Fiscal talk by the Fed: First cut

1. Creating a dictionary of fiscal-related expressions

We do not want to be missing keywords, especially since the vocabularies to discuss fiscal issues may have changed in the fifty years period. Therefore, to supplement the dictionary of keywords presented below, we decided to test using automated or semi-automated process of keyword selection, starting a “seed” with fiscal, deficit and budget which are undeniably relevant.

1.1 Embeddings and cosine similarity

The first step is word-embeddings + cosine similarity. First, we pre-process the whole corpus, removing numbers and stop-words, converting to lowercase, and grouping the sentences back at the intervention/speech level. Then, we create word vectors using GloVe (Global Vectors for Word Representation) which uses a skip-gram window of 5 words to capture context and generates 100-dimensional vectors representing word meanings. We then explore the top 100 words that are the closest to fiscal, the top 100 closest to deficit and the top 100 words for budget.

Fiscal Terms

Budget Terms

Deficit Terms

This only gives us the best single words (or 1-grams) matches to each of our three seed words. Our seed words always appear in the top 5 closest words of the other seed words, which is good indication that they are indeed very closely related. Some words appear in several lists, which suggests they are also highly relevant. In total, there are 203 unique words that appear in the top 100 of one, two or three of our seed words. These 203 unique words are listed below, sorted by order of summed proximity with our three seed-words

Combined Terms Analysis

Some of these 203 words like sequestration, taxation or austerity appear likely to be directly relevant for our research question. We could add them to our dictionary of relevant keywords directly. However, most words, despite having indeed some proximity to our seed concept, appear too broad. For example, “surplus” may often be used to talk about fiscal issues, but not always.

1.2 Inspecting bigrams

The obvious solution in the literature is to move away from single keywords to consider key expressions, or n-grams (expressions composed of a number n of words). Rather than using our 203 words as is, we should create expressions containing these words to make sure they are indeed used in a relevant context. For example, we may want to use “government spending” (rather than “government” or “spending”), or “federal expenditures” (rather than “federal” or “expenditures”).

Of course, there is a balance to strike. Keeping all words single would mean capturing all relevant discussion, but also (a lot of) false positives. Using only bigrams instead may lead to fewer (or even no) false positives, but risks missing relevant discussions, especially since some words like “budget” or “deficit” may often be used on their own to refer to the “federal budget” or the “fiscal deficit”.

To make an informed decision about this trade off with insight from our data, we follow three steps. First, we use the package quanteda to identify all the bigrams in our corpus, after having removed the stop words such as “the”, “have” and other very frequent words that do not carry meaning. Second, we zoom in on all the bigrams in which our 203 candidate keywords appear. Third, we count the number of times these potentially relevant bigrams appear in our corpus to order them by relevance. Last, we manually check the top 10 bigrams per candidate keyword, excluding those that include “fiscal”, “budget” or “deficit”, since those will be already captured. This process allows us to screen from 2030 likely relevant expressions, that are listed in the table below:

Terms

1.3 Creating a first dictionary

We complement this manual inspection of the 2030 terms above with checks of how these terms are used in context in our corpus, and end up with a first dictionary of relevant keywords and key expressions.

Single Keywords (Unigrams)

First, we keep a few single keywords that are either our seed words, are highly specific (minimizing risks of false positives), and/or are too often used solo, making a bigram approach impossible. For those single keywords, an exclusion step will be created later to minimize chances of extracting too many false positives.

• budget, budgetary (seed word)

• deficit, deficits (seed word)

• fiscal (seed word)

• austerity (highly specific)

• entitlement (highly specific)

• sequestration (highly specific)

• tax ; taxes, taxation (highly specific)

• defense (often solo)

Two-Word Expressions (Bigrams)

Then we add more specific bigrams, that were either found directly in the list bigrams (e.g. federal spending), or constructed by combining expressions of interest (e.g. welfare spending).

It should be noted that bigrams containing the keywords above are not included. For example, “fiscal spending” will already be captured thanks to “fiscal”, so we do not need to add this expression to our dictionary.

Spending-related terms:

• federal spending
• government spending
• governmental spending
• military spending
• welfare spending
• social spending
• state spending
• spending cuts
• spending programs

Expenditure-related terms:

• benefits expenditure
• federal expenditure
• government expenditure
• governmental expenditure
• military expenditure
• social expenditure
• state expenditure
• welfare expenditure

Revenue-related terms:

• benefits revenue
• federal revenue
• government revenue
• governmental revenue
• military revenue
• social revenue
• state revenue
• welfare revenue

Cutback-related terms:

• benefits cutback
• federal cutback
• government cutback
• governmental cutback
• military cutback
• social cutback
• state cutback
• welfare cutback

Outlay-related terms:

• benefits outlays
• federal outlays
• government outlays
• governmental outlays
• military outlays
• social outlays
• state outlays
• welfare outlays

Policy-related terms:

• policy stimulus
• stimulus package
• reform package

Benefits-related terms:

• social benefits
• benefits payments
• benefits transfers
• benefits increase
• benefits reform

To validate that those terms are indeed relevant, we select randomly 5 sentences per expression from our corpus, and extract them with context (one sentence before, one sentence after):

Five Random Examples per Term

Some of these bigrams have no match in the corpus, and are therefore removed from our final inclusion dictionary:

• benefits cutback
• benefits expenditure
• benefits outlays
• benefits payments because it is only there twice and captured by other
• benefits reform
• benefits revenue
• benefits spending
• benefits transfers
• governmental cutback
• military cutback
• social cutback
• social outlays
• social revenue
• welfare cutback
• welfare revenue

Therefore, our final dictionary contains 43 terms, including 11 single words and 32 bigrams, as listed below.

• benefits increase
• federal cutback
• federal expenditure
• federal outlays
• federal revenue
• federal spending
• government cutback
• government expenditure
• government outlays
• government revenue
• government spending
• governmental expenditure
• governmental outlays
• governmental revenue
• governmental spending
• military expenditure
• military outlays
• military spending
• policy stimulus
• reform package
• social benefits
• social expenditure
• social spending
• spending cuts
• spending programs
• state expenditure
• state spending
• stimulus package
• welfare expenditure
• welfare outlays
• welfare spending
• austerity
• budget
• budgetary
• defense
• deficit
• entitlement
• fiscal
• sequestration
• tax
• taxation
• taxes

1.5 Creating an exclusion list for single keywords

We do a last step to reduce the chance of getting false positives. For our 11 single keywords, we extract all the bigrams that include these terms and that appear at least 100 times in our corpus.

                   bigram frequency
1           fiscal policy      4411
2          federal budget      1685
3          budget deficit      1576
4         monetary fiscal      1390
5           trade deficit      1278
6         budget deficits      1129
7         account deficit      1026
8         fiscal stimulus      1009
9         fiscal monetary       932
10        fiscal policies       857
11            fiscal year       701
12        federal deficit       618
13                tax cut       587
14               tax cuts       585
15   congressional budget       578
16             income tax       511
17          budget office       501
18              tax rates       473
19       defense spending       378
20             tax credit       376
21       federal deficits       360
22           tax revenues       354
23      deficit reduction       350
24           fiscal cliff       347
25         trade deficits       343
26       account deficits       332
27             tax reform       317
28       fiscal restraint       299
29     fiscal authorities       291
30          tax increases       270
31         investment tax       257
32          policy fiscal       250
33        fiscal deficits       240
34             tax system       232
35         fiscal actions       226
36               tax rate       215
37    government deficits       214
38       fiscal situation       206
39               tax code       203
40              deficit $       202
41       deficit spending       202
42             u.s fiscal       202
43      fiscal discipline       201
44            fiscal drag       201
45          tax reduction       201
46           income taxes       197
47         unified budget       193
48        balanced budget       185
49            payroll tax       185
50            fiscal side       182
51             tax policy       181
52      government budget       170
53         current fiscal       166
54         billion fiscal       165
55         fiscal deficit       162
56            tax changes       160
57         tax reductions       160
58        billion deficit       159
59           higher taxes       159
60         federal fiscal       157
61         fiscal support       157
62         large deficits       157
63           large budget       156
64           tax receipts       155
65      fiscal regulatory       153
66           line defense       153
67         tax incentives       151
68      deficit financing       145
69            tax revenue       145
70           tax spending       143
71        fiscal problems       141
72    fiscal policymakers       138
73  fiscal sustainability       137
74            federal tax       136
75         fiscal package       130
76         budget surplus       129
77         budget process       124
78          strong fiscal       124
79               tax base       120
80           tax increase       120
81          tax structure       120
82      fiscal imbalances       118
83   fiscal consolidation       117
84          corporate tax       114
85         deficit fiscal       114
86            changes tax       113
87     government deficit       112
88          tax treatment       112
89   entitlement programs       111
90               tax laws       111
91             tax burden       109
92       budget surpluses       108
93         effects fiscal       107
94     nondefense capital       106
95            tax credits       106
96                tax law       106
97         reduce deficit       104
98          payroll taxes       103
99      fiscal challenges       102
100          fiscal years       101
101          tax policies       101

To help reach a decision, we also look at five random examples of these bigrams in context:

Examples of High-Frequency Bigrams

After manual inspection, we decide to exclude: * trade deficits, trade deficit * account deficits, account deficit * fiscal years, fiscal year

Of course, these bigrams are not excluded from the corpus, but only from our dictionary. It means that we may still read them in excerpts, provided that these excerpts also include another term of interest from the dictionary above.