Clear and efficient communication requires more than grammatical correctness to ensure fluency and semantic correctness, especially for non-native speakers. Thus, we propose a new task – Sentence Denoising, to go beyond Grammatical Error Correction (GEC). We define a rich and linguistics-inspired noise taxonomy consisting of 13 types of noise, and categorize them into vagueness, redundancy, and incoherence. We then generate and study 4 types of noise out of the 13 because they serve as building blocks. Methods are proposed to inject targeted noise into sentences for building datasets. We publish them and give benchmarks for denoising both individual noise and compound noise. Finally, an efficient training approach is designed for denoising combinations of noise.

In this paper we present our results from mining text to identify syntactic patterns to help discriminate between creative metaphorical expressions and non-metaphorical expressions, and we also present an application of our findings to the generation of novel metaphors. We trained an unsupervised LSTM model and use it in an inference engine to generate novel metaphors, where novelty is ensured in multiple ways. First, we use a weighted random choice with a "constraining factor" to select each word in our metaphor generation. Next, the inference engine checks for originality by ensuring that none of the generated sentence fragments match original fragments from the training data. Finally, the inference engine provides assurance that a metaphorical expression was generated, by checking against the identified syntactic patterns of metaphors that did not show up in the non-metaphorical language. We found that there are 611 repeated sentence patterns for metaphorical expressions that never appeared as the sentence pattern of non-metaphorical expressions. Furthermore, in a test set of 360 automatically generated metaphors, we observed 123 different syntactic patterns. This is a great deal more than the number of patterns used in the current state-of-the-art metaphor generators that rely on templates to generate the metaphors.

Misunderstanding happens all the time, especially when a non-native speaker is involved. To help recover the original meaning, we define categories of noise in an English sentence and differentiate our problem from Grammatical Error Correction (GEC). Methods are proposed to inject targeted noise into sentences for building training sets. Finally, a system comprised of two parts is designed for sentence correction (aka denoising) - One being fine-tuned BERT models for noise classification, and the other being Transformer models for "translating" noisy sentences into correct ones.

This paper addresses the effect of data reduction on speeding up training while keeping or improving the accuracy performance of classification. Since many studies have focused on feature selection, but did not adequately consider instance selection, our work focuses on both instance reduction and feature reduction, integrated into a whole reduction method. We examined in prior work Simple Random Sample Selection without Replacement, integrated with the Information Gain-based Feature Selection method, and compared its performance with the unintegrated instance selection and feature selection alone, using a single reduction rate. Our results proved that the integration of instance and feature selection performed much better than instance or feature selection alone. In this paper, we examine our approach in more depth, trying different reduction rates and different distributions of reduction rates between instance reduction and feature reduction. Our results show that for nearly all common classifiers, our integrated data reduction speeds up training significantly while keeping the accuracy unchanged (and sometimes even improved) at even high reduction rates. We also present the optimal feature-instance reduction-rates tradeoff.

The dramatic increase in dataset volumes available to train learning models has led to great advances in machine learning, but at the cost of slowing down training. This paper addresses the effect of data reduction on speeding up training while keeping or improving the accuracy performance of classification. Since many studies have focused on feature selection, but did not adequately consider instance selection, our work focuses on both instance reduction and feature reduction, integrated into a holistic reduction approach. We examined in prior work Simple Random Sample Selection without Replacement, integrated with the Information Gain-based Feature Selection method, and compared its performance with the unintegrated instance selection and feature selection individually, applied at various reduction rates. Our results proved that the integration of instance and feature selection performed much better than instance or feature selection alone, in terms of both training speedup and even accuracy improvement. In this paper, a novel transpose-based instance selection approach is introduced and integrated with feature selection, and its performance is investigated on classification and compared with the top-performing results of our prior work. Our results show that our new integrated method leads to significant increase in training speedup without impacting classification accuracy, and in fact, for some classifiers like Naive Bayes, the accuracy goes up considerably.

Restaurant servers are an example of an insider threat to the security of restaurant financial data. This paper applies machine learning to detect the digital representation of malevolent behavior of restaurant employees. The results of this research could be used to notify restaurant owners in real time when fraud is being committed. This paper applies machine learning (ML) techniques including neural networks, support vector machines, Random Forest, and Adaboost, to detecting insider fraud in restaurant point-of-sales data. By applying undersampling and oversampling class balancing techniques we show that ML techniques can improve fraud detection performance. In particular, detection with a Random Forest model using cross validation can be increased 55% by oversampling the minority class to the same size as the majority class. And results with a Neural Net model trained to detect fraud on the first year the restaurant opened, and tested on data from the following year can be improved by 50% by decreasing the majority class to be the same size as the minority class.

Mathematical formulae are essential in science, but face challenges of ambiguity, due to the use of a small number of identifiers to represent an immense number of concepts. Corresponding to word sense disambiguation in Natural Language Processing, we disambiguate mathematical identifiers. By regarding formulae and natural text as one monolithic information source, we are able to extract the semantics of identifiers in a process we term Mathematical Language Processing (MLP). As scientific communities tend to establish standard (identifier) notations, we use the document domain to infer the actual meaning of an identifier. Therefore, we adapt the software development concept of namespaces to mathematical notation. Thus, we learn namespace definitions by clustering the MLP results and mapping those clusters to subject classification schemata. In addition, this gives fundamental insights into the usage of mathematical notations in science, technology, engineering and mathematics. Our gold standard based evaluation shows that MLP extracts relevant identifier-definitions. Moreover, we discover that identifier namespaces improve the performance of automated identifier-definition extraction, and elevate it to a level that cannot be achieved within the document context alone.

Mathematical Information Retrieval concerns retrieving information related to a particular mathematical concept. The NTCIR-11 Math Task develops an evaluation test collection for document sections retrieval of scientific articles based on human generated topics. Those topics involve a combination of formula patterns and keywords. In addition, the optional Wikipedia Task provides a test collection for retrieval of individual mathematical formula from Wikipedia based on search topics that contain exactly one formula pattern. We developed a framework for automatic query generation and immediate evaluation. This paper discusses our dataset preparation, topic generation and evaluation methods, and summarizes the results of the participants, with a special focus on the Wikipedia Task.

Summary form only given. The goal is to gain a better understanding of the behavior of the image down/upsampling combinations, and find better down/upsampling methods. We examined existing down/upsampling methods and proposed new ones. We formulated a frequency response approach for understanding and evaluating down/upsampling combinations. The approach was validated experimentally by running the methods on various images and computing the signal to noise ratio (SNR) between the original and the down-then-upsampled images. The frequency response based evaluation correlates well with the experimental evaluation. Down/upsampling combinations were studied in a unified framework. Signals are pre-filtered then decimated by two, resulting in downsampling by two. Afterwards, signals are zero-upsampled by 2, i.e., inserting 0s between successive samples, and then post-filtering. Our analysis showed that for optimal performance, the pre-filter and the post-filter should both be low-pass filters with cutoff at /spl pi//2. We considered five classes of filters. The first corresponds to the simplest down/upsampling combination, decimation/duplication, where decimation is simply the skipping of every other row and every other column, and duplication (for upsampling) involves duplicating every row and every column. The second class corresponds to bilinear interpolation, for both upsampling and downsampling. The third class comprises the biorthogonal and orthogonal wavelets. The fourth class we termed binomial filters. The fifth class consists of least-square FIR filters.

Summary form only given. DPCM decoding is essentially the computation of a 2-indexed scalar recurrence relation; the two indices are: the row and column positions of the pixels. Although several logarithmic-time parallel algorithms for solving 1-indexed recurrence relations have been designed, no work has been reported on multi-indexed recurrence relations. Considering the importance of fast DPCM decoding of imagery, parallel algorithms for solving multi-indexed recurrence relations merit serious study. We designed novel parallel algorithms for solving 2-indexed recurrence relations, and identified the parallel architectures best suited for them. We developed three approaches: index sequencing, index decoupling, and dimension shifting. To solve a 2-indexed relation in DPCM decoding of an n/spl times/n image, index sequencing breaks down the relation into a sequence of n 1-indexed scalar recurrence relations that must be solved one after another. Each relation is then solved by a parallel O(nlogn) time algorithm on an n-processor hypercube or partitionable bus. Thus, the n equations take O(nlogn) time on n processors. Index decoupling, applicable in a common case of DPCM, breaks the 2-indexed relation into n independent 1-indexed recurrence relations, which are then solved simultaneously in O(logn) parallel time, using n/sup 2/ processors configured as a hypercube or a mesh of partitionable buses.