Statistical Machine Translation Model Based on Language

The disadvantage of word-based statistical machine translation is that it does not take contextual information and is based only on statistical analysis of words. The semantic-based statistical machine translation model is more advanced in that it processes on the semantics instead of words. This allows the system to translate phrases that avoid word-by-word translation.

In the semantic-based statistical machine translation system [26], sentences in the source language e are

polyglot​​

(is a sequence of words, not necessarily grammatically correct)

(The punctuation mark is also considered as a word.) Each 𝑒 𝑖 phrase is translated into

The corresponding 𝑣 𝑖 is based on the probability distribution 𝜙(𝑒 𝑖|𝑣 𝑖 ) . Then the terms 𝑣 𝑖

will be

The order conversion is based on the conversion model 𝑑(𝑎 𝑖 − 𝑏 𝑖−1 ) , where 𝑎 𝑖 is the starting position of phrase 𝑒 𝑖 and 𝑏 𝑖−1 is the ending position of phrase 𝑣 𝑖 . Therefore, the statistical translation system on the language will learn bilingual phrase pairs, especially idiomatic sentences.

Thus, the best translation that satisfies formula (2.1) will be rewritten as:

𝑚

𝑝 𝑒 𝑖𝑣 𝑖= 𝜙(𝑒 𝑖|𝑣 𝑖 ) × 𝑑(𝑎 𝑖 − 𝑏 𝑖−1 )

𝑖=1

(2.6)

takes a

small green box

.

little blue box

.

She

Figure 2.4 illustrates the process of statistical machine translation based on phrases. The English input sentence is broken down into many phrases, also known as phrases. The phrases are translated into the corresponding Vietnamese phrases, and the output Vietnamese phrases can change the order in the sentence to match Vietnamese.

She

take one

Figure 2.4. Example of semantic-based statistical translation

Because of the statistics on pairs of words, this translation system can convert the order between words in a sentence, but it still cannot automatically convert the order of phrases that are far apart in a sentence.

There are different ways to extract word pairs from bilingual corpora. [16] tested the following three methods:

i. Get the word based on the word alignment results

The author uses the GIZA++ tool to align words in the bilingual corpus. Then, Koehn uses some heuristics to further improve the alignment results and get all the pairs of words containing linked words. Then, 𝜙(𝑒 𝑖|𝑣 𝑖 ) is calculated as follows:

𝜙𝑒𝑣= 𝑐𝑜𝑢𝑛𝑡(𝑒 |𝑣 )

ii. Syntactic Separation

𝑖 𝑖

𝑒 𝑐𝑜𝑢𝑛𝑡(𝑒 |𝑣 )

(2.7)

First, the author aligns the words for the bilingual sentence pairs, then analyzes the sentence pairs into a syntax tree. The author extracts the bilingual language pairs by taking the word strings that are in the subtree of the syntax tree and have word alignment links. The translation probability of the language pair is calculated similarly to the above model.

iii. Using the combined model proposed by Marcu, D. and Wong, W: Direct phrase formation on bilingual corpus

Through experiments, the author concluded that the word alignment based model gives the best results among the three models.

Consider the pair of bilingual sentences:

That girl is the highest pupil in the class.

That girl is the tallest student in the class.

The model extracts the following pairs of words from the word alignment results: From the word alignment results of the sentence pair:

The extracted pairs of terms must be consistent as in figure (a), words that are linked to words in the source term are also included in the target term. The extraction in figure (b) is incorrect because the word “là” is linked to the word “is” but is not included in the term.

Initially, we can get the terms from the alignment links from

(That, that), (girl, cô bé), (is, là), (highest, cao nhất), (pupil, học sinh), (in, trong), (class, giá), (., .)

(That, that), (girl, Cô bé), (is, là), (highest, cao nhất), (pupil, học sinh), (in, trong), (class, lớp), (., .), (That girl, Cô bé mà), (is the, là), (highest pupil, học sinh cao nhất), (the class, trong học)

(That, that), (girl, Cô bé), (is, là), (highest, cao nhất), (pupil, học sinh), (in, trong), (class, lớp), (., .), (That girl, Cô bé mà), (is the, là), (highest pupil, học sinh cao nhất), (the class, trong học), (That girl is, Cô bé mà là), (highest pupil in, học sinh cao nhất trong), (in the class, trong học), (the class . , trong học .)

(That, that), (girl, Cô bé), (is, là), (highest, cao nhất), (pupil, học sinh), (in, trong), (class, lớp), (., .), (That girl, Cô bé mà), (is the, là), (highest pupil, học sinh cao nhất), (the class, trong học), (That girl is, Cô bé mà là), (highest pupil in, học sinh cao nhất trong), (in the class, trong học), (the class . , trong học .), (That girl is the, Cô bé mà là), (is the highest pupil, là học sinh cao nhất), (highest pupil in the, cao nhất trong), (in the class, trong học)