scale-score

Fast and Accurate Factual Inconsistency Detection Over Long Documents

Barrett Martin Lattimer, Patrick Chen, Xinyuan Zhang, Yi Yang

[e-mail]

EMNLP 2023

https://arxiv.org/abs/2310.13189

Overview

Introducing SCALE, an reference-free NLI based factual inconsistency detection method, and ScreenEval, the longest dialogue based dataset for factual inconsistency detection presently available. Both can be found in our paper Fast and Accurate Factual Inconsistency Detection Over Long Documents.

SCALE uses a novel chunking strategy to achieve state-of-the-art factual inconsistency deteciton performance across many NLG domains, tasks, and over long documents (>6k tokens). SCALE's chunking approach enables fast relevant source text retrival over long documents.

SCALE

This metrics outputs the estimated probablility that a hypothesis is supported by a given premise SCALE(premise, hypothesis). Commonly the hypothesis is generated text and the premise is some ground truth text. For example, a premise may be a document and the hypothesis may be a language model generated summary sentence. The score is bounded as follows 0≤SCALE(premise, hypothesis)≤1. A higher score signifies a higher probability the hypothesis is factually consistent with the premise. A lower score signifies the hypothesis is more likely to be factually inconsistent with the premise. It is recommended to use Flan_T5_XL as the base model for the best results.

Install

To use the evaluation metric, first pip install the python module.

pip install scale-score

or install from source

pip install -e .

Score

Running the Metric

Import the score function and load your premises, hypothesies. For scoring, the premise is a list of entire document strings while the hypothesis are single sentences represented as is a list of list of strings. Each premise has a list of associated hypothesis with a one to one mapping based on index (premise_0 -> ['hypothesis_0_0', 'hypothesis_0_1'], premise_1-> ['hypothesis_1_0', 'hypothesis_1_1', 'hypothesis_1_2']).

from scale_score import score

premise = [
    'premise_0',
    'premise_1',
]
hypothesis = [
    ['hypothesis_0_0', 'hypothesis_0_1'],
    ['hypothesis_1_0', 'hypothesis_1_1', 'hypothesis_1_2']
]

results = score(premise, hypothesis)

Where the results correspond to each hypothesis scored with it's respecitve premise

results = [
    SCALE(premise_0, hypothesis_0_0), 
    SCALE(premise_0, hypothesis_0_1), 
    SCALE(premise_1, hypothesis_1_0), 
    SCALE(premise_1, hypothesis_1_1),
    SCALE(premise_1, hypothesis_1_2),
]

You can also use the scorer object to prevent loading the model at every call like so,

from scale_score.scorer import SCALEScorer
scorer = SCALEScorer(size='small', device='cuda')
results = scorer.score(premise, hypothesis)

Arguments

These arguments are the exact same for both score and scorer.score functions except scorer.score does not take in a size or device as that is set up when building the scorer object.

Argument	Type	Default	Description
premise	List[str]	required	premise text, the ground truth
hypothesis	List[List[str]]	required	hypothesis text, usually the text predicted by a model being evaluated
chunk_size	int	1000	The size of the chunks used to perform chunking on the premise
window_size	float	0.25	The percentage of overlap between chunks. 0≤window_size<1
size	str	'xl'	Size of Flan-T5 model, options are 'small', 'base', 'large', 'xl', 'xxl'
device	str	'cuda'	torch device to send the model to.
model_path	str	None	Optional path to a Flan-T5 model to load. Note the corresponding size must be specified in the size argument.
model	T5ForConditionalGeneration	None	Optional model to use for scoring
tokenizer	T5Tokenizer	None	Optional tokenizer to use for scoring

Evaluation

After scoring, use the evaluate_scale function to evaluate the results.

from scale_score.eval import evaluate_scale
from scale_score.scorer import SCALEScorer
scorer = SCALEScorer(size='small', device='cuda')
results = scorer.score(premise, hypothesis)
metrics = evaluate_scale(results)

The arguments for evaluate_scale are as follows

Argument	Type	Default	Description
results	List[float]	required	Output from scale_score score or scorer run
incorrect	List[int]	required	List of labels for summary sentences, 1 for incorrect and 0 for correct
threshold	float	0.5	Threshold used to calculate binary, micro, macro, and weighted f1 scores
out_file	str	None	Optional json filepath to write the metrics to
print_outputs	bool	True	Whether to print the metrics

The metrics that are output are described below.

Metric	Description
pearson	Pearson correlation
spearman	Spearman correlation
kendalltau	Kendall Tau correlation
majority_class_accuracy	Accuracy if we always predict correct
best_accuracy	Best predicted accuracy possible after threshold tuning
best_detection_precision	Best predicted precision possible after threshold tuning f1 score
best_detection_recall	Best predicted recall possible after threshold tuning f1 score
best_detection_f1	Best predicted f1 possible after threshold tuning
accuracy@90%	Accuracy achieved if we want to keep 90% of all correct sentences
accuracy@70%	Accuracy achieved if we want to keep 70% of all correct sentences
threshold_f1	Threshold used to calculate best_detection_f1
threshold_@90%	Threshold used to calculate accuracy@90%
threshold_@70%	Threshold used to calculate accuracy@70%
f1_binary	F1 score of incorrect sentence detection
f1_macro	Average F1 score between correct and incorrect sentence detection
f1_micro	Calculate F1 globally by counting the total true positives, false negatives and false positives
f1_weighted	Calculate F1 for each label, and find their average weighted by support

Retrieve

Running Retrieval

Import the retrieve function and load your premises, hypothesies.

NOTE: Premises are lists of lists in retrieval. Both premises and hypothesis are split down to the sentence or utterance level.

Each premise list has an associated hypothesis list with a one to one mapping based on index.

from scale_score import retrieve

premise = [
    ['premise_0_utt_0', 'premise_0_utt_1', 'premise_0_utt_2'],
    ['premise_1_utt_0', 'premise_1_utt_1'],
]
hypothesis = [
    ['hypothesis_0_0', 'hypothesis_0_1'],
    ['hypothesis_1_0', 'hypothesis_1_1', 'hypothesis_1_2']
]

results = retrieve(premise, hypothesis)

Where the results correspond to a list which has the most relevant premise utterance/sentence and the corresponding score.

You can also use the scorer object to prevent loading the model at every call like so,

from scale_score.scorer import SCALEScorer
scorer = SCALEScorer(size='small', device='cuda')
results = scorer.retrieve(premise, hypothesis)

Arguments

These arguments are the exact same for both retrieve and scorer.retrieve functions except scorer.retrieve does not take in a size or device as that is set up when building the scorer object.

Argument	Type	Default	Description
premise	List[str]	required	premise text, the ground truth
hypothesis	List[List[str]]	required	hypothesis text, usually the text predicted by a model being evaluated
branches	int	2	The number of branches to have in the search tree
size	str	'xl'	Size of Flan-T5 model, options are 'small', 'base', 'large', 'xl', 'xxl'
device	str	'cuda'	torch device to send the model to.
model_path	str	None	Optional path to a Flan-T5 model to load. Note the corresponding size must be specified in the size argument.
model	T5ForConditionalGeneration	None	Optional model to use for scoring
tokenizer	T5Tokenizer	None	Optional tokenizer to use for scoring

ScreenEval

ScreenEval is located in the data folder stored as a json file. The following keys are important for the use of ScreenEval.

Key	Type	Description
original_convo	List[str]	The source document that is to be summarized as a string
convo	List[List[str]]	The source document that is to be summarized split into a list of utterances
inferred_summary	List[str]	The summary sentence that is paired with the given source document
summary_id	List[str]	The source model for the summary sentence
convo_id	List[int]	The ID of the source document
annotated_summary	List[str]	The entire associated summary, with the focus summary sentence surrounded by <mark><\mark>
prediction_annotated_source_doc	List[str]	Raw source document
agreement	List[float]	Annotator agreement on summary sentence facutal inconsistency label
agg_label	List[bool]	Factual inconsistency label (true -> factually consistent, false -> factually inconsistent)
rel_utt	List[List[int]]	The indices of related utterances in the corresponding convo list.