refactor: nlp processing unified into a class

Also removed surprise emotion from content endpoint

server/nlp.py

@@ -5,24 +5,37 @@ from transformers import pipeline
 from sentence_transformers import SentenceTransformer
 from sklearn.metrics.pairwise import cosine_similarity
 
+class NLP:
+    def __init__(self, df: pd.DataFrame, title_col: str, content_col: str, topics: dict):
+        self.df = df
+        self.title_col = title_col
+        self.content_col = content_col
+        self.device = 0 if torch.cuda.is_available() else 1
         
-model = SentenceTransformer("all-mpnet-base-v2", device=0 if torch.cuda.is_available() else 1)
+        # Topic model
+        self.topic_model = SentenceTransformer("all-mpnet-base-v2", device=self.device)
 
-def add_emotion_cols(
-        df: pd.DataFrame, 
-        content_col: str
-    ) -> None:
-    emotion_classifier = pipeline(
+        self.topic_labels = list(topics.keys())
+        self.topic_texts = list(topics.values())
+
+        self.topic_embeddings = self.topic_model.encode(
+            self.topic_texts,
+            normalize_embeddings=True,
+        )
+
+        # emotion model
+        self.emotion_classifier = pipeline(
             "text-classification",
             model="j-hartmann/emotion-english-distilroberta-base",
             top_k=None,
             truncation=True,
-        device=0 if torch.cuda.is_available() else -1
+            device=self.device
         )
 
-    texts = df[content_col].astype(str).str.slice(0, 512).tolist()
+    def add_emotion_cols(self) -> None:
+        texts = self.df[self.content_col].astype(str).str.slice(0, 512).tolist()
 
-    results = emotion_classifier(
+        results = self.emotion_classifier(
             texts,
             batch_size=64
         )
@@ -30,50 +43,32 @@ def add_emotion_cols(
         labels = [r["label"] for r in results[0]]
 
         for label in labels:
-        df[f"emotion_{label}"] = [
+            self.df[f"emotion_{label}"] = [
                 next(item["score"] for item in row if item["label"] == label)
                 for row in results
             ]
 
-def add_topic_col(
-        df: pd.DataFrame,
-        title_col: str,
-        content_col: str,
-        domain_topics: dict,
-        confidence_threshold: float = 0.3
-    ) -> None:
-
-    topic_labels = list(domain_topics.keys())
-    topic_texts = list(domain_topics.values())
-
-    topic_embeddings = model.encode(
-        topic_texts,
-        normalize_embeddings=True,
-    )
-
-    titles = df[title_col].fillna("").astype(str)
-    contents = df[content_col].fillna("").astype(str)
+    def add_topic_col(self, confidence_threshold: float = 0.3) -> None:
+        titles = self.df[self.title_col].fillna("").astype(str)
+        contents = self.df[self.content_col].fillna("").astype(str)
 
         texts = [
             f"{title}. {content}" if title else content
             for title, content in zip(titles, contents)
         ]
 
-    text_embeddings = model.encode(
+        text_embeddings = self.topic_model.encode(
             texts,
             normalize_embeddings=True,
         )
 
         # Similarity
-    sims = cosine_similarity(text_embeddings, topic_embeddings)
+        sims = cosine_similarity(text_embeddings, self.topic_embeddings)
 
         # Best match
         best_idx = sims.argmax(axis=1)
         best_score = sims.max(axis=1)
 
-    df["topic"] = [topic_labels[i] for i in best_idx]
-    df["topic_confidence"] = best_score
-
-    df.loc[df["topic_confidence"] < confidence_threshold, "topic"] = "Misc"
-
-    return df
+        self.df["topic"] = [self.topic_labels[i] for i in best_idx]
+        self.df["topic_confidence"] = best_score
+        self.df.loc[self.df["topic_confidence"] < confidence_threshold, "topic"] = "Misc"

server/stat_gen.py

@@ -5,7 +5,7 @@ import datetime
 
 from nltk.corpus import stopwords
 from collections import Counter
-from server.nlp import add_emotion_cols, add_topic_col
+from server.nlp import NLP
 
 DOMAIN_STOPWORDS = {
     "www", "https", "http",
@@ -21,7 +21,7 @@ nltk.download('stopwords')
 EXCLUDE_WORDS = set(stopwords.words('english')) | DOMAIN_STOPWORDS
 
 class StatGen:
-    def __init__(self, posts_df: pd.DataFrame, comments_df: pd.DataFrame, domain_topics: list) -> None:
+    def __init__(self, posts_df: pd.DataFrame, comments_df: pd.DataFrame, domain_topics: dict) -> None:
         posts_df["type"] = "post"
         posts_df["parent_id"] = None
 
@@ -30,6 +30,7 @@ class StatGen:
         self.domain_topics = domain_topics
 
         self.df = pd.concat([posts_df, comments_df])
+        self.nlp = NLP(self.df, "title", "content", domain_topics)
         self._add_extra_cols(self.df)
 
         self.original_df = self.df.copy(deep=True)
@@ -41,8 +42,8 @@ class StatGen:
         df["hour"] = df["dt"].dt.hour
         df["weekday"] = df["dt"].dt.day_name()
         
-        add_emotion_cols(df, "content")
-        add_topic_col(df, "title", "content", self.domain_topics)
+        self.nlp.add_emotion_cols()
+        self.nlp.add_topic_col()
 
     def _tokenize(self, text: str):
         tokens = re.findall(r"\b[a-z]{3,}\b", text)
@@ -192,10 +193,14 @@ class StatGen:
             .reset_index(drop=True)
         )
 
-        # avearge emotion by topic (excluding neutral)
+        emotion_exclusions = [
+            "emotion_neutral",
+            "emotion_surprise"
+        ]
+
         emotion_cols = [
             col for col in self.df.columns
-            if col.startswith("emotion_") and col != "emotion_neutral"
+            if col.startswith("emotion_") and col not in emotion_exclusions
         ]
 
         counts = (