Abstract

• 200K US congressional speeches + 5K presidential communications related to immigration from 1880 to the present

• political speech about immigration is much more positive on average than the past

  • shift largely between WW2 and the passage of Immigration and Nationality Act in 1965
  • since the late 1970s, political parties become polarized

• contextual embeddings of text

  • modern Republicans $\rightarrow$ suggestive of metaphors long associated with immigration (animals, cargo) and frames like "crimes" and "legality"

• nationality mentioned changed the tone of speeches (Mexican, Chinese) $\rightarrow$ still a major factor in how immigrants are spoken of in Congress

1. Introduction

• Recently, the attitude toward immigration became negative than ever before

  • anti-Chinese fearmongering in 1880s

  • Southern and Eastern European immigrants in 1920s

  • antiimmigration rhetoric of Trump (2017 to 2020)

    $\rightarrow$ "Certain types of immigrants can never truly join American society?"

• how have attitudes toward immigrants in US changed over the past century?
  • public opinion poll began in the 1960s
  • turned to Congressional Record
• Corpus
  • full corpus of more than 17M congressional speeches from 1880 to present
  • 200K speeches relevant to immigration
  • presidential communications
  • quantitative analysis
• Related works
  • qualitative approaches and historical archives
  • quantitative work on immigration used migration and census records
  • Rhetorical aspects of immigration debates $\rightarrow$ dehumanizing language (vermin and cargo) with qualitative analysis
  • NLP methods to cover in news media and Congress $\rightarrow$ not a long time span / not a comprehensive corpus with a consistent methodology
• Methods

  • identify relevant speeches

    • automated text classification based on extensive human annotations

  • curated and applied a set of lexicons for analyzing relevant frames with semi-automated method

  • neural contextual embedding models to quantify implicit dehumanizing metaphors

• Brief results and discussion

  • political speeches about immigration today are more positive than the past

    • the shift between WW2 and 1965 Immigration and Nationality Act

  • being net positive on average since early 1950s

  • Trump is the first president to express sentiment toward immigration more negative than the average member of his own party

  • two parties have become increasingly polarized over time

    • liniar increase in polarization on immigration since late 1970s

  • today, Democrats are unprecedentedly positive

  • generic political polarization observed in Gentzkow et al. by more than a decade

  • nationality of immigrants continues to matter greatly

    • Mexican $\rightarrow$ more negative than European
    • Mexican framed today is similar to Chinese framed during Chinese exclusion in 19th century
    • negative frame "crime", "labor", "legality" + dehumanizing metaphors

  • there remains a string and growing strain of antiimmigration speech among Replblicans

    • expressed opinions toward immigrants still vary greatly by country of origin
    • rhetorical strategies continue to be deployed

2. Results

Tone of Immigration Speeches

• 17M congressional speeches from 1880 to 2020

• human annotations and trained ML classifiers to detece immigration related speech with accompanying tone (pro, con, neutral)

• applied same models to all presidential communications by American Presidency Project (Bottom)

• Fig 1

  • average sentiment is negative throughout the late 19th and early 20th centuries (Chinese Exclusion Act(1882) to strint immigration quotas (1920s))

  • the attitude became more positive around the start of WW2

    • rising steadily from 1940 until the end of the Johnson administration (1969)
    • average tone has been pro since the beginning of the Eisenhower (1953)

  • beginning about a dacade after 1965, an overall decline in sentiment amone Republicans and incline among Democrats is observed

    • except for the early 1990s, this coincides with the end of the Cold War and NAFTA
    • Republican shows antiimmigration as 1920s

• Trends for presedential attitudes should be treated more cautiously as there is less text

  • involves a slight domain shift (the model is trained on congressional speeches)

  • found a similar pattern

    • early presidents were more antiimmigration

  • in recent years, presidents are uniformly more proimmigration even the Republican (Ronald Reagan) and the Democrats (Jimmy Carter)

    • Trump was a start exception (the most antiimmigration president over the past 140 years)

• Fig 2.

  • the tone was varied dramatically depending on which groups of immigrants are being discussed

  • Mexican and Chinese (Italian is Identifying Groups)

  • Speech mentioning Chinese immigrants were overwhelmingly negative during Chinese exclusion (1882 to 1943)

    • while the tone toward Italian was slightly more favorable

  • Attitute toward all groups improved from 1940 to 1970

    • mentioning China and Mexico remained relatively more negative overall

  • since the late 1970s, the gap between Italian and Mexican is large as the gap in tone that exists between Replublicans and Democrats today.

  • this pattern is mirrored in broader regional trends

    • Most European $\rightarrow$ referred to positively on average by the 1960s
    • Asian $\rightarrow$ by the 1980s
    • Caribbean $\rightarrow$ negative on average until the 2000s
    • few countries are mentioned as frequently as those three (Mexico, Italy, China)

Language, Framing, and Dehumanization

• trained interpretable logistic regression models to approximate the predictions of the contextual embedding models and determine feature importance using Shapley values

• Table 1

  • antiimmigration terms contains the words representing threats (dangerous, cheap), control (permit, violation), and the targets of early antiimmigration legislation (undesirable, Chinese)

    • by midcentury and beyond, another threats appear (subversive, terrorism) along with the themes of legality (aliens, illegal) and crime (criminals, smuggling)

  • proimmigration terms contain the words representing desirable characteristics (industrious), land (property, agriculture), and service (gave, served)

    • by post-WW2 era, humanitarian concerns (discriminatory, migrants) and community (citizens, families, children) appeared
    • this continued into the present (victoms, community) along with a celebration of once-vilified communities (Irish, Italian, heritage)

  • Despite the relatively negative tone toward Mexican in the modern period, Hispanic and Latino had strong positive associations

    • these are likely to be used by Democrats than Republicans $\rightarrow$ they are proimmigrations
    • but "Mexico" and "Mexican" are mentioned with very similar frequency by Democrat and Republicans $\rightarrow$ the tone difference is not simply a matter of Mexico
• To understand the rhetorical divergence between parties $\rightarrow$ they focused on the frame about immigration $\rightarrow$ built a series of lexicons
  • working upon the previous work, they developed 14 of these lexicons with automated/manual curation

• Fig 3

  • almost no difference in the frames by two parties in the earlier time period

  • today, they use strongly divergent use of different frames

    • Republicans : crime, legality, threats, deficiency, flood/tide $\rightarrow$ commonly heard antiimmigration comments
    • Democrats : family, victims, contributions, culture (positive)

  • these patterns are robust to the exclusion of any individual term as well as to automated lexicon expansion

  • the most salient aspects

    • earlier time period : difeciency, culture, labor
    • today : crime, legality (partly due to frequent mentions of legal and illegal immigrants + other legal terms and crime terms (laws, visas, criminals, terrorism))

  • economy is the most uncommon in speeches about immigration $\rightarrow$ the lease salient in both era

• measured more implicit dehumanizing metaphors

  • only flood and tide metaphor emerge dfrom the semiautomated frame construction process

  • measure the metaphors based on how probable such terms are as substitutes according to contextual embedding models

    • animals, cargo, disease, flood/tide, machines, vermin are drawn by this method
    • "dumping produces ~~~ " $\rightarrow$ cargo
    • "herding of ~~~" $\rightarrow$ animal

  • Republican used more dehumanizing metaphors

Differences by Country of Origin

• Fig 4
  • how Mexicans are framed today vs how Chinese were framed a centuryearlier
  • crime, labor, legality are deployed vastly
  • 4 most positive frames are all used far more in sentences mentioning European than the non-European groups (culture, victims, contributions, family)
• imploicit dehumanizing language is slightly but significantly more common for mentions of the non-European group in both cases

3. Discussion

• Congressional antagonism to immigration started much earlier than the quota period
  • China is mentioned in more than 20% of the speeches in 1880 to 1900
  • negative attitudes toward immigration remained from 1880 to 1940
• negative tone toward Chinese is consistent with the many pieces of anti-Chinese legislation introduced into Congress
  • 1875 Page Act
  • 1882 Chinese Exclusion
  • 1888 Scott
  • mentioning Chinese remained until the Chinese Exclusion Act was repealed in 1943
  • significantly greater use of implicit dehumanizing language to mention Chinese and emphasis on the threatening aspects
• The combination of frames (crime, threats) underscores the dual nature of immigrants (threatening vs cheap labor)
  • same pattern in the Mexican today
  • the frame of Europeans was more sympathetic although still negative until the middle of the 20th centurs
• gradual loosening of immigration laws from 1940s

  • this trend mirrored by congressional tone toward immigration

  • eventually becoming net positive on average in 1950s

  • possibly by the humanitarian concerns

    • signaling positive attitudes
    • increasing association with the 'victims' frame and decreasing the prominence of 'deficiency' and 'threats'
• nearly 30 years after the border reopened in 1965, the positive sentiment didn't fully erode even as the immigration from developing countries increased
  • partisnan divice on immigration cemerged in the late 1970s although the Republican showed neutral or positive stance until the election of Bill Clinton and NAFTA
  • this predates the previous work about polarization
• the results are consistent with the patterns in the previous work about the polarizations
  • polarization + overall tone toward immigration
  • beyond the sentiment, there are the 'framing' used in immigration debates
• understanding the causes of this polarizattion is beyond the scope of this paper
  • legislators' tone is weakly correlated with public opinion on the issue at the state level
  • noevidence of systematic differences in tone among House menbers in election vs nonelection years
• stark differences in framing between European and non-European groups
  • Chinese in the late 19th and early 20th century, Mexican today
  • more implicitly dehumanizing metaphors for non-European
  • also for the explicit frames (crime, labor, legality)
  • the gap between mentioning Mexico and European is equivalent to the modern gap between Democrats and Republicans
• modern immigration laws and the rhetoric of 'illegal' immegrants were crafted specifically to target immigration from Mexico
  • made associations with crime, legality and labor
• Mexico was also the target of early discrimination like China
  • Mexico was exempt from the quota system
  • Although the tone of speeches mentioning Mexican increased with other nationalities, these gains were largely eroded in the early 1970s $\rightarrow$ persistently nationality-based gap
• with the public opinion polls (by Gallup)
  • also shows the increase in proimmigrant sentiment from 1965 to present
  • in 2019, 77% answered as a positive
  • in 2002, it was 52% (after 9.11)
  • for asking the decrease of immigration, 65% answered it should be decreased in 1990s
  • in 2020, this fell to 28%
• the analysis of congressional and presidential speeches is more complicated
  • attitudes among Repuiblican are negative as the members of Congress were during the push for restrictive quotas
  • Chinese are still duscussed more negatively than European even the overall sentiment is positive today
  • recent years, COVID-19 made anti-Asian and hate crimes, anti-Chinese rhetoric
  • despite of the proimmigration among the general population, the tone differences in Congress based on nationality are strong as that between the parties

4. Materials and Methods

Data

• 43rd to 111th Congress : digitized copy of the Congressional Record

• 112th to 116th Congress : congressional-record tool by @unitedstates project

• data with speaker, party, state and date

• Procedural speeches were identified and excluded

• presidential communication : all presidential documents from The American Presidency Project

• Immigration statistics : Historical Statistics of the United States Millennial Edition Online + census data by the Migration Policy Institute

Classification

• Princeton University research assistants to label a speech
  • about immigration or not
  • proimmigration / antiimmigration
  • extensive set of keyword was used to select the candidate for annotation
  • 7626 segments annotated (3643 were judged relevant)
  • the judgements were aggregated with Bayesian item response model (to get a probability distribution over labels for each segment)
• trained RoBERTa
  • fine-tuned the pretrained roberta-base to congressional speeches with self-supervised
  • then fine-tuned it to be a classifier using annotated examples
  • ~90% accuracy on relevance and 65% on tone
  • major error in tone is between neutral and non-neutral
  • models trained on earlier and later parts of the data showed similar aggregate results in the intervening years
• the predictions on segments are used to predict the speeches
  • same predictor is used to presidential communication

Identifying Groups

• the mose prominent immigrant nationalities $\rightarrow$ historical data on the countries of origin of the foreign-born US population

• 45 countries that accounted for at least 1% of the foreitn-born population in at least 1 decade

• manually modified the country name and nationality

Measuring Impact

• used L1-regularized LR models to fit the predicted tone labels on all congressional segments classified as relevant

  • approximates the influence of individual words

• words in the vocab : at lease 20 times used / excluding numbers, punctuation, stop words / counts were binarized

• Shapley values computed (reflected in Table 1)

Curating Frames

• curated lexicons for 14 immigration frames
  • identified significantly frequently occurred terms mentioning of immigrants compared to terms mentioning generic people
  • considered initial exploration, annotators' comments, and prior literature to identified 14 relevant categories
  • which term should be selected in which frame is aggregated by majority votes

Identifyng Mentions

• collected direct mentions + group terms + more generic person references with nationality

• used to measure dehumanizing metaphorical language for each group

• included slang and derogatory terms to identify groups

Measuring Dehumanization

• introduce a method that is based purely on context $\rightarrow$ used BERT

  • trained on MLM task

  • fine-tuning to act as a classifier

  • to train implicit metaphorical language, began with the representative of that category

    • used static vector to fine similar terms
    • tried to find that kind of word in the BERT's vocabulary
• training procedure
  • for each sentence that mentions and immigrant of immigrant group
  • mask the mention with [MASK] token
  • to compute the probability of the candidate for the [MASK]
  • add up all the probabilities to get an overall score for each category for that sentence
  • showed log ratio of the mean probability for one set of mentions to the mean probability for the other
• validating procedure
  • collected human judgements on a sample of masked contexts
  • three of the auther independently rated whether a term would be a plausible replacement
  • resonably strong agreement (Krippendorff's alpha = 0.59)
  • correlated with the log probability by the model (r = 0.73)