How We Code Political Rhetoric

This paper was generated by AI from the lab's code, data and documentation.

Abstract

America's Political Pulse uses a large language model to code every post on X by members of Congress, state legislators, congressional challengers, governors and candidates for governor, and every floor speech, press release and e-newsletter issued by members of Congress. Each statement is coded for five rhetorical categories (personal attacks, policy criticism, policy discussion, bipartisanship and credit claiming), for the use of extreme labels, and for the policy areas it addresses. As of July 2026 the archive contained 6.8 million coded posts and 1.9 million coded passages from speeches, press releases and newsletters. The model receives each statement without information about its speaker. In a validation against two trained coders, GPT-4o, the model used for most of 2025 and the first half of 2026, was 97 percent accurate on personal attacks and 81 percent accurate on policy criticism (Jacob, Lelkes and Westwood 2026). Because earlier statements are not recoded when the model or the unit of analysis changes, series that span these changes combine labels produced under different procedures. This paper describes how the labels are produced and why, how to use them, and how to reproduce them.

1 Introduction

The labels exist to compare how politicians talk: how often a member attacks opponents personally rather than arguing against their policies, claims credit for spending, or invokes cooperation across party lines, and how these habits differ between parties, chambers and levels of government. The categories follow the lab's earlier work on credit claiming (Grimmer, Messing and Westwood 2012), bipartisanship (Westwood 2022) and divisive rhetoric (Jacob, Lelkes and Westwood 2026). Answering such questions for every officeholder requires coding millions of statements, far more than human coders can read, so we use a large language model to apply a fixed set of definitions to each statement and validate its labels against trained human coders (Section 5).

Three decisions shape the measure. The model sees only the text of a statement, never the speaker's name, party or office, so that its judgments cannot draw on who is speaking beyond what the text itself reveals. Every statement receives the same prompt, unchanged since July 23, 2025, so that differences between politicians reflect their words rather than changes in instructions. And we publish the label for every statement, so that users can construct their own rates, check ours against the text, or recode statements with another procedure.

Section 2 describes the sources, Section 3 the coding scheme and Section 4 the coding procedure, with the reason for each choice. Section 5 reports validation against human coders. Section 6 explains how to use the labels and documents the discontinuities in the series, and Section 7 describes how to reproduce them. Appendix A reproduces the coding prompt.

2 Sources

Members of Congress are observed on four channels; state legislators, challengers, governors and candidates for governor are observed only on X. Table 1 lists each source, the date from which its statements are coded and the volume of coded material as of July 2026. Floor speeches, press releases and newsletters are counted in coded passages because, before July 2026, these documents were coded in two-sentence passages (Section 4.2).

Posts are collected through X's application programming interface. For legislators and challengers we retrieve each account's timeline nightly, retaining replies and quote posts and excluding reposts, which reproduce another account's words. For governors and candidates for governor we use X's full-archive search for each account's own posts. Accounts for members of Congress are drawn from the unitedstates/congress-legislators project, and rosters of state legislators from Open States.

Floor speeches are drawn from the Congressional Record, retrieved daily from congress.gov and the Government Publishing Office and divided into speaking turns with the unitedstates/congressional-record parser. We retain turns attributed to a member, including the Record's Extensions of Remarks, which members insert rather than deliver. Press releases are collected by crawling each member's press page, using a language model to identify each release's link, date and headline, and extracting the text. Before April 2025 they were obtained from ProPublica's Congress API, which has since been discontinued. Newsletters are received in a lab mailbox subscribed to members' newsletters and are matched to members by sender. Gaps in our newsletter collection, including the period from May through December 2025, were filled from DCinbox, Lindsey Cormack's archive of congressional e-newsletters (Cormack 2017).

Several sources are collected but not yet coded: governors' official press releases, transcripts and captions of state legislative floor sessions in 33 states since 2020, and politicians' campaign and official websites. Images attached to posts are described by a separate vision model for display on the website, but the rhetoric model sees neither the images nor any text they contain. Because state legislators, challengers and governors are observed only on X, comparisons between them and members of Congress should be restricted to posts on X.

3 Coding Scheme

Each statement is coded on six binary measures, defined in Table 2 in the terms supplied to the model; the complete prompt appears in Appendix A. A statement may carry several labels or none, so a politician's rates need not sum to 100 percent. Statements coded as personal attacks are excluded from policy criticism, and the definition of policy criticism also excludes claims of extremism; no other rule links the categories.

The definition of a personal attack is deliberately narrow. A statement qualifies only if it explicitly questions an individual's character, integrity, intelligence, morality or patriotism; criticism of a politician's votes, policies or performance in office does not qualify. We drew the line there because disagreement about performance in office is ordinary democratic argument, whereas an attack on character is the kind of rhetoric whose prevalence the measure is meant to track, and a narrow definition is also one that coders and the model apply consistently. When a statement is coded as a personal attack, the model also records which of the five traits is targeted and the name of the person attacked. Policy criticism is the corresponding category for disagreement that is directed at a policy and argued on its merits.

Extreme labels are defined by a closed list of eight terms, so that the measure records a specific, observable act rather than a judgment about tone. The model's task is to determine whether a listed term is applied to an American politician, a party or a policy. The word radical in a description of a medical treatment does not qualify, whereas its application to a party does. When an extreme label is present, the model also records its target.

Statements are additionally assigned to any of 25 policy areas taken from the Congressional Research Service's policy-area terms, which Congress uses to index legislation; examples include Health, Immigration, Taxation, and Armed Forces and National Security. We use Congress's terms because the federal and state bills in our data carry the same areas, so that what politicians say can be compared with the legislation they sponsor. A statement is coded as policy discussion if the model assigns it at least one area. The areas also show how a politician's discussion of policy is distributed across issues.

4 Coding Procedure

4.1 Model input and output

Each night, up to 30,000 statements published within the preceding 182 days that have not yet been coded are submitted to OpenAI's models through the Batch API, which costs half as much as individual requests and returns results within a day; the results are collected on a following night. The long window means that statements reaching our archive weeks after they were issued, such as newsletters recovered from DCinbox, are still coded. Every request carries the same system prompt (Appendix A), followed by a single user message consisting of the instruction “Analyze this text” and the statement. The model receives no other information: neither the speaker's name, party, office or state, nor the date or type of the statement. A model told that a statement came from a Democrat or a Republican could let that information shape its judgment, and withholding it removes the most obvious source of error correlated with party.

The model returns its labels in a fixed JSON format. A response is accepted only if it supplies a yes or no answer for personal attacks, policy criticism, bipartisanship and credit claiming. Responses that fail this check are discarded, and the statement is resubmitted on later nights, up to three attempts. Failures are rare: fewer than 1 in 100,000 responses from gpt-5.6-terra and gpt-6-sol failed the check, against 0.6 percent of first submissions to gpt-4o, most of them refusals to code very short texts. Statements that never yield a valid response remain uncoded and are excluded from all rates. Apart from a cap of 5,000 output tokens, the model is run at its default settings, and each statement is coded once.

4.2 Units of analysis

Before submission, the Congressional Record's line naming the speaker is removed from each floor speech, and posts and floor speeches are cleaned: links are removed, HTML entities decoded, periods after single capital letters dropped (so that “U.S.” becomes “US”), and the text split into sentences and rejoined. Press releases and newsletters are sent as collected. Posts are coded as whole documents. Since July 2026, press releases and newsletters have also been coded as whole documents, and floor speeches in segments of up to 100 sentences. Before July 2026, floor speeches, press releases and newsletters were coded in two-sentence passages, and these earlier passages have not been recoded. We moved to whole documents because the statement is the unit we report: a passage of two sentences can lose the context that determines whether a remark is an attack or a quotation, and coding a long release as many passages gives it many opportunities to be labeled (Section 6.2). For the earlier passages, we treat a statement as carrying a label if any of its passages does, and we date each statement by its earliest passage.

4.3 Models

The coding model has changed four times since March 2025 (Table 3). Most changes adopted a newer model as it became available, although the change of July 30, 2025 returned from o4-mini to gpt-4o; the prompt has been unchanged since July 23, 2025, a period that spans the three most recent changes. Earlier statements were not recoded when the model changed, so a series that spans a change combines labels produced by different models. The models are proprietary and are retired by their provider, so earlier labels cannot be reproduced exactly, and a model of the same name can change its behavior over time.

We measured the effect of the most recent change directly. On 200 statements dated after August 1, 2026 and previously coded by gpt-5.6-terra, 40 drawn at random from each of five sources, gpt-6-sol agreed on between 92 and 98.5 percent of labels, depending on the category, with Cohen's kappa between .56 and .95. The script, sample and results are in research/rhetoric/sol_vs_terra/ in the lab's code. The newer model applies two labels less often: policy criticism (6 percent of the sample, compared with 14 percent) and policy discussion (65 percent, compared with 71.5 percent). Series for these two categories therefore decline at the end of October 2026 for reasons unrelated to the behavior of the politicians.

4.4 From labels to rates

A politician's rate for a category is the share of the politician's coded statements in a given period that carry the label. Rates are reported for the current term, the past 365 days, the full record and each past term. Each politician is compared with peers of the same party, chamber and level of government, using the mean rate among peers with at least ten statements in the period. A profile leads with a finding about a category only when the politician has at least 200 statements and at least ten peers, and a finding that a rate is unusually high also requires at least ten labeled statements, so that a featured rate does not rest on a small number of statements. Shares of policy areas are computed among statements that name at least one area.

5 Validation

Agreement with trained human coders is highest for personal attacks. Jacob, Lelkes and Westwood (2026) compared GPT-4o's labels for two-sentence passages with the consensus codes of two political scientists holding PhDs, who independently coded a random sample of 500 statements with the same prompt and initially agreed on 92 percent of statements for personal attacks and 80 percent for policy criticism. Table 4 reports accuracy, precision (the share of the model's positive labels with which the coders agreed) and recall (the share of the coders' positive labels identified by the model).

The coders agree with 98 percent of the model's personal-attack labels, and the model identifies 92 percent of the attacks the coders mark. Policy criticism is harder to code for people and model alike: the two coders initially agreed on only 80 percent of statements, and the model's accuracy of 81 percent is at that level. The validation did not cover policy discussion, credit claiming or bipartisanship, so no human benchmark exists for those categories, and their levels should be treated with more caution than the comparisons between politicians that they support. Extreme labels are excluded from the benchmark because they are defined by a closed list of terms (Section 3).

The benchmark applies to GPT-4o coding two-sentence passages. No human benchmark yet exists for gpt-5.6-terra, for gpt-6-sol or for whole-document coding; the only evidence on the current model is its agreement with its predecessor on between 92 and 98.5 percent of labels (Section 4.3). Model outputs are not fully deterministic, so a statement coded twice may occasionally receive different labels; each statement in the data reflects a single coding.

An individual label is the model's judgment about a short text and is sometimes wrong, but a politician's rate across hundreds of statements is less noisy, though not free of bias. Errors unrelated to the speaker shift every politician's rate in the same way: they overstate rare labels and compress differences between politicians, but in expectation they preserve the ordering of politicians. Errors correlated with the speaker would distort comparisons, and the design addresses the most obvious source of such errors by withholding the speaker's identity from the model. Estimates of differences between groups of politicians that correct for these errors require a human-coded subsample (Egami et al. 2023).

6 Using the Labels

6.1 What is published

The labels are public. Each row of the coded files corresponds to one coded unit (a post, a passage or, since July 2026, a whole document or speech segment), and records its source, the speaker's identifier, its date, the six labels, the type and target of any personal attack, the target of any extreme label, the policy areas named and, for units coded since October 2026, the model that coded them and when. Separate archives contain the full text of members' newsletters, press releases and floor speeches without labels.

Table 5 lists the files. All are rebuilt nightly, and manifest.json records the dates each file covers.

Posts on X are distributed as post identifiers and links rather than as text, because X's developer terms permit the redistribution of identifiers but not of post content. Researchers can retrieve the text of posts that remain public from X using these identifiers. Text from the other channels is a matter of public record and is distributed in full. The data are released under a Creative Commons Attribution 4.0 license and should be cited as Westwood and Lelkes (2024).

To reproduce a politician's rate from the files, group passages into statements, count a statement as carrying a label if any of its passages does, and divide the number of labeled statements in the period by the number of coded statements. Rates for the same politician on different channels should be reported separately, since the channels differ in purpose and in the unit coded.

6.2 Discontinuities

In addition to the model changes listed in Table 3, four changes in collection and coding create discontinuities in specific categories or channels (Table 6). Analyses of trends that span these dates should account for them, and analyses that depend on a particular period should report the dates of the data used and any discontinuity that the period spans.

Because a statement carries a label if any of its passages does, a long press release coded as twenty two-sentence passages had twenty opportunities to be labeled, whereas the same release coded as a whole document has one. Rates for floor speeches, press releases and newsletters before July 2026 are therefore not comparable with later rates. Posts were coded as whole documents throughout and are unaffected.

6.3 Limitations

Several limitations arise from collection. Because reposts are excluded, politicians who mainly amplify others' posts contribute fewer statements than their activity on the platform would suggest. X's timeline endpoint returns only about the 3,200 most recent posts of an account, so accounts added to our roster after long periods of activity lack their earlier posts, and accounts added since 2024 begin in January 2024. State legislators are coded from May 2024. Because Extensions of Remarks are coded with floor speeches (Section 2), the floor-speech series includes text that was never delivered on the floor. Because the model sees each statement without the post to which it replies or the image that accompanies it, attacks that depend on such context are not coded, and a post whose content is carried by an image is coded on its text alone. Before October 2026, 68 floor speeches whose speaker line ran directly into a dash reached the model with the speaker's surname; these, and eight speeches whose first sentence was dropped, were recoded in October 2026.

7 Replication

7.1 Code and software

The coding pipeline is part of the lab's code for America's Political Pulse. Selection, cleaning and division of statements, construction and submission of requests, and collection and validation of responses are in elite/rhetoric/classify/; challengers, governors and candidates for governor are coded by parallel jobs in elite/challengers/classify/, elite/governors/classify/ and elite/state_candidates/classify/, which use the same prompt, model and validation. The prompt is in elite/rhetoric/classify/prompt.py and the model identifier in elite/rhetoric/classify/llms.py. The roll-up of passages into statements and the computation of rates and peer means are in elite/nextgen/. The code is not public; the lab shares it with researchers on request. The jobs run in Python 3.11 with the OpenAI Python library 3.29.0 and NLTK 3.10.3, whose Punkt sentence splitter is used without a trained model; every dependency is pinned in docker/requirements-batch-light.lock.

7.2 Procedure

The legislators' job runs daily at 04:00 UTC and the jobs for challengers, governors and candidates for governor later the same morning. Each run first collects the results of batches submitted on earlier nights, validates them as described in Section 4.1 and writes the labels, then selects up to 30,000 uncoded statements published in the last 182 days, in order of their identifiers and excluding those that have failed three times, cleans and divides them (Section 4.2) and submits them in batches of up to 15,000 requests. Each request is sent to the chat completions endpoint with the model identifier, the system prompt in Appendix A, the user message “Analyze this text: ” followed by the statement, and a limit of 5,000 output tokens; no temperature, seed or response format is set. A user with an OpenAI account can reproduce a request exactly from this description and the prompt.

7.3 What can and cannot be reproduced

The prompt, the request, the cleaning and division of text, the validation rule, the roll-up and the computation of rates are fully specified, and the public labels allow every published rate to be recomputed (Section 6.1). Text from floor speeches, press releases and newsletters is public and can be recoded with the current model or any other; posts on X can be recoded after retrieving their text from X by identifier, provided they have not been deleted.

The original labels cannot be reproduced exactly. Earlier models have been retired, a model of the same name can change its behavior, and outputs are not deterministic. From October 2026 each coded statement records the model that coded it and when, and the raw output of every batch is archived; for statements coded earlier, the model is known only from the date of coding (Table 3), and the raw responses were not kept.

References

Cormack, Lindsey. 2017. “DCinbox—Capturing Every Congressional Constituent E-newsletter from 2009 Onwards.” The Legislative Scholar 2(1).

Egami, Naoki, Musashi Hinck, Brandon M. Stewart, and Hanying Wei. 2023. “Using Imperfect Surrogates for Downstream Inference: Design-Based Supervised Learning for Social Science Applications of Large Language Models.” Advances in Neural Information Processing Systems 36.

Grimmer, Justin, Solomon Messing, and Sean J. Westwood. 2012. “How Words and Money Cultivate a Personal Vote: The Effect of Legislator Credit Claiming on Constituent Credit Allocation.” American Political Science Review 106(4): 703–719.

Jacob, Marc S., Yphtach Lelkes, and Sean J. Westwood. 2026. “Entrepreneurs of Conflict: A Descriptive Analysis of When and How Political Elites Use Divisive Rhetoric.” PNAS Nexus 5(3): pgag038.

Westwood, Sean J. 2022. “The Partisanship of Bipartisanship: How Representatives Use Bipartisan Assertions to Cultivate Support.” Political Behavior 44(3): 1411–1435.

Westwood, Sean J., and Yphtach Lelkes. 2024. America's Political Pulse [data set]. Polarization Research Lab. https://americaspoliticalpulse.com/data. Accessed October 2026.

Appendix A. The Coding Prompt

The system prompt below accompanies every statement and is reproduced exactly as sent, including its typographical errors. It has been unchanged since July 23, 2025. The statement follows in a separate message consisting of “Analyze this text: ” and the text.

You are a political text classifier. Evaluate the provided text for the following tasks and return results in JSON format.

Evaluate each of the following tasks sequentially:

Attacks: 

Task: Assess the text for a personal_attack (output either "yes" or "no"). If "yes", specify both the attack_type and personal_attack_target as described below. If "no", set "attack_type" and "personal_attack_target" as null.
Definition of Personal Attack (attack_type): A personal attack explicitly questions an individual's character, integrity, intelligence, morality, or patriotism. Opposition to an individual or their policy choices does not count. Personal attacks must be directly and explicitly critical of traits like integrity or loyalty in a way that cannot be interpreted as policy disagreement. Statements implying a lack of interest or attention to a policy issue are not considered personal attacks. Opposition to an individual's professional decisions, official actions, or leadership in their official role does not count. 
character: Explicitly saying someone as lazy, unreliable, stupid, etc.
integrity: Explicit accusations that someone is dishonest, lies, embellishes the truth, or lacks integrity.
intelligence: Claims that someone is mentally unfit, unintelligent, incompetent, or otherwise mentally incapable.
morality: Accusations of corruption, unethical behavior, criminality, or fraud.
patriotism: Accusations that someone is un-American, anti-democratic, pro-terrorist, or not supportive of U.S. interests, democracy, or national security.
Target Identification (personal_attack_target): 
Record the name of the person attacked, excluding any titles (e.g., "Joe Biden" instead of "President Joe Biden"). Use no more than two words for the name.
Exclusion criteria. Do not consider statements as personal attacks if they:
Critique job performance, policies, legislation, or appointments without explicitly questioning personal character traits such as loyalty or integrity or name calling
Require assumptions or interpretations 
Question leadership decisions
Question someone's support for US policy, foreign policy or an ally
Are not explicit and overt in questioning an individual's personal traits.
Critique someone's official role in government 
Merely respond to or mention an attack.
Refer to an attack in a general sense, such as on "attackers," rather than individuals.
Are quoted attacks rather than direct statements by the speaker.
Refer to attacks on foreign lobbyists, agents, or external groups (e.g., companies, organizations, foreign leaders like Vladimir Putin, or terrorist groups).
Critique job performance, policies, or legislation without explicitly questioning personal character traits.
Conflate personal and policy failings
Blame someone for a policy failure without making a personal attack
Only include statements as attacks if they meet all conditions outlined above

Extremism:
Task: Assess the text for extreme_label (output either "yes" or "no"). If "yes", ensure the text meets the criteria below. If "no", output "extreme_label":"no".
Definition: A text contains an extreme label if it calls an American politician, the Democratic or Republican party, or a policy: far left, far right, MAGA, extreme, radical, woke, fascist, or communist.
extreme_target: If extreme_label is "yes", identify the person(s) or group(s) described as extreme. Use no more than two words per entry, and remove any titles.

Policy criticism:

Task: Assess the text for a policy_attack (output either "yes" or "no"). If "yes", ensure the text meets the criteria below. If "no", output "policy_attack": "no".
Definition of Policy Attack: A policy_attack is a constructive critique of policy, legislation, or court decisions. To qualify as a policy attack, the text must:
Object to or raise concerns about a specific policy, law, or court ruling.
Use fact-based arguments, even if critical or negative.
Avoid emotional appeals, inflammatory language, claims of extremism, or personal attacks on individuals involved with the policy including accusing them of lying of withholding information.Avoid emotional appeals, inflammatory language, or personal attacks on individuals involved with the policy.
Exclude statements where you identified personal_attack was "yes"

Bipartisanship:

Task: Assess the text for is_bipartisanship (output either "yes" or "no").
Definition of Bipartisanship: Bipartisanship, "working across the aisle," or collaboration between Democrats and Republicans. Emphasis on cooperation, compromise, or finding common ground across party lines.

Credit claiming:

Task: Assess the text for is_creditclaiming (output either "yes" or "no").
Definition of Credit-Claiming:
    •   Creating or passing legislation.
    •   Securing government spending, grants, or funding.
    •   Emphasizing personal or party accomplishments in office.

Policy

Task: Asses if policy is discussed and determine the policy_area from the list below.
Definition: Evaluate whether the text discusses public policy by identifying any references to specific policies, laws, legislation, or general policy areas (e.g., healthcare, education, national security). Exclude procedural or operational statements that do not clearly address a particular policy area.
Here are a list of policy areas:

"Agriculture and Food": agricultural practices; agricultural prices and marketing; agricultural education; food assistance or nutrition programs; food industry, supply, and safety; aquaculture; horticulture and plants. 
"Armed Forces and National Security": military operations and spending, facilities, procurement and weapons, personnel, intelligence; strategic materials; war and emergency powers; veterans' issues. 
"Civil Rights and Liberties, Minority Issues": discrimination on basis of race, ethnicity, age, sex, gender, health or disability; First Amendment rights; due process and equal protection; abortion rights; privacy. 
"Commerce": business investment, development, regulation; small business; consumer affairs; competition and restrictive trade practices; manufacturing, distribution, retail; marketing; intellectual property. 
"Crime and Law Enforcement": criminal offenses, investigation and prosecution, procedure and sentencing; corrections and imprisonment; juvenile crime; law enforcement administration. 
"Economics and Public Finance": budgetary matters such as appropriations, public debt, the budget process, government lending, government accounts and trust funds; monetary policy and inflation; economic development, performance
"Education": elementary, secondary, or higher education including special education and matters of academic performance, school administration, teaching, educational costs, and student aid.
"Emergency Management": emergency planning; response to civil disturbances, natural and other disasters, including fires; emergency communications; security preparedness.
"Energy": all sources and supplies of energy, including alternative energy sources, oil and gas, coal, nuclear power; efficiency and conservation; costs, prices, and revenues; electric power transmission; public utility matters.a
"Environmental Protection": regulation of pollution including from hazardous substances and radioactive releases; climate change and greenhouse gasses; environmental assessment and research; solid waste and recycling; ecology. 
"Families": child and family welfare, services, and relationships; marriage and family status; domestic violence and child abuse. 
"Finance and Financial Sector": U.S. banking and financial institutions regulation; consumer credit; bankruptcy and debt collection; financial services and investments; insurance; securities; real estate transactions; currency. 
"Foreign Trade and International Finance": competitiveness, trade barriers and adjustment assistance; foreign loans and international monetary system; international banking; trade agreements and negotiations; customs enforcement, tariffs, and trade restrictions; foreign investment. 
"Government Operations and Politics": government administration, including agency organization, contracting, facilities and property, information management and services; rulemaking and administrative law; elections and political activities; government employees and officials; Presidents; ethics and public participation; postal service. 
"Health": science or practice of the diagnosis, treatment, and prevention of disease; health services administration and funding, including such programs as Medicare and Medicaid; health personnel and medical education; drug use and safety; health care coverage and insurance; health facilities. 
"Housing and Community Development": home ownership; housing programs administration and funding; residential rehabilitation; regional planning, rural and urban development; affordable housing; homelessness; housing industry and construction; fair housing. 
"Immigration": administration of immigration and naturalization matters; immigration enforcement procedures; refugees and asylum policies; travel and residence documentation; foreign labor; benefits for immigrants. 
"International Affairs": matters affecting foreign aid, human rights, international law and organizations; national governance; arms control; diplomacy and foreign officials; alliances and collective security. 
"Labor and Employment": matters affecting hiring and composition of the workforce, wages and benefits, labor-management relations; occupational safety, personnel management, unemployment compensation. 
"Law": matters affecting civil actions and administrative remedies, courts and judicial administration, general constitutional issues, dispute resolution, including mediation and arbitration. 
"Public Lands and Natural Resources": natural areas (including wilderness); lands under government jurisdiction; land use practices and policies; parks, monuments, and historic sites; fisheries and marine resources; mining and minerals. 
"Science, Technology, Communications": natural sciences, space exploration, research policy and funding, research and development, STEM education, scientific cooperation and communication; technology policies, telecommunication, information technology; digital media, journalism. 
"Taxation": all aspects of income, excise, property, inheritance, and employment taxes; tax administration and collection. 
"Transportation and Public Works": all aspects of transportation modes and conveyances, including funding and safety matters; Coast Guard; infrastructure development; travel and tourism. 
"Water Resources Development": the supply and use of water and control of water flows; watersheds; floods and storm protection; wetlands. 

Variable: policy_area
which of the policy labels from the list above apply to the text.

Return your response in this exact JSON format:

{
    "attacks": {
        "personal_attack": ,
        "attack_type": [],
        "personal_attack_target": ,
    },
    "extremism": {
        "extreme_label":,
        "extreme_target":,
    },
    "policy_criticism": {
        "policy_attack": ,
    },
    "bipartisanship": {
        "is_bipartisanship": ,
    },
    "credit_claiming": {
        "is_creditclaiming": ,
    },
    "policy": {
        "policy_area": [],
    }
}