Measuring Attribution in Natural Language Generation Models

TL;DR

This paper introduces AIS, a new evaluation framework for assessing whether natural language generation models produce output supported by verifiable external sources, validated through human studies across multiple datasets.

Contribution

The paper presents AIS, a novel two-stage annotation pipeline and evaluation framework for measuring attribution in NLG outputs related to external sources.

Findings

AIS correlates well with human judgments of source support

Validated across conversational QA, summarization, and table-to-text datasets

Provides a standardized approach for attribution evaluation in NLG

Abstract

With recent improvements in natural language generation (NLG) models for various applications, it has become imperative to have the means to identify and evaluate whether NLG output is only sharing verifiable information about the external world. In this work, we present a new evaluation framework entitled Attributable to Identified Sources (AIS) for assessing the output of natural language generation models, when such output pertains to the external world. We first define AIS and introduce a two-stage annotation pipeline for allowing annotators to appropriately evaluate model output according to AIS guidelines. We empirically validate this approach on generation datasets spanning three tasks (two conversational QA datasets, a summarization dataset, and a table-to-text dataset) via human evaluation studies that suggest that AIS could serve as a common framework for measuring whether model-generated statements are supported by underlying sources. We release guidelines for the human evaluation studies.