src/modules/cohesion.py

import math
import re

import numpy as np
from sklearn.metrics.pairwise import cosine_similarity
from sentence_transformers import SentenceTransformer


class CohesionModule:
    def __init__(self):
        self.model = SentenceTransformer('all-MiniLM-L6-v2')

        self.conn_map_en = {
            "AdPos": ["also", "moreover", "furthermore", "in addition", "plus",
                      "equally", "likewise", "similarly", "as well", "alike",
                      "additionally", "what's more", "not only", "coupled with",
                      "along with", "including", "together with", "in conjunction with",
                      "along the same lines", "correspondingly"],
            "AdNeg": ["instead", "conversely", "rather", "on the contrary",
                      "alternatively", "nonetheless", "however", "nevertheless",
                      "by contrast", "on the other hand", "in contrast", "unlike",
                      "despite this", "notwithstanding", "even so", "still",
                      "be that as it may", "in spite of", "rather", "whereas"],
            "CausPos": ["therefore", "hence", "thus", "consequently", "accordingly",
                        "so", "as a result", "because of this", "for this reason",
                        "since", "given that", "inasmuch as", "seeing that", "seeing as",
                        "in view of", "in light of", "due to", "owing to", "stemming from", "insofar as"],
            "CausNeg": ["nevertheless", "however", "despite", "in spite of",
                        "notwithstanding", "still", "yet", "albeit",
                        "even though", "although", "whereas", "while",
                        "in contrast", "on the other hand", "conversely",
                        "be that as it may", "nonetheless", "regardless",
                        "irrespective of", "granted that"],
            "TempPos": ["then", "subsequently", "afterwards", "later", "next",
                        "thereafter", "following this", "after that", "meanwhile",
                        "subsequently", "instantly", "immediately", "directly",
                        "thereupon", "forthwith", "straightaway", "presently",
                        "quickly", "in due course", "momentarily"],
            "TempNeg": ["previously", "beforehand", "earlier", "prior",
                        "until", "before", "ahead of time",
                        "in advance", "beforehand", "previously", "heretofore",
                        "ere", "antecedently", "in anticipation", "preparatory",
                        "in preparation", "in readiness", "in prospect", "preemptively"],
            "LogPos": ["indeed", "actually", "in fact", "clearly",
                       "evidently", "obviously", "certainly", "surely",
                       "undoubtedly", "unquestionably", "without doubt",
                       "assuredly", "definitely", "absolutely", "positively",
                       "categorically", "decidedly", "indubitably", "unmistakably", "truly"],
            "LogNeg": ["conversely", "on the contrary", "inversely",
                       "contrariwise", "alternatively", "rather",
                       "by contrast", "in contrast", "on the other hand",
                       "oppositely", "counter to", "in opposition",
                       "the other way around", "in a different light",
                       "from another perspective", "viewed differently",
                       "seen from another angle", "to the contrary",
                       "in another sense", "constructively"]
        }

        self.conn_map_it = {
            "AdPos": ["inoltre", "in aggiunta", "ugualmente", "analogamente", "allo stesso modo",
                      "parimenti", "similmente", "per di più", "oltretutto", "altresì",
                      "anche", "pure", "persino", "perfino", "nonché",
                      "oltre a ciò", "a questo si aggiunge", "non solo", "insieme a",
                      "in più", "tanto quanto", "al pari di", "così come", "d'altronde"],
            "AdNeg": ["invece", "al contrario", "per contro", "d'altro canto", "viceversa",
                      "tuttavia", "però", "eppure", "nondimeno", "ciononostante",
                      "nonostante ciò", "al tempo stesso", "in contrasto", "a differenza di",
                      "diversamente", "contrariamente", "piuttosto", "in realtà", "anzi",
                      "bensì", "ma", "senonché", "fatta eccezione per", "a meno che"],
            "CausPos": ["quindi", "pertanto", "di conseguenza", "dunque", "perciò",
                        "per questo motivo", "per tale ragione", "così", "ecco perché",
                        "ne consegue che", "da ciò deriva", "in virtù di ciò", "a seguito di",
                        "per effetto di", "grazie a", "poiché", "dato che", "siccome",
                        "visto che", "dal momento che"],
            "CausNeg": ["nonostante", "sebbene", "benché", "malgrado", "pur",
                        "anche se", "quantunque", "per quanto", "a dispetto di",
                        "in barba a", "con tutto che", "pur tuttavia", "nel caso in cui",
                        "fermo restando che", "ciononostante", "neppure così", "eppure",
                        "tuttavia", "nondimeno", "ugualmente"],
            "TempPos": ["poi", "successivamente", "in seguito", "dopo", "dopodiché",
                        "in seguito a", "più tardi", "nel frattempo", "subito dopo",
                        "immediatamente", "a breve", "poco dopo", "contemporaneamente",
                        "nel contempo", "alla fine", "infine", "per ultimo", "da quel momento",
                        "a partire da allora", "in quello stesso istante"],
            "TempNeg": ["prima", "precedentemente", "in anticipo", "in precedenza", "già",
                        "tempo prima", "in passato", "fino a quel momento", "sino ad allora",
                        "antecedentemente", "prima ancora", "fin dall'inizio", "ab initio",
                        "precocemente", "fin da subito", "in anticipo su", "prima di",
                        "anteriormente", "per l'appunto prima", "al momento precedente"],
            "LogPos": ["infatti", "in effetti", "certamente", "ovviamente", "naturalmente",
                       "di certo", "senza dubbio", "evidentemente", "chiaramente",
                       "indubbiamente", "innegabilmente", "senz'altro", "con certezza",
                       "in verità", "invero", "difatti", "com'è noto", "come si sa",
                       "come è ovvio", "è indubbio che"],
            "LogNeg": ["al contrario", "per contro", "inversamente", "d'altra parte",
                       "in senso opposto", "in modo inverso", "ribaltando il ragionamento",
                       "al posto di", "piuttosto che", "contrariamente a quanto detto",
                       "in modo contrario", "a rovescio", "viceversa", "nel senso opposto",
                       "cambiando prospettiva", "vista da un altro angolo",
                       "in tutt'altro senso", "in senso contrario", "anziché", "in alternativa"]
        }

    def _get_jaccard(self, sets):
        """Calcola l'indice di Jaccard per una lista di N set."""
        if not all(sets):
            return 0.0
        intersection = set.intersection(*sets)
        union = set.union(*sets)
        return len(intersection) / len(union) if len(union) > 0 else 0.0

    def analyze(self, doc, sentences, tokens_text, token_text_SP, lang="en"):
        # Seleziona il dizionario connettivi in base alla lingua
        self.conn_map = self.conn_map_it if lang == "it" else self.conn_map_en

        # 1. RAGGRUPPAMENTO IN PARAGRAFI
        paragraphs = []
        current_para = []
        for sent in sentences:
            current_para.append(sent)
            if "\n" in sent.text_with_ws:
                paragraphs.append(current_para)
                current_para = []
        if current_para:
            paragraphs.append(current_para)

        # 2. ACCUMULATORI
        lex_scores_2, lex_scores_3 = [], []
        sem_scores_sent, sem_scores_para = [], []
        para_embeddings = []

        # 3. ANALISI PER PARAGRAFO
        for paragraph in paragraphs:
            if len(paragraph) < 2:
                para_embeddings.append(self.model.encode(paragraph[0].text))
                continue

            target_pos = {"NOUN", "VERB", "ADJ", "PROPN"}
            sent_lemmas = [
                {t.lemma_.lower() for t in s if t.pos_ in target_pos and not t.is_punct}
                for s in paragraph
            ]

            # --- Analisi Lessicale ---
            for i in range(len(sent_lemmas) - 1):
                lex_scores_2.append(self._get_jaccard([sent_lemmas[i], sent_lemmas[i + 1]]))

            if len(sent_lemmas) >= 3:
                for i in range(len(sent_lemmas) - 2):
                    lex_scores_3.append(self._get_jaccard([sent_lemmas[i], sent_lemmas[i + 1], sent_lemmas[i + 2]]))

            # --- Analisi Semantica (Livello Frase) ---
            texts = [s.text for s in paragraph]
            sent_embs = self.model.encode(texts)
            for i in range(len(sent_embs) - 1):
                sim = cosine_similarity(sent_embs[i].reshape(1, -1), sent_embs[i + 1].reshape(1, -1))[0][0]
                sem_scores_sent.append(float(sim))

            para_full_text = " ".join(texts)
            para_embeddings.append(self.model.encode(para_full_text))

        # --- 4. ANALISI SEMANTICA (Livello Paragrafo) ---
        if len(para_embeddings) > 1:
            for i in range(len(para_embeddings) - 1):
                sim = cosine_similarity(para_embeddings[i].reshape(1, -1), para_embeddings[i + 1].reshape(1, -1))[0][0]
                sem_scores_para.append(float(sim))

        full_text_lower = doc.text.lower()
        conns = {}
        for cat, words in self.conn_map.items():
            count = 0
            for phrase in words:
                pattern = rf"\b{re.escape(phrase.lower())}\b"
                count += len(re.findall(pattern, full_text_lower))
            conns[cat] = count

        wsum = (
            (conns.get("AdPos", 0) + conns.get("AdNeg", 0) * 1.5) +
            (conns.get("CausPos", 0) + conns.get("CausNeg", 0) * 2) +
            (conns.get("TempPos", 0) + conns.get("TempNeg", 0)) +
            (conns.get("LogPos", 0) + conns.get("LogNeg", 0) * 1.5)
        )
        valori = list(conns.values())
        stand_dev = np.std(valori) if valori else 0

        generalValue = math.log(stand_dev, wsum) if wsum > 0 and wsum != 1 and stand_dev > 0 else 0.0

        return {
            "connectors": {k: round((v / len(token_text_SP)) * 1000, 2) for k, v in conns.items()},
            "general_cohesion": round(generalValue, 4) if wsum > 0 else 0.0,
            "lexical_cohesion_local": np.mean(lex_scores_2) if lex_scores_2 else 0.0,
            "lexical_cohesion_global": np.mean(lex_scores_3) if lex_scores_3 else 0.0,
            "semantic_cohesion_sentences": np.mean(sem_scores_sent) if sem_scores_sent else 0.0,
            "semantic_cohesion_paragraphs": np.mean(sem_scores_para) if sem_scores_para else 0.0,
            "stats": {
                "num_paragraphs": len(paragraphs),
                "num_sentences": len(sentences)
            }
        }