models) document_store.add_document(document) else: raise gr.Error(f"Unsupported file format: {file_name}") except Exception as e: raise gr.Error(f"Failed to process uploaded files: {e}") return document_store.overview() def open_accordion(): return gr.Accordion(open=True) def close_accordion(): return gr.Accordion(open=False) def load_argumentation_model(model_name: str, revision: Optional[str] = None) -> Pipeline: try: model = AutoPipeline.from_pretrained( model_name, device=-1, num_workers=0, taskmodule_kwargs=dict(revision=revision), model_kwargs=dict(revision=revision), ) except Exception as e: raise gr.Error(f"Failed to load argumentation model: {e}") gr.Info(f"Loaded argumentation model: model_name={model_name}, revision={revision})") return model def load_embedding_model(model_name: str) -> Tuple[PreTrainedModel, PreTrainedTokenizer]: try: embedding_model = AutoModel.from_pretrained(model_name) embedding_tokenizer = AutoTokenizer.from_pretrained(model_name) except Exception as e: raise gr.Error(f"Failed to load embedding model: {e}") gr.Info(f"Loaded embedding model: model_name={model_name})") return embedding_model, embedding_tokenizer def load_models( model_name: str, revision: Optional[str] = None, embedding_model_name: Optional[str] = None ) -> Tuple[Pipeline, Optional[PreTrainedModel], Optional[PreTrainedTokenizer]]: argumentation_model = load_argumentation_model(model_name, revision) embedding_model = None embedding_tokenizer = None if embedding_model_name is not None and embedding_model_name.strip(): embedding_model, embedding_tokenizer = load_embedding_model(embedding_model_name) return argumentation_model, embedding_model, embedding_tokenizer def select_processed_document( evt: gr.SelectData, processed_documents_df: pd.DataFrame, document_store: DocumentStore, ) -> TextDocumentWithLabeledSpansBinaryRelationsAndLabeledPartitions: row_idx, col_idx = evt.index doc_id = processed_documents_df.iloc[row_idx]["doc_id"] gr.Info(f"Select document: {doc_id}") doc = document_store.get_document(doc_id) return doc def set_relation_types( models: Tuple[Pipeline, Optional[PreTrainedModel], Optional[PreTrainedTokenizer]], default: Optional[List[str]] = None, ) -> gr.Dropdown: arg_pipeline = models[0] if isinstance(arg_pipeline.taskmodule, PointerNetworkTaskModuleForEnd2EndRE): relation_types = arg_pipeline.taskmodule.labels_per_layer["binary_relations"] else: raise gr.Error("Unsupported taskmodule for relation types") return gr.Dropdown( choices=relation_types, label="Relation Types", value=default, multiselect=True, ) def download_processed_documents( document_store: DocumentStore, file_name: str = "processed_documents.json", ) -> str: file_path = os.path.join(tempfile.gettempdir(), file_name) with open(file_path, "w", encoding="utf-8") as f: json.dump(document_store.as_dict(), f, indent=2) return file_path def upload_processed_documents( file_name: str, document_store: DocumentStore, ) -> pd.DataFrame: with open(file_name, "r", encoding="utf-8") as f: processed_documents_json = json.load(f) for _, document_json in processed_documents_json.items(): document_store.add_document_from_dict(document_dict=document_json) return document_store.overview() def main(): example_text = "Scholarly Argumentation Mining (SAM) has recently gained attention due to its potential to help scholars with the rapid growth of published scientific literature. It comprises two subtasks: argumentative discourse unit recognition (ADUR) and argumentative relation extraction (ARE), both of which are challenging since they require e.g. the integration of domain knowledge, the detection of implicit statements, and the disambiguation of argument structure. While previous work focused on dataset construction and baseline methods for specific document sections, such as abstract or results, full-text scholarly argumentation mining has seen little progress. In this work, we introduce a sequential pipeline model combining ADUR and ARE for full-text SAM, and provide a first analysis of the performance of pretrained language models (PLMs) on both subtasks. We establish a new SotA for ADUR on the Sci-Arg corpus, outperforming the previous best reported result by a large margin (+7% F1). We also present the first results for ARE, and thus for the full AM pipeline, on this benchmark dataset. Our detailed error analysis reveals that non-contiguous ADUs as well as the interpretation of discourse connectors pose major challenges and that data annotation needs to be more consistent." print("Loading models ...") argumentation_model, embedding_model, embedding_tokenizer = load_models( model_name=DEFAULT_MODEL_NAME, revision=DEFAULT_MODEL_REVISION, embedding_model_name=DEFAULT_EMBEDDING_MODEL_NAME, ) default_render_kwargs = { "entity_options": { # we need to convert the keys to uppercase because the spacy rendering function expects them in uppercase "colors": { "own_claim".upper(): "#009933", "background_claim".upper(): "#99ccff", "data".upper(): "#993399", } }, "colors_hover": { "selected": "#ffa", # "tail": "#aff", "tail": { # green "supports": "#9f9", # red "contradicts": "#f99", # do not highlight "parts_of_same": None, }, "head": None, # "#faf", "other": None, }, } with gr.Blocks() as demo: document_store_state = gr.State(DocumentStore()) # wrap the pipeline and the embedding model/tokenizer in a tuple to avoid that it gets called models_state = gr.State((argumentation_model, embedding_model, embedding_tokenizer)) with gr.Row(): with gr.Column(scale=1): doc_id = gr.Textbox( label="Document ID", value="user_input", ) doc_text = gr.Textbox( label="Text", lines=20, value=example_text, ) with gr.Accordion("Model Configuration", open=False): model_name = gr.Textbox( label="Model Name", value=DEFAULT_MODEL_NAME, ) model_revision = gr.Textbox( label="Model Revision", value=DEFAULT_MODEL_REVISION, ) embedding_model_name = gr.Textbox( label=f"Embedding Model Name (e.g. {DEFAULT_EMBEDDING_MODEL_NAME})", value=DEFAULT_EMBEDDING_MODEL_NAME, ) load_models_btn = gr.Button("Load Models") load_models_btn.click( fn=load_models, inputs=[model_name, model_revision, embedding_model_name], outputs=models_state, ) predict_btn = gr.Button("Analyse") document_state = gr.State() with gr.Column(scale=1): with gr.Accordion("See plain result ...", open=False) as output_accordion: document_json = gr.JSON(label="Model Output") with gr.Accordion("Render Options", open=False): render_as = gr.Dropdown( label="Render with", choices=[RENDER_WITH_PRETTY_TABLE, RENDER_WITH_DISPLACY], value=RENDER_WITH_DISPLACY, ) render_kwargs = gr.Textbox( label="Render Arguments", lines=5, value=json.dumps(default_render_kwargs, indent=2), ) render_btn = gr.Button("Re-render") rendered_output = gr.HTML(label="Rendered Output") # add_to_index_btn = gr.Button("Add current result to Index") upload_btn = gr.UploadButton( "Upload & Analyse Documents", file_types=["text"], file_count="multiple" ) with gr.Column(scale=1): with gr.Accordion( "Indexed Documents", open=False ) as processed_documents_accordion: processed_documents_df = gr.DataFrame( headers=["id", "num_adus", "num_relations"], interactive=False, ) with gr.Row(): download_processed_documents_btn = gr.DownloadButton("Download") upload_processed_documents_btn = gr.UploadButton( "Upload", file_types=["json"] ) with gr.Accordion("Selected ADU", open=False): selected_adu_id = gr.Textbox(label="ID", elem_id="selected_adu_id") selected_adu_text = gr.Textbox(label="Text") with gr.Accordion("Retrieval Configuration", open=False): min_similarity = gr.Slider( label="Minimum Similarity", minimum=0.0, maximum=1.0, step=0.01, value=0.8, ) top_k = gr.Slider( label="Top K", minimum=2, maximum=50, step=1, value=20, ) retrieve_similar_adus_btn = gr.Button("Retrieve similar ADUs") similar_adus = gr.DataFrame(headers=["doc_id", "adu_id", "score", "text"]) relation_types = set_relation_types( models_state.value, default=["supports", "contradicts"] ) # retrieve_relevant_adus_btn = gr.Button("Retrieve relevant ADUs") relevant_adus = gr.DataFrame( label="Relevant ADUs from other documents", headers=[ "relation", "adu", "reference_adu", "doc_id", "sim_score", "rel_score", ], interactive=False, ) render_event_kwargs = dict( fn=render_annotated_document, inputs=[document_state, render_as, render_kwargs], outputs=rendered_output, ) predict_btn.click(fn=open_accordion, inputs=[], outputs=[output_accordion]).then( fn=wrapped_process_text, inputs=[doc_text, doc_id, models_state, document_store_state], outputs=[document_json, document_state], api_name="predict", ).success( fn=lambda document_store: document_store.overview(), inputs=[document_store_state], outputs=[processed_documents_df], ) render_btn.click(**render_event_kwargs, api_name="render") document_state.change( fn=lambda doc: doc.asdict(), inputs=[document_state], outputs=[document_json], ).success(close_accordion, inputs=[], outputs=[output_accordion]).then( **render_event_kwargs ) upload_btn.upload( fn=open_accordion, inputs=[], outputs=[processed_documents_accordion] ).then( fn=process_uploaded_files, inputs=[upload_btn, models_state, document_store_state], outputs=[processed_documents_df], ) processed_documents_df.select( select_processed_document, inputs=[processed_documents_df, document_store_state], outputs=[document_state], ) download_processed_documents_btn.click( fn=download_processed_documents, inputs=[document_store_state], outputs=[download_processed_documents_btn], ) upload_processed_documents_btn.upload( fn=upload_processed_documents, inputs=[upload_processed_documents_btn, document_store_state], outputs=[processed_documents_df], ) retrieve_relevant_adus_event_kwargs = dict( fn=partial(DocumentStore.get_relevant_adus_df, columns=relevant_adus.headers), inputs=[ document_store_state, selected_adu_id, document_state, min_similarity, top_k, relation_types, ], outputs=[relevant_adus], ) selected_adu_id.change( fn=partial(get_annotation_from_document, annotation_layer="labeled_spans"), inputs=[document_state, selected_adu_id], outputs=[selected_adu_text], ).success(**retrieve_relevant_adus_event_kwargs) retrieve_similar_adus_btn.click( fn=DocumentStore.get_similar_adus_df, inputs=[ document_store_state, selected_adu_id, document_state, min_similarity, top_k, ], outputs=[similar_adus], ) models_state.change( fn=set_relation_types, inputs=[models_state], outputs=[relation_types], ) # retrieve_relevant_adus_btn.click( # **retrieve_relevant_adus_event_kwargs # ) js = """ () => { function maybeSetColor(entity, colorAttributeKey, colorDictKey) { var color = entity.getAttribute('data-color-' + colorAttributeKey); // if color is a json string, parse it and use the value at colorDictKey try { const colors = JSON.parse(color); color = colors[colorDictKey]; } catch (e) {} if (color) { entity.style.backgroundColor = color; entity.style.color = '#000'; } } function highlightRelationArguments(entityId) { const entities = document.querySelectorAll('.entity'); // reset all entities entities.forEach(entity => { const color = entity.getAttribute('data-color-original'); entity.style.backgroundColor = color; entity.style.color = ''; }); if (entityId !== null) { var visitedEntities = new Set(); // highlight selected entity const selectedEntity = document.getElementById(entityId); if (selectedEntity) { const label = selectedEntity.getAttribute('data-label'); maybeSetColor(selectedEntity, 'selected', label); visitedEntities.add(selectedEntity); } // highlight tails const relationTailsAndLabels = JSON.parse(selectedEntity.getAttribute('data-relation-tails')); relationTailsAndLabels.forEach(relationTail => { const tailEntity = document.getElementById(relationTail['entity-id']); if (tailEntity) { const label = relationTail['label']; maybeSetColor(tailEntity, 'tail', label); visitedEntities.add(tailEntity); } }); // highlight heads const relationHeadsAndLabels = JSON.parse(selectedEntity.getAttribute('data-relation-heads')); relationHeadsAndLabels.forEach(relationHead => { const headEntity = document.getElementById(relationHead['entity-id']); if (headEntity) { const label = relationHead['label']; maybeSetColor(headEntity, 'head', label); visitedEntities.add(headEntity); } }); // highlight other entities entities.forEach(entity => { if (!visitedEntities.has(entity)) { const label = entity.getAttribute('data-label'); maybeSetColor(entity, 'other', label); } }); } } function setReferenceAduId(entityId) { // get the textarea element that holds the reference adu id let referenceAduIdDiv = document.querySelector('#selected_adu_id textarea'); // set the value of the input field referenceAduIdDiv.value = entityId; // trigger an input event to update the state var event = new Event('input'); referenceAduIdDiv.dispatchEvent(event); } const entities = document.querySelectorAll('.entity'); entities.forEach(entity => { const alreadyHasListener = entity.getAttribute('data-has-listener'); if (alreadyHasListener) { return; } entity.addEventListener('mouseover', () => { highlightRelationArguments(entity.id); setReferenceAduId(entity.id); }); entity.addEventListener('mouseout', () => { highlightRelationArguments(null); }); entity.setAttribute('data-has-listener', 'true'); }); } """ rendered_output.change(fn=None, js=js, inputs=[], outputs=[]) demo.launch() if __name__ == "__main__": # configure logging logging.basicConfig() main()