PROSTLE commited on
Commit
ce198f2
Β·
unverified Β·
1 Parent(s): ebb3ef6

Update sentiment.py

Browse files
Files changed (1) hide show
  1. sentiment.py +1 -28
sentiment.py CHANGED
@@ -1,8 +1,3 @@
1
- """
2
- News sentiment analysis using NLTK VADER + ProsusAI/FinBERT ensemble.
3
- VADER provides fast rule-based scoring; FinBERT provides deep financial context.
4
- The final score is a weighted blend (40% VADER, 60% FinBERT).
5
- """
6
 
7
  import torch
8
  from transformers import BertTokenizer, BertForSequenceClassification
@@ -11,31 +6,24 @@ import yfinance as yf
11
  import nltk
12
  from nltk.sentiment.vader import SentimentIntensityAnalyzer
13
 
14
- # ── Download VADER lexicon if missing ──────────────────────────
15
  try:
16
  nltk.data.find("sentiment/vader_lexicon.zip")
17
  except LookupError:
18
  nltk.download("vader_lexicon", quiet=True)
19
 
20
- # ── Load models once at import ─────────────────────────────────
21
- # VADER
22
  _vader = SentimentIntensityAnalyzer()
23
 
24
- # FinBERT
25
  MODEL_NAME = "ProsusAI/finbert"
26
  _tokenizer = BertTokenizer.from_pretrained(MODEL_NAME)
27
  _finbert = BertForSequenceClassification.from_pretrained(MODEL_NAME)
28
  _finbert.eval()
29
  FINBERT_LABELS = ["negative", "neutral", "positive"]
30
 
31
- # Ensemble weights
32
  VADER_WEIGHT = 0.4
33
  FINBERT_WEIGHT = 0.6
34
 
35
 
36
- # ── Individual scorers ─────────────────────────────────────────
37
  def _score_vader(text: str) -> dict:
38
- """VADER compound score mapped to label + confidence."""
39
  scores = _vader.polarity_scores(text)
40
  compound = scores["compound"]
41
 
@@ -51,7 +39,6 @@ def _score_vader(text: str) -> dict:
51
 
52
 
53
  def _score_finbert(text: str) -> dict:
54
- """FinBERT probability-based label + confidence."""
55
  tokens = _tokenizer(text, return_tensors="pt", padding=True,
56
  truncation=True, max_length=128)
57
  with torch.no_grad():
@@ -59,8 +46,7 @@ def _score_finbert(text: str) -> dict:
59
  probs = torch.nn.functional.softmax(logits, dim=-1).squeeze().tolist()
60
  idx = int(np.argmax(probs))
61
 
62
- # Convert to a signed score: positive=+1, neutral=0, negative=βˆ’1
63
- signed_score = probs[2] - probs[0] # positive_prob - negative_prob
64
  return {
65
  "label": FINBERT_LABELS[idx],
66
  "score": round(signed_score, 4),
@@ -69,14 +55,9 @@ def _score_finbert(text: str) -> dict:
69
 
70
 
71
  def _score_text(text: str) -> dict:
72
- """
73
- Ensemble: blend VADER and FinBERT scores.
74
- Returns per-headline detail with both individual and combined scores.
75
- """
76
  vader = _score_vader(text)
77
  finbert = _score_finbert(text)
78
 
79
- # Weighted blend of signed scores
80
  combined_score = (VADER_WEIGHT * vader["score"]) + (FINBERT_WEIGHT * finbert["score"])
81
 
82
  if combined_score > 0.1:
@@ -98,14 +79,11 @@ def _score_text(text: str) -> dict:
98
  }
99
 
100
 
101
- # ── News fetcher ───────────────────────────────────────────────
102
  def fetch_news_headlines(ticker: str, max_items: int = 10) -> list[str]:
103
- """Fetch recent news headlines from Yahoo Finance."""
104
  stock = yf.Ticker(ticker)
105
  news = stock.news or []
106
  headlines = []
107
  for item in news[:max_items]:
108
- # yfinance v0.2.36+ nests data under "content"
109
  content = item.get("content", {})
110
  title = content.get("title", "") or item.get("title", "")
111
  if title:
@@ -113,11 +91,7 @@ def fetch_news_headlines(ticker: str, max_items: int = 10) -> list[str]:
113
  return headlines
114
 
115
 
116
- # ── Public API ─────────────────────────────────────────────────
117
  def analyze_sentiment(ticker: str) -> dict:
118
- """
119
- End-to-end: fetch headlines ➜ score each with VADER + FinBERT ➜ return aggregate.
120
- """
121
  headlines = fetch_news_headlines(ticker)
122
  if not headlines:
123
  return {
@@ -131,7 +105,6 @@ def analyze_sentiment(ticker: str) -> dict:
131
 
132
  details = [_score_text(h) for h in headlines]
133
 
134
- # Aggregate combined scores
135
  avg = float(np.mean([d["combined_score"] for d in details]))
136
 
137
  if avg > 0.15:
 
 
 
 
 
 
1
 
2
  import torch
3
  from transformers import BertTokenizer, BertForSequenceClassification
 
6
  import nltk
7
  from nltk.sentiment.vader import SentimentIntensityAnalyzer
8
 
 
9
  try:
10
  nltk.data.find("sentiment/vader_lexicon.zip")
11
  except LookupError:
12
  nltk.download("vader_lexicon", quiet=True)
13
 
 
 
14
  _vader = SentimentIntensityAnalyzer()
15
 
 
16
  MODEL_NAME = "ProsusAI/finbert"
17
  _tokenizer = BertTokenizer.from_pretrained(MODEL_NAME)
18
  _finbert = BertForSequenceClassification.from_pretrained(MODEL_NAME)
19
  _finbert.eval()
20
  FINBERT_LABELS = ["negative", "neutral", "positive"]
21
 
 
22
  VADER_WEIGHT = 0.4
23
  FINBERT_WEIGHT = 0.6
24
 
25
 
 
26
  def _score_vader(text: str) -> dict:
 
27
  scores = _vader.polarity_scores(text)
28
  compound = scores["compound"]
29
 
 
39
 
40
 
41
  def _score_finbert(text: str) -> dict:
 
42
  tokens = _tokenizer(text, return_tensors="pt", padding=True,
43
  truncation=True, max_length=128)
44
  with torch.no_grad():
 
46
  probs = torch.nn.functional.softmax(logits, dim=-1).squeeze().tolist()
47
  idx = int(np.argmax(probs))
48
 
49
+ signed_score = probs[2] - probs[0]
 
50
  return {
51
  "label": FINBERT_LABELS[idx],
52
  "score": round(signed_score, 4),
 
55
 
56
 
57
  def _score_text(text: str) -> dict:
 
 
 
 
58
  vader = _score_vader(text)
59
  finbert = _score_finbert(text)
60
 
 
61
  combined_score = (VADER_WEIGHT * vader["score"]) + (FINBERT_WEIGHT * finbert["score"])
62
 
63
  if combined_score > 0.1:
 
79
  }
80
 
81
 
 
82
  def fetch_news_headlines(ticker: str, max_items: int = 10) -> list[str]:
 
83
  stock = yf.Ticker(ticker)
84
  news = stock.news or []
85
  headlines = []
86
  for item in news[:max_items]:
 
87
  content = item.get("content", {})
88
  title = content.get("title", "") or item.get("title", "")
89
  if title:
 
91
  return headlines
92
 
93
 
 
94
  def analyze_sentiment(ticker: str) -> dict:
 
 
 
95
  headlines = fetch_news_headlines(ticker)
96
  if not headlines:
97
  return {
 
105
 
106
  details = [_score_text(h) for h in headlines]
107
 
 
108
  avg = float(np.mean([d["combined_score"] for d in details]))
109
 
110
  if avg > 0.15: