=Paper=
{{Paper
|id=Vol-2410/paper9.pdf
|storemode=property
|title=Multi-Candidate Ranking Algorithm Based Spell Correction
|pdfUrl=https://ceur-ws.org/Vol-2410/paper9.pdf
|volume=Vol-2410
|authors=Chao Wang,Rongkai Zhao
|dblpUrl=https://dblp.org/rec/conf/sigir/WangZ19
}}
==Multi-Candidate Ranking Algorithm Based Spell Correction==
https://ceur-ws.org/Vol-2410/paper9.pdf
Multi-Candidate Ranking Algorithm Based Spell Correction
Chao Wang Rongkai Zhao
The Home Depot The Home Depot
Atlanta, GA Atlanta, GA
chao_wang1@homedepot.com rongkai_zhao@homedepot.com
ABSTRACT error hypothesis, these hypothesis are based on the observation of
Spell correction is an important component in Natural Language common user mistakes. Unlike deep learning based approaches, we
Processing (NLP). In the context of a product search engine, an found the deterministic approach has much higher accuracy in a
effective spell correction system can improve the accuracy of the close domain. Common user mistakes including wrong spelling of
search results and reduce the occurrence of No Results Found (NRF). a word, fat finger typing error, failing to break a composite word,
Conversely, a sub-optimal spell correction has negative effects, e.g., unnecessarily creating a composite word, aggressive device native
failing to correct misspelled queries, modifying correct queries into word level spell corrector introduced error, phonetic error, foreign
wrong ones. In this paper, three novel components / algorithms language input, keyboard malfunction, etc. For each hypothesis,
currently used in The Home Depot (THD) spell correction service correction candidates will be generated and ranked by the cluster
is presented: (1) word embedding based dictionary construction; density in a high dimensional word embedding space. The final
(2) multi-path candidates generation; (3) high dimensional cluster result is the most probable set ordered by the rank.
analysis based ranking model. The dictionary provides data about Categorized by the correction context, most standalone spell
the inner relationships among the words for a given corpus. With corrections (e.g., Jazzy spell correction [4], Aspell spell correction
the dictionary, the candidate generation algorithm recommends [5] and Hunspell spell correction [6]) are word-level approaches,
a set of correction candidates for several misspelling hypothesis, which correct the misspelled words without considering the context
e.g., word editing error, word breaking error, word concatenation information. The other spell corrections (Lucene spell correction
error, fat finger typing error, and so on. Then the ranking model [7], Grammarly spell correction [8], Microsoft Bing spell correction
projects the candidates into a high dimensional space and sorts them [9] and Google spell correction [10]) adopt context information. In
based on cluster density analysis. In the experiment, the new THD other words, they are context based solution.
spell correction is compared with the old version (without these Categorized by the correction methods, most of the spell correc-
features), Lucene spell correction and Grammarly spell correction. tion algorithms [1; 4; 6; 7] use Edit Distance (e.g., insertion, dele-
The evaluation results indicated the THD spell correction has higher tion, substitution) [11] and Phonetics Matching [12] as objective
correction accuracy than the other widely used implementations. function to find closely related words. Some other spell correc-
tions [13; 14] construct noisy channel models which recover the
KEYWORDS intended correction c in word set C from a misspelled word w to
Spell correction, Spell checker, Spell corrector, Word embedding, maximize Pr (c)Pr (w |c) where c ∈ C; Pr (c) is a prior model of word
Dictionary Building, Multi-candidate generation, Ranking model, probabilities; Pr (w |c) is a model of the noisy channel for word
Similarity context transformations from c to w due to Edit Distance. Another spell cor-
rections [15; 16] adopt Deep Neural Networks [17] based language
ACM Reference format: models. Xie et al. [15] proposed to use char-level encoder-decoder
Chao Wang and Rongkai Zhao. 2019. Multi-Candidate Ranking Algorithm
recurrent neural network [18] to learn the character relationships
Based Spell Correction. In Proceedings of the SIGIR 2019 Workshop on
within the words and phrases, which can avoid the problem of
eCommerce (SIGIR 2019 eCom), 8 pages.
out-of-vocabulary words. Chollampatt and Ng [16] presented a
multi-layer convolutional encoder-decoder neural network [19] for
1 INTRODUCTION this char-level correction. Based on their analysis, the proposed
The spell correction [1; 2] has been widely used in search engines, network can cover more grammatical errors than recurrent neural
works as a "gate keeper" for query parsing [3]. Generally, spell network.
correction has two components: spell checker and spell corrector. Before continuing onto further detail, here are some examples
The spell checker is used to check the validity of the queries at of common spelling problems and corrections:
word level as well as phrase level. If a query is valid, no action is (1) Single word error correction, e.g. "garage dor opener" ->
needed from the spell corrector, otherwise, the query is passed onto "garage door opener";
spell corrector for revision. The spell corrector has a set of common (2) Multi-word errors correction, e.g. "garge dor opener" -> "garage
Copyright © 2019 by the paper’s authors. Copying permitted for private and academic
door opener";
purposes. (3) Word breaking issue, e.g. "kohlertoilet" -> "kohler toilet";
In: J. Degenhardt, S. Kallumadi, U. Porwal, A. Trotman (eds.): (4) Word breaking issue with spelling errors, e.g. "kholertiolet"
Proceedings of the SIGIR 2019 eCom workshop, July 2019, Paris, France, published at
http://ceur-ws.org -> "kohler toilet";
(5) Word concatenation issue, e.g. "tom cat mouse trap" -> "tomcat
mouse trap";
SIGIR 2019 eCom, July 2019, Paris, France Wang et al.
3 WORD EMBEDDING BASED DICTIONARY
CONSTRUCTION
3.1 Traditional Dictionary Structure
Most of the spell correction dictionaries [6; 7] are just composed of
unique English words. However, it lacks of the word relationship
information for the context-based spell correction algorithms [8–
10; 13–16]. In order to extract the context information, the following
methods can be adopted: Given a specific corpus, e.g., user query
log, calculate the occurrence of every unique valid bigram words to
construct a bigram co-occurrence model; feed the corpus into some
char-level network models [15; 16] to learn the inner relationships
among the characters. However, these methods still have some
drawbacks. The co-occurrence model is only limited to bigram
coverage, which is hard to be extended to n-gram (n > 2) words due
to the limited memory space. In addition, the construction rules of
n-gram (n > 2) words model are much restricted, which can easily
lead to an overfitting issue. Conversely, the char-level model is
much flexible to generate some unreasonable correction results.
3.2 Multi-Source Dictionary Construction
Using Word Embedding
For our spell correction, in addition to bigram co-occurrence model,
word embedding model based dictionary construction is proposed
Figure 1: Spell Correction Architecture to extract the context information from the corpuses. We use user
query log, product catalog and selected Wikipedia documents for
context extraction. Wikipedia documents that are not related to
our product catalog are pruned.
(6) Word concatenation issue with spelling errors, e.g. "replace In order to better capture and utilize the word relationships in
ment light bulb" -> "replacement light bulb"; the context sensitive environment for our spell correction purpose,
(7) Word correction containing digits and special characters, e.g. Word2Vec [20] is chosen as our word embedding model [21], it
"door ;ocks" -> "door locks"; converts words into vectors and project them into a n-dimensional
(8) Real word errors correction, e.g. "mug knife" -> "mud knife". vector space, which could reveal word relationships in terms of
Where the real word error is that the individual words ("mug", geometric orientations.
"knife") of the query are valid but the phrase ("mug knife") does not The three datasets (product catalog dataset, user query log dataset
make sense. and related Wikipedia dataset [22]) are integrated based on word
embedding model as the following diagram (see Figure 2). Initially,
2 ARCHITECTURE the dictionary model is built with product catalog dataset as the
In this paper, three main components / algorithms are presented: kernel part. The catalog dataset has stronger correlations than the
(1) word embedding based dictionary construction; (2) multi-path query log dataset and Wikipedia dataset. In addition, it is much
candidates generation; (3) high dimensional cluster analysis based more accurate (containing few misspellings) than the query log
ranking model. dataset. Therefore, it can be used to build the initial model and be
As shown in the architecture (see Figure 1), the input query is helpful for the fast and accurate training convergence. However, it
first passed into Language Identifier to check which language the also has some drawbacks, e.g., limited contents (small coverage) and
query belongs to, e.g., Spanish or English (English spell correction is finite semantic expressions (low flexibility), because they are mainly
mainly focused in this paper). After that, the query is put into Spell provided by the product vendors. The query log dataset can com-
Checker to verify the validity. If valid, directly return the original plement these missed contents. Therefore, the query log dataset is
query as final result; if not, pass the query into Spell Corrector. In incrementally feed into the initial model so that the model can cap-
Spell Corrector, the invalid query will be corrected by all of means ture more information. Along with the increase of embedding model
synchronously, e.g., word correction, word breaking, word con- coverage, additional noise (misspelled words) is also introduced.
catenation, fat finger typing error, and so on. This procedure may The two datasets, especially the user query log, contain sizable
generate multiple candidates based on different correction compo- misspelling errors. The occurrences of some error bigram words
nents. Then these candidates are ranked based on the closeness of are very high (very popular), which is hard to be removed based on
their word vectors clustering. Finally, the best candidate is returned threshold mechanisms. Wikipedia dataset based embedding model
as the correction result. In addition, the proposed dictionary works is adopted to trim the incorrect contents from the already trained
through the whole architecture. model due to its higher coverage and diluted percentage of those
Multi-Candidate Ranking Algorithm Based Spell Correction SIGIR 2019 eCom, July 2019, Paris, France
2. Word Breaking: It is to break a misspelled word into multiple
valid words. In addition, the combination of the words should make
sense. For example, "cordlessdrill" -> "cordless drill".
3. Word Concatenation: It is to concatenate multiple valid or
invalid words into one valid word. Maybe the original words are
all valid, but their combination does not make sense. The invalid
combination is called real word error. For example, "dish washer"
-> "dishwasher".
4. Fat Finger Typing Error: Similar to Word Corrector, it adopts
keyboard layout to retrieve all of the possible correction words
rather than using phonetics matching. For example, "gloor" ->
"floor" where "g" is near to "f" on the keyboard.
5. Digit & Special Character Error Corrector: Similar to alphabetic
characters word corrector, it is to identify if the digits or special
characters of the query are useless or not. Then go for different
correction methods based on the identification. For example, "drill1"
-> "drill".
6. Unit Word Corrector: It is used to correct misspelled unit words,
most of the unit words in the query are followed by a numeric token.
For example, "18 voult drill" -> "18 volt drill".
4.2 Structures of Different Spell Corrector
Components
Generally, these spell corrector components can be organized as
two kinds of structures: Cascade Structure and Parallel Structure.
4.2.1 Cascade Structure. As shown in Figure 3, all of the correc-
Figure 2: Integration Diagram of Different Datasets Based tor components are cascaded one by one. Every component has a
Embedding Models user determined threshold (passing occurrence). For any specific
component, if the occurrence of the correction result is higher than
the threshold, it will be returned as the final result; if not, go to
errors. In addition, the Wikipedia model can also be used to validate the next corrector component. The cascade structure has lower
whether a given bigram is a proper phrase or not in the bigram CPU runtime complexity. However, it may be stuck with some sub-
co-occurrence model. optimal correction result. For example, suppose the best correction
Based on the constructed model, all of feature vectors of the result of a query should be gotten from Word Concatenation (Com-
existing words can be generated by it. So the similarity score of the ponent NO. 3). But the occurrence of the correction result from
bigram words can be calculated via Equation (1). Word Corrector (Component NO. 1) has been higher than the given
threshold. Then the final result is not the best one. So it depends on
v®a · v®b the cascade order of the correction components and the threshold
S(w a , wb ) = cos θ = (1)
va vb of every component. However, it is hard to change or tune them to
where S(w a , wb ) represents the similarity score of the bigram get the best performance, which varies for different queries.
words w a and wb ; θ represents the angle between the words w a
and wb ; va represents the feature vector of word w a ; vb represents 4.2.2 Parallel Structure. As shown in Figure 4, all of the cor-
the feature vector of word wb . rection components can also be organized in a parallel structure.
Different from the cascade structure, all of the possible correction
results of the corrector components can be obtained. Then these
4 MULTI-CANDIDATE GENERATION
candidates can be put into a ranking system (Described in Section
4.1 Introduction of Different Spell Corrector 5) to get the best one. The advantage of the parallel structure is
Components the global optimal feature of the correction results. In addition, the
In the proposed spell correction, multiple correction candidates can defined thresholds are not required. However, it needs more CPU
be generated by means of different correction components listed as power than the cascade structure. In the proposed spell correction,
follows: the parallel structure is adopted to get the best correction result.
1. Word Corrector: It is mainly referred to single misspelled
word correction. Given a misspelled word, all of the pronunciation 5 RANKING MODEL
similar words are retrieved based on phonetics matching. Among In this paper, the ranking model of spell correction acts as a sorter
the retrieved words, the one with the smallest edit distance is chosen and selector of the generated correction candidates. In this section,
as the best correction result. For example, "garadge" -> "garage". two kinds of ranking models are presented: unigram word & bigram
SIGIR 2019 eCom, July 2019, Paris, France Wang et al.
5.1 Unigram Word & Bigram Words Occurrence
Based Ranking Model and The Problems
For a given corpus, the occurrence of every existing unigram word
& bigram words can be gotten through occurrence accumulation.
The occurrence can reflect the popularity of the unigram word
and the bigram words. For single word candidate, its ranking score
can be defined as the occurrence of the unigram word. Similarly,
for two-word candidate, its ranking score can be defined as the
occurrence of the bigram words. However, not all of the candidates
are only composed of single word or two words. As the number of
words increasing from 2 to more, it is impossible to record every n-
gram occurrence for computation. As for the candidates containing
more than three words, the following equation is used to calculate
the occurrence of the candidate.
n−1
Õ
Oq = O(w i , w i+1 ) (2)
i=0
where Oq represents the occurrence of the candidate; n rep-
resents the number of the words in the candidate; w i represents
the ith word; O(w i , w i+1 ) represents the occurrence of the bigram
words w i and w i+1 .
Based on the above method, all of the candidates have the occur-
rence value. Therefore, for multiple correction candidates of a given
query, these candidates can be ranked based on their occurrences.
Figure 3: Cascade Structure
However, the method also has three drawbacks:
1. The occurrence depends on the quality of the source data. If
the source data contains much noises, the occurrences of some
wrong bigram words are also very big.
2. Suppose there are more than one kind of data sources (e.g.,
user query log and product catalog), for a given candidate, every
data source may have an occurrence. However, it is hard to combine
the occurrences together for the same candidate because different
data sources have different noise level and different signal coverage.
3. In Equation (2), the occurrence of the candidate is defined as
the mean value of the bigram words occurrences. But the variance
of the bigram words occurrences is not considered. Suppose there
are two candidates, one of them has big mean and big variance; the
other one has small mean and small variance. It is hard to conclude
which one is better.
Therefore, a novel ranking model is proposed to solve the above
problems.
5.2 Word Embedding Based Ranking Model
As described in Section 3, a novel word embedding based dictionary
is generated. The dictionary is combined with the product catalog
dataset, the query log dataset and the related Wikipedia dataset
based on word embedding mechanism. They works as different roles
in the integration. The product catalog dataset acts as an initial
Figure 4: Parallel Structure model builder; the user query log dataset acts as an incremental data
feeder; the related Wikipedia dataset acts as a data validator. In other
words, the first two datasets are used to build a Core Embedding
Model(CEM) and the Wikipedia dataset is used to build an Auxiliary
Embedding Model(AEM), where all of the noises and incorrectly
words occurrence based ranking model, and word embedding based paired words in CEM are removed by cross validation with the
ranking model. information contained in the AEM. This integration mechanism
Multi-Candidate Ranking Algorithm Based Spell Correction SIGIR 2019 eCom, July 2019, Paris, France
Figure 5: The Generation of The Centroid Vector
Figure 6: The Centroid Vector Based On The Perpendicular
solves Problem 1 and Problem 2 in Section 5.1. In order to solve Distance
Problem 3, different ranking metrics are proposed.
In order to best describe the validity of a given phrase contain-
ing multiple words, a good ranking metric should be robust to the
Ín
i=1 v®i
number of words and represent the closeness of the word vectors.
Based on the Word2Vec models, the valid n-grams words tends to Ĉ® = . (4)
n
cluster closely in the embedding space. Therefore, for any given
multi-word candidate, the spreadness of the set of embedded vec- This centroid vector is pointing to the center of the polygon or
tors corresponding to those words could be a good indicator of how polytope formed by the given set of word vector projections on the
likely this candidate is valid. The less spread the word vectors, the surface of the unit n-sphere centered at the origin. Furthermore,
more likely the candidate is valid. Then, by calculating the average the distances from all the word vectors for the given phrase to it
distance of all the vectors to their centroid vector, a ranking score could describe the spreadness of the words. Intuitively, less spread
can be generated to define the validity of the candidate. Mathemat- means better. The solution vector C usually lies in the interior of
ically, it is a Least Squares optimization problem in n-dimensional the polygon or polytope as shown in Figure 5, which is not a unit
vector space. As shown in Figure 5, the centroid vector C needs to vector, thus not on the surface of the n-sphere. The solution is
be fitted by a set of word vectors A1 , A2 and A3 in a given phrase. unique.
Goal and Metric: Another choice could be the centroid obtained from the spherical
1. For a given phrase, the centroid vector should satisfy that its K-means method, where K = 1 in this case. The centroid vector
distances to all the word vectors could be calculated and aggregated C needs to minimize the sum of squared perpendicular distances
to describe the spreadness/closeness of word vectors. from all other vectors to itself, as shown in Figure 6. The objective
2. The metric need to capture the outliers such that if any words function is
are more distant to the center than the others, then all the words in
n
the phrase as a whole should not be considered coherent enough Õ
than the case where all the word vectors have similar distances to f (v, C) = ArдminC ∈R d [sin(v®i , C) ∗ ∥v®i ∥]2 . (5)
the center. Squared distance is a good choice here. i=1
Objective Function Choices: vi is an unit vector, it can be transformed as
Based on the above requirements, an objective function is pro-
posed, which is in terms of the Euclidean distances of the center to n
Õ
all the tips of the polytope formed by the word vectors (d dimen- f v, C) = ArдminC ∈R d [1 − cos2 (v®i , C)]
sional space): i=1
n
(6)
Õ
2
n
Õ 2
= ArдminC ∈R d [1− < v®i , C® > ].
f (v, C) = ArдminC ∈R d v®i − C® , (3) i=1
i=1 The solution to the equation is not unique and the iterative ap-
where the center C should have the property such that the sum proximation method is not guaranteed to converge. This method is
of the squared distances from all the words in a candidate should more closely related to the plane fitting method in high dimensional
be minimized. Since this is a quadratic equation, by taking the first spaces. One other method closely related to the spherical K-means
order derivative and setting it to 0, the solution can be obtained: has the following objective function:
SIGIR 2019 eCom, July 2019, Paris, France Wang et al.
(10) Dimension word related queries (13,498 terms), e.g., "4in.x
n
Õ 4in. wall tile" for correct one "4in. x 4in. wall tile", occupies about
f (v, C) = ArдminC ∈R d [1 − cos(v®i , C)], (7) 0.64% of total queries.
i=1 All of the queries are labelled by Microsoft Bing spell correction
which seeks to minimize the loss of the similarities and their API [23] as ground truth. Since Google does not provide with API
maximum value 1. This method has the same drawbacks as the service for spell correction, it cannot be used for the evaluation.
previous one does. Therefore, Equation (3) is chosen as the objective
function of the ranking metric in this paper. 6.2 Evaluation Method
After comparisons, the ranking metric is proposed as the mean
squared distance of all the tips of the polytope formed by word The testing dataset is divided into correct queries and different
vectors to the center C : error types. The correction accuracy (see Equation (9)) can be used
to describe the evaluation performance.
Ín ® 2
i=1 ∥v®i − Ĉ ∥ Ncor r ect ion
D̂ = , (8) Paccur acy =
n Ntot al
(9)
where D̂ is the spreadness score of the ranking metric; the center
vector C is obtained from the Euclidean distance based objective where Paccur acy represents the correction accuracy; Ncor r ect ion
functions as shown in Equation (3). It could correctly capture the represents the number of algorithm-modified (e.g., THD spell cor-
spreadness of the word vectors and has the range [0, n]. Smaller rection) results which are the same with ground truth; Ntot al rep-
resents the number of correct queries or the number of queries in
value D̂ indicates better result (more closeness of the word vectors
some specific error type.
in the phrase).
For the correct queries, the correction accuracy can also be called
true positive rate or recall. For different error types, true negative
6 EXPERIMENT
rate can be derived from the correction accuracy.
In this experiment, different kinds of evaluations are implemented
on a collected testing dataset. It involves the comparisons between
6.3 Result
THD spell correction with the presented algorithms (word em-
bedding based dictionary generation, multi-candidate generation,
word embedding based ranking model) and that without them. The 2 million Bing-labelled queries dataset has totally 10 differ-
In addition, the proposed THD spell correction is also compared ent correction types. As shown in Table 1, (1) correct queries are
with Lucene spell correction [7]. All of the testing are specified in represented by "Correct"; (2) brand name errors are represented by
different correction types. "Brand"; (3) non-word errors are represented by "NWE"; (4) real
word errors are represented by "RWE"; (5) word breaking errors
6.1 Evaluation Dataset are represented by "Break"; (6) word concatenation errors are rep-
The testing dataset contains correct queries and 9 error types of resented by "Concatenate"; (7) phonetics errors are represented by
queries, collecting from THD query log. It has 2,095,028 terms "Phonetic"; (8) product type errors are represented by "Product"; (9)
totally. unit word errors are represented by "Unit"; (10) dimension errors
(1) Correct queries (1,559,534 terms), occupies about 74.44% of are represented by "Dim".
total queries. In addition, the following correction algorithms are compared
(2) Brand name related errors (4,678 terms), e.g., "ryoby drill" for based on the dataset:
correct spelling "ryobi drill", occupies about 0.22% of total queries. (1) THD spell correction with bigram co-occurrence model (de-
(3) Non-word errors (10,083 terms), e.g., "garage door openr" for scribed in Section 3.1), and cascade structure of corrector compo-
correct spelling "garage door opener", occupies about 0.48% of total nents (described in Section 4.2.1), is the first version of THD spell
queries. correction and is represented as "THD SC v1".
(4) Real word errors (1,528 terms), e.g., "mug knife" for correct (2) THD spell correction with word embedding based context
one "mud knife", occupies about 0.07% of total queries. dictionary (described in Section 3), parallel structure based multi-
(5) Word breaking errors (10,594 terms), e.g., "firepit" for correct candidate generation (4), and word embedding based ranking model
one "fire pit", occupies about 0.51% of total queries. (described in Section 5), is the second version of THD spell correc-
(6) Word concatenation errors (89,762 terms), e.g., "night light tion and is represented as "THD SC v2".
replace ment bulbs" for correct one "night light replacement bulbs", (3) Lucene spell correction [7], is represented as "Lucene SC".
occupies about 4.28% of total queries. As shown in Table 1, THD spell correction v1 performs better
(7) Phonetics related errors (345,323 terms), e.g., "foto frame" for than Lucene spell correction on overall correction accuracy. For
correct one "photo frame", occupies about 16.48% of total queries. different types of queries, THD spell correction v1 also has higher
(8) Product types related errors (9,490 terms), e.g., "hamer drill" accuracy than Lucene spell correction except real word error type.
for correct one "hammer drill", occupies about 0.45% of total queries. The biggest difference between THD spell correction v1 and Lucene
(9) Unit word related errors (50,538 terms), e.g., "12 vot cordless spell correction is that the first method adopts bigram co-occurrence
drill" for correct one "12 volt cordless drill", occupies about 2.41% model to get the context information among the words of a given
of total queries. phrase. The experiment results prove that adopting the context
Multi-Candidate Ranking Algorithm Based Spell Correction SIGIR 2019 eCom, July 2019, Paris, France
Table 1: Spell Correction Comparison on 2 Million Bing-Labelled Queries Dataset
Correct Brand NWE RWE Break Concatenate Phonetic Product Unit Dim Overall
Count 1,559,534 4678 10,083 1,528 10,594 89,762 345,323 9,490 50,538 13,498 2,095,028
THD SC v1 91.57% 64.71% 42.94% 15.05% 22.71% 25.25% 40.96% 62.17% 33.19% 23.76% 77.71%
THD SC v2 95.36% 67.36% 43.18% 17.24% 23.07% 24.29% 40.03% 64.98% 32.25% 23.37% 80.32%
Lucene SC 81.24% 43.50% 26.10% 21.92% 4.18% 13.97% 20.72% 48.23% 10.76% 5.72% 65.26%
information is a huge benefit to increase correction accuracy. How- data can be feed into a deep neural network for the real word errors
ever, due to some noises issues existing in the bigram co-occurrence checking and correcting.
model, some wrong bigram words also have high occurrences, e.g.,
popular error "dish washer". It results in a low performance on the
correction of the real word errors. REFERENCES
The highest accuracy values of different correction types are [1] James L Peterson. Computer programs for detecting and correcting spelling
errors. Communications of the ACM, 23(12):676–687, 1980.
marked with Bold in the tables. As shown in Table 1, the proposed [2] Neha Gupta and Pratistha Mathur. Spell checking techniques in nlp: a survey.
THD spell correction v2 has the highest overall accuracy among International Journal of Advanced Research in Computer Science and Software
Engineering, 2(12), 2012.
the three algorithms. For the comparisons on most of the correction [3] James F Allen. Natural language processing. 2003.
types ("Correct", "Brand", "NWE", "Break" and "Product"), THD spell [4] Fei Liu, Fuliang Weng, Bingqing Wang, and Yang Liu. Insertion, deletion, or sub-
correction v2 is also the best one. It proves that word embedding stitution?: normalizing text messages without pre-categorization nor supervision.
In Proceedings of the 49th Annual Meeting of the Association for Computational
based dictionary, multi-candidate based ranking model and their Linguistics: Human Language Technologies: short papers-Volume 2, pages 71–76.
integration can improve spell correction. The correction accuracy Association for Computational Linguistics, 2011.
of real word errors is still lower than that of Lucene spell correction. [5] Daniel Naber et al. A rule-based style and grammar checker. Citeseer, 2003.
[6] Tommi Pirinen and Krister Lindén. Creating and weighting hunspell dictionar-
But it is much better than that of THD spell correction v1. As for iesas finite-state automata. Investigationes Linguisticae, 21:1–16, 2010.
the correction types "Concatenate", "Phonetic", "Unit" and "Dim", [7] Manu Konchady. Building Search Applications: Lucene, LingPipe, and Gate. Lulu.
com, 2008.
THD spell correction v1 still gets the highest accuracy but has very [8] Genaro V Japos. Effectiveness of coaching interventions using grammarly soft-
minimal differences compared with THD spell correction v2. ware and plagiarism detection software in reducing grammatical errors and
plagiarism of undergraduate researches. JPAIR Institutional Research, 1(1):97–109,
2013.
7 CONCLUSION [9] Youssef Bassil and Mohammad Alwani. Post-editing error correction algo-
In summary, the proposed algorithms of word embedding based rithm for speech recognition using bing spelling suggestion. arXiv preprint
arXiv:1203.5255, 2012.
multi-source (product catalog, query log and Wikipedia) dictionary [10] Youssef Bassil and Mohammad Alwani. Ocr post-processing error correction
construction and multi-candidate ranking model can improve spell algorithm using google online spelling suggestion. arXiv preprint arXiv:1204.0191,
correction much more than the other context based algorithms, e.g., 2012.
[11] Eric Sven Ristad and Peter N Yianilos. Learning string-edit distance. IEEE
just using bigram co-occurrence model, especially which performs Transactions on Pattern Analysis and Machine Intelligence, 20(5):522–532, 1998.
much better than that only considering edit distance and phonetics [12] Justin Zobel and Philip Dart. Phonetic string matching: Lessons from information
retrieval. In Proceedings of the 19th annual international ACM SIGIR conference on
matching. Adopting word embedding, multiple data sources can be Research and development in information retrieval, pages 166–172. ACM, 1996.
integrated together without considering the weight mechanism to [13] Eric Brill and Robert C Moore. An improved error model for noisy channel
balance them. In addition, the merits of these data sources are also spelling correction. In Proceedings of the 38th Annual Meeting on Association
for Computational Linguistics, pages 286–293. Association for Computational
effectively exerted, e.g., using Wikipedia data to trim the invalid Linguistics, 2000.
terms due to the low noises performance. With parallel structure [14] Mark D Kernighan, Kenneth W Church, and William A Gale. A spelling cor-
based multi-candidate generation, all of the possible candidates rection program based on a noisy channel model. In Proceedings of the 13th
conference on Computational linguistics-Volume 2, pages 205–210. Association for
can be gotten synchronously. Then the proposed word embedding Computational Linguistics, 1990.
based ranking model can be used to select the best one. [15] Ziang Xie, Anand Avati, Naveen Arivazhagan, Dan Jurafsky, and Andrew Y
Ng. Neural language correction with character-based attention. arXiv preprint
arXiv:1603.09727, 2016.
8 FUTURE WORK [16] Shamil Chollampatt and Hwee Tou Ng. A multilayer convolutional encoder-
decoder neural network for grammatical error correction. In Thirty-Second AAAI
Through all of the comparisons, real word error is regarded as the Conference on Artificial Intelligence, 2018.
weakest point for the proposed THD spell correction algorithms. [17] Ronan Collobert and Jason Weston. A unified architecture for natural language
Most of the real word errors are blocked by spell checker compo- processing: Deep neural networks with multitask learning. In Proceedings of the
25th international conference on Machine learning, pages 160–167. ACM, 2008.
nent of spell correction, especially for some popular errors, e.g., [18] Tomáš Mikolov, Martin Karafiát, Lukáš Burget, Jan Černockỳ, and Sanjeev Khu-
"coffee mud" for correct one "coffee mug". Even searching with danpur. Recurrent neural network based language model. In Eleventh annual
Google and Microsoft Bing, they cannot do anything with them conference of the international speech communication association, 2010.
[19] Nal Kalchbrenner, Edward Grefenstette, and Phil Blunsom. A convolutional
as well. Observed from some search engines behaviors, when en- neural network for modelling sentences. arXiv preprint arXiv:1404.2188, 2014.
countering these search queries, they would present some results [20] Yoav Goldberg and Omer Levy. word2vec explained: deriving mikolov et al.’s
only searching "coffee" or only searching "mud". Among them, the negative-sampling word-embedding method. arXiv preprint arXiv:1402.3722,
2014.
users should click the really correct ones they want. The clicked [21] Omer Levy and Yoav Goldberg. Dependency-based word embeddings. In Proceed-
content can be used as a user feedback to help identify the error ings of the 52nd Annual Meeting of the Association for Computational Linguistics
(Volume 2: Short Papers), volume 2, pages 302–308, 2014.
"coffee mud" and correct it into "coffee mug". The user behavior
SIGIR 2019 eCom, July 2019, Paris, France Wang et al.
[22] David Milne and Ian H Witten. Learning to link with wikipedia. In Proceedings
of the 17th ACM conference on Information and knowledge management, pages
509–518. ACM, 2008.
[23] Srikanth Machiraju and Ritesh Modi. Azure cognitive services. In Developing
Bots with Microsoft Bots Framework, pages 233–260. Springer, 2018.
ACKNOWLEDGMENTS
We thank our colleagues, Yuemeng Li, Mingzi Cao, Felipe Castrillon,
Nagaraj Palanichamy for assistance with the software development
and experiments, and Surya Kallumadi for comments that greatly
improved this paper.