In the field of data-driven machine translation, it is desirable to obtain as much parallel data as possible about the language pair for which the translation system is built. Mutual translations of source and target language texts and text fragments are used as data to feed a learning engine, which builds models that are then used by an actual translation engine. Parallel texts, i.e., texts and text fragments that are mutual translations of each other, are an important resource in these applications.
Unfortunately, parallel texts are a scarce resource. Parallel texts are often limited in size, coverage and language. The parallel texts that do exist are usually from one domain, which may be problematic because certain machine translation systems trained in a first domain will not perform well in a second, different domain.
Certain textual resources which are not parallel may still be related in that they contains information about the same subject. Examples of such resources include the multilingual newsfeeds produced by several news agencies. Examples of these news agencies may include Agence France Presse, Xinhua News, and others. The same or similar news stories are often found in different languages. Therefore, while the texts may not be parallel—an Aljazera story about president Bush's visit to Europe may be written independently from a CNN story about the same visit, much information can be obtained from these comparable stories that can be useful in the context of developing translation systems.
A parallel text discovery system attempts to discovers pairs of sentences or segments which are translations of one another starting from collections of non-parallel documents. Previous research efforts have attempted to discover parallel sentences in parallel text. These techniques assume the parallel texts to be mutual, complete translations of each other and attempt to align all the sentences in these related text.
Zhou et al, “Adaptive parallel sentences mining from Web bilingual news collection” 2002 IEEE international conference on data mining, use a generative model for discovering parallel sentences between Chinese and English sentences. This and other comparable systems define a sentence alignment score and use dynamic programming to find the best sentence alignment between a pair of documents. Performance depends heavily on the degree to which the input consists of true parallel documents. If the method is applied, for example, to 2 documents of 20 sentences each that share only one sentence or sentence fragment, the techniques will not be likely to obtain useful information from sentences that convey the same meaning.
The present system allows judging each of a plurality of sentence pairs in a “nonparallel” comparable corpus individually, without using context, and without assuming that the texts which contain the sentences are in any way related. Throughout this document, the term “nonparallel” refers to a collection of texts or other information that is not necessarily parallel—it may include both parallel and nonparallel portions, but basically the relationship between the sentences in the articles in the database is unknown.
An embodiment uses a classifier which accurately labels sentence pairs as parallel or nonparallel by inspecting different kinds of correspondences between the words. This enables extracting parallel sentences from very large comparable databases of “unrelated” texts The texts are “unrelated” in the sense that they are not necessarily related. However, of course, they must include related information in order to be useful as a training resource.
An embodiment may use document pair selection as a filter.
Another embodiment may carry out a iterative training, or bootstrapping using the information in a database. Small amounts of parallel data are used in combination with a massive nonparallel Corpus to extract a larger set of parallel sentences. These parallel sentences can then be used as additional training data to improve the performance of the sentence extraction program. The procedure can be applied repetitively until no the improvements reach a certain level.
These and other aspects will now be described in detail with respect to the accompanying drawings, wherein:
A statistical machine translation system relies on training data that comes from various sources. The training data represents translation information between first language and second language information.
The present system uses unrelated texts. The texts include comparable text in addition to unrelated text. The term “text” is used to refer to any machine readable data in any language, and may include electronic documents of various forms, newsfeeds, web documents, scanned documents, and machine readable information of any kind and from any source. The system obtains its training information from that data. It does this by finding data fragments (sentences or sentence fragments) that are mutual translations of each other and then uses these mutual translations to train a data-driven machine translation system.
While the present specification describes some parameters that can be used to determine parallel phrases and sentences, it should be understood that other parameters can be similarly used.
The domain from which the translations are taken may be used as a parameter in the present system. For example, parallel training data from one domain such as parliamentary proceedings may not perform well on another domain such as news reporting. In another aspect, initial training is carried out using comparable domains, and supplemental training is carried out using other data. This aspect extracts parallel information, e.g., parallel sentences or phrases, from comparable corpora from the domain of interest.
In an embodiment, two different bodies of monolingual texts are obtained. One of the bodies of data uses in-domain data, and the other is referred to as out of domain data. A technique is provided to show how end to end performance of a statistical machine translation system may be improved using both the in domain, and out of domain, data.
The processor 150 may be any computer driven device, including a general-purpose processor, a computer, a digital signal processor, or a dedicated hardware device that can operate based on programming instructions.
The corpora are divided into articles at 204 and pairs of comparable articles are selected at 210 as another parameter indicative of training. Each article pair is analyzed to determine possible sentence pairs at 215. The candidate sentence pairs obtained at 120 are analyzed using a dictionary 220 and maximum entropy classifier 125 which produces a determination indicative of whether the sentences in each pair are mutual translations of one another.
The output is parallel sentences 230 which can be used for training.
This system may operate with minimal resources, e.g. a dictionary, and/or a small amount of parallel data. In an embodiment, the system may operate with only a small amount of parallel data, from which a dictionary can be learned automatically.
Document selection may be performed using the IR engine inquiry described in Callen et al, “TREC and Tipster experiments with InQuery”, Information Processing and Management, 31(3): 327-343. All the English documents are indexed into a database, and a query is created for each Arabic document. A probabilistic dictionary is formed from the queries. The top translations of each word in the document are obtained, e.g. the top 5 translations for each word. Each word translation is then used to find sentences in the other document that includes that word. A query is created using InQuery's weighted sum, or wsum operator, using the translation probabilities as weights. The query is then run and use to retrieve the top 100 English documents with individual words that match to the Arabic document. This is shown as 500 in
Extrinsic information is also used at 505 as another parameter. For example, it is likely that documents with similar content will have publication dates that are close to one another. Thus, the top 100 English documents may be further filtered using this extrinsic information. In an embodiment, only those documents published within a window of five days around the publication date of the Arabic query document may be maintained.
Once article pairs have been selected, the candidate sentence pair selection 215 takes all possible sentence pairs in each document pair and passes them through a word overlap filter 510. The filter verifies information to determine the likelihood of a sentence match. For example, the filter may run a ratio check, in which it checks to determine if the ratio of lengths of the two sentences is not greater than two. The filter may then run a word percentage check, e.g. using common words as a parameter, and to check that at least half the words in each sentence have a translation in the other sentence. Any sentence that does not fulfill these two conditions are discarded. The other sentences are passed on to the parallel sentence selection stage as parallel candidate sentence pairs at 120.
The sentence pair candidate selection at 115 reduces much of the noise introduced by the recall oriented document selection procedure. While it may also remove good pairs, many of those good pairs could not have been handled reliably anyway. Therefore, the overall effect of this filter is to improve the precision and robustness of the system.
The candidate sentence pairs at 120 are further analyzed to determine whether the two sentences in a pair are mutual translations. A maximum entropy classifier 125 may be used for this purpose. The pairs that are classified as being mutual translations form the output 130 of the system as parallel sentences.
The maximum entropy statistical modeling framework imposes constraints on the model of the data by defining so-called feature functions. The feature functions emphasize the properties of the data in most useful for the modeling task. For any sentence pair sp, the number of words in either sentence that have a translation in the other sentence, or word overlap, is a useful indicator of whether the sentences are parallel. A feature function f(sp) is defined whose value is a log linear combination of the functions, representing the word overlap of the sentences in sp.
where c is the class, meaning parallel or not parallel, Z(sp) is a normalization factor, and fi are the feature functions.
The resulting model has free parameters λj, the so-called feature weights. Parameter values that maximize the likelihood of a given training corpus can be computed using techniques such as the GIS algorithm or the IIS algorithm described in Darroch, et. al.
The present system attempts to find feature functions that distinguish between parallel and nonparallel sentence pairs. The pairs are determined by computing and exploiting word level alignments between sentences in each pair. A word alignment between two sentences in two different languages is used as a parameter to specify words in one sentence which are exact translations of words in the other.
In a correct alignment between two nonparallel sentences, as shown in
This is addressed by defining the fertility of a word in an alignment as the number of words that the word is connected to. In an automatically computed alignment, the presence between a pair of sentences of words of high fertility is indicative of non-parallelism. For example, the English word at in
Another parameter is the presence of long contiguous spans. Contiguous spans are defined as pairs of bilingual substrings in which the words in one sub string are connected only two words in the other sub string. A span may contain a few words without any connection (a small percentage of the length of the span) but no word with a connection outside the span. For example, the spans may include 3-10 words in a span, that directly translate to corresponding spans.
For a probabilistic dictionary, the alignment score can be defined as the normalized product of the translation probabilities of the connected word pairs. This score is indicative of how parallel the sentences may be. For example, a pair of nonparallel sentences should have connections of lower probabilities.
Parameters indicative of alignment features may also be tested as part of the alignment score. These may include:
All of this information can be combined to form a score.
This can be computed using the IBM model
Where the source sentence is f, the target sentence is e, and the alignment is a. In the equation, m is the length of the source sentence, l the is the length of the target sentence, and e-P(m|e).
This basically corresponds to the normalized product of the translation probabilities of all the links in the alignment. The best alignment is computed according to this model, as the alignment that maximizes the product term
One alignment is computed for each translation direction, that is, a first alignment is computed from f to e, and a second alignment e to f. The alignments are then combined. Three different combination methods may be used, known as intersection, Union and refine, where refine is a form of intersection expanded with certain additional neighboring links.
In the embodiment therefore, alignments are computed for each sentence pair. One set of general features and five set of alignment features are extracted. The parameters of the model are trained on instances obtained from a small (e.g., 30K-200K word) parallel corpus. All possible bilingual sentence pairs from the corpus are obtained, placed through the word overlap filter, and used to create training instances.
In this embodiment, the classifier is used repetitively, each repetitive use of the classifier improving the results. This re-use of the database and classifier effectively bootstraps the system to improve its performance. The classifier's performance is affected by the dictionary coverage, and the similarity between the domains of the training and test instances. In the embodiment, all the dictionaries are automatically learned from parallel data. This enables creation of dictionaries of various coverage by learning them from parallel corpuses of different sizes. In the test, five dictionaries are used, to learn from five initial out of domain parallel corpuses whose sizes are 100K, 1 Million, 10 M, 50 M and 95 M word corpuses (on the English side). Two training sets were used, one generated from an in domain parallel corpus and another from an out of domain parallel corpus.
Each initial out of domain corpus is used to learn a dictionary. The training and test corpuses are obtained as above, and sentence pairs are generated, passed through the word overlap filter using the dictionary, to obtain two sets of training instances, and one set of test instances. Two different classifiers are respectively trained; one on each training set. Both of them are evaluated on the test set.
The parallel corpora used for generating training and test instances have about 5000 sentence pairs each or approximately 50,000 English tokens. This generates around 10,000 training instances for each training sets.
The comparable corpora used for parallel sentence extractions are collections of news stories. Each language pair is analyzed from these collections to create an in domain comparable corpus by piecing together articles from the same agency around the same time. Another aspect uses this data along with data mined from the Web, obtaining comparable corpora using bilingual news web sites and downloading news articles in each language independently. Web sites and search engines may be used to obtain lists of articles and their URLs to obtain the data.
All of this data may be used to improve the quality of a machine translation system. The main goal extracts parallel training data from an in domain comparable corpus that improves the performance of an out of domain trained system. Thus, the extracted corpus is added to the out of domain training data to improve its performance.
The system evaluates the extracted corpora presented the above. The extraction system that was used to obtain each of those corpora made use of a certain initial out of domain parallel corpus. The baseline system is trained based on that initial corpus. Another system which will be called “plus extracted” is trained on the initial corpus plus the extracted corpus. Results show that the automatically extracted additional training data yields significant improvements in performance over the initial training corpora.
The bootstrapping embodiment described herein allows obtaining a small amount of parallel training data, e.g. in domain data. That data is used to learn a new dictionary. After that, the system uses that dictionary to extract again.
As a test, bootstrapping iterations are carried out starting from two very small corpora, 100,000 English tokens and a million English tokens respectively. After each iteration, the system is trained and evaluated. After each iteration, the system uses the new dictionary to evaluate the database again. This bootstrapping system therefore allows starting with the machine translation from very little parallel data, and using a large amount of comparable nonparallel data to bootstrap into an improved translation.
Iteration may be continued until there is no further improvement in machine translation performance based on the development data or until some point of diminishing returns is reached, e.g., that the improvement is less than some amount. As an example, the iterative system may start with as little as 100,000 English tokens of parallel data. This can iteratively be continued until it may become comparable to parallel training on three orders of magnitude more data.
Although only a few embodiments have been disclosed in detail above, other modifications are possible, and this disclosure is intended to cover all such modifications, and most particularly, any modification which might be predictable to a person having ordinary skill in the art. For example, while the above has described certain models and parameters, it should be understood that the system may use different models and different parameters for the basic functions of article selection and sentence pair extraction. Moreover, while the above has described certain languages, it should be understood that this system may be used with other languages and with other units. Moreover, while the above has described certain size databases, it should be understood that databases of other sizes can alternatively be used.
Also, only those claims which use the words “means for” are intended to be interpreted under 35 USC 112, sixth paragraph. Moreover, no limitations from the specification are intended to be read into any claims, unless those limitations are expressly included in the claims.
This application claims the benefit of the priority of U.S. Provisional Application Ser. No. 60/555,807, filed Mar. 23, 2004 and entitled “Parallel Text Discovery System”, the disclosure of which is hereby incorporated by reference.
Number | Date | Country | |
---|---|---|---|
60555807 | Mar 2004 | US |