AlignScoreCS

A MultiTask multilingual model is developed to assess factual consistency in context-claim pairs across various Natural Language Understanding (NLU) tasks, including Summarization, Question Answering (QA), Semantic Textual Similarity (STS), Paraphrase, Fact Verification (FV), and Natural Language Inference (NLI). AlignScoreCS is fine-tuned on a vast multi-task dataset consisting of 7 million documents, encompassing these NLU tasks in both Czech and English languages. Its multilingual pre-training enables its potential utilization in various other languages. The architecture is capable of processing tasks using regression, binary classification, or ternary classification, although for evaluation purposes, we recommend employing the AlignScore function.

This work is influenced by its English counterpart AlignScore: Evaluating Factual Consistency with a Unified Alignment Function. However, we employed homogeneous batches instead of heterogeneous ones during training and utilized three distinct architectures sharing a single encoder. This setup allows for the independent use of each architecture with its classification head.

Evaluation

As in the paper AlignScore, we use their AlignScore function which chunk context into roughly 350 tokens and splits claim into sentences each context chunk is evaluated against each claim sentence and aggregated one consistency score

AlignScoreCS model is built on three XLM-RoBERTa architectures sharing one encoder

MultiTask multilingual model for assessing facticity in various NLU tasks in Czech and English language. We followed the initial paper AlignScore https://arxiv.org/abs/2305.16739. We trained a model using a shared architecture of checkpoint xlm-roberta-large xlm-roberta with three linear layers for regression, binary classification and ternary classification.

Usage

  # Assuming you copied the attached Files_and_versions/AlignScore.py file for ease of use in transformers.
  from AlignScoreCS import AlignScoreCS
  alignScoreCS = AlignScoreCS.from_pretrained("krotima1/AlignScoreCS")
  # put the model to cuda to accelerate
  print(alignScoreCS.score(context="This is context", claim="This is claim"))

Results

Training datasets

The following table shows datasets that has been utilized for training the model. We translated these english datasets to Czech using seamLessM4t.

NLP Task	Dataset	Training Task	Context (n words)	Claim (n words)	Sample Count
NLI	SNLI	3-way	10	13	Cs: 500k
					En: 550k
	MultiNLI	3-way	16	20	Cs: 393k
					En: 393k
	Adversarial NLI	3-way	48	54	Cs: 163k
					En: 163k
	DocNLI	2-way	97	285	Cs: 200k
					En: 942k
Fact Verification	NLI-style FEVER	3-way	48	50	Cs: 208k
					En: 208k
	Vitamin C	3-way	23	25	Cs: 371k
					En: 371k
Paraphrase	QQP	2-way	9	11	Cs: 162k
					En: 364k
	PAWS	2-way	-	18	Cs: -
					En: 707k
	PAWS labeled	2-way	18	-	Cs: 49k
					En: -
	PAWS unlabeled	2-way	18	-	Cs: 487k
					En: -
STS	SICK	reg	-	10	Cs: -
					En: 4k
	STS Benchmark	reg	-	10	Cs: -
					En: 6k
	Free-N1	reg	18	-	Cs: 20k
					En: -
QA	SQuAD v2	2-way	105	119	Cs: 130k
					En: 130k
	RACE	2-way	266	273	Cs: 200k
					En: 351k
Information Retrieval	MS MARCO	2-way	49	56	Cs: 200k
					En: 5M
Summarization	WikiHow	2-way	434	508	Cs: 157k
					En: 157k
	SumAug	2-way	-	-	Cs: -
					En: -