FrameNet Semantic Frame Classification

Sentence embedding-based semantic frame retrieval for Turkish lexical concepts using English FrameNet

This Python project explores semantic frame identification for Turkish lexical concepts by matching Turkish synsets and definitions to English semantic frames from FrameNet.

View the Project on GitHub

Unlike traditional classification approaches that require a dedicated classifier, this project uses sentence embeddings and semantic similarity to retrieve candidate frames. It embeds English FrameNet frame definitions and translated Turkish lexical entries into a shared vector space, then ranks frames according to their semantic similarity to each query.

Embedding Models

The project implements two sentence embedding models from Sentence Transformers:

  • MiniLM (all-MiniLM-L6-v2): A lightweight model for efficient semantic similarity search.
  • MPNet (all-mpnet-base-v2): An alternative embedding model for comparing retrieval results.

Both models are used to identify candidate frames for each input record, allowing their ranked predictions to be compared.

Input Format

The file-prediction scripts expect a JSON array containing Turkish lexical entries and their English translations. Each record includes the fields synset_english and definition_english, which are combined to form the semantic search query.

For example:

{
  "synset_english": "run",
  "definition_english": "To move swiftly on foot."
}

The system compares this query against the embedded FrameNet frame definitions to identify semantically related frames.

Semantic Retrieval & Prediction

The prediction scripts calculate semantic similarity between the query embedding and the precomputed FrameNet frame embeddings.

  1. Frame Embedding: FrameNet frame definitions are encoded using a selected Sentence Transformer model.
  2. Query Representation: The English synset and definition are combined and converted into an embedding.
  3. Similarity Ranking: Candidate frames are ranked by semantic similarity to the query.
  4. Candidate Generation: The file-prediction scripts return the top 10 candidate frames for each record, while the interactive terminal interface returns the top five candidates for a user-provided phrase or sentence.

The resulting predictions are stored in JSON format, preserving the original input records alongside the ranked frame candidates.

The project also provides an interactive terminal interface through prediction_terminal.py, allowing users to explore FrameNet without preparing an input file.

python prediction_terminal.py

Enter an English phrase or sentence to retrieve the five most semantically similar FrameNet frames. Type exit or quit to terminate the session.

Technologies

Python · Sentence Transformers · MiniLM · MPNet · PyTorch · FrameNet · Semantic Similarity · Sentence Embeddings

The project investigates how embedding-based retrieval can support semi-automatic semantic frame annotation by leveraging English semantic resources for Turkish lexical-semantic analysis. The retrieved frames serve as candidate annotations for further examination rather than definitive predictions validated against gold-standard labels.