Unsupervised alignment of comparable data and text resources

Abstract

In this paper we investigate automatic datatext alignment, i.e. the task of automatically aligning data records with textual descriptions, such that data tokens are aligned with the word strings that describe them. Our methods make use of log likelihood ratios to estimate the strength of association between data tokens and text tokens. We investigate datatext alignment at the document level and at the sentence level, reporting results for several methodological variants as well as baselines. We find that log likelihood ratios provide a strong basis for predicting data-text alignment.