Semi-supervised training for statistical word alignment

Abstract

A system and method for aligning words in parallel segments is provided. A first probability distribution of word alignments within a first corpus comprising unaligned word-level parallel segments according to a model estimate is calculated. The model estimate is modified according to the first probability distribution. One or more sub-models associated with the modified model estimate are discriminatively re-ranked according to word-level annotated parallel segments. A second probability distribution of the word alignments within the first corpus is calculated according to the re-ranked sub-models associated with the modified model estimate.

Description

BACKGROUND OF THE INVENTION

1. Field of the Invention

The present invention relates generally to statistical machine translation, and more particularly to systems and methods for statistical word alignment.

2. Description of Related Art

Word alignment is used in statistical machine translation (SMT) to generate improved translations of documents in two or more foreign languages. SMT may align sentences to extract parallel sentences from parallel documents. After determining sentence alignments, SMT typically includes further aligning words or fragments of the sentences. Conventionally, word alignment in SMT is performed to determine whether a specific word or phrase in one language (e.g., English) corresponds to a specific word or phrase in another language (e.g., French). More specifically, word alignment is a process in which a large collection of parallel documents is used to automatically identify word-to-word or word-to-phrase correspondences.

The Expectation-Maximization (E-M) algorithm is commonly used to perform a word alignment in SMT. In the expectation step of the E-M algorithm, the hypothetical dictionary is used to induce word alignments in a large corpus containing millions of sentences. Based on the induced word alignments, the hypothetical dictionary is modified in the maximization step. The modified dictionary is then used to induce better word alignments by repeating the expectation step. This process is repeated as needed until the hypothetical dictionary remains substantially unmodified from cycle to cycle.

More recently, SMT performs an additional step after the E-M Algorithm is completed. The additional step uses a small corpus comprising manual annotations to indicate word alignments. The additional step estimates another dictionary based on the small corpus and combines this dictionary with the hypothetical dictionary generated by the E-M Algorithm. The combined dictionary is then used to correct word alignments in the large corpus in one final step. However, further improvements to increase the accuracy of SMT are still desired by users of SMT.

SUMMARY OF THE INVENTION

The present invention provides a system and method for aligning words in parallel segments. According to one method, a first probability distribution of word alignments within a first corpus comprising unaligned parallel segments according to a model estimate is calculated. The model estimate used to generate the word alignments is modified according to the first probability distribution of the word alignments. One or more sub-models associated with the modified model estimate are discriminatively re-ranked according to word-level annotated parallel segments. A second probability distribution of the word alignments within the first corpus is calculated according to the re-ranked sub-models associated with the modified model estimate.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 illustrates an exemplary word alignment environment in which the invention may be practiced;

FIG. 2 illustrates a schematic diagram of an exemplary statistical word alignment engine;

FIG. 3 illustrates a flowchart showing an exemplary process for word alignment; and

FIG. 4 illustrates a flow chart showing an exemplary process for discriminatively re-ranking the model estimate.

DETAILED DESCRIPTION

A system and method for word alignment in statistical machine translation (SMT) is provided. The system and method compares parallel segments to produce word alignments indicating a translational correspondence between the words in each of the parallel segments. Segments may comprise parallel text of any length such as documents, sections of documents, paragraphs, sentences, or sentence fragments.

A probability distribution of word alignments in a first corpus of parallel segments may be calculated according to a first model estimate to create a modified model estimate. The modified model estimate may comprise an N-best list where “N” is a constant indicating the number of sub-models comprising the list. The N-best list comprises a list of the top N hypothesized word alignments according to the model estimate for each parallel segment pair. The N-best list may be used to approximate of the full probability distribution of word alignments for these segment pairs according to the model estimate.

A second N-best list based on word alignments in annotated parallel documents may be used to discriminatively re-weight or re-rank one or more sub-models within the modified model estimate. If the N-best list associated with the modified model estimate contains sub-models that are different from an N-best list associated with the first model estimate, the word alignments have not converged. A second probability distribution within the first corpus may be calculated according to the modified model estimate to generate a third model estimate.

FIG. 1 illustrates an exemplary environment 100 in which word alignment may be performed. The environment 100 comprises a word alignment server 102, a network 104, and a client 106. The word alignment server 102 communicates with the client 106 via the network 104. The word alignment server 102 is configured to store a first corpus and a second corpus used to generate word alignments in the first corpus and may comprise a word alignment engine such as word alignment engine 108. The network 104 may comprise a public network (e.g., the Internet) or a private network. The client 106 may comprise storage, a display, a word alignment engine 108 and/or additional functionality not relevant to the scope of this implementation.

In operation, the client 106 accesses the first corpus and/or the second corpus in the word alignment server 102 via the network 104. The word alignment engine 108 processes the first corpus and the second corpus to generate word alignments in the first corpus. Additionally, the word alignment server 102 may receive generated word alignments from the client 106 via the network 104.

FIG. 2 illustrates a schematic diagram of an exemplary statistical word alignment engine 200. The statistical word alignment engine 200 is configured to calculate statistical probabilities of word alignments in a first corpus, according to exemplary embodiments. The statistical word alignment engine 200 comprises a probability module 202, a discrimination module 204, and an error module 206. The probability module 202 is configured to calculate probabilities of word alignments and probability distributions of the word alignments for each sentence pair within a first corpus and to modify a model estimate.

The first corpus may comprise parallel segments. Parallel segments are translations of at least one segment in at least two languages. The parallel segments include “sentence pairs.” Sentence pairs may comprise one or more sentences in a first translation that correspond to one or more sentences in a second translation. The first translation and/or the second translation, however, do not include word alignments, according to some embodiments. Word alignments comprise annotations indicating a correspondence of words and/or phrases in one language to words and/or phrases in another language. The first corpus may comprise millions of sentences.

The model estimate may comprise one or more sub-models. The sub-models can be utilized to calculate a probability of a word-to-word alignment or a word-to-phrase alignment. For example, a sub-model may indicate that if the first four letters of words in two translations match, a word alignment between these words is more probable than other possible word alignments. A second sub-model may indicate that the left-most words in a sentence pair are likely to result in an accurate word alignment.

The word alignments are based on a linear-logarithmic model, according to exemplary embodiments. The linear-logarithmic model may include at least five sub-models h_m (e.g., IBM Model 4), for example. Each sub-model may have an associated weight λ_m. The probability, p, of a word alignment a, may be represented as:

$p_{\lambda }\left(f,a|e\right)=\frac{\exp \left(\sum _i{\lambda }_i{h}_i\left(a,e,f\right)\right)}{\sum _{a^{\prime },f^{\prime }}\exp \left(\sum _i{\lambda }_i{h}_i\left(a^{\prime },e,f^{\prime }\right)\right)}$ where e and f are words. The best word alignment, â, of the words may be specified as:

$\hat{a}=\underset{a}{\arg \max }\sum _i{\lambda }_i{h}_i\left(f,a,e\right)$ where a is an alignment of words e and f. The probability module 202 calculates the probability distribution of the word alignments, a, generated according to the model estimate comprising the sub-models, h_m. The probability distribution may represent a specification of the probabilities, p, for each possible random word alignment, a.

After calculating the probability distribution throughout the first corpus, the probability module 202 modifies the model estimate according to the probability distribution. In alternative embodiments, the probability module 202 creates a new model estimate based on the probability distribution of the model estimate. For example, the probability module 202 may add or delete sub-models. The probability module 202 may additionally alter various parameters within the sub-models. For example, a modified sub-model may indicate that if the first five letters rather than the first four letters of words in two translations match, a word alignment between these words is more probable than other possible word alignments. Further, a second sub-model may indicate that the second from left-most words rather than the left-most words in a sentence pair are likely to result in an accurate word alignment.

The probability module 202 uses the sub-models in the model estimate to evaluate the suitability of a hypothesized alignment. The probability module 202 may, for example, compare a suitability of a first sub-model to a suitability of a second sub-model for each hypothesized word-level alignment. The probability module 202 may additionally determine an ability of the sub-models to discriminate between hypothesized word alignments.

The probability module 202 may modify the model estimate by adjusting a weighting factor of each of the sub-models within the model estimate. A weighting factor may comprise a multiplier indicating the likelihood that the sub-model will yield an accurate translation. For example, a first sub-model having a weighting factor of 0.85 is more likely to be accurate than a second sub-model having a weighting factor of 0.35. Because the weighting factor of the first sub-model is greater, the first sub-model is more favored during a second calculation of the probability distribution based on the modified model estimate.

The model estimate can also be modified by generating new sub-models, removing indiscriminative sub-models, and/or combining sub-models. For example, an indiscriminative sub-model may comprise a sub-model that has a low probability of generating an accurate word alignment. For example, a sub-model that is indiscriminative may specify that a first word in a first translation is likely to correspond to the last word of the source.

If the model estimate has been modified, the discrimination module 204 performs a discriminative process. The discriminative process steers the modification of the model estimate or corrects a new model estimate. To steer the model estimate, the discrimination module 204 generates an N-best list comprising sub-models and sub-model weights according to word alignment annotations within a second corpus.

The second corpus comprises a smaller set of parallel segments translated into the same two or more languages as the first corpus. The second corpus may comprise a section of the first corpus or the second corpus may be completely separate from the first corpus. The second corpus may include approximately one hundred to ten thousand sentence pairs. However, any number of sentence pairs may be included in the second corpus, according to some embodiments.

The sentence pairs in the second corpus include annotations to indicate word-to-word alignments and word-to-phrase alignments within each of the sentence pairs. The annotations may be generated by a human translator. The word alignments may be referred to as “gold standard word alignments.” Alternatively, the annotations may be initially generated using SMT or any other automated process and corrected by a human translator.

Based on the annotations in the small corpus, the discrimination model creates an N-best list where “N” is a constant indicating the size of the list. The N-best list includes a specified number of sub-models that are each associated with a weighting factor based on the gold standard word alignments, discussed herein. After generating a list of sub-models based on the annotations within the second corpus, the discrimination module 204 removes sub-models or word alignments that have a weighting factor that is less than the weighting factor of N number of sub-models. For example, the constant N may be 128. The N-best list includes the 128 most heavily weighted sub-models derived from the annotations in the second corpus. In alternative implementations, the probability module 202 may generate the N-best list.

The discrimination module 204 may check whether the model estimate generated based on the first corpus has converged with the model estimate based on the second corpus. If these model estimates have converged such that there are no new N-best entries in the N-best list, the word alignment is complete.

If there are new N-best entries, the discrimination module 204 then enriches the model estimate generated by the probability module 202 based on the first corpus with the N-best list based on the second corpus. The enrichment adds the weighting factors in the N-best list based on the second corpus to the weighting factors in the model estimate based on the first corpus, compares the manual alignments to the hypothesized alignments, and weights one or more sub-models within the model estimate based on the first corpus according to the comparison. For example, a sub-model associated with a weighting factor of 0.15 in the model estimate may have a weighting factor of 0.45 in the N-best list. As a result of the enrichment, the sub-model may be re-weighted to a weight of 0.30. Having a higher weight, the sub-model is more likely to generate word alignments in a subsequent iteration. In this example, the weights are evenly weighted. However, in other embodiments, the weighting factor in the model estimate may be greater than the weight of the sub-model in the N-best list, and vice-versa.

To re-weight the model estimate, the discrimination module 204 compares the gold-standard word alignments to hypothesized word alignments in the first corpus. The hypothesized word alignments in the first corpus are the word alignments generated according to the model estimate. The discrimination module 204 compares the weighting factor of each sub-model to randomly generated weighting factors. In one implementation, the discrimination module 204 generates 999 weighting factors. To determine the best weighting factor for each sub-model in the model estimate, an error for each weighting factor is calculated.

A one-dimensional error minimization is performed on the selected weighting factor. The one-dimensional error-minimization is the weight that results in increased error reduction if the weights associated with the other sub-models are held constant. The one-dimensional error-minimization is repeated until there is no sub-model weight that can be modified to reduce error. The weighting factor with the lowest error is selected and a corresponding component of the sub-model weight vector may be modified. The discrimination module 204 repeats the discriminative step until substantially all of the hypotheses in the N-best list are in the model estimate. Other error-minimization methods, such as Powell's method, that are apparent to one skilled in the art may be used.

The error module 206 determines whether the model estimate has converged with the N-best list. The model estimate has converged if the error was not decreased from a previous iteration. If there is no convergence, the error module 206 instructs the probability module 202 to perform another iteration comprising calculating a second probability distribution, creating a second modified model estimate, and/or performing the discriminative step. The error module 206 may additionally transmit a command to display and/or record the word alignments generated by the model estimate or a final version of the model estimate. The F-measure is used to measure the error of the word alignments using the value 1—F-measure. The F-measure may be calculated as:

$F\left(A,S,\alpha \right)=\frac{1}{\frac{\alpha }{\mathrm{precision}\left(A,S\right)}+\frac{\left(1-\alpha \right)}{\mathrm{recall}\left(A,S\right)}}$ where α is a relative weighting ratio of the recall value to the precision value.

In alternative embodiments, the error module 206 may determine a finite number of iterations to perform. In these embodiments, the error module 206 may not calculate the F-measure. For example, if the model estimate typically converges at iteration number ten, the error module may perform subsequent word alignments according to a default of ten iterations. Although various modules are illustrated as comprising the statistical word alignment engine 200 in FIG. 2, fewer or more modules may comprise the statistical word alignment engine 200 and still fall within the scope of various embodiments.

FIG. 3 illustrates a flowchart showing an exemplary process 300 for word alignment. In some embodiments, the exemplary process 300 may be implemented as a computer program embodied on a computer-readable medium having instructions comprising the exemplary process 300. In each of the iterations of the process 300 for word alignment, the discriminative step steers the model estimate to produce a more accurate model estimate and, therefore, determines more accurate word alignments in the first corpus. The process 300 may comprise step 302, initializing the model estimate; step 304, inducing word alignments in a first corpus according to the model estimate; step 306, calculating a probability distribution of all word alignments in the first corpus; step 308, replacing the model estimate according to a probability distribution of the word alignments to generate an N-best list; step 310, determining whether the word alignments have converged; step 312, using a second corpus to discriminatively re-rank the model estimate if the word alignments have not converged; and step 314, if the word alignments have converged, recording the word alignments generated by the model estimate. However, fewer or more steps may comprise the process 300 and still fall within the scope of various embodiments.

Step 302 includes initializing the model estimate that induces word alignments in the first corpus. The initial model estimate may comprise an assumption that all translations are equally probable. In alternative embodiments, the initial model estimate may comprise one or more default sub-models associated with one or more default weightings.

Step 304 includes inducing word alignments in the first corpus. The word alignments are induced according to the sub-models within the model estimate. In the initial inducing step using the initial model estimate, the sub-models may be equally weighted.

Step 306 includes calculating the probability distribution of all or substantially all of the induced word alignments in the first corpus. The probability distribution indicates the likelihood of each induced word alignment according to the model estimate. Based on the probability distribution associated with the first corpus, the sub-models having a higher probability will be assigned a higher weighting factor than the less probable sub-models in the model estimate.

In step 308, according to the probability distribution of the first corpus, a new model estimate replaces the former model estimate. In some embodiments, the former model estimate may be simply modified by, for example, changing a weighting factor of a sub-model within the model estimate. In some embodiments, Viterbi training may approximate the hypothesized word alignments to update the new model estimate using an assumed word alignment.

Step 310 includes determining whether the hypothesized word alignments in the first corpus have converged. Convergence occurs if there are no new N-best entries in the modified model estimate. If the hypothesized word alignments have not converged, the process 300 continues to step 312. In alternative embodiments, step 310 may comprise determining whether a default number of iterations have been performed.

In step 312, the sub-models within the model estimate are re-ranked using the small corpus of annotated parallel segments. The sub-models are re-ranked by modifying the weighting factors associated with each sub-model. This process is discussed in greater detail in connection with FIG. 2 and FIG. 4.

If the word alignments have converged in step 310 or if the default number of iterations has been performed, the process 300 continues to step 314. Step 314 includes recording the word alignments in the first corpus. The word alignments may be recorded in an electronic media such as ROM, a hard disk, or similar media.

FIG. 4 illustrates a flow chart showing an exemplary process 400 for discriminatively re-ranking the model estimate. The discrimination module 204 in the statistical word alignment engine 200 may perform the process 400, for example. Discriminatively re-ranking the model estimate associated with the first corpus in each of the iterations improves word alignment in a first corpus by steering the weighting of the sub-models towards the manual word alignments.

In step 402, manual word alignments within the second corpus are added to the hypothesized word alignments within the first corpus according to sub-model weighting factors. More specifically, the model estimate derived from the first corpus comprises at least one sub-model that is associated with a weighting factor. The sub-models in the N-best list associated with the first corpus correspond to one or more weighted sub-models within an N-best list associated with the second corpus. The weighting factor of each sub-model within the N-best list associated with the first corpus is processed with the weighting factor in the N-best list associated with the second corpus to generate an updated weighting factor.

In step 404, to compare the manual word alignments in the second corpus to the hypothesized word alignments in the first corpus, the updated weighting factor for each sub-model derived using the first corpus is compared to randomly generated weighting factors. The weighting factor that generates the least amount of error is selected. By comparing the weighting factors of the sub-models, the accuracy of the word alignments generated by those sub-models is also compared. If the weighting factors associated with the sub-models have a smaller amount of error, the model estimate associated with the first corpus will produce more accurate word alignments because the model estimate associated with the first corpus is more likely to apply an accurate sub-model when performing each word alignment within the first corpus.

In step 406, weighting the one or more sub-models according to the comparison, the selected weighting factor is further refined using a one-dimensional error minimization until there is no further error reduction. An example of a one-dimensional error minimization algorithm includes calculating a piecewise constant function which evaluates the error of the word alignments which are selected by the “best word alignment”, â, equation if all weights associated with the sub-models remain constant except for the weight of the sub-model on which the error is evaluated.

In step 408, determining whether to change a weight associated with a sub-model to reduce error, a weighting factor associated with one or more of the sub-models may be selected. The selected weighting factors are modified to reduce overall error generated by the N-best list associated with the first corpus. If no further error reduction is possible by adjusting the weighting factors, the process returns to step 312, depicted in FIG. 3.

In step 410, re-weighting word alignment sub-models to steer model estimate associated with the first corpus, the weighting factor is associated with the sub-model in the model estimate is modified to decrease word alignment error in the first corpus. In some embodiments, a component of the sub-model weight may be updated. The new weighting factor for each sub-model within the model estimate re-ranks the model estimate. If the weighting factor is selected, the process 400 returns to step 406 to further refine the new weighting factor until there is substantially no further error reduction.

While various embodiments have been described above, it should be understood that they have been presented by way of example only, and not limitation. For example, any of the elements associated with the statistical word alignment engine 200 may employ any of the desired functionality set forth hereinabove. Thus, the breadth and scope of a preferred embodiment should not be limited by any of the above-described exemplary embodiments.

Claims

1. A method for aligning words in parallel segments, the method comprising: calculating a first probability distribution, utilizing a processor and a memory, according to a model estimate of word alignments within a first corpus comprising word-level unaligned parallel segments, the model estimate comprising an N-best list of one or more sub-models;modifying the model estimate according to the first probability distribution;discriminatively re-ranking one or more sub-models associated with the modified model estimate according to word-level annotated parallel segments; andcalculating a second probability distribution of the word alignments within the first corpus according to the re-ranked sub-models associated with the modified model estimate;wherein discriminatively re-ranking one or more sub-models within the modified model estimate according to manual alignments further comprises:adding manual alignments to hypothesized alignments within the first corpus;comparing the manual alignments to the hypothesized alignments; andweighting the one or more sub-models according to the comparison; andwherein the comparing of the manual alignments to the hypothesized alignments comprises:comparing an updated weighting factor for each sub-model derived using the first corpus to randomly generated weighting factors; andselecting one of the updated weighting factor and the randomly generated weighting factor that generates a least amount of error.

2. The method recited in claim 1, wherein the word-level annotated parallel segments comprise annotations indicating manual alignments.

3. The method recited in claim 1, further comprising determining whether a first error associated with the re-ranked modified model estimate converges with a second error associated with the model estimate.

4. The method recited in claim 1, further comprising determining a number of iterations to perform, the iterations comprising the steps of: calculating a third probability distribution according to the re-ranked modified model estimate within the first corpus;further modifying the re-ranked modified model estimate according to the third probability distribution; andfurther discriminatively re-ranking one or more sub-models associated with the re-ranked modified model estimate according to the word-level annotated parallel segments.

5. The method recited in claim 1, wherein the first corpus is larger than a second corpus.

6. The method recited in claim 1, further comprising initializing the model estimate.

7. The method recited in claim 1, wherein: the weighting of the one or more sub-models according to the comparison is according to at least one weighting factor; andthe discriminative re-ranking of the one or more sub-models within the modified model estimate according to manual alignments further comprises refining at least one of the at least one weighting factors using a one-dimensional error minimization until there is no further error reduction.

8. The method recited in claim 7, wherein the refining of the at least one weighting factor further comprises calculating a piecewise constant function that evaluates an error of the word alignments selected by a best word alignment equation keeping the at least one weighting factor for each of the one or more sub-models constant except for one of the at least one weighting factor for the sub-model being evaluated.

9. A computer program embodied on a non-transitory computer readable medium having instructions for aligning words in parallel segments comprising: calculating a first probability distribution of word alignments within a first corpus comprising unaligned parallel segments according to a model estimate, the model estimate comprising an N-best list of one or more sub-models;modifying the model estimate according to the probability distribution;discriminatively re-ranking one or more sub-models within the modified model estimate according to annotated parallel segments; andcalculating a second probability distribution of the word alignments within the first corpus according to the re-ranked modified model estimate;wherein discriminatively re-ranking one or more sub-models within the modified model estimate according to manual alignments further comprises:adding manual alignments to hypothesized alignments within the first corpus;comparing the manual alignments to the hypothesized alignments; andweighting the one or more sub-models according to the comparison;wherein:the weighting of the one or more sub-models according to the comparison is according to at least one weighting factor; andthe discriminative re-ranking of the one or more sub-models within the modified model estimate according to manual alignments further comprises refining at least one of the at least one weighting factors using a one-dimensional error minimization until there is no further error reduction; andwherein the refining of the at least one weighting factor further comprises calculating a piecewise constant function that evaluates an error of the word alignments selected by a best word alignment equation keeping the at least one weighting factor for each of the one or more sub-models constant except for one of the at least one weighting factor for the sub-model being evaluated.

10. The computer program recited in claim 9, wherein the annotated parallel segments comprise annotations indicating manual alignments.

11. The computer program recited in claim 9, further comprising an instruction for determining whether a first error due to the re-ranked modified model estimate converges with a second error due to the model estimate.

12. The computer program recited in claim 9, further comprising an instruction for determining a number of iterations to perform, the iterations comprising the steps of: calculating a third probability distribution according to the re-ranked modified model estimate within the first corpus;further modifying the re-ranked modified model estimate according to the third probability distribution; andfurther discriminatively re-ranking one or more sub-models associated with the re-ranked modified model estimate according to the word-level annotated parallel segments.

13. The computer program recited in claim 9, wherein the first corpus is larger than a second corpus.

14. The computer program recited in claim 9, further comprising an instruction for initializing the model estimate.