1 Introduction

Word alignment is a process that is used to determine the equivalent words in a bilingual sentence pair. The bilingual sentence pair contains two different languages. One is the source language which is the language the translation starts from and another one that is known as the target language is the language the translation ends in. In the literature, two different types of word alignment applications have been considered. One main goal is to produce lexical data for bilingual dictionaries, while another goal can be providing data for MT (Wang 2004). A number of word alignment applications are multi lingual lexicography, word sense disambiguation, translation connections in computational lexicography (Vulić and Moens 2010), patent retrieval that is a branch of information retrieval (Jochim et al. 2011), and cross-lingual information retrieval (Nie 2010). Word alignment is also a necessary step for almost all state-of-the-art translation systems including syntax-based machine translation (SBMT), statistical MT (SMT), hierarchical phrase-based systems (Brunning 2010), example-based MT (Vulić and Moens 2010) and many other multi lingual applications. Therefore, the task of word alignment is interesting in itself for plenty of applications (Koehn 2010).

From one point of view, there are two general approaches for alignment models including statistical and heuristic ones (Och and Ney 2003). Och and Ney (2003) have shown that the statistical model is better than the heuristic one, but Vulić and Moens (2010) have explained some advantages and disadvantages of these approaches. In the statistical model, no linguistic knowledge is required and only a large corpus of the bilingual sentence pair known as sentence-aligned parallel corpus is needed. The sentence-aligned parallel corpus is including bilingual sentence pairs of the source and the target languages so that each sentence pair of the two different languages are aligned sentence by sentence. The goal is to find a correspondence between the words in each sentence pair, i.e., constructing a word-aligned parallel corpus. To do this, firstly, an alignment function from target words positions into source words positions is defined (Koehn 2010). Thus, one-to-one or one-to-many alignment between the words are obtained such that it is possible to link each source word with none, one or several target words. In fact, many-to-many and many-to-one alignments are impossible and impose a limitation on this model. Och and Ney (2003) have proposed the symmetrization of word alignment so that the first training is performed in both translation directions (source-to-target and target-to-source), and then these are combined as various methods (union, intersection, or grow-diag) in order to intensify the quality of the word alignment and overcome this restriction. Therefore, this symmetric word alignment includes many-to-one and many-to-many correspondences.

In this paper a new R package named word.alignment (Daneshgar and Sarmad 2017) is described. In the package, a statistical word alignment model has been considered. It includes several functions to do one-to-one, one-to-many and symmetric alignments for each sentence pair, building an automatic dictionary based on two input languages (the source and the target languages) and measuring the quality of the word alignment using Precision, Recall, Alignment Error Rate (AER), and F-measure metrics. More details can be found in Fraser and Marcu (2007), Koehn (2010), Och and Ney (2003).

Two main functions (align.ibm1 and align.symmet) in the package have been written in R using S3 classes and methods. These functions store the results in a specific class: align.

While there are a few packages like quanteda (Benoit et al. 2018), tm (Feinerer and Hornik 2015), and NLP (Hornik 2015) in the R environment (R Core Team 2015) for NLP tasks, they do not separately compute word alignment. Of course, some functions in the packages like NLP and tm do not work on Arabic Script (e.g., Persian) properly.

Meanwhile, there are some other open-source tools to compute word alignment. In Table 1, some advantages and disadvantages of some tools have been shown:

It should be noted that none of the tools of Table 1 except word.alignment has been written in R which seemed to be necessary in this environment. For example, pialign, UNL Aligner, Anymalign and Nile have been written in Python.

As pointed out in Table 1, NATools and the produced package compute some operations, simultaneously.

NATools builds a sentene-aligned parallel corpus, computes word alignment for the bilingual sentence-aligned parallel corpus and it creates a probabilistic dictionary.

The word.alignment package implements one-to-many and symmetric alignments. It also constructs an automatic bilingual dictionary. The package measures the quality of the word alignments as well by a new function evaluation introduced in Section 5. It is noteworthy that evaluating the quality of the word alignment is a very important issue that none of the previous mentioned tools carry out this case.

In Sect. 6.4, using the function evaluation, a brief comparison has been made between some tools with the package. The package’s results are general better than some other tools.

Table 1 Comparison of some word alignment tools
Full size table

This paper uses by word.alignment version 1.1 which is available on the Comprehensive R Archive Network. Note that Some of previous versions are available in R-Forg in https://r-forge.r-project.org/. In addition, there are two dependent packages data.table (Dowle 2017), and openxlsx (Walker 2017) which should be installed along with the package. The new version does not require the quanteda package which it is one of the dependent packages for some previous versions. which is the most important difference with the previous ones.

The rest of the article is organized as follows. In Sect. 2, a general concept of word alignment is defined. Statistical alignment model is briefly outlined in Sect. 3.1. In Sect. 3.2, symmetric word alignment is explained. The concept evaluation of the quality of the word alignment is defined in Sect. 4. The detailed usages are described in Sect. 5. Some examples of the functions are provided in Sect. 6. Finally, conclusion and future works are presented in Sect. 7.

2 The concept of word alignment

The concept of word alignment between two sentence pairs can be quite complicated. Typically, a word alignment is associated with reorderings, droppings, insertions, and one-to-many alignments. Formally, the following definition is considered in this paper.

Assume that we have a sentence pair (f,e) that we want to align. \({{\varvec{f}}} = f_{1}, \ldots , f_{i}, \ldots , f_{I}\) is a source language string and \({{\varvec{e}}} = e_{1}, \ldots , e_{j}, \ldots , e_{J}\) is a target language string. A word alignment A can be defined as a subset of the Cartesian product of the word positions acquired by the following relation:

$$\begin{aligned} A\subseteq \{ (j, i): j = 1, \ldots , J; i = 1, \ldots , I \} \end{aligned}$$
(1)

This model is quite general as an arbitrary connection between source and target languages positions and of course it is computationally hard as there are \(2^{IJ}\) alignments. Hence, often additional constraints have been supposed on word alignment models (Och and Ney 2003).

A usual restriction is letting each target word be linked to exactly one source word. This rule does not hold in the other direction, i.e., a source word can be assigned to multiple target words or none of them. Therefore, the alignment mapping in such models is a map from a target word at position j to a source word at position i with a function \(a: j \longrightarrow i\) . In other words, we consider alignment variable \({{\varvec{a}}} = a_{1}, \ldots , a_{j}, \ldots , a_{J}\) as \( j\longrightarrow i = a _{j}\). This alignment may contain some alignments \(a_{j} = 0\). It is used to align the target word \(e_{j}\) to the “empty” word \(f_{0}\) which is called the “null token” (Koehn 2010).

In the next section, statistical word alignment will be presented.

3 Statistical word alignment

MT is an automated translation of a source language transcript to a target language one with or without human assistance (Chéragui 2012). SMT is a statistical approach to MT that is specified by using statistical models and does not need any prior knowledge of linguistics. It was proposed by Brown et al. (1990) for the first time. In fact, the concept of statistical word alignment is associated with SMT. Statistical word alignment using IBM Model 1 and symmetric word alignment will be demonstrated in the following section.

3.1 Statistical word alignment using IBM model 1

The concept of statistical word alignment is associated with SMT as mentioned before. The objective of SMT is to find the most probable target sentence \({{\varvec{e}}} = e_{1}, \ldots , e_{j}, \ldots , e_{J}\) of a given source sentence \({{\varvec{f}}} = f_{1}, \ldots , f_{i}, \ldots , f_{I}\) as the following.

$$\begin{aligned} \hat{{\varvec{e}}} = \underset{{\varvec{e}}}{\mathrm {argmax}} P ({{\varvec{e}}} | {{\varvec{f}}}) \end{aligned}$$
(2)

The \( P ({{\varvec{e}}} | {{\varvec{f}}}) \) is determined by training a statistical model using the sentence-aligned parallel corpus (Brunning 2010). The relationship between this probability and the alignment model is given by

$$\begin{aligned} P ({{\varvec{e}}} | {{\varvec{f}}}) = \sum _{{{\varvec{a}}}} P ({{\varvec{e}}}, {{\varvec{a}}} | {{\varvec{f}}}) \end{aligned}$$
(3)

where \({{\varvec{a}}} = a_{1}, \ldots , a_{j}, \ldots , a_{J}\) is the hidden alignment variable which specifies the position of a source word aligned with each target word.

One the other hand, \(P ({{\varvec{e}}}, {{\varvec{a}}}| {{\varvec{f}}})\) is decomposed to three components \(P(J|{{\varvec{f}}})\), \(P({{\varvec{a}}}|J, {{\varvec{f}}})\), and \( P({{\varvec{e}}}|{{\varvec{a}}},J,{{\varvec{f}}})\). This is the story of generative mode, the simplest and most primary one with the following assumptions which is the so-called IBM Model 1.

  1. 1.

    \(P(J|{{\varvec{f}}})\) is as a constant like \(\epsilon \).

  2. 2.

    \(P({{\varvec{a}}}|J, {{\varvec{f}}}) = \frac{1}{(I+1)^{J}}\).

  3. 3.

    \(P({{\varvec{e}}}|{{\varvec{a}}}, J, {{\varvec{f}}}) = \prod _{j=1}^{J} t(e_{j}|f_{a_{j}})\) where \(t(e_{j}|f_{a_{j}})\) is the probability of translating a source word \(f_{i}\) into a target word \(e_{j}\) with a map \( j\longrightarrow i = a _{j}\).

The IBM Model 1 parameters are \(t(e_{j}|f_{i})\) which can be estimated using the EM algorithm with the following steps:

  1. 1.

    Initialize \(t(e_{j}|f_{i})\), usually with uniform distributions.

  2. 2.

    E-step: Compute the probability of an alignment given a sentence pair in the training data, i.e., \(P({{\varvec{a}}}|{{\varvec{e}}}, {{\varvec{f}}})\) using the current parameters.

  3. 3.

    M-step: “Compute a count function \(c(e_{j}|f_{i};\mathbf{e} ,\mathbf{f} )\) that collects evidence from a sentence pair \(({{\varvec{f}}},{{\varvec{e}}})\) that a particular input word \(f_{i}\) translates into the output word \(e_{j}\)” (Koehn 2010). Then, new parameter values are estimated using the following equation given these counts:

    $$\begin{aligned} t(e_{j}|f_{i}; ({{\varvec{f}}}, {{\varvec{e}}})) = \frac{\sum _{{{\varvec{e}}},{{\varvec{f}}}}{c(e_{j}|f_{i};({{\varvec{f}}},{{\varvec{e}}}))}}{\sum _{e}{\sum _{{{\varvec{e}}},{{\varvec{f}}}}{c(e_{j}|f_{i};({{\varvec{f}}},{{\varvec{e}}}))}}} = \frac{count(e|f)}{total(f)} \end{aligned}$$
    (4)
  4. 4.

    The E and M steps are iterated until the algorithm is converges.

More details and the pseudo code for “the EM algorithm for IBM Model 1” can be found in Koehn (2010).

Now, the best alignment can be obtained by using these estimated parameters . In the IBM models (Brown et al. 1993), the best alignment \(\hat{{{\varvec{a}}}}\) for a given sentence pair (f,e) is acquired by

$$\begin{aligned} \hat{{{\varvec{a}}}} = \underset{{{\varvec{a}}}}{\mathrm {argmax}}P ({{\varvec{a}}}, {{\varvec{e}}} | {{\varvec{f}}}) \end{aligned}$$
(5)

The alignment \(\hat{{{\varvec{a}}}}\) is also known as the Viterbi alignment of the sentence pair (Och and Ney 2003). For Model 1, it can be simply summarized as shown in the following equation.

$$\begin{aligned} \hat{a}_{j} = \underset{i}{\mathrm {argmax}}t(e_{j} | f_{i}); \quad j=1, \ldots , J \end{aligned}$$
(6)

The pseudo code for the best alignment is provided by Jochim et al. (2011).

Note that this alignment model only includes one-to-one and one-to-many alignments while many-to-one or many-to-many alignments between languages are usual and this fact imposes a major restriction on IBM models.

3.2 Symmetric word alignment

In the preceding section, it was mentioned that the statistical alignment model has some limitations. One way to lift these limitations is the symmetrization of word alignments. First, we run IBM model training in both directions and then we combine the results in various ways. Although several methods of symmetrization are proposed (Koehn 2010; Och and Ney 2004; Wu and Wang 2007), only three of them have been considered in this paper.

Let \(A_{1}\) and \(A_{2}\) be alignments according to source-to-target and target-to-source translation directions, respectively. The matrices of \(A_{1}\) and \(A_{2}\) named alignment matrix can be created, separately. The rows are related to the initial target words and the columns are referred to the initial source words. Note that in both matrices, rows and columns are alike. The alignment between words are represented by points in the matrix.

The above-mentioned symmetrization methods are defined in the following:

  1. 1.

    Intersection: \(A = A_{1} \cap A_{2}\),

  2. 2.

    Union: \(A = A_{1} \cup A_{2}\),

  3. 3.

    Grow-diag: It starts with the alignments in the intersection set.

Then it grows by adding adjacent alignment points in the union sets but not at the intersection. These neighbors include left, right, top, or bottom and also diagonally adjacent points in the alignment matrix.

4 Evaluation of the word alignment quality

Several methods have been presented in the literature to compute word alignments (Jochim et al. 2011). Therefore, the measurement of alignment quality is an important issue. For this purpose, firstly a small sentence-aligned parallel corpus is considered as a test set. Note that a few number of these test sets for the variety of paired European languages is available at the Europarl website in http://statmt.org/europarl/. Then, each sentence pair should be aligned word by word by human annotators. It means that, in each sentence pair the source language words is linked to the corresponding target language words by some experts. The final result of these aligned words is based on integration of at least the two experts’ comments. This result is called “gold standard alignment” or “reference alignment” (Vulić and Moens 2010). It should be noted that there are different alignment guidelines to create a good gold standard (Vulić and Moens 2010).

To evaluate the quality of the word alignment, the alignments computed by an algorithm will be compared to the gold standard. The comparison would be based on one or more desired metrics. In this work, the words of the test set has been aligned based on the model made by a training set. After that, the result is compared to the gold standard with respect to the following metrics.

Precision and Recall are two common metrics in clustering, binary classification and various branches of NLP like information retreival (IR). Precision is the ratio of true relevant samples among all retrieved ones. While, Recall is the ratio of true relevant samples among all the relevant ones (There is a similarity concept between these two metrics and type I and II in hypothesis testing). In Table 2, these concepts have been displayed.

Table 2 Precision and recall concepts
Full size table

Now, their harmonic average (\(\frac{2PR}{P+R} \)) can be considered as a combined measure named F-measure (Sasaki 2007). Therefore, \(1-\frac{2PR}{P+R}\) can be as a measure of error called Alignment Error Rate (AER).

Note that usually a weighted F-measure as the following is considered:

$$\begin{aligned} F{-}measure = \frac{1}{\alpha \frac{1}{P} + (1- \alpha ) \frac{1}{R}}\quad (0 \le \alpha \le 1) \end{aligned}$$
(7)

For \(\alpha = \frac{1}{2}\), AER = \(1- (F{-}measure)\).

Now, these measures can be utilized for the evaluation of the quality of the word alignment as follows.

$$\begin{aligned} Precision(A, A_{r}) = \frac{|A \cap A_{r}|}{|A|} \end{aligned}$$
(8)
$$\begin{aligned} Recall(A, A_{r}) = \frac{|A\cap A_{r}|}{|A_{r}|} \end{aligned}$$
(9)

Here, A is the result of the word alignment using the assumed algorithm and \(A_{r}\) is the alignment in the gold standard which is determined by the experts. Note that A, \(A_{r}\) and \(A\cap A_{r}\) have been taken as the retrived samples, relevant samples and true positive ones, respectively.

During constructing of the gold standard (making links between word pairs of the sentences in the test set), there is possibility that the experts are not sure about their decision. Therefore, Och and Ney (2000) proposed using “Sure alignment points” and “Possible alignment points” (referred to as S and P) to align the word pairs. S and P are related to unambiguous and ambiguity links, respectively. In fact, P is corresponding to the aligned words in the idiomatic expressions, free translation and etc.

Now, Precision, Recall, AER and F-measure can be used to evaluate the quality of the word alignment as follows.

$$\begin{aligned} Precision(A, P)= & {} \frac{|A \cap P|}{|A|}\end{aligned}$$
(10)
$$\begin{aligned} \nonumber \\ Recall(A, S)= & {} \frac{|A\cap S|}{|S|}\end{aligned}$$
(11)
$$\begin{aligned} \nonumber \\ AER(A, P, S)= & {} 1- \frac{|A \cap P|+|A \cap S|}{|S|+|A|}\end{aligned}$$
(12)
$$\begin{aligned} \nonumber \\ F{-}measure(A,P,S,\alpha )= & {} \frac{1}{\frac{\alpha }{Precision(A,P)} + \frac{(1 - \alpha )}{Recall(A, S)}} \end{aligned}$$
(13)

In Eq. 13, the trade-off between Precision and Recall is assigned by \(\alpha \) \((0<\alpha <1)\). Koehn (2010) had shown that the best results can be obtained for \(\alpha \) between 0.2 and 0.4.

If we have two experts to create the gold, then we have S1, S2, P1 and P2. The final S is the intersection of S1 and S2, while final P is the union of P1 and P2.

Fraser and Marcu (2007) introduced a more precise measure as follows.

$$\begin{aligned} F{-}measure(A,S,\alpha )= & {} \frac{1}{\frac{\alpha }{Precision(A,S)} + \frac{(1 - \alpha )}{Recall(A, S)}} \end{aligned}$$
(14)

Note that, it is defined only by S.

All the five above measures can be obtained by the package for a computed word alignment.

5 The main functions in the word.alignment packages

The current section offers a brief description of word.alignment to make it more clear for the end-user. The package aims to provide a general framework for computing the statistical word alignment, symmetric alignments, building an automatic bilingual dictionary and evaluating the quality of the word alignment.

Table 3 presents a brief summary of the seven major functions in the package.

Table 3 Summary of functions in the word.alignment package
Full size table

6 An example

In the word.alignment package, there is a memory limitation for the number of sentence pairs. It depends to the number of words in sentence pairs as well as the computer’s specifications (RAM, CPU, cache, etc.). In the following, there is a comparison between the ability to do word alignment for three different computers in terms of the minimum of the number of words in the used sentence pairs (Table 4).

Table 4 Comparison between the ability of aligning for three different computers
Full size table

Several corpus like Bulgarian–English, Swedish–English, French–English and Persian–English [the three first have been taken form the Europal copus and the last has been taken of Mizan (Supreme Council of Information and Communication Technology 2013)] have been used to compare the three computers. Then, the minimum of the number of words in the used sentence pairs which the computer can be align have been reported. Notice that the difference between the various corpus is the difference in the number of prepositions in them.

6.1 align.ibm1 example

As described in Sect. 1, a large corpus of aligned sentences is required in two separated files as the input. Here, the Swedish-English corpus has been considered. Note that it is selected because its gold standard is available in Holmqvist and Ahrenberg (2011) while one can easily set each desired language pairs as a sentence-aligned parallel corpus.

In order to compute the word alignment using the function align.ibm1 for Swedish-English, it is sufficient to write the code lines in the following. The arguments file.sorc and file.trgt have been acheived from Swedish–English Europarl corpus.

figure a

Note that as a default, min.len is 5 and max.len is 40. Because, typically according to Déchelotte et al. (2007), Okita (2009) sentences with length longer than X1 and shorter than X2 are removed for word alignment, especially in its application for SMT. X1 and X2 are optionally minimum and maximum limitation to the length of a sentence, respectively. This mechanism is known as Sentence Cleaning Algorithm (Okita 2009). Meanwhile, here, this function is called for the first time. Therefore, dtfile.path is assigned to NULL. Then, the constructed data.table including all of the combinations of words in each sentence pair and their corresponding probability is saved with the name ‘sv.en.10000.5.RData’. In future calls to the function, the saved file name can be set in the dtfile.path argument. In fact, the function align.ibm1 picks up the results in ‘sv.en.10000.5.RData’ and will not construct it again, so this would save us some time.

figure b

The results are as follows:

figure c
figure d
figure e

The results are stored as an object of align class except the two first lines. They are defined during the program.

Now, the above outputs have been explained:

  • The stored associated data.table path.

  • The main outputs filename (stored). In case of asking to access all the results, they will easily be found with the given address.

  • Time difference is the program runtime.

  • A model that utilizes the word align. Here IBM Model 1 is implemented.

  • The number of input sentence pairs which is 10,000 here.

  • The remained sentence pairs after removing the sentences with shorter than min.len and loger than max.len. In the example, 8000(?) sentence pairs remain.

  • The number of EM algorithm iterations in order to estimate the IBM Model 1 parameters.

  • The word alignment results for the first three and the last three sentence pairs. For instance, i in English sentence is aligned to jag in Swedish sentence in the first sentence pair.

The align class has been utilized for the sake of better and more suitable output illustration of the functions align.ibm1() and align.symmet(). When the class is not defined, in addition to the above components and a series of other outputs, all alignments performed for all sentence pairs are displayed which is not proper.

6.2 align.symmet example

Here, the symmetric word alignment for Swedish-English has been run by calling align.symmet(), e.g., ‘intersection’ method. This method and other methods have been explained in Sect. 3.2.

figure f

The results are as follows:

figure g
figure h

The above outputs have been demonstrated as follows:

  • The stored associated data.table path for one-sided alignment (source-to-target).

  • The main outputs filename (stored) for one-sided alignment (source-to-target).

  • The stored associated data.table path for one-sided alignment (target-to-source).

  • The main outputs filename (stored) for one-sided alignment (target-to-source).

  • The main outputs filename (stored) for symmetric alignment (here; intersection).

  • The other results are stored as an object of align class like the function align.ibm1().

6.3 bidictionary example

The following results have been obtained by running the function on the 10,000 first sentence pairs of the Swedish–English Europarl corpus with prob = 0.8 and iter = 15 as an example:

figure i

The results are as follows (the brief results have been displayed):

figure j
figure k

The outputs of the function bidictionary() are as a list of the following components:

  • As the previous functions, the corresponding data.table path.

  • The number of input sentence pairs.

  • The minimum probability of word translation which is used to build the dictionary. The default is 0.8.

  • The number of EM algorithm iterations to perform IBM Model 1. The default is 15.

  • The source language name.

  • The target language name.

  • A dictionary of the source and the target languages. It is found based on the above minimum probability criterion.

6.4 evaluation example

In this subsection, an applied example is explained to show how the function evaluation has been used. As explained in the previous sections, the training set is the Swedish–English Europarl corpus and the source and the target languages are Swedish and English, respectively.

In order to evaluate the quality of the any word alignment, the gold standard which is explained in Sect. 4 is required. It will be constructed with the following steps:

  1. 1.

    Using the function cross.table, an excel file with n (number of sentences in the test set explained in Sect. 4) sheets of the two languages is built. Each sheet includes a cross table of the two language words related each sentence pair.

  2. 2.

    Then, the file is given to two annotators to be filled by Sure/Possible (1/2). The final gold will be the intersection of the Sure alignments and the union of the Possible alignments provided by the two annotators.

  3. 3.

    Then using the function excel2rdata, it is converted to rdata format.

For the current example, as previously mentioned, the gold standard for Swedish–English was available in Holmqvist and Ahrenberg (2011) named ep-ensv-alignref.v2015-10-12/test/test.ensv.naacl. Therefore, instead of the above step 2, we fill up the cross table based on the gold results in test.ensv.naacl.

The gold has been already created (the created gold standard has been located in ‘http://www.um.ac.ir/~sarmad/word.a/finalgold.sv.en.RData’). Now, two modes are considered for the alignments. In the first one, it is assumed that the alignments between words in each sentence pairs have been computed using the function align.ibm1 in the package. In this case, using the function align.ibm1, the alignment between words are constructed with the following code:

figure l

The result of the above command line is as follows:

figure m

therefore, having the gold and the alignment files, the code for the above example is:

figure n

The second mode is when the word alignments have been computed by another method (symmetric alignments described in Sect. 3.2) or another word alignment tool. For example, pialign tool. In this issue:

  • Concatenating the traing set and the test set.

  • Run pialign over both of them.

  • Extracting word alignments.

  • Selecting the part corresponding to the test set and evaluating the quality of the word alignment. Notice that using the function cross.table, an excel file with n sheets of the two languages have been created. then, we fill up the cross table based on pialign results for the test set (it has been located in ‘http://www.um.ac.ir/~sarmad/word.a/result.pialign.RData’). The same steps are taken for the rest of the methods. The code for measuring the quality of the word alignment is:

    figure o

    The results of these evaluations are summarized in Table 5.

The most precise numbers in each column have been bolded.

Table 5 Evaluation of the quality of the word alignment by five metrics
Full size table

From Table 5, it can be seen that if the optimal criteria explained in Sect. 4 have been intended, the grow-diag is the best. According to the results, in general, word.alignment shows the best overall performance in comparison to the other packages tested.

7 Future work

We introduced a new R package named word.alignment for computing the statistical word alignment, symmetric alignment, as well as constructing the automatic bilingual dictionary and evaluating the quality of the word alignment by considering five metrics.

The current version of word.alignment provides the most applicable statistical word alignment based on IBM Model 1 and the package can be used for it. Note that there are other statistical word alignments like IBM Models 2 to 5 which have other applications in NLP such as SMT. The package word.alignment can be considered as an appropriate starting point in R for this family. Ultimately, it is our plan to solve the memory restriction in the next versions of this package.