Automated Measures of Specific Vocabulary Knowledge

Automated Measures of Specific Vocabulary Knowledge from Constructed
Responses (“Use These Words to Write a Sentence Based on this Picture”)
Swapna Somasundaran
Educational Testing Services
660 Rosedale Road,
Princeton, NJ 08541, USA
[email protected]
Martin Chodorow
Hunter College and the Graduate Center
City University of New York,
New York, NY 10065, USA
[email protected]
In this paper we present a system for automated
scoring of targeted vocabulary knowledge based
on short constructed responses in a picture description task. Specifically, we develop a system
for scoring a vocabulary item type that is in operational use in English proficiency tests for nonnative speakers. Each task prompt in this item type
consists of two target key words, for which the vocabulary proficiency is tested, and a picture that
provides the context for the sentence construction.
The task is to generate a single sentence, incorporating both key words, consistent with the picture.
Presumably, a test-taker with competent knowledge of the key words will be able to use them in a
well-formed grammatical sentence in the context
of the picture.
Picture description tasks have been employed in
a number of areas of study ranging from second
language acquisition to Alzheimer’s disease (Ellis, 2000; Forbes-McKay and Venneri, 2005). Pictures and picture-based story narration have also
been used to study referring expressions (Lee et
al., 2012) and to analyze child narratives in order
to predict language impairment (Hassanali et al.,
2013). Evanini et al. (2014) employ a series of
pictures and elicit (oral) story narration to test English language proficiency. In our task, the picture
is used as a constraining factor to limit the type
and content of sentences that can be generated using the given key words.
In the course of developing our system, we examined existing features that have been developed
for essay scoring, such as detectors of errors in
grammar, usage and mechanics, as well as collocation features, to see if they can be re-used
for scoring short responses. We also developed
new features for assessing the quality of sentence
construction using Pointwise Mutual Information
(PMI). As our task requires responses to describe
the prompt pictures, we manually constructed detailed textual descriptions of the pictures, and de-
We describe a system for automatically
scoring a vocabulary item type that asks
test-takers to use two specific words in
writing a sentence based on a picture. The
system consists of a rule-based component
and a machine learned statistical model
which uses a variety of construct-relevant
features. Specifically, in constructing the
statistical model, we investigate if grammar, usage, and mechanics features developed for scoring essays can be applied to
short answers, as in our task. We also explore new features reflecting the quality of
the collocations in the response, as well as
features measuring the consistency of the
response to the picture. System accuracy
in scoring is 15 percentage points greater
than the majority class baseline and 10
percentage points less than human performance.
It is often said that the best way to see if a person knows the meaning of a word is to have that
person use the word in a sentence. Despite this
widespread view, most vocabulary testing continues to rely on multiple choice items (e.g. (Lawless et al., 2012; Lawrence et al., 2012)). In
fact, few assessments use constructed sentence responses to measure vocabulary knowledge, in part
because of the considerable time and cost required
to score such responses manually. While much
progress has been made in automatically scoring writing quality in essays (Attali and Burstein,
2006; Leacock et al., 2014; Dale et al., 2012),
the essay scoring engines do not measure proficiency in the use of specific words, except perhaps
for some frequently confused homophones (e.g.,
its/it’s, there/their/they’re, affect/effect).
Proceedings of the Ninth Workshop on Innovative Use of NLP for Building Educational Applications , pages 1–11,
Baltimore, Maryland USA, June 26, 2014. 2014
Association for Computational Linguistics
veloped features that measure the overlap between
the content of the responses and the textual description. Our automated scoring system is partly
based on deterministic scoring criteria and partly
statistical. Overall, it achieves an accuracy of
76%, which is a 15 percentage point improvement
over a simple majority class baseline.
The organization of this paper is as follows:
Section 2 describes the picture description task
and the scoring guide that is used to manually
score the picture description responses operationally. It also considers which aspects of scoring may be handled best by deterministic procedures and which are more amenable to statistical
modeling. Section 3 details the construction of a
reference corpus of text describing each picture,
and Section 4 presents the features used in scoring. Section 5 describes our system architecture
and presents our experiments and results. Detailed
analysis is presented in Section 6, followed by related work in Section 7 and a summary with directions for future research in Section 8.
the form of the word(s) may not be accurate),
AND (c) are consistent with the picture.
score = 1 The response: (a) has errors that interfere with meaning, (b) omits one or both
key words, OR (c) is not consistent with the
score = 0 The response is blank, written in
a foreign language, or consists of keystroke
Our decisions about scoring system design are
based on the scoring guide and its criteria. Some
aspects of the scoring can be handled by simple
pattern matching or lookup, while others require
machine learning. For example, score 0 is assigned to responses that are blank or are not in
English. This can be detected and scored in a
straightforward way. On the other hand, the determination of grammaticality for the score points
3, 2 and 1 depends on the presence and severity
of grammatical errors. A wide variety of such errors appear in responses, including errors of punctuation, subject-verb agreement, preposition usage
and article usage. The severity of an error depends
on how problematic the error is, and the system
will have to learn this from the behavior of the
trained human scorer(s), making this aspect of the
scoring more amenable to statistical modeling.
Similarly, statistical modeling is more suitable
for determining the consistency of the response
with respect to the picture. According to the scoring guide, a response gets a score of 0 or 1 if it is
not consistent with the picture, and gets a score of
2 or 3 if it is consistent. Thus, this aspect cannot
solely determine the score of a response – it influences the score in conjunction with other language
proficiency factors. Further, measures of how relevant a response is to a picture are likely to fall on
a continuous scale, making a statistical modeling
approach appropriate.
Finally, although there are some aspects of the
scoring guide, such as the number of sentences
and the presence of the key words, that can be
measured trivially, they do not act as sole determinants of the score. For example, having more
than one sentence can result in the response receiving a score of 2 or 1. The number of sentences
works in conjunction with other factors such as
severity of grammar errors and relevance to the
picture. Hence its contribution to the final score
is best modeled statistically.
Task Description and Data
The picture description task is an item type that is
in actual operational use as part of a test of English. It consists of a picture, along with two key
words, one or both of which may be in an inflected
form. Test-takers are required to use the two words
in one sentence to describe the picture. They may
change the inflections of the words as appropriate
to the context of their sentence, but they must use
some form of both words in one sentence. Requiring them to produce a response based on the picture constrains the variety of sentences and words
that they are likely to generate.
Trained human scorers evaluate the responses
based on appropriate use of grammar and the relevance of the sentence to the picture. The operational scoring guide is as follows:
score = 3 The response consists of ONE sentence that: (a) has no grammatical errors, (b)
contains forms of both key words used appropriately, AND (c) is consistent with the picture.
score = 2 The response consists of one or
more sentences that: (a) have one or more
grammatical errors that do not obscure the
meaning, (b) contain BOTH key words, (but
they may not be in the same sentence and
As a result of the heterogeneous nature of the
problem, our system is made up of a statistical
learning component as well as a non-statistical
overall setting (restaurant), as well as the objects that make up the picture (e.g. man, table,
food). These are generally (but not necessarily) nouns and noun phrases. Some pictures
can have many items, while some have only a
few. The goal is to list 10-15 items and to capture as many items as possible, *starting with
the most obvious ones*.
If the picture is too sparse, and you are not
able to list at least 10 items, please indicate
this as a comment.
Part:2 Describe the picture
Describe the scene unfolding in the picture.
The scene in the picture may be greater than
the sum of its parts (many of which you will
list in part-1). For example, the objects in a
picture could be “shoe” “man” “chair”, but the
scene in the picture could be that of a shoe
purchase. The description tries to recreate the
scene (or parts of the scene) depicted in the
Generate a paragraph of 5-7 sentences describing the picture. Some of these sentences
will address what is going on, while some may
address relations between items. The proportions of these will differ, based on the picture.
Make sure that you generate at least one sentence containing the two key words.
If the picture is too simple, and you are not
able to generate at least 5 sentences, please
indicate this as a comment.
The data set consists of about 58K responses to
434 picture prompts. The mean response length
was 11.26 words with a standard deviation of 5.10.
The data was split into 2 development sets (consisting of a total of about 2K responses) and a final train-test set (consisting of the remaining 56K
responses) used for evaluation. All 58K responses
were human scored using the scoring rubric discussed in Section 2. About 17K responses were
double annotated. The inter-annotator agreement,
using quadratic weighted kappa (QWK), was 0.83.
Score point 3, the most frequent class, was assigned to 61% of the responses, followed by score
point 2 (31%), score point 1 (7.6%) and score
point 0 (0.4%).
Reference Picture Descriptions
The pictures in our task vary in their complexity.
A typical prompt picture might be a photograph
of an outdoor marketplace, the inside of an airport
terminal, a grocery store, a restaurant or a store
room. Because consistency with respect to the picture is a crucial component in our task, we needed
a reliable and exhaustive textual representation of
each picture. Therefore, we manually constructed
a reference text corpus for each of our 434 picture prompts. We chose to use manual creation of
the reference corpus instead of trying automated
image recognition because automated methods of
image recognition are error prone and would result
in a noisy reference corpus. Additionally, automated approaches would, at best, give us a (noisy)
list of items that are present in the picture, but not
the overall scene or event depicted.
Two annotators employed by a company that
specializes in annotation created the reference corpora of picture descriptions. The protocol used for
creating the reference corpus is shown below:
The human annotator was given the picture and
the two key words. The protocol for creating each
reference corpus asked the annotator to first exhaustively list all the items (animate and inanimate) in the picture. Then, the annotator was
asked to describe the scene in the picture. We used
this two step process in order to capture, as much
as possible, all objects, relationships between objects, settings and events depicted in the pictures.
The size of the reference corpus for each prompt
is much larger than the single sentence test-taker
response. This is intentional as the goal is to make
the reference corpus as exhaustive as possible. We
used a single annotator for each prompt. Double
annotation using a secondary annotator was done
in cases where we felt that the coverage of the corpus created by the primary annotator was insuffi-
Part-1: List the items, setting, and events
in the picture.
List, one by one, all the items and events you
see in the picture. These may be animate objects (e.g. man), inanimate objects (e.g. table)
or events (e.g. dinner). Try to capture both the
cient1 .
In order to test coverage, we used a small development set of essays from each prompt and compared the coverage of the generated reference corpus over the development essays. If the coverage (proportion of content words in the responses
that were found in the reference corpus) was less
than 50% (this was the case for about 20% of
the prompts), we asked the secondary annotator to
create a new reference corpus for the prompt. The
two reference corpora for the prompt were then
simply combined to form a single reference corpus.
grammar, usage and mechanics features that detect and quantify different types of English usage
errors in essays. Examples of some of these error
types are: Run-on Sentences, Subject Verb Agreement Errors, Pronoun Errors, Missing Possessive Errors, Wrong Article Errors, Missing Article Errors, Preposition Errors, Non-standard Verb
or Word Form Errors, Double Negative Errors,
Fragment or Missing Comma Errors, Ill-formed
Verb Errors, Wrong Form of Word Errors, Spelling
Errors, Wrong Part of Speech Errors, and Missing
Punctuation Errors .
In addition to these, essay scoring engines often also use as features the Number of Sentences
that are Short, the Number of Sentences that are
Long, the Number of Passive Sentences, and other
features that are relevant only for longer texts such
as essays. Accordingly, we selected, from e-rater
113 grammar, word usage, mechanics and lexical
complexity features that could be applied to our
short response task. This forms our grammar feature set.
Features for automated scoring
Because the score points in the scoring guide conflate, to some degree, syntactic, semantic, and
other weaknesses in the response, we carried out
a scoring study on a second small development
set (comprising of a total of 80 responses from 4
prompts, picked randomly) to gather insight into
the general problems in English language proficiency exhibited in the responses. For the study,
it was necessary to have test-taker responses rescored by an annotator using an analytic scheme
which makes the types and locations of problems explicit. This exercise revealed that, in addition to the factors stated explicitly in the scoring guide, there is another factor that results in
low comprehension (readability) of the sentence
and that reflects lower English proficiency. Specifically, the annotator tagged many sentences as being “awkward”. This awkwardness was due to
poor choice of words or to poor construction of the
sentence. For example, in the sentence “The man
is putting some food in bags while he is recording for the payment”, “recording for the payment”
was marked as an awkward phrase. Based on our
annotation of the scores and on the descriptions in
the scoring guide, we selected features designed to
capture grammar, picture relevance and awkward
usage. We discuss each of our feature sets in the
following subsections.
Features for Measuring Content
We generated a set of features that measure the
content overlap between a given response and the
corresponding reference corpus for the prompt.
For this, first the keywords and the stop words
were removed from the response and the reference
corpus, and then the proportion of overlap was calculated between the lemmatized content words of
the response and the lemmatized version of the
corresponding reference corpus, as follows:
|Response ∩ Corpus|
It is not always necessary for the test-taker to
use exactly the same words found in the reference
corpus. For example, the annotator might have
referred to a person in the picture as a “lady”,
while a response may refer to the same person
as a “woman” or “girl” or even just “person”.
Thus, we needed to go beyond simple lexical
match. In order to account for synonyms, we expanded the content words in the reference corpus
by adding their synonyms, as provided in Lin’s
thesaurus (Lin, 1998) and then compared the expanded reference to each response. Along the
same lines, we also used expansions from WordNet synonyms, WordNet hypernyms and WordNet
hyponyms. The following is the list of our content
Features for Grammatical Error
Essay scoring engines such as e-rater
and Burstein, 2006) typically use a number of
We do not conduct inter-annotator agreement studies as
the goal of the double annotation was to create a diverse description.
this intuition in constructing our PMI-based features.
We find the PMI of all adjacent word pairs
(bigrams), as well as all adjacent word triples
(trigrams) in the Google 1T web corpus (Brants
and Franz, 2006) using the TrendStream database
(Flor, 2013).
PMI between word pairs (bigram AB) is defined
relevance features. Each measures the proportion
of overlap as described by the equation above between the lemmatized response and
1. lemmas: the lemmatized reference corpus.
2. cov-lin: the reference corpus expanded using
Lin’s thesaurus.
3. cov-wn-syns: the reference corpus expanded
using WordNet Synonyms.
4. cov-wn-hyper: the reference corpus expanded using WordNet Hypernyms.
and between word triples (trigram ABC) as
5. cov-wn-hypo: the reference corpus expanded using WordNet Hyponyms.
The higher the value of the PMI, the more common is the collocation for the word pair/triple in
well formed texts. On the other hand, negative
values of PMI indicate that the given word pair (or
triple) is less likely than chance to occur together.
We hypothesized that this would be a good indicator of awkward usage, as suggested in (Chodorow
and Leacock, 2000).
The PMI values for adjacent words obtained
over the entire response are then assigned to bins,
with 8 bins for word pairs and another 8 for word
triples. Each bin represents a range for PMI p taking real values R as follows:
6. cov-all: the reference corpus expanded using
all of the above methods.
Mean proportions of overlap ranged from 0.65
for lemmas to 0.97 for cov-all.
The 6 features listed above, along with the
prompt id give a total of 7 features that form our
relevance feature set. We use prompt id as a feature because the extent of overlap can depend on
the prompt. Some pictures are very sparse, so,
the description of the picture in the response will
be short, and will not vary much from the reference corpus. For these, a good amount of overlap
between the response and reference corpus is expected. Other pictures are very dense with a large
number of objects and items shown. In this case,
any single response may describe just a small subset of the items and satisfy the consistency criteria,
and consequently, even a small overlap between
the response and the reference corpus may be sufficient.
bin1 = {p ∈ R | p > 20}
bin2 = {p ∈ R | 10 < p ≤ 20}
bin3 = {p ∈ R | 1 < p ≤ 10}
bin4 = {p ∈ R | 0 < p ≤ 1}
bin5 = {p ∈ R | − 1 < p ≤ 0}
Features for Awkward Word Usage
bin6 = {p ∈ R | − 10 < p ≤ −1}
In order to measure awkward word usage, we explored PMI-based features, and also investigated
whether some features developed for essay scoring can be used effectively for this purpose.
bin7 = {p ∈ R | − 20 < p ≤ −10}
bin8 = {p ∈ R | p ≤ −20}
Once the PMI values for the adjacent word pairs
in the response are generated, we generate two sets
of features. The first set is based on the counts
of word pairs falling into each bin (for example,
Number of pairs falling into bin1 , Number of pairs
falling into bin2 and so on). The second set of features are based on percentages (for example Percentage of pairs falling into bin1 , Percentage of
pairs falling into bin2 etc.). These two sets result
in a total of 16 features. We similarly generate
16 more features for adjacent word triples. We
4.3.1 PMI-based ngram features
Non-native writing is often characterized by inappropriate combinations of words, indicating the
writer’s lack of knowledge of collocations. For example, “recording for the payment” might be better expressed as “entering the price in the cash register”. As “recording for the payment” is an inappropriate construction, it is not likely to be common, for example, in a large web corpus. We use
the prompt present in the response? Is there more
than one sentence in the response?
Table 1 provides a list of feature types and the
corresponding number of features of each type.
use percentages in addition to raw counts to account for the length of the response. For example,
it is possible for a long sentence to have phrases
that are awkward as well as well formed, giving
the same counts of phrases in the high-PMI value
bins as that of a short sentence that is entirely well
In addition to binning, we also encode as features the maximum, minimum and median PMI
value obtained over all word pairs. The first two
features capture the best and the worst word collocations in a response. The median PMI value
captures the overall general quality of the response
in a single number. For example, if this is a low
number, then the response generally has many bad
phrasal collocations. Finally a null-PMI feature is
used to count the number of pairs that had zero
entries in the database. This feature is an indicator that the given words or word collocations were
not found even once in the database. Given the
size of the underlying database, this usually happens in cases when words are misspelled, or when
the words never occur together.
All features created for bigrams are also created
for trigrams. We thus have a total of 40 features,
called the pmi feature set.
Feature set type
Table 1: Feature sets and the counts of features in
each set
System and Evaluation
Features from essay scoring
A number of measures of collocation quality have
been proposed and implemented (e.g. (Futagi et
al., 2008; Dahlmeier and Ng, 2011)). We use erater’s measure of the density of ‘good’ collocations found in the response. Another source of
difficulty for non-native writers is the selection of
appropriate prepositions. We use the mean probability assigned by e-rater to the prepositions in the
response. These two measures, one for the quality of collocations and the other for the quality of
prepositions, are combined in our colprep feature
Number of Features
Figure 1: System Architecture
As noted earlier, the system is partly rule-based
and partly statistical. Figure 1 illustrates the system architecture. The rule-based part captures
the straightforward deterministic scoring criteria
while the machine learning component encodes
features described in Section 4 and learns how to
weight the features for scoring based on humanscored responses.
As described in Section 2, detection of conditions that result in a score of zero are straightforward. Our rule-based scorer (shown as “Foreign Language Detector” in Figure 1) assigns a
zero score to a response if it is blank or nonEnglish. The system determines if the response is
non-English based on the average of PMI bigram
scores over the response. If the average score is
less than a threshold value, the system tags it as
Scoring Rubric-based Features
As seen in Section 2, some of the criteria for scoring are quite straightforward (e.g. “omits one or
both key words”). While these are not sole determinants of a score, they are certainly strong influences. Thus, we encode four criteria from the scoring guide. These form our final feature set, rubric,
and are binary values, answering the questions: Is
the first key word from the prompt present in the
response? Is the second key word from the prompt
present in the response? Are both key words from
The precision P rule of the rule-based system is
calculated as
a non-English sentence. The threshold was determined by manually inspecting the PMI values obtained for sentences belonging to English and nonEnglish news texts. Responses given zero scores
by this module are filtered out and do not go to the
next stage.
Responses that pass the rule-based scorer are
then sent to the statistical scorer. Here, we encode
the features discussed in Section 4. Spell checking
and correction are carried out before features for
content relevance and PMI-based awkward word
usage are computed. This is done in order to prevent misspellings from affecting the reference corpus match or database search. The original text
is sent to the Grammar feature generator as it creates features based on misspellings and other word
form errors. Finally, we use all the features to train
a Logistic Regression model using sklearn. Note
that the statistical system predicts all 4 scores (0
through 3). This is because the rule-based system
is not perfect; that is, it might miss some responses
that should receive zero scores, and pass them over
to the next stage.
P0rule =
Similarly, Recall is calculated as
R0rule =
= 87.6%
The corresponding F-measure is 85.2%
The remaining responses pass to the next stage
where machine learning is employed. We performed 10 fold cross-validation experiments using Logistic Regression as well as Random Forest
learners. As the results are comparable, we only
report those from logistic regression.
Accuracy in %
Agreement (QWK)
Table 2: Overall system and human accuracy
(in percentage) and agreement (using Quadratic
Weighted Kappa)
We report our results using overall accuracy,
quadratic weighted kappa (QWK) and score-level
precision, recall and f-measure. The precision P
of the system is calculated for each score point i
|Si ∩ Hi |
Pi =
|Si |
where |Si | is the number of responses given a
score of i by the system, and |Si ∩ Hi | is the number of responses given a score of i by the system
as well as the human rater.
Similarly, recall, R is calculated for each score
point i as
|Si ∩ Hi |
Ri =
|Hi |
F-measure Fi is calculated as the harmonic
mean of the precision Pi and recall Ri at each
score point i. Accuracy is the ratio of the number of responses correctly classified over the total
number of responses.
= 82.9%
Table 2 reports the results. The system achieves
an accuracy of 76.23%, which is more than a 15
percentage point improvement over the majority
class baseline of 61%. The majority class baseline always predicts a score of 3. Compared to human performance, system performance is 10 percentage points lower (human-human agreement
is 86%). Quadratic weighted kappa for systemhuman agreement is also lower (0.63) than for
human-human agreement (0.83).
Table 3 reports the precision, recall and Fmeasure of the system for each of the score points.
Score point
Table 3: Overall system performance at each score
point using all features
All of the responses in the train-test set were
passed through the rule-based zero-scorer. A total
of 210 responses had been scored as zero by the
human scorer. The rule-based system scored 222
responses as zeros, of which 184 were correct.
In order to understand the usefulness of each feature set in scoring the responses, we constructed
systems using first the individual features alone,
and then using feature combinations. Table 4 reports the accuracy of the learner using individual
features alone. We see that, individually, each feature set performs much below the performance of
the full system (that has an accuracy of 76.23%),
which is expected, as each feature set represents
a particular aspect of the construct. However, in
general, each of the feature-sets (except colprep)
shows improvement over baseline, indicating that
they contribute towards performance improvement
in the automated system.
Grammar features are the best of the individual feature sets at 70% accuracy, indicating that
grammatical error features developed for longer
texts can be applied to single sentences. The PMIbased feature set is the second best performer, indicating its effectiveness in capturing word usage
issues. While colprep and pmi both capture awkward usage, pmi alone shows better performance
(67.44%) than colprep alone (61.26%). Also,
when rubric is used alone, the resulting system
produces a four percentage point improvement
over the baseline, with 65% accuracy, indicating
the presence of responses where the test-takers are
not able to incorporate one or both words in a single sentence. The relevance feature set by itself
does not show substantial improvement over the
baseline. This is not surprising, as according to
the scoring guide, a response gets a score of 0 or 1
if it does not describe the picture, and gets a score
of 2 or 3 if it is relevant to the picture. Hence, this
feature cannot solely and accurately determine the
Feature Set
pmi (colprep+pmi, row 2), there is a slight drop
in performance as compared to using pmi-based
features alone. These results indicate that colprep,
while being useful for larger texts, does not transfer well to the simple single sentence responses in
our task.
Further, in Table 5 we see that the system using
a combination of the pmi feature set and the relevance feature set (pmi+relevance) achieves an accuracy of 69%. Thus, this feature combination is
able to improve performance over that using either
feature set alone, indicating that while content relevance features by themselves do not create an impact, they can improve performance when added
to other features. Finally, the feature combination
of all new features developed for this task (pmi +
relevance+ rubric) yields 73% accuracy, which is
again better than each individual feature set’s performance, indicating that they can be synergistically combined to improve system performance.
Feature Set
(i) grammar + colprep
(ii) colprep + pmi
(iii) pmi + relevance
(iv) pmi + relevance + rubric
Accuracy in %
Table 5: System performance for feature combinations (i) typically used in essay scoring, (ii) that
measure awkwardness, (iii) newly proposed here,
(iv) newly proposed plus rubric-specific criteria
Related Work
Most work in automated scoring and learner language analysis has focused on detecting grammar
and usage errors (Leacock et al., 2014; Dale et al.,
2012; Dale and Narroway, 2012; Gamon, 2010;
Chodorow et al., 2007; Lu, 2010). This is done
either by means of handcrafted rules or with statistical classifiers using a variety of information.
In the case of the latter, the emphasis has been on
representing the contexts of function words, such
as articles and prepositions. This work is relevant inasmuch as errors in using content words,
such as nouns and verbs, are often reflected in the
functional elements which accompany them, for
example, articles that indicate the definiteness or
countability of nouns, and prepositions that mark
the cases of the arguments of verbs.
Previous work (Bergsma et al., 2009; Bergsma
et al., 2010; Xu et al., 2011) has shown that mod-
Accuracy in %
Table 4: System performance for individual features
Table 5 reports accuracies of systems built using feature set combinations. The first feature set
combination, grammar + colprep, is a set of all
features obtained from essay scoring. Here we see
that addition of colprep does not improve the performance over that obtained by grammar features
alone. Further, when colprep is combined with
paraphrasing corpus.
In a vast body of related work, automated
methods have been explored for the generation
of descriptions of images (Kulkarni et al., 2013;
Kuznetsova et al., 2012; Li et al., 2011; Yao et
al., 2010; Feng and Lapata, 2010a; Feng and Lapata, 2010b; Leong et al., 2010; Mitchell et al.,
2012). There is also work in the opposite direction, of finding or generating pictures for a
given narration. Joshi et al. (2006) found the
best set of images from an image database to
match the keywords in a story. Coyne and Sproat
(2001) developed a natural language understanding system which converts English text into threedimensional scenes that represent the text. For a
high-stakes assessment, it would be highly undesirable to have any noise in the gold-standard reference picture descriptions. Hence we chose to use
manual description for creating our reference corpus.
els which rely on large web-scale n-gram counts
can be effective for the task of context-sensitive
spelling correction. Measures of ngram association such as PMI, log likelihood, chi-square, and
t have a long history of use for detecting collocations and measuring their quality (see (Manning
and Sch¨utze, 1999) and (Leacock et al., 2014)
for reviews). Our application of a large n-gram
database and PMI is to detect inappropriate word
Our task also differs from work focusing on
evaluating content (e.g. (Meurers et al., 2011;
Sukkarieh and Blackmore, 2009; Leacock and
Chodorow, 2003)) in that, although we are looking for usage of certain content words, we focus
primarily on measuring knowledge of vocabulary.
Recent work on assessment measures of depth
of vocabulary knowledge (Lawless et al., 2012;
Lawrence et al., 2012), has argued that knowledge of specific words can range from superficial
(idiomatic associations built up through word cooccurrence) to topical (meaning-related associations between words) to deep (definitional knowledge). Some of our features (e.g. awkward word
usage) capture some of this information (e.g., idiomatic associations between words), but assigning the depth of knowledge of the key words is not
the focus of our task.
Summary and Future Directions
We investigated different types of features for automatically scoring a vocabulary item type which
requires the test-taker to use two words in writing a sentence based on a picture. We generated a
corpus of picture descriptions for measuring the
relevance of responses, and as a foundation for
feature development, we performed preliminary
fine-grained annotations of responses. The features used in the resulting automated scoring system include newly developed statistical measures
of word usage and response relevance, as well as
features that are currently found in essay scoring
engines. System performance shows an overall
accuracy in scoring that is 15 percentage points
above the majority class baseline and 10 percentage points below human performance.
There are a number of avenues open for future
exploration. The automated scoring system might
be improved by extending the relevance feature
to include overlap with previously collected highscoring responses. The reference corpus could
also be expanded and diversified by using a large
number of annotators, at least some of whom are
speakers of the languages that are most prominently represented in the population of test-takers.
Finally, one particular avenue we would like to explore is the use of our features to provide feedback
in low stakes practice environments.
Work that is closely related to ours is that of
King and Dickinson (2013). They parse picture
descriptions from interactive learner sentences,
classify sentences into syntactic types and extract
the logical subject, verb and object in order to recover simple semantic representations of the descriptions. We do not explicitly model the semantic representations of the pictures, but rather our
goal in this work is to ascertain if a response is
relevant to the picture and to measure other factors that reflect vocabulary proficiency.
We employ human annotators and use word
similarity measures to obtain alternative forms of
description because the proprietary nature of our
data prevents us from releasing our pictures to
the public. However, crowd sourcing has been
used by other researchers to collect human labels
for images and videos. For example, Rashtchian
et al. (2010) use Amazon Mechanical Turk and
Von Ahn and Dabbish (2004) create games to entice players to correctly label images. Chen and
Dolan (2011) use crowd sourcing to collect multiple paraphrased descriptions of videos to create a
Keelan Evanini, Michael Heilman, Xinhao Wang, and
Daniel Blanchard. 2014. Automated scoring for
TOEFL Junior comprehensive writing and speaking.
Technical report, ETS, Princeton, NJ.
Yigal Attali and Jill Burstein. 2006. Automated essay
scoring with e-rater v. 2.0. Journal of Technology,
Learning, and Assessment, 4:3.
Yansong Feng and Mirella Lapata. 2010a. How many
words is a picture worth? Automatic caption generation for news images. In Proceedings of the 48th Annual Meeting of the Association for Computational
Linguistics, ACL ’10, pages 1239–1249, Stroudsburg, PA, USA. Association for Computational Linguistics.
Shane Bergsma, Dekang Lin, and Randy Goebel.
2009. Web-scale n-gram models for lexical disambiguation. In IJCAI.
Shane Bergsma, Emily Pitler, and Dekang Lin. 2010.
Creating robust supervised classifiers via web-scale
n-gram data. In Proceedings of the 48th Annual
Meeting of the Association for Computational Linguistics, pages 865–874. Association for Computational Linguistics.
Yansong Feng and Mirella Lapata. 2010b. Topic
models for image annotation and text illustration.
In Human Language Technologies: The 2010 Annual Conference of the North American Chapter of
the Association for Computational Linguistics, HLT
’10, pages 831–839, Stroudsburg, PA, USA. Association for Computational Linguistics.
Thorsten Brants and Alex Franz. 2006. Web 1T 5gram Version 1. In Linguistic Data Consortium,
David L Chen and William B Dolan. 2011. Collecting highly parallel data for paraphrase evaluation.
In Proceedings of the 49th Annual Meeting of the
Association for Computational Linguistics: Human
Language Technologies-Volume 1, pages 190–200.
Association for Computational Linguistics.
Michael Flor. 2013. A fast and flexible architecture for
very large word n-gram datasets. Natural Language
Engineering, 19(1):61–93.
KE Forbes-McKay and Annalena Venneri. 2005. Detecting subtle spontaneous language decline in early
Alzheimers disease with a picture description task.
Neurological sciences, 26(4):243–254.
Martin Chodorow and Claudia Leacock. 2000. An unsupervised method for detecting grammatical errors.
In Proceedings of the Conference of the North American Chapter of the Association of Computational
Linguistics (NAACL), pages 140–147.
Yoko Futagi, Paul Deane, Martin Chodorow, and Joel
Tetreault. 2008. A computational approach to detecting collocation errors in the writing of non-native
speakers of English. Computer Assisted Language
Learning, 21(4):353–367.
Martin Chodorow, Joel R Tetreault, and Na-Rae Han.
2007. Detection of grammatical errors involving
prepositions. In Proceedings of the fourth ACLSIGSEM workshop on prepositions, pages 25–30.
Association for Computational Linguistics.
Michael Gamon. 2010. Using mostly native data to
correct errors in learners’ writing: A meta-classifier
approach. In Human Language Technologies: The
2010 Annual Conference of the North American
Chapter of the Association for Computational Linguistics, pages 163–171. Association for Computational Linguistics.
Bob Coyne and Richard Sproat. 2001. Wordseye: an
automatic text-to-scene conversion system. In Proceedings of the 28th annual conference on Computer
graphics and interactive techniques, pages 487–496.
Khairun-nisa Hassanali, Yang Liu, and Thamar
Solorio. 2013. Using Latent Dirichlet Allocation
for child narrative analysis. ACL 2013, page 111.
Daniel Dahlmeier and Hwee Tou Ng. 2011. Correcting semantic collocation errors with L1 induced
paraphrases. In Proceedings of the Conference on
Empirical Methods in Natural Language Processing, EMNLP ’11, pages 107–117, Stroudsburg, PA,
USA. Association for Computational Linguistics.
Dhiraj Joshi, James Z. Wang, and Jia Li. 2006. The
story picturing engine—a system for automatic text
illustration. ACM Trans. Multimedia Comput. Commun. Appl., 2(1):68–89, February.
Robert Dale and George Narroway. 2012. A framework for evaluating text correction. In LREC, pages
Levi King and Markus Dickinson. 2013. Shallow semantic analysis of interactive learner sentences. In
Proceedings of the Eighth Workshop on Innovative
Use of NLP for Building Educational Applications,
pages 11–21, Atlanta, Georgia, June. Association
for Computational Linguistics.
Robert Dale, Ilya Anisimoff, and George Narroway.
2012. HOO 2012: A report on the preposition and
determiner error correction shared task. In Proceedings of the Seventh Workshop on Building Educational Applications Using NLP, pages 54–62. Association for Computational Linguistics.
Girish Kulkarni, Visruth Premraj, Vicente Ordonez,
Sagnik Dhar, Siming Li, Yejin Choi, Alexander C.
Berg, and Tamara L. Berg. 2013. Babytalk: Understanding and generating simple image descriptions.
IEEE Transactions on Pattern Analysis and Machine
Intelligence, 99(PrePrints):1.
Rod Ellis. 2000. Task-based research and language
pedagogy. Language teaching research, 4(3):193–
Detmar Meurers, Ramon Ziai, Niels Ott, and Stacey M
Bailey. 2011. Integrating parallel analysis modules to evaluate the meaning of answers to reading
comprehension questions. International Journal of
Continuing Engineering Education and Life Long
Learning, 21(4):355–369.
Polina Kuznetsova, Vicente Ordonez, Alexander C
Berg, Tamara L Berg, and Yejin Choi. 2012. Collective generation of natural image descriptions. In
Proceedings of the 50th Annual Meeting of the
Association for Computational Linguistics: Long
Papers-Volume 1, pages 359–368. Association for
Computational Linguistics.
Margaret Mitchell, Xufeng Han, Jesse Dodge, Alyssa
Mensch, Amit Goyal, Alex Berg, Kota Yamaguchi,
Tamara Berg, Karl Stratos, and Hal Daum´e III.
2012. Midge: Generating image descriptions from
computer vision detections. In Proceedings of the
13th Conference of the European Chapter of the Association for Computational Linguistics, pages 747–
756. Association for Computational Linguistics.
Ren´e Lawless, John Sabatini, and Paul Deane. 2012.
Approaches to assessing partial vocabulary knowledge and supporting word learning: Assessing vocabulary depth. In Annual Meeting of the American Educational Research Association, April 13-17,
2012, Vancouver, CA.
Joshua Lawrence, Elizabeth Pare-Blagoev, Ren´e Lawless, and Chen Deane, Paul andLi. 2012. General vocabulary, academic vocabulary, and vocabulary depth: Examiningpredictors of adolescent reading comprehension. In Annual Meeting of the American Educational Research Association.
Cyrus Rashtchian, Peter Young, Micah Hodosh, and
Julia Hockenmaier. 2010. Collecting image annotations using Amazon’s Mechanical Turk. In Proceedings of the NAACL HLT 2010 Workshop on Creating
Speech and Language Data with Amazon’s Mechanical Turk, pages 139–147. Association for Computational Linguistics.
Claudia Leacock and Martin Chodorow. 2003. C-rater:
Automated scoring of short-answer questions. Computers and the Humanities, 37(4):389–405.
Jana Zuheir Sukkarieh and John Blackmore. 2009.
C-rater: Automatic content scoring for short constructed responses. In FLAIRS Conference.
Claudia Leacock, Martin Chodorow, Michael Gamon,
and Joel Tetreault. 2014. Automated Grammatical
Error Detection for Language Learners. Synthesis
Lectures on Human Language Technologies. Morgan & Claypool.
Luis Von Ahn and Laura Dabbish. 2004. Labeling
images with a computer game. In Proceedings of the
SIGCHI conference on Human factors in computing
systems, pages 319–326. ACM.
Wei Xu, Joel Tetreault, Martin Chodorow, Ralph Grishman, and Le Zhao. 2011. Exploiting syntactic
and distributional information for spelling correction with web-scale n-gram models. In Proceedings
of the Conference on Empirical Methods in Natural Language Processing, pages 1291–1300. Association for Computational Linguistics.
Choonkyu Lee, Smaranda Muresan, and Karin
Stromswold. 2012. Computational analysis of referring expressions in narratives of picture books.
NAACL-HLT 2012, page 1.
Chee Wee Leong, Rada Mihalcea, and Samer Hassan.
2010. Text mining for automatic image tagging. In
Proceedings of the 23rd International Conference on
Computational Linguistics: Posters, pages 647–655.
Association for Computational Linguistics.
Benjamin Z Yao, Xiong Yang, Liang Lin, Mun Wai
Lee, and Song-Chun Zhu. 2010. I2t: Image parsing to text description. Proceedings of the IEEE,
Siming Li, Girish Kulkarni, Tamara L Berg, Alexander C Berg, and Yejin Choi. 2011. Composing
simple image descriptions using web-scale n-grams.
In Proceedings of the Fifteenth Conference on Computational Natural Language Learning, pages 220–
228. Association for Computational Linguistics.
Dekang Lin. 1998. Automatic retrieval and clustering
of similar words. In Proceedings of the 17th international conference on Computational linguisticsVolume 2, pages 768–774. Association for Computational Linguistics.
Xiaofei Lu. 2010. Automatic analysis of syntactic
complexity in second language writing. International Journal of Corpus Linguistics, 15(4).
Christopher D. Manning and Hinrich Sch¨utze. 1999.
Foundations of Statistical Natural Language Processing. MIT Press.