Temporal trends and presidential traits in the Italian end-of-year addresses: comparing and contrasting KBS and STM results

Abstract This study compares and contrasts the results of two lexical-based methods aimed at identifying content temporal trends in diachronic text corpora. A corpus of end-of-year addresses of the presidents of the Italian Republic constitutes a relevant case of political speech useful to understand how the temporal evolution of topics can be represented and whether a downward (ex post) or an upward (ex ante) extraction of topics is more effective for the identification of presidents’ distinctive traits and trends. The first method is a knowledge-based system (KBS), which identifies clusters of words sharing a similar temporal pattern through a three-step statistical learning procedure. The second is a structural topic model (STM), which identifies main topics by probing the possible effect of the year and president factors on the speech-topic and the topic-word distributions. In KBS clusters, the individual trait of the president stands out as one of the most relevant elements and determines the contents of speeches; moreover, topic trends can also be discerned ex post while interpreting the results. On the other hand, STM directly achieves the whole topic structure but seems not as powerful as expected in portraying the life cycle of words and detecting groups of words that distinguish the speeches of a specific president. As most presidential speeches are rich and cover a wide range of topics, the results suggest that, in this case, the interpretative tool offered by STM brings out more challenges than strengths. Conversely, direct observation of the temporal trajectory of individual words allows for more detailed analyses and meaningful results, thanks to the flexible and adaptive KBS approach.