Deep neural networks for worker injury autocoding

Deep neural networks for worker injury autocoding
Alexander Measure
U.S. Bureau of Labor Statistics
[email protected]
Draft as of 9/18/2017
Each year millions of American workers are injured on the job. A better understanding of the
characteristics of these incidents can help us prevent them in the future, but collecting and analyzing the
relevant information is challenging. Much of the recorded information exists only in the form of short
text narratives written on OSHA logs and worker’s compensation documents.
The largest ongoing effort to collect and analyze this information is the Bureau of Labor Statistics’ (BLS)
Survey of Occupational Injuries and Illnesses (SOII), an annual survey of U.S. establishments that collects
approximately 300,000 descriptions of these injuries each year. A typical narrative might resemble the
following:
Narrative collected through survey

Codes assigned by BLS

Job title: RN

Occupation: 29-1141 (registered nurse)

What was the worker doing? Helping patient
get into wheel chair

Nature: 1233 (strain)

What happened? Patient slipped and employee
tried to catch her
What was the injury or illness? Strained lower
back

Part: 322 (lower back)
Event: 7143 (overexertion in catching)
Source: 574 (patient)
Secondary source: None

What was the source? Patient
To facilitate aggregation and analysis, BLS assigns 6 detailed codes to each narrative; an occupation code
assigned according to the Standard Occupational Classification system (SOC), and 5 injury codes
assigned according to version 2.01 of the Occupational Injuries and Illnesses Classification System
(OIICS). Historically BLS has relied entirely on staff to read and code these narratives, but this has
changed in recent years. Starting with just over a quarter of occupation codes in 2014, BLS is on pace to
automatically assign nearly half of all non-secondary SOII codes for 2016 data.

Portion of SOII codes automatically assigned by survey year
100%
80%

60% 57%

60%
40%

75% 70%
68%
42%

26%

20%

0%

0%

0%

0%

0%

0%

2014

14%

0%

2015
occupation

nature

part

25%

2016
event

source

This advance has been made possible by the use of machine learning algorithms that learn from
previously coded data. Although previous research suggests we can use these algorithms to
automatically assign all SOII codes at near or better than human accuracy (Measure, 2014), it also shows
that autocoders still make plenty of errors. Motivated by this, and by recent advances in machine
learning, we briefly review the limitations of SOII’s existing autocoders and then introduce a new neural
network autocoder that demonstrates substantial improvements.

Background
SOII autocoding is currently performed by 5 regularized multinomial logistic regression models, one for
each of the 5 main coding tasks; occupation, nature, part, event, and source1. Each was created using
the free and open source scikit-learn library (Pedregosa, et al., 2011) with bindings to Liblinear (Fan,
Chang, Hsieh, Wang, & Lin, 2008) and is closely related to the logistic regression models described in
(Measure, 2014) with only minor tweaks to the input features (see Appendix A for a complete
description). Mathematically each is described by:
= (

+ )

where is an -dimensional vector of predicted code probabilities, is the number of codes in our
classification system, is a -dimensional input vector where each position contains a value that
indicates some potentially relevant information about a case requiring classification (such as the
occurrence of a particular word or pair of words), is a learnable bias vector, is a matrix of weights
indicating how strongly each element of contributes to the probability of each possible code
represented in , and ( ) is the multinomial logistic function which constrains each element of to be
between 0 and 1, and the sum of all elements to total 1.
Given a collection of previously coded SOII cases, the optimal weight and bias values are
approximated by minimizing the L2 regularized cross entropy loss of the model’s predictions on the
data. More precisely, if:



1

takes a value of 1 when the th training example has code and 0 otherwise
th
training example has code
( , ) is the model’s estimated probability that the
( , ) is the weight specific to code and input
(, )

We do not currently have a logistic regression autocoder for secondary source, in part because most cases do not
indicate a secondary source.



is a constant controlling the tradeoff between regularization loss and empirical loss

then the L2 regularized cross-entropy loss is given by:
−

+

( , )

−

1

( , ) ln( ( , ) )

+ 1−

(, )

ln(1 −

( , ))

The bracketed expression on the left is the L2 regularization loss intended to discourage overfitting and
the bracketed expression on the right is the empirical loss designed to fit the model to the available
data.
Once the optimal weight and bias values have been calculated cases can then be autocoded by, for a
given input, calculating the probability of all possible classifications and assigning the one with the
highest probability. When only partial autocoding is desired we set a threshold and only assign codes
with predicted probabilities exceeding that threshold.
We can divide the errors made by these autocoders into three broad categories: those due to
fundamental ambiguities in the task being performed, those due to errors and limitations in the
available training data, and those due to failures in fitting or specifying the autocoding model.
Consider, for example, an occupation described simply as “firefighter/paramedic.” The SOC defines two
separate classifications, one for firefighter, one for paramedic, but in this case there is no obvious single
answer. As a result different coders often come to different conclusions, at least one of which is wrong
according to coding guidelines. We have some evidence that this is common. For example, when we
asked a select group of regional coding experts to code a random sample of 1000 SOII cases we found
that their average pair-wise agreement rate was only 70.4%. Unfortunately, any ambiguities responsible
for this disagreement can only be resolved by changing the data collection or classification systems in
non-trivial ways.
Errors due to limitations in the training data are also difficult to fix. Our current training dataset amounts
to more than 1.3 million coded cases and there is no obvious or simple way to significantly expand it nor
to find and correct any errors it may contain.
We instead focus on the third source of errors, those due to failures in the machine learning algorithm
itself. Here, recent research in deep neural networks suggests many potential improvements.

Overview of Deep Neural Networks
In recent years, deep neural networks have been responsible for advances in a wide variety of fields
including natural language processing. A full review of this work is well beyond the scope of this paper,
we refer the reader to (Goldberg, 2016) and (Goodfellow, 2016) for broader and more detailed
overviews. Instead, we briefly introduce the basic concepts upon which we build.

Artificial Neurons
The basic building block of the deep neural network is the artificial neuron which is typically modeled as
some variation of the simple linear model = (
+ ) where y is a real number, is an input
vector, is a weight vector, is a bias value, and ( ) is a nonlinear function, typically either the
logistic function, the hyperbolic tangent, or the rectified linear function defined as:

( )=

>0
0 ℎ

Although these neurons can be organized into any pattern, for convenience and computational
efficiency they are often organized into layers with the outputs of some layers forming the inputs to
others. A single layer of artificial neurons in which each neuron is connected to all outputs from the
previous layer is called a densely connected layer. When such a layer also uses the multinomial logistic
function as its nonlinear activation it is equivalent to the multinomial logistic regression model which we
use for our existing SOII autocoders.

Convolutional Layer
Although fully connected layers are powerful models, they allow all inputs to interact with each other
which can often be ruled out beforehand. By constraining our model to consider only relationships we
know to be more likely we can often improve the model’s ability to learn the correct relationship. One
way to accomplish this is with a convolutional layer, which applies spatial constraints to the interactions
of the inputs.
A convolutional layer consists of filters, each typically modeled as = (
+ ), which are applied to
varying subsets of the input, typically each contiguous subset of a predetermined size. For example,
consider an input consisting of the word “manager”. A convolutional layer consisting of 2 filters, applied
to each contiguous 5 letter subsequence of the word would produce 6 outputs which might resemble
the following:
Step

Input to filter

1
2
3

manag
anage
nager

Filter 1
output
1.8
-.2
-.5

Filter 2
output
-.001
.01
1.5

Because there is often significant redundancy in the output of a convolutional layer, it is common to
follow a convolutional layer with a pooling operation which aggregates or discards some of it. We use a
global max pooling layer which simply discards all but the largest output of each filter. In our example
this would leave us only with the outputs 1.8 and 1.5 for filters 1 and 2, respectively.

LSTM Layer
Another approach, which is particularly well suited for modeling sequential structure, is Long Short Term
Memory (LSTM) (Hochreiter & Schmidhuber, 1997). Like the other layers we have discussed, an LSTM
layer is composed of smaller computational units. Unlike the others, it is designed specifically to operate
over a sequence of inputs and uses, at each step of the sequence, its output from the previous step as
one of its inputs. As a result, it performs a recurrent computation which allows it, in theory, to efficiently
model interactions between inputs appearing at different steps in the input sequence. This is
particularly attractive for modeling language which has clear sequential structure and many interactions
between inputs at various steps.
Suppose, for example, that we have a sequence of 4 words “She tripped and fell” and an LSTM layer
consisting of 2 LSTM units. Let denote the vector representing the th word, let
denote the vector
valued output of the LSTM at step . For each word in the sequence the LSTM will accept as input and
the output from the previous step
and produce a new output . In our example, it might produce
the following:

Step ( )
1
2
3
4

Input from sequence ( )
She
tripped
and
fell

Input from previous output (
[0.0, 0.0]
[0.1, 0.4]
[2.4, -1.3]
[0.4, 1.5]

)

Output ( )
[0.1, 0.4]
[2.4, -1.3]
[0.4, 1.5]
[3.2, 2.2]

At each step the output of the LSTM layer is controlled by 4 interconnected structures known as the
input gate vector , the forget gate vector , the cell state vector , and the output gate vector .
Each of these structures is in turn associated with its own weight matrices, and , and bias vector .
We use subscripts , , , and to indicate the specific structure (forget gate, input gate, output gate, or
cell state, respectively) that each weight matrix and bias vector belongs to. Each is calculated as shown
below, where ∘ denotes the element wise product:
=

=

=

(

(

+

+

+

+

+

+

)

)

=

∘

=

+

∘

(

∘

( )

+

+

)

A common modification to the LSTM which often improves performance (and which we use in our work)
is to combine the outputs of 2 LSTM layers, each operating on the same input in opposite directions.
This is known as a Bidirectional LSTM.

Highway Layer
Although not a recurrent layer, the highway layer none the less draws inspiration from the LSTM. A
highway layer is a densely connected layer with LSTM style gating operations added to encourage the
flow of information through networks with many layers (Srivastava, Greff, & Schmidhuber, 2015). Let
be the vector of outputs from a highway layer, let denote a transform gate vector, let denote the
carry gate vector, and let ∘ denote the dot product. For a given input vector , the output of the
highway layer is calculated as follows:
=

(

= (

+

=1−
=

∘ +

+

∘

)

)

When the transform gate is fully closed, i.e. every element is 0, the highway layer simply outputs the
original input vector. When the transform gate is fully open, the highway layer operates as a dense
logistic layer. When it is partially open, it interpolates between the two.

Dropout
Although not a layer with learnable parameters, dropout is another important technique for regularizing
neural networks that is sometimes conceptualized as a layer (Srivastava, Hinton, Krizhevsky, Sutskever,
& Salakhutdinov, 2014). During training, dropout is a binomial mask that randomly sets a fraction of the
inputs to the specified layer to 0. This has the effect of discouraging the following layer from becoming
overly reliant on any individual input or combination of inputs. With some modifications, dropout has
also been successfully applied to the recurrent connections of the LSTM (Gal, 2016).

Our neural network
The neural network techniques described above provide a rich set of building blocks for modeling a
variety of phenomena but still leave open the question of how best to apply them to our task. In
attempting to answer this, we are guided in particular by the following:
1. Our previous work on logistic regression based SOII autocoders found that sub-word
information was often useful for addressing the misspellings and unusual abbreviations that
frequently occur in SOII data. A successful neural network autocoder must be able to address
similar issues. We draw particular inspiration from the work of (Kim, Jernite, Sontag, & Rush,
2016), which showed that character level convolutions followed by highway layers produce
effective word representations for translation tasks.
2. SOII autocoding is a text classification task in which the context in which words appears is
frequently important. It should be important, therefore, that the model be able to efficiently

capture this context. Our model is particularly inspired by the LSTM with hierarchical attention
proposed by Yang, which is, to our knowledge, the current state of the art for text classification
(Yang, et al., 2016).
3. SOII coding is inherently multitask, each case receives 6 codes and there are clear relationships
between these codes. Previous work by (Collobert & Weston, 2008) has shown the benefits of
jointly learning a single model for multiple tasks, we seek to create a neural network that can do
this as well.
With these considerations in mind, we propose a single neural network autocoder for all SOII coding
tasks consisting of the following 4 major components:
1. a single word encoder, which, for each word in a SOII text field generates a vector
representation of that word based on the characters it contains
2. 3 field encoders, specifically:
o a narrative encoder, which generates context aware vector representations for each
word embedding in a narrative field
o an occupation encoder, which generates context aware vector representations for each
word vector embedding in the “job title” or “other text” fields
o a company encoder, which generates context aware vector representations for each
word vector in a company name or secondary name field
3. a NAICS encoder, which generates a vector representation of a NAICS code
4. 6 task specific output modules, one for each of the coding tasks (occupation, nature, part, event,
source, and secondary source). Each module consists of:
o an attention layer which weights and aggregates the concatenated outputs of the field
encoders to form a vector representation of the document
o two highway layers, which accept as input the concatenation of the NAICS embedding
and the document embedding
o a softmax layer, which produces predicted probabilities for each possible code from the
output of the last highway layer
We describe the operation of the network and its components in more detail below.

Input and Preprocessing
The key inputs for SOII coding consist of the following information collected for each case:
Name of field
occupation_text
other_occupation_text
primary_estab_name
secondary_estab_name
nar_activity
nar_event
nar_nature

Description
The worker’s job title
An optional field indicating the worker’s job
category.
The primary name of the worker’s establishment.
The secondary name of the worker’s establishment.
A narrative answering “What was the worker doing
before the incident occurred?”
A narrative answering “What happened?”
A narrative answering “What was the injury or
illness?”

Example
Registered nurse
Elder care
ACME hospitals inc.
ACME holding corp
Helping patient get
out of bed
The patient slipped
Employee strained
lower back

nar_source
naics

A narrative answering “What object or substance
directly harmed the employee?”
The 2012 North American Industry Classification
System (NAICS) code for the establishment

Patient and floor
622110

Before this data is fed into our neural network, we pre-process it as follows:






Each text field (all inputs except naics) is separated into a sequence of its component words
using the NLTK TreebankTokenizer (Bird, Klein, & Loper, 2009). Each sequence is then right
padded (or, when necessary, truncated) to a fixed length of 15, for occupation and
establishment fields, and 60 for the others.
Each word in each text field is further separated into its component characters. Special
and tokens are added to the start and end of the sequence. A special
token is also added to the start of the sequence to indicate the field from which the word came.
The character sequence is then right padded (or truncated) to a length of 15.
Each of the 6 digits of the NAICS code is mapped to one-hot 10 dimensional vectors. These are
then concatenated to create a 60 dimensional representation of the NAICS code.

Word encoder
The word encoder operates at the lowest level of our network and converts each sequence of characters
representing a word into a 700 dimensional vector. This is accomplished by mapping each character to a
15 dimensional embedding and then applying convolutional layers with input lengths ranging from 1-7
across the input. Each convolutional layer is given a number of filters equal to its input length times 25.
The output of each convolutional layer is then max pooled, concatenated with the output of other
convolution layers, and fed through a highway network. Our word encoder differs from the small
version proposed in (Kim, Jernite, Sontag, & Rush, 2016) in only a few respects. Instead of using
character level convolutions of lengths 1-6 we use lengths of 1-7. We also found our encoder performs
better when the convolutional layers have no nonlinear activation and when we follow them with only
one highway layer.
After being processed by the word encoder each text field is represented by a sequence of vectors, each
corresponding to a word (or padding-word) it contains. The occupation_text field, for example, is
represented by a 15 x 700 dimension matrix.

Field encoders
Each of our three field encoders is a bidirectional LSTM with 512 cells in each direction (1024 total),
applied to the sequence of word embeddings generated by applying our word encoder to each word in
an input field. Each field encoder operates on the embeddings from the fields indicated in the table
below.
Field encoder
narrative encoder
occupation encoder
company encoder

Input fields
nar_activity, nar_event, nar_nature, nar_source
occupation_text, other_occupation_text
primary_estab_name, secondary_estab_name

Sharing the encoders across fields in this way is motivated partly by experimentation and partly by
intuition. The nar_activity, nar_event, nar_nature, and nar_source fields frequently contain similar
information, often a sentence or two describing some aspect of the injury or a few key words. The
occupation_text and other_text fields normally contain only job titles and the primary_estab_name and
secondary_estab_name fields generally contain names of establishments.
Each field encoder also uses Gal (Gal, 2016) style dropout of 50% between the input to hidden and
hidden to hidden connections. After processing a field the outputs produced by the bidirectional LSTM
at each time step are concatenated and then output to following layers.

NAICS encoder
The NAICS encoder accepts as input the 60 dimensional vector representation of the NAICS code. This is
then passed through 2 highway layers with rectified linear activations and finally a linear dense layer to
produce a 60 dimensional embedding intended to map similar NAICS codes to similar vector
representations.

Attention layer
The attention layer is a multilayer neural network in the style of (Bahdanau, 2014) applied to the
concatenated outputs of the field encoders. For each step in the concatenated outputs it calculates an
attention weight between 0 and 1 such that the total attention across steps sums to 1. The attention
weighted average of these outputs becomes the output of the attention layer. Our use of attention was
inspired by the work of (Yang, et al., 2016), who found attention layers were useful for text
classification. In contrast to their work however, we did not find any benefit from using a hierarchical
system of attention that operates at both the word and sentence level. This is likely because SOII
narratives are rarely longer than two sentences.
The attention layers for OIICS coding operate only over the concatenated outputs of the narrative
encoder. The attention layer for SOC coding operates over the concatenated outputs of the occupation,
company, and narrative encoders.

Implementation
We implement our neural network in Python using the Keras library with the Tensorflow backend
(Chollet, 2017; Abadi, et al., 2016).

Evaluation
To evaluate our approach we use the approximately 1.3 million SOII cases collected, coded, and deemed
usable for estimation between 2011 and 2015. We divide this data into 2 non-overlapping sets, a test set
consisting of a random sample of 1000 SOII cases from survey year 2011, and the rest, which we use for
training and validation.
Although each of the codes in these cases was assigned by trained BLS coders (or reviewed by BLS
coders, for those that were automatically assigned), and each is subject to a variety of automatic
consistency checks and reviews in the regional and national offices, we know that errors none the less
make it through. To determine the best possible codes for the test set we hid the codes that were
originally assigned and then distributed the cases to coding experts in regional offices in such a way that
each case would be given to 3 different experts, each in a different region. We then instructed the
experts to recode the cases from scratch as accurately as possible without referencing previously coded

data or discussing code assignments with others. After the cases were recoded, the codes were
reviewed by experts in the national office who resolved any coding disagreements. During this process
one of our national office experts determined that 2 cases could not receive a reliable SOC code, these
cases were excluded from the SOC portion of our evaluation. We refer to the remaining codes as our
gold standard and use it as our measure of truth.
To train our neural network autocoder we randomly divide the remaining data into a training set of
1,200,000 and a validation set of 100,000. We then train the neural network on the training set using
the Adam optimizer with an initial learning rate of .004 and a 50% decay in learning rate after 2 epochs
with no decrease in cross-entropy loss as measured against the validation set. Training was stopped
after loss on the validation set failed to decrease for 5 consecutive epochs.
For comparison, we also evaluate the current SOII logistic regression autocoders and the coding
produced by the manual process, as it existed in 2011. The logistic regression autocoders were trained
and validated on the same 1,300,000 cases used for the neural network, although they have the slight
benefit of, after validation, being retrained on the full set of 1,300,000 cases. We use the codes
originally assigned to our gold standard cases to assess the quality of our manual coding process, which
consists of human coding followed by automated consistency checks and review at the state and
national level.
We rely on two metrics to assess the quality of coding: accuracy, and macro-F1 score. Accuracy is
calculated as the fraction of cases receiving the correct code when all cases are coded. Macro-F1-score
is calculated as the average of the code specific F1-scores, which can be thought of as code-specific
accuracy measures. Unlike accuracy, which is more reflective of performance on more frequently
occurring codes, the macro-F1 score gives each code specific F1-score equal weight, regardless of how
frequently that code occurs. Each code specific F1-score is calculated as follows:

where

1

∗
+

=2∗

=

=

ℎ

ℎ

One ambiguity in calculating the F1-score is that precision and recall are sometimes 0, making the F1score undefined. When this occurs we adopt the convention used in scikit-learn and simply treat the F1score as 0 (Pedregosa, et al., 2011).
Comparison of the accuracy and macro-F1-scores shows that the neural network autocoder outperforms
the alternatives across all coding tasks and makes an average of 24% fewer errors than our logistic
regression autocoders (excluding secondary source, for which we do not currently have a logistic
regression autocoder), and an estimated 39% fewer errors than our manual coding process. On each
task the neural network’s accuracy is statistically greater than the next best alternative at a p-value of
0.001 or less.

Accuracy
100

78.6
68.3 75.0

80.4

91.9
84.4 88.4

88.3 91.1

52.4

50

61.3

69.8

62.8 66.7

75.8

79.4

85.0

0
soc

nature

part
human

LR autocoder

event

source

secondary source

NN autocoder

Macro F1-score
100
50

56.0
46.4 53.6

38.2

73.4
62.0 67.5

46.8 54.0

32.2 34.1

44.4

40.7 41.4

52.4
18.2

26.2

0
soc

nature

part
human

LR autocoder

event

source

secondary source

NN autocoder

Evaluation at the 2 digit level of coding detail, which is a popular level of coding aggregation, was
accomplished by truncating codes to the 2 left most digits and shows better performance for the neural
network autocoder here as well, except for the macro-F1-score on occupation coding. The macro-F1
scores are quite noisy at the 2 digit level of detail however as even a few data points can have a large
impact on the score, so we urge caution in interpreting them. On each 2-digit task the neural network’s
accuracy is statistically greater than the next best alternative at a p-value of 0.001 or less, except when
compared to logistic regression on SOC coding, where the p-value is approximately 0.015.

Accuracy at the 2 digit level
100

83.4 88.2 90.2

89.1 92.5 94.4

91.8 92.7 95.8

soc

nature

part

82.1 81.4 86.7

78.7 77.1 84.1

event

source

81.6

86.0

50
0
human

LR autocoder

NN autocoder

secondary source

Macro F1-score at the 2 digit level
100

76.6

86.2 81.0
63.2

74.1 77.8

71.0 75.5

85.8
59.0 59.8

71.5

58.0 56.2

66.4

50

34.5

41.0

0
soc

nature

part
human

LR autocoder

event

source

secondary source

NN autocoder

Model analysis and future challenges
Although improved coding performance is an important goal of our research, SOII autocoding also faces
other challenges which we hope to partially address with our use of neural networks. One of these is the
transition to new SOC and OIICS classification systems, which is scheduled to occur within the next few
years. When this happens SOII autocoders will need to be retrained for the new systems but we may not
have sufficient training data coded under these new systems, requiring a return to manual coding until
such data can be acquired.
Although there are already a variety of techniques for transferring machine learning knowledge across
tasks, including some that are algorithm agnostic such as (Daumé III, 2009), neural networks have
unique advantages. Unlike SOII’s logistic regression autocoders, for example, which essentially learn a
direct mapping from inputs (i.e. words) to code probabilities, our neural network learns a variety of
intermediate computations which may be useful for other coding tasks. To the extent this is the case,
the layers responsible for these computations can be readily transferred directly into autocoders
intended for new tasks, reducing the amount of data needed to learn an effective model. Similar
approaches are in fact already widely used in text processing, most extensively in the form of word
embeddings (see (Mikolov, Chen, Corrado, & Dean, 2013) and (Pennington, Socher, & Manning, 2014)),
but there is evidence supporting the transferability of other components as well. For examples, see (Lili,
et al., 2016) and (Lee, Dernoncourt, & Szolovits, 2017).
Although we cannot fully determine how useful the intermediate computations of our network will be
for classification systems that do not yet exist, a closer examination provides some clues and sheds light
on the operation of the neural network more generally. In particular, we examine the behavior of our
word encoder, NAICS encoder, and attention mechanisms.
To better understand our NAICS and word encoders we take the 500 most frequently occurring inputs to
these encoders, calculate the vector representations they produce, then project those representations
to 2 dimensional space using the t-SNE algorithm (Maaten & Hinton, 2008) which is designed to
preserve the local distance relationships between vectors. Subsets of the results for the job title and
NAICS fields, are plotted below. The complete visualizations appear in the Appendix.

Job title embeddings

Our visualization of the 500 most frequently occurring words in the job title field indicates that the word
encoder has learned to perform a variety of normalizations. For example, in the upper left portion of our
visualization we find a cluster of words that all represent some variation of the word “nurse”, including
“NURSE”, “nursing”, “Nurses”, and “RN”, which is a popular acronym for “registered nurse”. The
embeddings also reflect similarities purely in meaning. In the top right, variations of the words “food”
and “kitchen” appear next to each other.

NAICS embeddings

Our visualizations also indicate that the NAICS encoder has learned to capture similarities in meaning.
Although the NAICS embeddings are mostly projected into sector level (2 digit) and subsector level (3
digit) clusters, there are interesting and informative exceptions. In the top middle of the visualization,
for example, we see a cluster of 4 NAICS codes, 511110 (Newspaper Publishers), 511130 (Book
Publishers), 323117 (Commercial Screen Printing), and 323111 (Books Printing). These are clearly related
industries and the proximity of the associated NAICS embeddings accurately reflects this, but
interestingly, the official NAICS hierarchy does not. In fact, the official hierarchy separates “Publishing”

and “Printing” into 2 entirely separate super sectors, the largest possible separation. In this sense, the
NAICS encoder has learned a more useful representation than the official manual, suggesting that
similar embedding techniques might be well suited for creating improved hierarchies.

Attention
The neural network’s attention layers provide another window into its function. Recall that for each
word in an input field, the field encoder produces a vector output. This results in a sequence of vectors
which are then fed to the attention layer. The purpose of the attention layer is to weight and then
aggregate (i.e. sum) these vectors so that the following highway layers and softmax layer can accurately
predict the correct code. A side effect is that the attention weights provide clues as to what inputs the
model is focusing on. In the visualization below, we align the word inputs of our model with the
attention weights produced by the “nature code” attention layer and then highlight those words
according to the weight they receive. To make the attention more visible, we normalize the alpha of our
yellow highlighting so that 0 is the lowest attention weight assigned and 1 is the highest.

Nature code attention

In this example, our attention layer has placed extra weight on the vectors coinciding with the words,
“strained” and “Spondylothesi” [sic], both of which indicate potential injury natures. We find that this
holds in general, the part of body attention layer focuses on parts of body, the event layer focuses on
words associated with events, and so on. Unexpectedly, we also find that the attention mechanisms
place a lot of weight on the vectors produced immediately after the word inputs stop and padding
(denoted by “_”) begins. We notice this phenomenon in all of our attention layers on all of the narrative
fields, but not the occupation and company name fields. This likely reflects some sort of summary
computation being performed in the field encoder, which is capable of storing information collected
during the reading of a narrative and then releasing that information when it detects that the word
sequence has ended. Earlier LSTM research often relied entirely on the summary vector computed by
the LSTM. Our work suggests that our model, despite having access to the intermediate steps in the
LSTM computation, still does this to some extent.

Conclusion
We have introduced a deep neural network autocoder that makes 24% fewer coding errors than our
logistic regression autocoders and 39% fewer errors than our manual coding process. Our analysis
suggests the neural network accomplishes this in part by performing a variety of intermediate
computations including text and NAICS code normalization and attention, which are likely to be useful

for a variety of tasks and future SOII classification systems. Our work adds to the growing evidence that
automatic coding is not merely a tool for easing workload but also for improving the overall quality of
coding. This has clear implications for a wide variety of similar coding tasks performed in the health
economic statistics communities.

Acknowledgements
The opinions presented in this paper are those of the author and do not represent the opinions or
policies of the Bureau of Labor Statistics or any other agency of the U.S. government.

References
Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z., Citro, C., . . . Ghemawat, S. (2016). TensorFlow:
Large-scale machine learning on heterogeneous systems. Retrieved from http://tensorflow.org
Bahdanau, D. C. (2014). Neural Machine Translation by Jointly Learning to Align and Translate. Retrieved
from arXiv preprint: https://arxiv.org/pdf/1409.0473.pdf
Bird, S., Klein, E., & Loper, E. (2009). Natural Language Processing with Python. O'Reilly Media Inc.
Retrieved from http://www.nltk.org/
Chollet, F. a. (2017). Keras. Retrieved from https://github.com/fchollet/keras
Collobert, R., & Weston, J. (2008). A unified architecture for natural language processing: Deep neural
networks with multitask learning. Proceedings of the 25th international conference on Machine
learning (ICML) (pp. 160-167). ACM. Retrieved from
https://ronan.collobert.com/pub/matos/2008_nlp_icml.pdf
Daumé III, H. (2009). Frustratingly Easy Domain Adaptation. Retrieved from arXiv preprint:
https://arxiv.org/pdf/0907.1815.pdf
Fan, R.-E., Chang, K.-W., Hsieh, C.-J., Wang, X.-R., & Lin, C.-J. (2008). LIBLINEAR: A Library for Large Linear
Classification. Journal of Machine Learning Research, 1871-1874. Retrieved from
http://www.jmlr.org/papers/volume9/fan08a/fan08a.pdf
Gal, Y. a. (2016, October 5). A Theoretically Grounded Application of Dropout in Recurrent Neural
Networks. Retrieved from arXiv preprint: https://arxiv.org/pdf/1512.05287.pdf
Goldberg, Y. (2016). A Primer on Neural Network Models for Natural Langauge Processing. J. Artif. Intell.
Res.(JAIR), 353-420. Retrieved from https://www.jair.org/media/4992/live-4992-9623-jair.pdf
Goodfellow, I. B. (2016). Deep Learning. MIT press. Retrieved from http://www.deeplearningbook.org/
Hochreiter, S., & Schmidhuber, J. (1997). Long Short-Term Memory. Journal Neural Computation, 17351780.
Kim, Y., Jernite, Y., Sontag, D., & Rush, A. M. (2016). Character-Aware Neural Language Models. AAAI,
(pp. 2741-2749). Retrieved from
https://www.aaai.org/ocs/index.php/AAAI/AAAI16/paper/viewFile/12489/12017
Lee, J., Dernoncourt, F., & Szolovits, P. (2017, May 17). Transfer Learning for Named-Entity Recognition
with Neural Networks. Retrieved from arXiv preprint: https://arxiv.org/pdf/1705.06273v1.pdf
Lili, M., Zhao, M., Rui, Y., Ge, L., Yan, X., Lu, Z., & Zhi, J. (2016, October 13). How Transferable are Neural
Networks in NLP Applications? Retrieved from arXiv preprint:
https://arxiv.org/pdf/1603.06111.pdf
Maaten, L., & Hinton, G. (2008). Visualizing High-Dimensional Data. Journal of Machine Learning
Research, 2579-2605. Retrieved from
http://www.jmlr.org/papers/volume9/vandermaaten08a/vandermaaten08a.pdf

Measure, A. (2014). Automated Coding of Worker Injury Narratives. JSM. Retrieved from
https://www.bls.gov/iif/measure_autocoding.pdf
Mikolov, T., Chen, K., Corrado, G., & Dean, J. (2013, September 7). Efficient estimation of word
representations in vector space. Retrieved from arXiv preprint:
https://arxiv.org/pdf/1301.3781.pdf
Pedregosa, F., Varoquax, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., . . . Duchesnay, É. (2011).
Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research, 2825-2830.
Retrieved from http://www.jmlr.org/papers/volume12/pedregosa11a/pedregosa11a.pdf
Pennington, J., Socher, R., & Manning, C. (2014). Glove: Global Vectors for Word Representation.
Proceedings of the 2014 conference on empirical methods in natural language processing
(EMNLP), (pp. 1532-1543). Retrieved from http://www.aclweb.org/anthology/D14-1162
Srivastava, N., Hinton, G. E., Krizhevsky, A., Sutskever, I., & Salakhutdinov, R. (2014). Dropout: A Simple
Way to Prevent Neural Networks from Overfitting. Journal of machine learning research, 19291958. Retrieved from Journal of Machine Learning Research:
http://www.jmlr.org/papers/volume15/srivastava14a/srivastava14a.pdf
Srivastava, R., Greff, K., & Schmidhuber, J. (2015). Highway networks. Retrieved from arXiv preprint:
https://arxiv.org/pdf/1505.00387.pdf
Yang, Z., Yang, D., Dyer, C., He, X., Smola, A. J., & Hovy, E. H. (2016). Hierarchical Attention Networks for
Document Classification. HLT-NAACL, (pp. 1480-1489). Retrieved from
http://www.aclweb.org/anthology/N16-1174

Appendix A
Description of features used in SOII’s logistic regression autocoders
All features are represented by a position in the feature vector for each unique value they take in the
training data. When the feature is present in an input it is represented by a value of 1 in the
corresponding position of the feature vector, otherwise it takes a value of 0
Feature name
occ_char4_un
full_company_wrd
occ_wrd
other_wrd
cat
naics
naics2
fips_state_code
occ_n2
occ_bi
nar
nar_bi
nar_nature
nature_char4
nature_conj
nature_list
nar_source
source_char4

Description
4 letter substrings in words in the occupation_text field
Word unigrams in the primary_estab_name and secondary_estab_name fields
Word unigrams in the occupation_text field
Word unigrams in the other_occupation_text field
1 of 12 job categories
6 digit NAICS code for the establishment
2 digit NAICS code for the establishment
2 digit FIPS state code for the establishment
Concatenation of occupation_text unigrams and the 2 digit NAICS code of the
establishment
Word bigrams in the occupation_text field
Word unigrams in the nar_activity, nar_event, nar_nature or nar_source fields
Word bigrams in the nar_activity, nar_event, nar_nature, or nar_source fields
Word unigrams in the nar_nature field
4 letter substrings in words in the nar_nature field
Alphabetically sorted concatenation of words appearing in the nar_nature field
that are separated by ‘and’ or ‘/’.
Alphabetically sorted concatenation of words appearing in the nar_nature field
that are separated by commas.
Word unigrams in the nar_source field.
4 character substrings in words in the nar_source field.

Subset of features used in each task-specific logistic regression model
Task specific model
occupation
nature
part
event
source

Features
occ_char4_un, full_company_wrd, occ_wrd, other_wrd, cat, naics,
fips_state_code, occ_n2, occ_bi
nar_nature, nature_char4, nar_bi, naics
nar_nature, nature_char4, nar, nature_conj, nature_list,
nar, nar_bi, naics
nar_source, source_char4, nar_bi, naics2, fips_state_code

Appendix B
Job title word embeddings
Our t-SNE visualization of job title word embeddings suggests that the word encoder has learned to
produce similar vectors for words and symbols with similar meanings. For example, the words “&” and
the “and” appear very close together, as do other closely related words such as “Associate” and “Aide”,
“Administrator” and “Secretary”, and “Manager” and “Director”.

Appendix C
NAICS code embeddings
Our visualization of NAICS embeddings suggests they closely match the hierarchy defined by NAICS, with
codes from similar sectors and subsectors mostly grouped together. The embeddings also reflect
similarities not captured by the official NAICS 2012 hierarchy however. For example, the embeddings for
511110 (Newspaper Publishers) and 511130 (Book Publishers), are close to 323113 (Commercial Screen
Printing) and 323117 (Books Printing). Similarly, the embeddings for 311 (Animal Product
Manufacturing) are near 112 (Animal Production) and 1152 (Support Activities for Animal Production),
but 11310 (Logging) is located far away, near 337 (Furniture) and 321 (Wood Product Manufacturing).