System for Text Correction Adaptive to the Text Being Corrected

This invention relates to the correction of text, and more particularly to an adaptive system for text correction in which a training set is utilized that includes text which is assumed to be mostly correct.

In the past, especially with respect to text editing, user mistakes such as spelling errors were corrected utilizing conventional spell-checking systems utilizing lookup tables. Subsequently, more sophisticated spell-checking systems were developed in which the context in which the word occurred was taken into account. These systems traditionally involved the utilization of so-called training corpora which contain examples of the correct use of the words in the context of the sentences in which they occurred.

One of the major problems with such context-sensitive spelling-correction systems is the inability of these systems to take into account situations in which the corpus on which they were trained is dissimilar to the target text to which they are applied. This is an important problem for text correction because words can be used in a wide variety of contexts; thus there is no guarantee that the particular contextual uses of the words seen in the target text will also have been seen in the training corpus.

Consider, for example, an algorithm whose job is to correct context-sensitive spelling errors; these are spelling errors that happen to result in a valid word of English, but not the word that was intended--for example, typing "to" for "too", "casual" for "causal", "desert" for "dessert", and so on. It is very difficult to write an algorithm to do this by hand. For instance, suppose we want to write an algorithm to correct confusions between "desert" and "dessert". We could write rules such as: "If the user types `desert` or `dessert`, and the previous word is `for`, then the user probably meant `dessert`". This rule would allow the algorithm to fix the error in: "I would like the chocolate cake for desert". The rule would not work, however, for many other cases in which "desert" and "dessert" were confused, for instance: "He wandered aimlessly through the dessert", where "desert" was probably intended. To fix this particular sentence, a different rule is needed, such as: "If the user types `desert` or `dessert`, and the previous two words are a preposition followed by the word `the`, then the user probably meant `desert`". In general, it is extremely difficult to write a set of rules by hand that will cover all cases.

This difficulty of writing a set of rules by hand is the motivation for moving to adaptive algorithms--algorithms that learn to correct mistakes by being trained on examples. Instead of writing rules by hand, it is much easier to provide a set of examples of sentences that use "desert" and "dessert" correctly, and let the algorithm automatically infer the rules behind the examples.

A wide variety of techniques have been presented in the Machine Learning literature for training algorithms from examples. However, what they all have in common is that they make the assumption of representativeness; that is, they assume that the set of examples that the algorithm is trained on is representative of the set of examples that the algorithm is asked to correct later. Put another way, they assume that the examples in the training and test sets are drawn, in an unbiased way, from the same population. It follows that whatever rules the algorithm learns from the training set will apply correctly to the test set. For example, if the training set contains examples illustrating that the word "for" occurs commonly before "dessert", but rarely before "desert", then, by the assumption of representativeness, the same distributional property of "for" should hold in the test set. If this assumption is violated, the algorithm's performance on the test set will degrade, because the rules it learned from the training set will not necessarily carry over to the test set. Existing machine learning techniques are therefore effective only to the extent that the training set is representative of the test set. This is a serious limitation, since, in general, there is no way to guarantee representativeness.

In order to provide for a more flexible method of correcting words in a text, in the subject system not only is a conventional training corpus used, but also the target text is analyzed to ascertain how the words to be corrected are used elsewhere in the text. Assuming that, on the whole, the words are used properly throughout the text, then when checking a particular occurrence of one of the words, the suggestion is made that the word take on the form of its usage in similar contexts that appear throughout the text.

The subject system therefore provides a training procedure called sup/unsup which utilizes both a conventional training corpus and the target text to overcome this limitation of existing techniques, thereby enabling adaptive correction algorithms to work well even when the training set is only partially representative of the test set.

The sup/unsup procedure works simply by training the algorithm, whichever adaptive correction algorithm is of interest, on both the traditional training set and the test set. It may seem, at first glance, to be counterproductive to train on the test set, as this set will presumably contain errors. However, the sup/unsup procedure is based on the assumption that although the test set contains errors, they will tend to be distributed sporadically through the test set; thus the learning procedure should still be able to extract correct rules, despite the presence of this noise. An example will clarify this concept.

Returning to the problem of correcting context-sensitive spelling errors for the case of "desert" and "dessert", suppose the test set is a news article containing 24 occurrences of the phrase "Operation Desert Storm" and 1 occurrence of the incorrect phrase "Operation Dessert Storm". Following the sup/unsup procedure, the algorithm is trained on its usual training set plus this test set. The objective of training is to extract rules about when to use "desert" versus "dessert". The test set suggests the rule: "If the user types `Desert` or `Dessert`, and the previous word is `Operation`, and the following word is `Storm`, then the user probably meant `Desert`". This rule will work for 24 out of the 25 examples to which it applies. Although it is not perfect, it is reliable enough that the system can safely learn this rule. It can then apply the rule to the very same test set from which it was learned to detect the 1 occurrence of "Operation Dessert Storm" that violates the rule.

The strength of this procedure is that it can detect sporadic errors, such as the single incorrect spelling "Operation Dessert Storm" in a test document, even if there are no relevant occurrences in the training set--that is, even if the training set is unrepresentative of the test set. One way to look at this procedure is that it is checking for inconsistencies in the test set, rather than "errors" with respect to some training set that is deemed error-free. The procedure is, of course, still able to detect errors that are illustrated by the original training set. The weakness of the procedure is that it cannot detect systematic errors in the test set. For example, if the user types "Operation Dessert Storm" every time, the system will be unable to find the error.

It should be borne in mind that the effectiveness of the sup/unsup procedure depends on two factors. The first is the size of the test set; the larger the test set, the easier it will be to detect inconsistencies. For instance, in the example above, there were 24 correct occurrences of "Operation Desert Storm"; if instead the test set were much smaller and had only 2 occurrences, the algorithm might not have enough information to learn the rule about "Operation" and "Storm" implying "Desert". The second factor affecting the effectiveness of sup/unsup is the percentage of mistakes in the test set. In the example above, the user made 1 mistake in 25 occurrences of "Operation Desert Storm"; if instead the level were 10 or 15 out of 25, it would become difficult for the algorithm to learn the appropriate rule.

It will be appreciated that the sup/unsup training procedure applies to any adaptive correction algorithm, regardless of the means used for adaptation, and regardless of the correction task under consideration.

The term "adaptive correction algorithm" as used herein refers to algorithms that correct users' mistakes, e.g., context-sensitive spelling errors, and that learn to do their job of correcting mistakes by being trained on examples that illustrate correct answers and/or mistakes.

More particularly, in one embodiment, the specific algorithm utilized for target text analysis involves scanning the full collection of training texts, which in this embodiment includes both a conventional training corpus, and the target text to which the system is being applied, so as to ascertain the features that characterize the context in which each word that is being corrected may appear. By features is meant two types of text patterns. The first type is called context words, and refers to the presence of a particular keyword within some fixed distance of the target word that is being corrected. For instance, if the words being corrected are "desert" and "dessert", then useful context words might include "hot", "dry", and "sand" on the one haled, and "chocolate", "cake", and "sweet" on the other hand. The presence of words in the former group within, for instance, 10 words on either side of the target word tends to indicate that "desert" was intended as the target word; whereas the presence of words in the latter group tends to imply that "dessert" was intended.

The second type of feature is called collocations, and refers to the pattern of part-of-speech tags and specific words in the immediate context of the target word. For instance, if the words being corrected are again "desert" and "dessert", then one useful collocation might be "`PREPOSITION the` occurs immediately to the left of the target word". This collocation matches any sentence in which the target word, which is either "desert" or "dessert", is directly preceded by the word "the", which in turn is directly preceded by a word that has been tagged as "PREPOSITION". For instance, the collocation would match the sentence "He went to the desert", in which the target word, "desert", is immediately preceded by the word "the", which is immediately preceded by the word "to", which is tagged as a preposition. This collocation, when matching a sentence, tends to imply that "desert" was intended as the target word, and not "dessert". In one embodiment, the part-of-speech tags needed for this analysis are derived by a lookup procedure that utilizes a dictionary which lists, for any given word, its set of possible part-of-speech tags. A collocation is considered to match a sentence if each specific word in the collocation matches the corresponding word in the sentence, and if each part-of-speech tag in the collocation is a member of the set of possible part-of-speech tags of the corresponding word in the sentence.

The subject system derives a set of features of the two types described above by scanning through the training texts for all occurrences of the words being corrected. For each such occurrence, it proposes as candidate features all context words and collocations that match that occurrence. After working through the whole set of training texts, it collects and returns the set of features proposed. In one embodiment, pruning criteria are applied to this set of features to eliminate features that are based on insufficient data, or that are ineffective at discriminating among the words being corrected.

Having derived a set of features that characterize the contexts in which each of the words being corrected tends to occur, then by a conventional Bayesian method, these features are used as evidence to ascertain the probability that each of the words being corrected is in fact the word that the user intended to type. Having derived these probabilities, the word selected as correct is that word having the highest probability. For instance, if the word "desert" is used 100 times in a particular context, and if the word "dessert" is used either not at all or a limited number of times in the same context, then the subject system will select "desert" as the correct word in future occurrences of this context, even if the user typed "dessert". Bayesian analysis for context-sensitive spelling correction is described in the paper, "A Bayesian hybrid method for context-sensitive spelling correction", by Andrew R. Golding, in the Proceedings of the Third Workshop on Very Large Corpora, at the June 1995 conference of the Association for Computational Linguistics, pages 39-53.

Thus both a conventional training corpus and a special training corpus involving the target text are utilized to ascertain the correct spelling of a target word in the text. The conventional training corpus is merged with the target-text corpus so that context-sensitive spelling correction is based on both corpora. The result is that improved spelling correction is achieved through the analysis of the use of the word throughout the target text, as well as in the conventional training corpus.

In summary, a system is provided for correcting users' mistakes including context-sensitive spelling errors and the like in which an adaptive correction algorithm is utilized which is trained on not only a conventional training corpus, but also on the text which is being corrected, thus to permit the correction of words based on the particular usages of the words in the text being corrected, taking advantage of the fact that the text to be corrected is by and large already mostly correct.

These and other features of the Subject Invention will be better understood taken in conjunction with the Detailed Description taken in conjunction with the Drawings, of which:

FIG. 1 is a flowchart indicating training and run-time processing to accomplish word correction in a target text in which a combined corpus of the target text and a conventional training corpus are utilized followed by a probability-based spelling-correction system utilizing the combined corpus;

FIG. 2 is a flowchart of the feature-extraction system of FIG. 1 illustrating the utilization of a dictionary of sets of part-of-speech tags utilized in the proposal of all possible context words and collocations as candidate features, followed by the counting of occurrences of all candidate features in the combined corpus, in turn followed by pruning features that have insufficient data or are uninformative discriminators; and

FIG. 3 is a flowchart of the probability-based spelling-correction system of FIG. 1 illustrating a sequence of finding successive occurrences of the target word in the target text, and for each such occurrence, finding in turn all features that match the occurrence, and combining the evidence from these features using Bayes' rule so as to select the word that has the highest probability of being the word that was intended for the target word.

Referring now to FIG. 1, a system 10 for correcting text includes a training phase 12, followed by a run-time processing phase 14. In the training phase, target text 16 is combined with a conventional training corpus 18 to provide a combined corpus 20 in which the corpus utilized includes the target text and therefore is exceedingly useful in tailoring the text-correction system to the text in question.

The combined corpus is utilized in a feature-extraction step 22 in which selected features are culled from the combined corpus as illustrated at 24. These features include context words and collocations, so that it is these features which provide evidence that allows the computation of the probability of each word that could have been intended for the target word in the target text.

During run-time processing, a probability-based spelling-correction system 30 is provided with the target text and determines first a target word and secondly the probability of the correctness of this word based on the combined corpus and the selected features. For instance, the list of selected features is matched against each occurrence of the target word in the target text, so as to collect evidence about the likely intended identity of the target word. The evidence is combined into a single probability for each word that could have been intended for the target word using Bayes' rule, in one embodiment. Other spelling-correction systems for use in determining the identity of the correct word include a system based on the Winnow algorithm which employs a multiplicative weight-updating scheme as well as a variant of weighted-majority voting. This technique is described in a paper entitled, "Applying Window to Context-Sensitive Spelling Correction" by Andrew R. Golding and Dan Roth, in Machine Learning: The Proceedings of the 13th International Conference, Lorenza Saitta, ed., Morgan Kaufmann, San Francisco, Calif., 1996.

It can be seen that the spelling correction in step 30 is provided, in one embodiment, by a conventional Bayesian method which, rather than using a conventional training corpus alone, utilizes a corpus which includes the target text. The advantage of so doing is that the system for finding the probability of the correct word is enhanced by inspecting the target text for similar occurrences of the word so that a more powerful technique for obtaining the correct word is achieved.

The result of the probability-based spelling correction is the suggestion of a word to be inserted in the text in place of the target word if the target word needs changing yielding a corrected text 32 as illustrated.

Referring now to FIG. 2, feature extraction 22 is detailed such that in one embodiment, the output of the combined corpus 20 is coupled to a module 34 which lists all possible context words and collocations as candidate features, utilizing a dictionary 36 of sets of part-of-speech tags. For instance, for the word "walks", the dictionary would give the set of part-of-speech tags consisting of "PLURAL NOUN" and "THIRD PERSON SINGULAR VERB". The process of listing all possible context words and collocations is illustrated by the sentence, "John lives in the desert", in which the target word is "desert", for which the user could intend either "desert" or "dessert". In this sentence, the set of possible part-of-speech tags for "in" consists of the single tag "PREPOSITION", while that for the word "the" consists of the single tag "DETERMINER". In this case, four context words and four collocations are proposed as candidate features. The context words are the words "John", "lives", "in", and "the", each of which is a word that occurs nearby the target word. The collocations are: "`in the` occurs immediately to the left of the target word"; "`PREPOSITION the` occurs immediately to the left of the target word"; "`in DETERMINER` occurs immediately to the left of the target word"; and "`PREPOSITION DETERMINER` occurs immediately to the left of the target word". These four collocations represent all ways of expressing the nearby context of the target word in terms of specific words and part-of-speech tags.

Having provided a list of all possible context words and collocations as candidate features, as seen at 36, a module counts the occurrences of all candidate features in the combined corpus 20, followed by a pruning step at 38 to prune features that have insufficient data or are uninformative discriminators. By insufficient data is meant the number of occurrences of the feature in the training corpus is below a prespecified threshold, which, in one embodiment, is set to 10. By uninformative discriminator is meant that the presence of the feature fails to be significantly correlated with the identity of the target word, as determined by a chi-square statistical test, which, in one embodiment, is set to the 5% level of significance. The result is a list of features 40 to be utilized in the run-time processing for probability-based spelling correction.

Referring now to FIG. 3, the system 30 for performing probability-based spelling correction from target text 16 and features 40 includes, as a first step, finding the next occurrence of a given target word as illustrated at 42. Upon a determination at 44 of the occurrence of the target word in the target text, module 46 initializes the probability of each word, w.sub.i, that could have been intended for the target word. In one embodiment, the probability of each word, w.sub.i, is set to the ratio of the probability that the word occurred in the combined corpus to the total probability that any of the words w.sub.i occurred in the corpus.

Having initialized the probability of each word, w.sub.i, module 48 finds the next feature that matches the occurrence of the target word. The inputs to this module are features 40 and the dictionary of sets of part-of-speech tags 36. What is happening is that having found the occurrences of the target word, the system now finds which features in its list of possible features match each occurrence of the target word. The utilization of features enhances the robustness of the system in that not only must the probability be based on the target word occurrence, it must also be based on the presence of features that match that target word occurrence.

Assuming that the next feature that matches the occurrence is found, then module 58 updates the probability of each word, w.sub.i, that could have been intended for the target word, using Bayes' rule. The probability update is performed to adjust the probability for each word, w.sub.i, so as to take into account the evidence about the likely identity of the target word that is provided by the feature that matched the target word occurrence. As a result, when all features that match the target occurrence have been processed, a probability will have been calculated, based on the context of the target word, that measures the probability that each word w.sub.i is the word that was intended for the target word given the context in which it occurred. This is done by gathering the evidence from the features that matched the occurrence of the target word.

In the illustrated embodiment, Bayes' rule is utilized to update the probability of each word, w.sub.i, based on a feature match with the target text. Bayes' rule is given by the following formula: ##EQU1## where F is the set of features that matched the occurrence of the target word; P(w.sub.i .vertline.) is the probability that word w.sub.i was intended, given that the set F of features has been found to match the target occurrence; P(w.sub.i) is the so-called prior probability of word w.sub.i, which is the value to which the probability for word w.sub.i was initialized in 46; P() is a scaling factor that need not be used as it does not affect the final outcome of the computation; and each term P(.function..vertline.w.sub.i) is a so-called likelihood term that is used to update, at 58, the probabilities of the words w.sub.i given that feature .function. has been found to match the target occurrence, and is calculated from the number of occurrences of feature .function. in the combined corpus.

If the result of the test at 50 is that no feature that matches the target occurrence was found, then, as illustrated at 52, the correct word selected is that word, w.sub.i, with the highest probability. It will be appreciated that when a word is selected as being correct by module 52, this word is entered into a text buffer 56 into which has previously been loaded the original target text. Following the selection by module 52 of the word, w.sub.i, with the highest probability, the system iteratively cycles back to find the next occurrence of a target word in the target text, with the iteration continuing until all such occurrences have been processed. At this time, when no more occurrences are found, the text buffer 56 will contain the original target text, as modified by all spelling corrections made during the spelling-correction procedure. Corrected text 54 is therefore generated through the readout of the text buffer 56 as illustrated at 60 at the end of the sequence.

In summary, a combined corpus is formed in the training phase in which the target text is an integral part of the combined corpus and specially tailors the corpus to the text under consideration. Having formed the combined corpus, selected features are extracted to enable a more robust calculation of the probability that a given word is the correct word in the context of the target occurrence. By considering those words that could have been intended for the target word, as well as features, one can provide more accurate probabilities in a run-time sequence for text correction. Utilization of context words and collocations as features is by virtue of example only, as other features are within the scope of this invention.

Moreover, the method for selecting probabilities, while being described in terms of a Bayes' rule iterative process, is but one of a number of techniques for determining the probability of a word as being correct in a particular context.

The C code for the subject system follows, in two parts: feature extraction, and spelling correction. Code for feature extraction: ##SPC1##

Having above indicated several embodiments of the Subject Invention, it will occur to those skilled in the art that modifications and alternatives can be practiced within the spirit of the invention. It is accordingly intended to define the scope of the invention only as indicated in the following claims: