Models

As described in the paper, the model consists of three pipeline components: abstract retrieval, rationale selection, and label prediction. We also run "oracle" experiments where pipeline components are replaced with oracles that always make correct predictions given correct outputs. The examples below show how to make oracle and model predictions, evaluate each pipeline component.

Outline

• Abstract Retrieval
• Rationale Selection
• Label Prediction
• Full pipeline

Abstract Retrieval

Oracle

python verisci/inference/abstract_retrieval/oracle.py \
    --dataset data/claims_dev.jsonl \
    --output abstract_retrieval.jsonl

Additional args:

• --include-nei (default: False) include documents for NOT_ENOUGH_INFO claims.

TF-IDF

python verisci/inference/abstract_retrieval/tfidf.py \
    --corpus data/corpus.jsonl \
    --dataset data/claims_dev.jsonl \
    --k 3 \
    --min-gram 1 \
    --max-gram 2 \
    --output abstract_retrieval.jsonl

Evaluate

python verisci/evalaute/abstract_retrieval.py \
    --dataset data/claims_test.jsonl \
    --abstract-retrieval abstract_retrieval.jsonl

---

Rationale Selection

Oracle

python verisci/inference/rationale_selection/oracle.py \
    --dataset data/claims_dev.jsonl \
    --abstract-retrieval abstract_retrieval.jsonl \
    --output rationale_selection.jsonl

Oracle + TF-IDF

Use oracle if gold rationale for that document is present, otherwise use tfidf.

python verisci/inference/rationale_selection/oracle_tfidf.py \
    --corpus data/corpus.jsonl \
    --dataset data/claims_dev.jsonl \
    --abstract-retrieval abstract_retrieval.jsonl \
    --output rationale_selection.jsonl

First Sentence

python verisci/inference/rationale_selection/first.py \
    --abstract-retrieval abstract_retrieval.jsonl \
    --output rationale_selection.jsonl

Last Sentence

python verisci/inference/rationale_selection/last.py \
    --corpus data/corpus.jsonl \
    --abstract-retrieval abstract_retrieval.jsonl \
    --output rationale_selection.jsonl

TF-IDF

python verisci/inference/rationale_selection/tfidf.py \
    --corpus data/corpus.jsonl \
    --dataset data/claims_dev.jsonl \
    --abstract-retrieval abstract_retrieval.jsonl \
    --min-gram 1 \
    --max-gram 1 \
    --k 2 \
    --output rationale_selection.jsonl

Transformer

python verisci/inference/rationale_selection/transformer.py \
    --corpus data/corpus.jsonl \
    --dataset data/claims_dev.jsonl \
    --abstract-retrieval abstract_retrieval.jsonl \
    --model PATH_TO_MODEL \
    --output-flex rationale_selection_flex.jsonl \
    --output-k2 rationale_selection_k2.jsonl \
    --output-k3 rationale_selection_k3.jsonl \
    --output-k4 rationale_selection_k4.jsonl \
    --output-k5 rationale_selection_k5.jsonl

Additional args:

• --threshold (default: 0.5) the threshold for flex output
• --only-rationale only pass abstract sentences to the model without claim
• --output-* are optional. Only provide them if you need them.

Evaluate

python verisci/evaluate/rationale_selection.py \
    --corpus data/corpus.jsonl \
    --dataset data/claims_dev.jsonl \
    --rationale-selection rationale_selection.jsonl

---

Label Prediction

Oracle

python verisci/inference/label_prediction/oracle.py \
    --dataset data/claims_dev.jsonl \
    --rationale-selection rationale_selection.jsonl \
    --output label_prediction.jsonl

Transformers

python verisci/inference/label_prediction/transformer.py \
    --corpus data/corpus.jsonl \
    --dataset data/claims_dev.jsonl \
    --rationale-selection rationale_selection.jsonl \
    --model PATH_TO_MODEL \
    --output label_prediction.jsonl

• --mode dictates what to pass in the model, options: claims_and_rationale (default), only_claim, only_rationale

Evaluate

python verisci/evaluate/label_prediction.py \
    --corpus data/corpus.jsonl \
    --dataset data/claims_dev.jsonl \
    --label-prediction label_prediction.jsonl

---

Full Pipeline

Merge label and rationale predictions

python verisci/inference/merge_predictions.py \
    --rationale-file prediction/rationale_selection.jsonl \
    --label-file prediction/label_prediction.jsonl \
    --result-file prediction/merged_predictions.jsonl

Evaluate the merged predictions

python verisci/evaluate/pipeline.py \
    --gold data/claims_dev.jsonl \
    --corpus data/corpus.jsonl \
    --prediction prediction/merged_predictions.jsonl \
    --output predictions/metrics.json   # If no `output`, provided, print to console.