app.py

import json

import gradio as gr
from pie_modules.models import *  # noqa: F403
from pie_modules.taskmodules import *  # noqa: F403
from pytorch_ie.annotations import LabeledSpan
from pytorch_ie.auto import AutoPipeline
from pytorch_ie.documents import TextDocumentWithLabeledSpansBinaryRelationsAndLabeledPartitions
from pytorch_ie.models import *  # noqa: F403
from pytorch_ie.taskmodules import *  # noqa: F403

def render_pretty_table(
    document: TextDocumentWithLabeledSpansBinaryRelationsAndLabeledPartitions, **render_kwargs
):
    from prettytable import PrettyTable

    t = PrettyTable()
    t.field_names = ["head", "tail", "relation"]
    t.align = "l"
    for relation in list(document.binary_relations) + list(document.binary_relations.predictions):
        t.add_row([str(relation.head), str(relation.tail), relation.label])

    html = t.get_html_string(format=True)
    html = "<div style='max-width:100%; max-height:360px; overflow:auto'>" + html + "</div>"

    return html


def render_spacy(
    document: TextDocumentWithLabeledSpansBinaryRelationsAndLabeledPartitions,
    style="ent",
    inject_relations=True,
    **render_kwargs,
):
    from spacy import displacy

    spans = list(document.labeled_spans) + list(document.labeled_spans.predictions)
    spacy_doc = {
        "text": document.text,
        "ents": [
            {"start": entity.start, "end": entity.end, "label": entity.label} for entity in spans
        ],
        "title": None,
    }

    html = displacy.render(
        spacy_doc, page=True, manual=True, minify=True, style=style, **render_kwargs
    )
    html = "<div style='max-width:100%; max-height:360px; overflow:auto'>" + html + "</div>"
    if inject_relations:
        print("Injecting relation data")
        binary_relations = list(document.binary_relations) + list(
            document.binary_relations.predictions
        )
        sorted_entities = sorted(spans, key=lambda x: (x.start, x.end))
        html = inject_relation_data(
            html, sorted_entities=sorted_entities, binary_relations=binary_relations
        )
    else:
        print("Not injecting relation data")
    return html


def inject_relation_data(html: str, sorted_entities, binary_relations) -> str:
    from bs4 import BeautifulSoup

    # Parse the HTML using BeautifulSoup
    soup = BeautifulSoup(html, "html.parser")

    # Add unique IDs to each entity
    entities = soup.find_all(class_="entity")
    entity2id = {}
    for idx, entity in enumerate(entities):
        entity["id"] = f"entity-{idx}"
        entity["data-original-color"] = (
            entity["style"].split("background:")[1].split(";")[0].strip()
        )
        entity_annotation = sorted_entities[idx]
        # sanity check
        if str(entity_annotation) != entity.next:
            raise ValueError(f"Entity text mismatch: {entity_annotation} != {entity.text}")
        entity2id[sorted_entities[idx]] = f"entity-{idx}"

    # Generate prefixed relations
    prefixed_relations = [
        {
            "head": entity2id[relation.head],
            "tail": entity2id[relation.tail],
            "label": relation.label,
        }
        for relation in binary_relations
    ]

    # Create the JavaScript function to handle mouse over and mouse out events
    script = (
        """
    <script>
    function highlightRelations(entityId, relations) {
        // Reset all entities' styles
        const entities = document.querySelectorAll('.entity');
        entities.forEach(entity => {
            entity.style.backgroundColor = entity.getAttribute('data-original-color');
            entity.style.color = '';
        });

        // If an entity is hovered, highlight it and its related entities with different colors
        if (entityId !== null) {
            const selectedEntity = document.getElementById(entityId);
            if (selectedEntity) {
                selectedEntity.style.backgroundColor = '#ffa';
                selectedEntity.style.color = '#000';
            }

            relations.forEach(relation => {
                if (relation.head === entityId) {
                    const tailEntity = document.getElementById(relation.tail);
                    if (tailEntity) {
                        tailEntity.style.backgroundColor = '#aff';
                        tailEntity.style.color = '#000';
                    }
                }
                if (relation.tail === entityId) {
                    const headEntity = document.getElementById(relation.head);
                    if (headEntity) {
                        headEntity.style.backgroundColor = '#faf';
                        headEntity.style.color = '#000';
                    }
                }
            });
        }
    }

    // Event listeners for mouse over and mouse out on each entity
    document.addEventListener('DOMContentLoaded', (event) => {
        const relations = %s;
        const entities = document.querySelectorAll('.entity');
        entities.forEach(entity => {
            entity.addEventListener('mouseover', () => {
                highlightRelations(entity.id, relations);
            });
            entity.addEventListener('mouseout', () => {
                highlightRelations(null, relations);
            });
        });
    });
    </script>
    """
        % prefixed_relations
    )

    # Inject the script into the HTML
    soup.body.append(BeautifulSoup(script, "html.parser"))

    # Return the modified HTML as a string
    return str(soup)


def predict(text, render_as, render_kwargs_json):
    document = TextDocumentWithLabeledSpansBinaryRelationsAndLabeledPartitions(text=text)

    # add single partition from the whole text (the model only considers text in partitions)
    document.labeled_partitions.append(LabeledSpan(start=0, end=len(text), label="text"))

    # execute prediction pipeline
    pipeline(document)

    render_kwargs = json.loads(render_kwargs_json)
    if render_as == "Pretty Table":
        html = render_pretty_table(document, **render_kwargs)
    elif render_as == "spaCy":
        html = render_spacy(document, **render_kwargs)
    else:
        raise ValueError(f"Unknown render_as value: {render_as}")

    return html


if __name__ == "__main__":

    model_name_or_path = "ArneBinder/sam-pointer-bart-base-v0.3"
    # local path
    # model_name_or_path = "models/dataset-sciarg/task-ner_re/v0.3/2024-03-01_18-25-32"

    pipeline = AutoPipeline.from_pretrained(model_name_or_path, device=-1, num_workers=0)
    re_pipeline = AutoPipeline.from_pretrained(
        model_name_or_path,
        device=-1,
        num_workers=0,
        # taskmodule_kwargs=dict(create_relation_candidates=True),
    )

    default_render_kwargs = {
        "style": "ent",
        "options": {
            "colors": {"own_claim": "#009933", "background_claim": "#0033cc", "data": "#993399"}
        },
    }

    iface = gr.Interface(
        fn=predict,
        inputs=[
            gr.Textbox(
                lines=20,
                value="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.",
            ),
        ],
        additional_inputs=[
            gr.Dropdown(
                label="Render as",
                choices=["Pretty Table", "spaCy"],
                value="spaCy",
            ),
            gr.Textbox(
                label="Render Arguments",
                lines=5,
                value=json.dumps(default_render_kwargs, indent=2),
            ),
        ],
        additional_inputs_accordion=gr.Accordion(label="Render Options", open=False),
        outputs=["html"],
    )
    iface.launch()