pull down to refresh

I am building a minimal behavior prediction engine called TiresiasIQ (v2) that:
  • Logs user emotions/actions via a Streamlit GUI
  • Extracts keywords, sentiment (TextBlob), and vector embeddings (spaCy)
  • Combines them with normalized temporal cues (hour, weekday)
  • Trains a simple 2-layer FFN using Keras on the sqlite3 db created
all through python Then, given a natural query (e.g. “Will I cry at 11PM?”), it extracts the verb “cry,” embeds it, and predicts likelihood.
                    ┌──────────────────────┐
                    │    User Log Input    │ ◄───── "I feel empty after the breakup"
                    └─────────┬────────────┘
                              │
                ┌────────────▼────────────┐
                │  NLP Feature Extractor  │ (spaCy + TextBlob)
                └────────────┬────────────┘
                             │
      ┌──────────────────────┼────────────────────────────┐
      │                      │                            │
┌─────▼──────┐       ┌───────▼────────┐         ┌─────────▼────────┐
│ Keywords   │       │ Sentiment      │         │ Action Verb      │
│ (Entities, │       │ (Polarity &    │         │ Extraction via   │
│ Lemmas)    │       │ Subjectivity)  │         │ VERB/NOUN Lemmas │
└─────┬──────┘       └───────┬────────┘         └─────────┬────────┘
      │                      │                            │
      └──────────────┬───────┴───────┬────────────────────┘
                     │               │
                     ▼               ▼
           ┌───────────────────────────────┐
           │  Temporal Features Extractor  │ (Hour, Weekday)
           └────────────────┬──────────────┘
                            │
                            ▼
             ┌────────────────────────────┐
             │   Feature Vector Builder   │
             │ [keywords + sentiment +    │
             │  action embedding + time]  │
             └────────────┬───────────────┘
                          ▼
          ┌─────────────────────────────────┐
          │  Feedforward Neural Network     │
          │ keras.Sequential([              │
          │   Dense(32, relu) → Dense(1, σ) │
          │ ])                              │
          └──────────────┬──────────────────┘
                         ▼
               ┌────────────────────┐
               │ Prediction Score % │ ◄──── "Cry: 78% at 11PM"
               └────────────────────┘
(yeah mermaid didn't help)
It’s flat-featured for now (FFN only), but I’m running into some conceptual design questions:
How do you model semantic drift in action verbs over time of day?
E.g. “run” at 6AM = jog, “run” at 11PM = escape, “run” during a breakup = avoid
Given the evolving semantic drift of verbs over time and polysemy in user prompts, I want an idea for a good embedding strategy to capture temporal-semantic shifts in the meaning of actions, especially when grounded in subjective time-of-day behavior patterns like in the example. Any ideas would help but I'll pick up the pace after this year 🙏