ACT Selection: Answer Candidate Type Selection

Answer Candidate Type Selection: Text-To-Text Language Model for Closed Book Question Answering Meets Knowledge Graphs
Mikhail Salnikov, Maria Lysyuk, Pavel Braslavski, Anton Razzhigaev, Valentin Malykh, Alexander Panchenko
Proceedings of KONVENS 2023, pp. 155–164

📌 Overview

ACT Selection is a lightweight post-processing method that improves Knowledge Graph Question Answering (KGQA) by filtering and re-ranking answer candidates generated by pre-trained Text-to-Text models (e.g., T5, BART).

The key insight: even when a closed-book LM generates an incorrect answer, it often predicts the correct answer type. We leverage Wikidata's instance_of (P31) property to extract candidate types and rerank answers using a simple scoring function.

🔧 Pipeline Components

1. Candidate Generation: Diverse Beam Search over a fine-tuned seq2seq model produces an initial list of answer candidates.
2. Answer Type Extraction: Aggregate instance_of types from candidates; merge semantically similar types using Sentence-BERT.
3. Entity Linking: Extract question entities via fine-tuned spaCy NER + mGENRE; enrich candidates with one-hop Wikidata neighbors.
4. Candidate Scoring: Rank candidates using a weighted sum of four signals:
   • S_type: Intersection between candidate types and predicted answer types
   • S_neighbour: Binary score if candidate is a 1-hop neighbor of question entities
   • S_t2t: Rank from the original Text-to-Text model output
   • S_property: Cosine similarity between question and candidate property (Sentence-BERT)

🚀 Usage

The full implementation is integrated into the M3M demo:

🔗 Main pipeline code:
app/pipelines/act_selection.py

Available Endpoints (FastAPI)

Endpoint	Description
POST /pipeline/act_selection/ner	NER + sentence insertion
POST /pipeline/act_selection/mgenre	Entity linking via mGENRE
POST /pipeline/act_selection/seq2seq	Raw Text-to-Text generation
POST /pipeline/act_selection/main	Full ACT Selection pipeline (with all scores)
POST /pipeline/act_selection/simple_type_selection/	Lightweight version (type + neighbor scores only)
POST /pipeline/act_selection/simple_with_description_qustion_similarity_type_selection/	Extended version with description-question similarity

Example Request

import requests

response = requests.post(
    "http://localhost:8000/pipeline/act_selection/main",
    json={"text": "Who published Neo Contra?"}
)
result = response.json()
print(result["answers"][:5])  # Top-5 ranked Wikidata entities

Cite Bibtex

@inproceedings{salnikov-etal-2023-answer,
    title = "Answer Candidate Type Selection: Text-To-Text Language Model for Closed Book Question Answering Meets Knowledge Graphs",
    author = "Salnikov, Mikhail  and
      Lysyuk, Maria  and
      Braslavski, Pavel  and
      Razzhigaev, Anton  and
      Malykh, Valentin A.  and
      Panchenko, Alexander",
    editor = "Georges, Munir  and
      Herygers, Aaricia  and
      Friedrich, Annemarie  and
      Roth, Benjamin",
    booktitle = "Proceedings of the 19th Conference on Natural Language Processing (KONVENS 2023)",
    month = sep,
    year = "2023",
    address = "Ingolstadt, Germany",
    publisher = "Association for Computational Lingustics",
    url = "https://aclanthology.org/2023.konvens-main.16/",
    pages = "155--164"
}