• Digital Humanities and Cultural Heritage
• Publications close
• French National Research Agency (AN... close
• Archivio Istituzionale close

Data-to-Text Generation (DTG) is a subfield of Natural Language Generation aiming at transcribing structured data in natural language descriptions. The field has been recently boosted by the use of neural-based generators which exhibit on one side great syntactic skills without the need of hand-crafted pipelines; on the other side, the quality of the generated text reflects the quality of the training data, which in realistic settings only offer imperfectly aligned structure-text pairs. Consequently, state-of-art neural models include misleading statements - usually called hallucinations - in their outputs. The control of this phenomenon is today a major challenge for DTG, and is the problem addressed in the paper. Previous work deal with this issue at the instance level: using an alignment score for each table-reference pair. In contrast, we propose a finer-grained approach, arguing that hallucinations should rather be treated at the word level. Specifically, we propose a Multi-Branch Decoder which is able to leverage word-level labels to learn the relevant parts of each training instance. These labels are obtained following a simple and efficient scoring procedure based on co-occurrence analysis and dependency parsing. Extensive evaluations, via automated metrics and human judgment on the standard WikiBio benchmark, show the accuracy of our alignment labels and the effectiveness of the proposed Multi-Branch Decoder. Our model is able to reduce and control hallucinations, while keeping fluency and coherence in generated texts. Further experiments on a degraded version of ToTTo show that our model could be successfully used on very noisy settings. 20 pages, 6 figures, 5 tables (excluding Appendix). Source code: https://github.com/KaijuML/dtt-multi-branch

European populations display low genetic differentiation as the result of long-term blending of their ancient founding ancestries. However, it is unclear how the combination of ancient ancestries related to early foragers, Neolithic farmers, and Bronze Age nomadic pastoralists can explain the distribution of genetic variation across Europe. Populations in natural crossroads like the Italian peninsula are expected to recapitulate the continental diversity, but have been systematically understudied. Here, we characterize the ancestry profiles of Italian populations using a genome-wide dataset representative of modern and ancient samples from across Italy, Europe, and the rest of the world. Italian genomes capture several ancient signatures, including a non-steppe contribution derived ultimately from the Caucasus. Differences in ancestry composition, as the result of migration and admixture, have generated in Italy the largest degree of population structure detected so far in the continent, as well as shaping the amount of Neanderthal DNA in modern-day populations. ispartof: SCIENCE ADVANCES vol:5 issue:9 ispartof: location:United States status: published