Training for a text-to-text application which uses string to tree conversion for training and decoding

Abstract

Training and translation using trees and/or subtrees as parts of the rules. A target language is word aligned with a source language, and at least one of the languages is parsed into trees. The trees are used for training, by aligning conversion steps, forming a manual set of information representing the conversion steps and then learning rules from that reduced set. The rules include subtrees as parts thereof, and are used for decoding, along with an n-gram language model and a syntax based language mode.

Description

BACKGROUND

Modern machine translation systems use word to word and phrase to phrase probabilistic channel models as well as probabilistic n-gram language models.

A conventional way of translating using machine translation is illustrated in FIG. 1. FIG. 1 illustrates the concept of Chinese and English as being the language pair, but it should be understood that any other language pair may be alternatively used.

Training is shown as 150, where a training corpora 153 is used. The corpora has an English string 151 and a Chinese string 152. An existing technique may be used to align the words in the training corpora at a word level. The aligned words are input to a training module 155 which is used to form probabilities 165 based on the training corpora. A decoding module 167 is used that maximizes the argument argmax/e P(e)*P(f|e), and maximizes the probability of e, given certain languages in the corpora, where e and f are words or phrases in the training corpora. The decoding module 167, which may simply be a module within the same unit as the training module. The decoder thus takes a new Chinese string such as 160, and uses the probabilities 165 along with a language model 161 which may be an n-gram language model. The decoder outputs English strings which correspond to the highest scores based on the probabilities and the language model.

Phrase based systems may sometimes yield the most accurate translations. However, these systems are often too weak to encourage long-distance constituent reordering when translating the source sentences into a target language, and do not control for globally grammatical output.

Other systems may attempt to solve these problems using syntax. For example, certain reordering in certain language pairs can be carried out. One study has shown that many common translation patterns fall outside the scope of the Child reordering model of Yamada & Knight, even for similar language pairs such as English/French. This led to different possible alternatives. One suggestion was to abandon syntax on the grounds that syntax was a poor fit for the data. Another possibility is to maintain the valid English syntax while investigating alternative transformation models.

SUMMARY

The present application describes carrying out statistical analysis using trees created from the strings. In training, trees are created and used to form rules in addition to the probabilities. In application, trees are used as output, and either the trees, or information derived from the trees, may be output. The system may input strings of source symbols, and outputs target trees.

In an embodiment, transformation rules that condition on larger fragments of tree structure are created. These rules can be created manually, or automatically through corpus analysis to form a large set of such rules. Specific cases of crossing and divergence may be used to motivate the algorithms to create better explanation of the data and better rules.

The present description describes string to tree translation. Different aspects are described which enable a direct translation between the string and the syntax tree.

BRIEF DESCRIPTION OF THE DRAWINGS

These and other aspects will now be described in detail with reference to the accompanying drawings, wherein:

FIG. 1 shows a block diagram of a translation system;

FIG. 2 shows an embodiment using tree portions as parts of the rules;

FIG. 3A-3C show formation of trees and alignment of the steps;

FIG. 4 shows derivation steps and the induced rules therefrom;

FIG. 5 shows an alignment graph and

FIGS. 6 and 7 show minimal fragments derived from the FIG. 5 alignment graph;

FIG. 8 shows how the minimal fragments are combined;

FIG. 9 shows a flowchart, run on the computer of FIG. 10;

FIGS. 11 and 12 show crossing and reordering; and

FIG. 13 shows a decoding rule.

DETAILED DESCRIPTION

The general structure and techniques, and more specific embodiments which can be used to effect different ways of carrying out the more general goals are described herein.

FIG. 2 illustrates an overall block diagram of an embodiment. In an embodiment, the rule learning is used for learning rules for a text to text application. The rule learning and the text to text application may each be carried out on a computer 1000 such as shown in FIG. 10, which includes an associated memory 1001 storing the translation rules, probabilities and/or models. The computers described herein may be any kind of computer, either general purpose, or some specific purpose computer such as a workstation. The computer may be a Pentium class computer, running Windows XP or Linux, or may be a McIntosh computer. The programs may be written in C, or Java, or any other programming language. The programs may be resident on a storage medium, e.g., magnetic or optical, e.g. the computer hard drive, a removable disk or other removable medium. The programs may also be run over a network.

In this embodiment, the English string 151 and Chinese string 152 are first word aligned by alignment device 251. The English string is parsed by a parser 250, as described herein, into an English tree 255 that represents the contents of the English string. The English tree is used along with the Chinese string 152 by a string based training module 260. The translation module 260 produces probabilities shown as 265, and also produces subtree/sub string rules indicative of the training, and shown as 270. Thus, the training device produces rules with probabilities, where at least a portion of at least some of these rules are in the form of trees.

The rules and probabilities are used by the decoding module 267 for subsequent decoding of a new Chinese string 160. Decoding module 267 also uses multiple language models, here an n-gram language model 161, and also a syntax based language model 262. The output 280 of the decoding module 267 corresponds to all possible English trees that are translations of the Chinese string according to the rules. The highest scoring English trees are displayed to the user. Alternatively, information that is based on those trees may be displayed, for example, string information corresponding to those trees.

Some advantages of the embodiment include the following. The use of information from trees within the rules can allow the model to learn what the different parts represent. For example, the FIG. 1 machine translation system has no idea what a noun is, but the embodiment can learn that as part of the translation. In addition, the present embodiment provides tree/string rules, as compared with the phrase substitution rules which are produced by the FIG. 1 system. The use of trees enables the use of the syntax based language model 262, which is not conventional in the prior art.

According to another embodiment, the training information in both languages may be parsed into trees prior to the training.

Tree outputs produce outputs which are well formed, having a verb in the right place, for example, and other parts also in the right places. In addition, tree/string rules capture information about when reordering may be useful. Tree/string rules control when to and when not to use function words. However, many of the tree string rules may be simple word to phrase substitutions.

The training is described herein with reference to FIGS. 3-9.

FIG. 3 a shows a French sentence, (il ne va pas) and a parse tree 300 of its translation into English. The parse tree includes the conventional parsing parts, the sentence S, noun phrase (NP), verb phrase (VP) and other conventional sentence parts.

An embodiment defines determining rules using a string from a source alphabet that is mapped to a rooted target tree. Nodes of that rooted target tree are labeled from a target alphabet. In order to maintain this nomenclature, symbols from the source alphabet are referred to as being “source symbols”. Symbols from the target alphabet are referred to as being “target symbols”. A symbol tree is defined over an alphabet Δ as a rooted directed tree. The nodes of this alphabet are each labeled with a symbol of Δ. In an embodiment, a process by which the symbol tree is derived from the string of source signals, over the target language, is captured. The symbol tree to be derived is called the target tree, since it is in the target language. Any subtree of this tree is called a target subtree.

A derivation string S is derived as an ordered sequence of elements, where each of the elements is either a source symbol or a target subtree.

The following is a formal definition of the derivation process. Given a derivation string S, a derivation step replaces the substring S of S with a target subtree T that has the following properties:

1. Any target subtree in S′ is also a subtree of T,

2. Any target subtree in S that is not in S′ does not share nodes with T, and

3. A derivation from a string S of source symbols to the target tree T is a sequence of derivation steps that produces T from S.

Consider the specific example of the alignment in FIG. 3a. FIG. 3B illustrates three different derivations of the target tree 300 from the source French string. The three derivations are labeled 201, 202 and 203. Each of these derivations are consistent with the definitions 1 through 3 above.

However, analysis of these derivations shows that at least one of the derivations is more “wrong” then the others. In the second derivation 202, for example, the word “pas” has been replaced by the English word “he”, which is incorrect.

Alignment allows the training system to distinguish between a good derivation and a bad derivation. Alignment between S and T can be carried out in order to improve the possible derivations. If S is a string of source symbols, and T is a target tree, then the definitions would lead to the conclusion that each element of S is replaced at exactly one step in the derivation and, and to each node of T is created at exactly one step in the derivation. Thus, for each element s of s₁ a set called replaced(s, D) is created at the step of the derivation D during which s is replaced. This set keeps track of where in the derivation, different parts are replaced.

At 201, the word “va” is replaced in the second step of the derivation.

Each of the different derivations includes a number of “steps”, each step, therefore, doing certain things. The derivation 201, for example, includes the steps 210, 211, 212, 213. In 201, for example, the French word “va” is replaced during the second step, 211, of the derivation. Thus, in notation form, files can be created which indicate the step at which the words are replaced. For example, here,

Replaced(s,D)=2

Analogously, each node t of T can have a file defined called created (T,D) to be the step of derivation D during which t is created. In 201, the nodes labeled by auxiliary and VP (verb phrase) are created during the third step 212 of the derivation. Thus, created (AUX, D)=3 and created(VP,D)=3.

Given a string S of source symbols and a target tree T, an alignment A with respect to S and T forms a relation between the leaves of T and the elements of S. If derivation D between S and T is selected, then the alignment induced by D is created by aligning an element s of S with a leaf node t of T, but if and only if the replaced(s, D) is equal to the created(T, D). In other words, a source word is “aligned” with the target word if the target word is created during the same step as that in which the source word is replaced.

FIG. 3C illustrates alignments. The tree 301 in FIG. 1 corresponds to the derivation 201 in FIG. 3B. Analogously, 302 corresponds to 202 and 303 corresponds to 203. A rule to analyze the derivations is described. The set of “good” derivations according to an alignment A is precisely that set of derivations that induce alignments A′, such that A is a sub alignment of A′. The term sub alignment as used herein requires that A⊂A′. Since alignments are simple mathematical relationships, this is relatively easy to determine. Another words, A is a sub alignment of A′ if A aligns 2 elements only if A′ also aligns those two elements. This is intuitively understandable from FIGS. 3B and 3C. The two derivations that seem correct at a glance include derivations 201 and 203. These are superalignments of the alignment given in FIG. 3A. The derivation 202 which is clearly wrong is not such a super alignment.

Notationally speaking, the derivation is admitted by an alignment A if it induces a super alignment of A. The set of derivations between source string S and target string T that are admitted by the alignment A can be denoted by

δ_A (S, T)

In essence, each derivation step can be reconsidered as a rule. This, by compiling the set of derivation steps used in any derivation of δA(S, T), the system can determine all relevant rules that can be extracted from (S, T, A). Each derivation step is converted into a usable rule according to this embodiment. That rule can be used for formation of automated training information.

Derivation step 212 in derivation 201 begins with a source symbol “ne”, which is followed by a target subtree that is rooted at VB and followed by another source symbol “pas”. These three elements of the derivation are replaced, by the derivation, with a target subtree rooted at VP that discards the source symbols and contains the started target subtree rooted at VB.

FIG. 4 illustrates how this replacement process can be captured by a rule. 401 shows the derivation step on the left, where the elements are replaced with other elements. 402 shows the induced rule that is formed. The input to the rule 402 include the roots of the elements in the derivation string that are being replaced. Here, the root of the symbol is defined as being the symbol itself. The output of the rule is a symbol tree. The tree may have some of its leaves labeled with variables rather than symbols from the target alphabet. The variables in the symbol tree correspond to the elements of the input to the rule. For example, the leaf labeled x2 in the induced tree means that when this rule is applied, x2 is replaced by the target subtree rooted at VB, since VB is the second element of the input. The two induced rules 403 and 404 are obtained from the respective derivations. Thus this rule format may be a generalization of CFG rules. Each derivation step can use this system to map to a rule in this way.

Hence, given a source string S, a target string T, and an alignment A, the set δA(S, T) can be defined as the set of rules in any derivation DεδA(S, T). This set of rules is the set of rules that can be inferred from the triple (S, T, A)

In an embodiment, the set of rules δA(S, T) can be learned from the triple (S, T, A) using a special alignment graph of the type shown in FIG. 5. The alignment graph is a graph that depicts the triple (S, T, A) as a rooted, directed, acyclic graph. FIG. 5 is shown with direction as being top-down, but it should be understood that this can alternatively very easily be turned upside down. In an embodiment, certain fragments of the alignment graph are converted into rules of δA(S, T). A fragment is defined herein as being a directed acyclic graph and G as a nontrivial subgraph G′ if a node A is in G′. Here, nontrivial means that the graph has more than just a single mode. The subgraph G′ is such that if the node n is in G′ then either n is a sink node of G′ (a node with no children) or all of n's children are in G′ and connected to all of the nodes thereof. FIG. 6 illustrates graph fragments formed from the alignment graph of FIG. 5.

The span of the node N of the alignment graph constitutes the subset of nodes from S that are reachable from n. A span is defined as being contiguous if it contains all the elements in a contiguous sub string of S. The closure of span (n) is the shortest contiguous span which is a superset of span (n) for example, the closure of (s2, s3, s5, s7) would be (s2, s3, s4, s5, s6, s7). The alignment graph of FIG. 5 is annotated with the span of each node. For example, each node, such as 500, has an annotation 502 that represents the span of that node.

One aspect is to determine the smallest set of information from these graphs that can form the set of rules. According to this aspect, first smaller parts of the rules are found, and then the rules are put together to form larger parts. The chunk can be defined in different ways—in an embodiment, certain fragments within the alignment graph are defined as being special fragments called frontier graph fragments. Frontier sets of the alignment graph include the set of nodes n in which each node n′ of the alignment graph, that is connected to n but is neither an ancestor nor a descendent of n, span(n′) ∩ closure(span(n))=0. The frontier set in FIG. 5 is shown in bold face and italics.

The frontier graph fragment of an alignment graph is the graph fragment where the root and all sinks are within the frontier set. Frontier graph fragments have the property that the spans of the sinks of the fragment are each contiguous. These spans form a partition of the span of the root, which is also contiguous. A transformation process between spans and roots can be carried out according to the following:

1) first, the sinks are placed in the order defined by the partition. The sink whose span is the first part of the span of the root goes first. This is followed by the Se whose span is the second part of the span of the root. This forms the input of the rule.

2) Next, the sink nodes of the fragment are replaced with a variable corresponding to their position in the input. Then, the tree part of the fragment is taken, for example by projecting the fragment on T. This forms the output of the rule.

FIG. 6 illustrates certain graph fragments, and the rules: both input and output, that are generated from those graph fragments. Rules constructed according to the conversion between the alignment graph and the rules are within a subset which is called ρ_A (S, T).

A number of rule extraction techniques are also described herein.

In a first embodiment, rules of ρ_A(S, T) are extracted from the alignment graph by searching the space of graph fragments for frontier graph fragments. One conceivable problem with this technique, however, is that the search space of all fragments of a graph becomes exponential to the size of the graph. Thus, this procedure can take a relatively long time to execute. The technique can be improved by taking the following simplifications.

The frontier set of an alignment graph can be identified in a time that is linear to the size of the graph. The second simplification is that for each node N of the frontier set, there is a unique minimal frontier graph fragment rooted at n. Because of the definition of the frontier set, any node n′ that is not in the frontier set can not have a frontier graph fragment rooted at n′. The definition of a minimal fragment requires that the frontier graph fragment is a subgraph of every other frontier graph fragment that has the Se route.

For an alignment graph that has k nodes, there are at most k minimal frontier graph fragments.

FIG. 7 shows the seven minimal frontier graph fragments from the alignment graph of FIG. 5. All of the other frontier graph fragments can be created by composing two or more minimal graph fragments. FIG. 8 illustrates how the other frontier graph fragments can be created in this way.

Thus, the entire set of frontier graph fragments, as well as all the rules derivable from those fragments, can be computed systemically according to the flowchart of FIG. 9. The flowchart of FIG. 9 can be run on the computer system of FIG. 10, for example. At 900, the set of minimal frontier graph fragments is computed for each training pair. More generally, any minimal set of information that can be used as a training set can be obtained at this operation.

At 910, the set of graph fragments resulting from composing the minimal graph fragments is computed. This allows the rules derived from the main minimal frontier graph fragments to be regarded as a basis for all of the rules that are derivable from the frontier graph fragments.

The rules are actually derived at 920. These rules have been derived from the minimal fragments. The rules include trees, or information derived from those trees.

At 930, the rules from the minimal fragments are combined to form “composed” rules.

Thus, the extracting of rules becomes a task of finding the set of minimal frontier graph fragments of any given alignment graph.

This is carried out by computing the frontier set of the alignment graph. For each node of the frontier set, the minimal frontier graph fragment rooted at the node is determined. The computing of the frontier set can be computed in a single pass through the alignment graph. The frontier set is computed as the union of each node with its span and also with its complement span, which is the union of the complement span of its parents and the span of all its siblings. Here, siblings are nodes that share the same parent.

A node n is in the frontier set if and only if its complement span (n) ∩ closure(span(n)) is equal to 0. Thus, the complement span nearly summarizes the spans of all nodes that are neither ancestors nor descendents of n. This step requires only a single traverse through the graph and thus runs in linear time.

The second step of computing the minimal frontier graph fragment rooted at the node is also relatively straightforward. For each node n of the frontier set, n is expanded. As long as there is some sink node n′ of the resulting graph fragment that is not in the frontier set, n′ needs to be expanded also. After computing the minimal graph fragment rooted at the node of the frontier set, every node of the alignment graph has thus been expanded at most once. Hence, this operation can also run in linear time.

The above has simplified certain aspects; for example, unaligned elements are ignored. However, processes to accommodate these unaligned elements can be determined. This system computes all derivations corresponding to all ways of accounting for unaligned words, and collects rules from all the derivations. Moreover, these techniques can include derivations where sub strings are replaced by sets of trees rather than by one single tree.

This corresponds to allowing rules that do not require the output to be a single rooted tree. This generalization may allow explaining linguistic phenomena such as immediately translating “va” into “does go”, instead of delaying the creation of the auxiliary word “does” until later in the derivation.

The above has been tested with a number of observations. The quality of alignment plays an important role in this derivation. Moreover, the technique which simplifies to running in linear time is barely affected by the size of the rules of abstracts, and produces good effects.

FIG. 11 identifies one cause of crossing between English and French which can be extended to other language pairs. Adverbs and French often appear after the verb, but this is less common in English. A machine parser creates a nested verb phrase when the adverbs are present. This prevents child reordering from allowing the verb and adverbs should be permeated. Multilevel reordering as shown in FIG. 11 can prevent or reduce these kinds of crossings.

One solution, initially suggested by Fox, may be to flatten the verb phrases. This constitutes a solution for this sentence pair. It may also account for adverb-verb reorderings. Flattening the tree structure is not necessarily a general solution since it can only apply to a very limited number of syntactic categories. Sometimes, however, flattening the tree structure does not resolve the crossing in the node reordering malls. In these models, a crossing remains between MD and AUX no matter how VPs are flattened.

The transformation rule model creates a lexical rule as shown in FIG. 12 as 1200. This lexical rule allows transformation of “will be” into -sera-, as the only way to resolve the crossing.

These techniques can also be used for decoding, as described herein. This embodiment describes automatic translation of source natural language sentences into target natural language sentences using complex probabilistic models of word to word, phrase to phrase, syntactic and semantic rule translation. This also describes probabilistic word, syntax and semantic language models.

This second embodiment forms trees directly from the string based information, here, the input information being the information to be translated. The translation is constructed by automatically deriving a number of target language parse trees from the source language sentence that is given as input. Each tree is scored by a weighted combination between the probabilistic models, as well as an additional set of language features. The tree of maximum probability provides the translation into the target language.

This embodiment defines a cross-lingual parsing framework that enables developing statistical translation systems that use any type of probabilistic channel or target language model: any of word based, phrase based, syntax based or semantic based.

The channel and target language models can be trained directly from a parallel corpus using traditional parameter estimation techniques such as the expectation maximization algorithm. The models can alternatively be estimated from word or phrase aligned corpora that have been aligned using models that have no knowledge of syntax. In addition, this enables exploring a much larger set of translation possibilities.

In this embodiment, a target language parse tree is created directly from the source language string. All channel operations are embodied as one of the different type of translation rules. Some of these operations are of a lexical nature, such as the word to word or phrase to phrase translation rules. Other rules are syntactic.

Table 1 illustrates rules that are automatically learned from the data.

TABLE 1
1. DT(these) →
2. VBP(include) →
3. VBP(includes) →
4. NNP(France) →
5. CC(and) →
6. NNP(Russia) →
7. IN(of) →
8. NP(NNS(astronauts)) →
9. PUNC(.)→ .
10. NP(x0:DT, CD(7), NNS(people) → x0, 7
11. VP(VBG(coming), PP(IN(from), x0:NP)) →  , x0
12. IN(from) →
13. NP(x0:NNP, x1:CC, x2:NNP) → x0, x1, x2
14. VP(x0:VBP, x1:NP) → x0, x1
15. S(x0:NP, x1:VP, x2:PUNC) → x0, x1, x2
16. NP(x0:NP, x1:VP) → x1,  , x0
17. NP(DT(“the”), x0:JJ, x1:NN) → x0, x1

These translation rules fall into a number of different categories.

Lexical simple rules are rules like numbers 1-7 that have one level syntactic constituents that dominate the target language part. These rules include a type of the word, the word itself, and the translation.

Lexical complex rules are rules like number 8, where there are multiple levels of syntactic constituents that dominate the target language part.

Rules 10, 11, 16 and 17 are lexically anchored complex rules. These rules explain how complex target syntactic structures should be constructed on top of mixed inputs. The mixed inputs can be lexical source language items and syntactic constituent target language constituents. For example, rule 16 says that if the Chinese particle occurs between two syntactic constituents x1 x0, then the resultant target parse tree is an NP with NP:x0 and X1:VP. In other words, this rule stores order information for the syntactic constituents between the languages.

The syntactic simple rules are rules like rule 13 which enable target syntactic structures to be derived. Finally, syntactic complex rules enable multiple level target syntactic structures to be derived. This technique can use cross lingual translation rules such as 11 and 16 that make reference to source language lexical items and target language syntactic components or constituents. Note that many of these rules include features that are actually tree based information written in string form. NP(DT (“the”), x0: . . . for example represents tree based information.

FIG. 13 illustrates a syntactic tree form derivation for the input sentence. A top down traversal of this derivation enables the creation of the target sentence because each node in the derivation explicitly encodes the order in which the children need traversal in the target language.

The decoding is carried out using clusters of decoding according to different levels. At a first step, each of the rules is applied first to the individual words within the phrase 1300. Note that existing software has already divided the new Chinese string 160 into its individual words. Each word such as 1302 is evaluated against the rules set to determine if any rule applies to that word alone. For example, the word 1302 has an explicit rule 1304 (rule 1) that applies to that single word. This forms a first level of rules shown as rule level 1; 1310.

At level 2, each pair of words is analyzed. For example, the pair 1302, 1312 is analyzed by rule 1314. Similarly, the pair 1312, 1316 is analyzed to determine if any rules apply to that pair. For example, the rule 1314 applies to any word that is followed by the word 1312. Accordingly, rule 1314 applies to the word pair 1302, 1312. These dual compound rules form level 2; 1320 analogously, triplets are analyzed in level 3, and this is followed by quadruplets and the like until the top level rule shown as level x is executed.

Each of these rules includes strings for string portions within the rule. For example, rule 13 shows the information of a specific tree which is written in text format. The tree portion may include variables within the tree.

When this is all completed, the English tree is output as the translation, based on the tree that has the highest score among all the trees which are found.

Although only a few embodiments have been disclosed in detail above, other embodiments are possible and the inventor (s) intend these to be encompassed within this specification. The specification describes specific examples to accomplish a more general goal that may be accomplished in other way. This disclosure is intended to be exemplary, and the claims are intended to cover any modification or alternative which might be predictable to a person having ordinary skill in the art. For example, different rules and derivation techniques can be used.

Also, the inventor(s) intend that 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.

Claims

1. A computer implemented method, comprising: executing, by a processor, instructions stored in memory to use information that is based on corpora of string-based training information to create a plurality of rules that are based on the training information; andperforming source language string to target language tree translation using an n-gram language model, a syntax-based language model, and the plurality of rules for an executable text to text application stored in memory, the plurality of rules including syntactic translation rules that are each associated with a probability for a translation, wherein a syntactic translation rule is determined by analyzing an alignment graph that includes a source string, a target tree, and an alignment of the source string and the target tree.

2. A computer implemented method as in claim 1, further comprising: obtaining a string to be translated;executing instructions stored in memory to compile sets of different possible translation trees using the rules; andexecuting instructions stored in memory to determine which of those translation trees represents probable translations.

3. A computer implemented method as in claim 2, further comprising executing instructions stored in memory to align a target word with a source word when the target word was created during a same process as that in which the source word is replaced.

4. A computer implemented method as in claim 2, wherein the compiling comprises: executing instructions stored in memory to find individual words;executing instructions stored in memory to find rules that apply to the individual words;executing instructions stored in memory to find combinations of the individual words; andexecuting instructions stored in memory to find rules that apply to the combinations of individual words.

5. A computer implemented method as in claim 2, wherein the compiling comprises compiling a complete set of derivation steps in any derivation of source string, target tree, and alignment.

6. A computer implemented method as in claim 5, wherein at least some elements in the trees are variables whose contents are defined by other trees.

7. A computer implemented method as in claim 6, wherein the alignment graph is used to determine an alignment by aligning source parts of a training corpora with target parts of the training corpora when the source part is created during the same step as that in which the source part is replaced.

8. A computer implemented method as in claim 7, wherein the alignment graph is analyzed to determine a smallest set of information that can form the set of rules.

9. A computer implemented method as in claim 8, wherein the smallest set of information includes frontier information.

10. A computer implemented method as in claim 1, further comprising executing instructions stored in memory to use parameter estimation techniques.

11. A computer implemented method as in claim 1, further comprising executing instructions stored in memory to flatten the trees to enable reordering of phrases.

12. A computer implemented method as in claim 1, further comprising: executing instructions stored in memory to form an alignment graph that represents a conversion between nodes of the source string, leaves of the target tree, and alignment; andexecuting instructions stored in memory to convert fragments of the alignment graph into rules.

13. A computer implemented method as in claim 12, wherein the fragments include substrings of the source string, and a span for the substring.

14. A computer implemented method as in claim 1, wherein the rules are formed by determining operations at which source symbols are replaced by target subtrees and by forming rules from the replacement process.

15. A computer implemented method as in claim 14, wherein an output of the rule is a symbol tree with at least some of its leaves labeled with variables rather than symbols from a target alphabet.

16. A computer implemented method as in claim 1, wherein using the information comprises extracting translation rules from word aligned pairs.

17. A computer implemented method as in claim 16, wherein the word aligned pairs comprise a tree in a first language and a string in a second language.

18. The method according to claim 1, further comprising resolving a syntactic crossing by: generating a frontier set for the alignment graph by determining minimal frontier fragments for the alignment graph;assembling together at least a portion of the minimal frontier fragments to form one or more frontier graph fragments; andgenerating a lexical rule by performing multi-level reordering of nodes within at least one frontier graph fragment to resolve the syntactic crossing.

19. A computer implemented method, comprising: executing, by a processor, instructions stored in memory to align items of information in first and second different languages to form aligned information, wherein at least the information in the first language is in a tree form; andexecuting, by a processor, instructions stored in memory to extract rules from the aligned information, the rules utilizable in conjunction with an n-gram model and a syntax based language model, the rules configured for use with performing source language string to target language tree translation using an n-gram language model, a syntax-based language model, and the plurality of rules for an executable text to text application stored in memory, the plurality of rules including syntactic translation rules that are each associated with a probability for a translation, wherein a syntactic translation rule is determined by analyzing an alignment graph that includes a source string, a target tree, and an alignment of the source string and the target tree.

20. A computer implemented method as in claim 19, wherein the information in both the first and second languages are in the tree form.

21. A computer implemented method as in claim 19, further comprising: executing instructions stored in memory to form tree based information into an alignment graph that aligns between a string in the first language and a tree in the second language; andexecuting instructions stored in memory to extract rules from the alignment graph.

22. A computer implemented method as in claim 21, further comprising executing instructions stored in memory to analyze a reduced set of fragments of the alignment graph prior to extracting the rules, wherein the rules are utilized to resolve crossings between a source string and a target tree.

23. A computer implemented method, comprising: obtaining a string in a source language to be translated into a target language; andexecuting, by a processor, instructions stored in memory to translate the string into the target language using at least one rule set, an n-gram language model, and a syntax based language model, wherein the at least one rule set comprises both rules that include at least parts of subtrees and probabilities, a rule set including translation rules in a subtree to substring rule form for a machine translation, the translation rules being associated with probabilities for the rules, wherein a translation rule is determined by analyzing an alignment graph that includes a source string, a target tree, and an alignment of the source string and the target tree.

24. A computer implemented method as in claim 23, wherein the translating comprises first applying rules to individual words, and then applying rules to combinations of words.

25. A computer implemented method as in claim 23, further comprising executing instructions stored in memory to output trees as the translation.

26. A computer implemented method as in claim 25, wherein the system produces a plurality of different trees as possible translations, and selects the best tree according to a highest probability.

27. A system comprising: a training part executable by a processor and stored in memory, the training part receiving a corpora of string-based training information to create a plurality of rules that are based on the training information, the rules including parts of trees as components of the rules; anda text to text application portion that uses an n-gram language model, a syntax-based language model, and the rules for a text to text application performing source language string to target language tree translation, the rules including translation rules in a subtree to substring rule form for a machine translation, the translation rules being associated with probabilities for the rules, wherein a translation rule is determined by analyzing an alignment graph that includes a source string, a target tree, and an alignment of the source string and the target tree.

28. A system as in claim 27, further comprising a memory that stores the rules including parts of trees that are translation rules in a subtree to substring rule form for a machine translation, and also stores probabilities for the rules.

29. A system as in claim 28, wherein the application portion obtains a string to be translated, compiles sets of different possible translation trees using the rules, and determines which of those translation trees represents probable translations.

30. A system as in claim 27, wherein the training part forms an alignment graph that represents a conversion between source, target, and alignment, and converts fragments of the alignment graph into rules.

31. A system as in claim 30, wherein the rules are formed by determining operations at which source symbols are replaced by target subtrees and forming rules from the replacement process.

32. A system as in claim 30, wherein the alignment graph is analyzed to determine a smallest set of information that can form the set of rules.

33. A system, comprising: a training module, executable by a processor and stored in a memory, that aligns items of information in first and second different languages to form aligned information and extracts rules from the aligned information,wherein at least the information in the first language is in a tree form, and the rules are utilizable in conjunction with an n-gram model and a syntax based language model, the tree form utilized in a source language string to target language tree translation, the rules including translation rules in a subtree to substring rule form for a machine translation, the translation rules being associated with probabilities for the rules, wherein a translation rule is determined by analyzing an alignment graph that includes a source string, a target tree, and an alignment of the source string and the target tree.

34. A system as in claim 33, wherein the information in both the first and second languages are in the tree form.

35. A system as in claim 33, wherein the training part forms tree based information into an alignment graph that aligns between the first language and the second language, and extracts rules from the alignment graph.

36. A system as in claim 33, wherein the training part forms a reduced set of fragments of the alignment graph prior to extracting the rules.

37. A system, comprising: a memory that stores at least one rule set that comprises both rules that include at least parts of subtrees and probabilities; anda decoding part that obtains a string in a source language to be translated into a target language, receives the at least one rule set, and uses the at least one rule set, an n-gram language model, and a syntax based language model to translate the string into the target language, the decoding part performing source language string to target language tree translation, a rule set including translation rules in a subtree to substring rule form for a machine translation, the translation rules being associated with probabilities for the rules, wherein a translation rule is determined by analyzing an alignment graph that includes a source string, a target tree, and an alignment of the source string and the target tree.

38. A system as in claim 37, wherein the decoding part first applies rules to individual words, and then applies rules to combinations of words.

39. A system as in claim 37, wherein the decoding part outputs trees as the translation.

40. A system as in claim 39, wherein the decoding part produces a plurality of different trees as possible translations, and selects the best tree according to a highest probability.