TL;DR
This paper uses computational linguistics to analyze 140 years of US political speeches, revealing that while overall sentiment towards immigration has become more positive since WWII, partisan polarization has dramatically increased. Notably, modern Republican rhetoric mirrors the negativity of the 1920s, and framing of Mexican immigrants today strongly echoes historical anti-Chinese sentiment.
What Problem Does This Paper Address?
Public discourse on immigration is often characterized by heated rhetoric and strong opinions, leading to the perception that attitudes are more negative and polarized than ever before. However, historical data on public opinion regarding immigration is limited, especially for periods before the 1960s. This makes it difficult to objectively assess how attitudes have evolved over time and to understand the nuances of political framing. The authors aim to fill this gap by using computational methods to analyze a vast corpus of political speeches, providing a quantitative, long-term perspective on the evolution of immigration discourse in the US. They specifically want to understand how the tone, framing, and dehumanizing language used in political speeches have changed, and how these changes correlate with party affiliation and the nationality of immigrants being discussed.
Methods
The researchers employed a multi-pronged computational approach:
- Automated Text Classification: They trained machine learning classifiers (specifically, a fine-tuned RoBERTa model) on a manually annotated subset of speeches to identify which speeches were relevant to immigration and to classify their tone as pro-immigration, anti-immigration, or neutral. This allowed them to process over 200,000 congressional speeches and 5,000 presidential communications.
- Lexicon-Based Frame Analysis: They developed and utilized 14 lexicons (sets of keywords) to quantify the prevalence of specific frames used to discuss immigration (e.g., "crime," "legality," "family," "contributions").
- Contextual Embeddings for Dehumanization: A novel approach was developed using neural contextual embedding models (BERT) to measure implicit dehumanizing metaphors. This method assesses how likely certain words (associated with metaphors like "animals," "cargo," "disease") are to appear in sentences discussing immigrants, indicating subtle dehumanizing language.
Datasets
- Congressional Record: The primary dataset comprised approximately 17 million congressional speeches from 1880 to 2020, from which ~200,000 relevant speeches were identified.
- Presidential Communications: Approximately 5,000 presidential communications from the same period were also analyzed.
- Historical Immigration Statistics: Data from the Historical Statistics of the United States and the Migration Policy Institute were used to contextualize findings with demographic information.
- Annotated Speech Data: A manually annotated sample of speeches was used to train and validate the classification models. The authors make this data and their replication code publicly available.
Key Insights and Ideas
- The "Great Shift" to Positivity: While anti-immigration rhetoric is prominent today, the overall tone of political speeches about immigration has become significantly more positive on average since World War II, with a notable shift occurring between WWII and the 1965 Immigration and Nationality Act.
- Partisan Divergence: Despite the overall positive shift, a dramatic partisan divide has emerged since the late 1970s. Republican speeches today exhibit a negativity comparable to the average congressional speech in the 1920s, a period of strict immigration quotas.
- Echoes of the Past in Present Framing: The language used to describe Mexican immigrants today bears a striking resemblance to the language used for Chinese immigrants during the Chinese Exclusion era (late 19th century). This includes a greater emphasis on frames like "crime" and "legality," and increased use of dehumanizing metaphors.
- Nationality Still Matters: Even with the formal elimination of race-based restrictions, the nationality of immigrants remains a significant factor in how they are framed in political discourse, with non-European groups consistently facing more negative framing.
- Subtle Dehumanization: The study highlights the power of implicit dehumanizing metaphors (e.g., "animals," "cargo") and finds they are used more frequently by Republicans, particularly when discussing non-European immigrants.
Key Results
- Overall Tone: Congressional speeches shifted from consistently negative (1880-1940) to net positive on average from the mid-1950s onwards.
- Partisan Polarization: Since the late 1970s, Democrats have become increasingly positive, while Republicans have become as negative as the average legislator in the 1920s.
- Nationality-Based Framing: Speeches mentioning Mexican immigrants are consistently more negative than those mentioning European immigrants. This gap in tone between Mexican and European immigrants is comparable to the partisan gap between Democrats and Republicans.
- Frame Usage: Republicans disproportionately use frames like "crime," "legality," and "threats," while Democrats emphasize "family," "victims," and "contributions."
- Dehumanizing Metaphors: Republicans use implicit dehumanizing metaphors about 1.6 times more often than Democrats in their speeches.
Limitations and Future Work
- Nuance in Classification: The authors acknowledge that many speeches contain nuanced positions that are difficult to categorize perfectly as purely pro- or anti-immigration. Their classifiers aim to capture the extent to which a speech leans in one direction.
- Presidential Data Limitations: The analysis of presidential communications is treated with more caution due to less available text and a slight domain shift from congressional speeches.
- Causality of Polarization: While the study documents the rise of polarization, understanding the precise causes (top-down elite-driven vs. bottom-up public opinion) is beyond its scope.
- Future Work: The authors suggest further investigation into the correlation between elite rhetoric and public opinion, and exploring how different chambers (House vs. Senate) or election cycles might influence framing. They also note that while their dehumanization metric shows a pattern for Latin American immigrants broadly, further analysis might be needed for more specific regional groups.
Summary updated