Detection of Common English Grammar Usage Errors

Abstract

Our research aims to provide writers with automated tools to detect grammatical usage errors and thus improve their writing. Correct English usage is often lacking in scientific and industry papers. [16] has compiled 130 common English usage errors. We address the automated detection of these errors, and their variations, that writers often make. Grammar checkers, e.g., [9] and [11], also implement error detection. Other researchers have employed machine learning and neural networks to detect errors. We parse only the part of speech (POS) tags using different levels of generality of POS syntax and word-sense semantics. Our results provide accurate error detection and are feasible for a wide range of errors. Our algorithm specifies precisely the ability to increase or decrease the generality in order to prevent a large number of false positives. We derive this observation as a result of using The Brown corpus, which consists of 55, 889 untagged sentences, covering most genres of English usage, both fiction and non-fiction. This corpus was much larger than any corpus employed by related researchers. We implemented 80 of Swan’s most common 130 rules; and detected 35 true positives distributed among 15 of Swan’s rules. Such a low true positive rate, 35/55889, had been expected. No false positives were detected. We employed a separate, smaller, test suite of both true positive and true negative examples. Our system, as expected, correctly detected errors in all the true positive examples, and ignored all the true negative ones. The Language-Tool system had a detection rate of 28/130 = 22%; Grammarly had a detection rate of 60/130 = 46%. Our results show significant improvement in the detection of common English usage errors.

Appendices

Appendix

A Swan Error 326_1.

B Our Own Related Neural Network

Lets first review the mechanics of our off-the-shelf neural network, [13], that we employed.

First, the network’s input and output nodes are chosen with values—these are concrete examples with and without errors. Second, the network is trained using these examples. Finally, the network is tested with similar (but not identical) examples. If the network does not converge during training, use a different architecture—hidden nodes, but retain with the original training examples.

Swan’s errors (634 of them) can be listed in any numerical order, and our subset (at the time) of 52 errors was mapped into binary digits. 000000—0—constitutes no error in the input string. Six binary digits were enough to cover 52 errors—11110 would be Swan’s error #30. Text strings were converted, internally, into decimal numbers; padding to the right occurred to make the sentences of equal length. These converted text strings constitute our training input. Our final training input and output is in Table 1. The last error example, 463.2 (Swan), involves the weather—the simple present tense should not be used for talking about a temporary condition.

The training algorithm [13] was back-propagation; we chose a slow training rate to minimize errors–0.01. While training, the network failed to converge. Because of this failure, we tried separate digits for each error (using significantly more nodes)—to no avail. Hidden layers were varied from three to eight layers; the number of nodes per layer varied from nine to 100. In other words, many combinations were tried—none of the architectures converged for our given (training) inputs and outputs. Convergence would decrease and then stop, and stay constant. Thus, no testing could be performed on the neural network.

However, testing the error detection for exactly two rules, and a non-error state did work as expected. We are not surprised—neural networks require many examples and counter examples—but each Swan rule usually has on average five examples, and at most ten examples which is not nearly enough for training purposes.

Table 1. Neural Network training inputs; 0=OK, outputs