A desirable property of interpretable models is small size, so that they are easily understandable by humans. This leads to the following challenges: (a) small sizes typically imply diminished accuracy, and (b) bespoke levers provided by model families to restrict size, e.g., L1 regularization, might be insufficient to reach the desired size-accuracy trade-off. We address these challenges here. Earlier work has shown that learning the training distribution creates accurate small models. Our contribution is a new technique that exploits this idea. The training distribution is encoded as a Dirichlet Process to allow for a flexible number of modes that is learnable from the data. Its parameters are learned using Bayesian Optimization; a design choice that makes the technique applicable to non-differentiable loss functions. To avoid the challenges with high dimensionality, the data is first projected down to one-dimension using uncertainty scores of a separate probabilistic model, that we refer to as the uncertainty oracle. We show that this technique addresses the above challenges: (a) it arrests the reduction in accuracy that comes from shrinking a model (in some cases we observe $\sim 100\%$ improvement over baselines), and also, (b) that this maybe applied with no change across model families with different notions of size; results are shown for Decision Trees, Linear Probability models and Gradient Boosted Models. Additionally, we show that (1) it is more accurate than its predecessor, (2) requires only one hyperparameter to be set in practice, (3) accommodates a multi-variate notion of model size, e.g., both maximum depth of a tree and number of trees in Gradient Boosted Models, and (4) works across different feature spaces between the uncertainty oracle and the interpretable model, e.g., a GRU might act as an oracle for a decision tree that ingests n-grams.
CNNs, RNNs, GCNs, and CapsNets have shown significant insights in representation learning and are widely used in various text mining tasks such as large-scale multi-label text classification. However, most existing deep models for multi-label text classification consider either the non-consecutive and long-distance semantics or the sequential semantics, but how to consider them both coherently is less studied. In addition, most existing methods treat output labels as independent methods, but ignore the hierarchical relations among them, leading to useful semantic information loss. In this paper, we propose a novel hierarchical taxonomy-aware and attentional graph capsule recurrent CNNs framework for large-scale multi-label text classification. Specifically, we first propose to model each document as a word order preserved graph-of-words and normalize it as a corresponding words-matrix representation which preserves both the non-consecutive, long-distance and local sequential semantics. Then the words-matrix is input to the proposed attentional graph capsule recurrent CNNs for more effectively learning the semantic features. To leverage the hierarchical relations among the class labels, we propose a hierarchical taxonomy embedding method to learn their representations, and define a novel weighted margin loss by incorporating the label representation similarity. Extensive evaluations on three datasets show that our model significantly improves the performance of large-scale multi-label text classification by comparing with state-of-the-art approaches.
Active learning is a popular methodology in text classification - known in the legal domain as "predictive coding" or "Technology Assisted Review" or "TAR" - due to its potential to minimize the required review effort to build effective classifiers. In this study, we use extensive experimentation to examine the impact of popular seed set selection strategies in active learning, within a predictive coding exercise, and evaluate different active learning strategies against well-researched continuous active learning strategies for the purpose of determining efficient training methods for classifying large populations quickly and precisely. We study how random sampling, keyword models and clustering based seed set selection strategies combined together with top-ranked, uncertain, random, recall inspired, and hybrid active learning document selection strategies affect the performance of active learning for predictive coding. We use the percentage of documents requiring review to reach 75% recall as the "benchmark" metric to evaluate and compare our approaches. In most cases we find that seed set selection methods have a minor impact, though they do show significant impact in lower richness data sets or when choosing a top-ranked active learning selection strategy. Our results also show that active learning selection strategies implementing uncertainty, random, or 75% recall selection strategies has the potential to reach the optimum active learning round much earlier than the popular continuous active learning approach (top-ranked selection). The results of our research shed light on the impact of active learning seed set selection strategies and also the effectiveness of the selection strategies for the following learning rounds. Legal practitioners can use the results of this study to enhance the efficiency, precision, and simplicity of their predictive coding process.
We propose a new approach to address the text classification problems when learning with partial labels is beneficial. Instead of offering each training sample a set of candidate labels, we assign negative-oriented labels to the ambiguous training examples if they are unlikely fall into certain classes. We construct our new maximum likelihood estimators with self-correction property, and prove that under some conditions, our estimators converge faster. Also we discuss the advantages of applying one of our estimator to a fully supervised learning problem. The proposed method has potential applicability in many areas, such as crowdsourcing, natural language processing and medical image analysis.
Attention mechanisms have recently boosted performance on a range of NLP tasks. Because attention layers explicitly weight input components' representations, it is also often assumed that attention can be used to identify information that models found important (e.g., specific contextualized word tokens). We test whether that assumption holds by manipulating attention weights in already-trained text classification models and analyzing the resulting differences in their predictions. While we observe some ways in which higher attention weights correlate with greater impact on model predictions, we also find many ways in which this does not hold, i.e., where gradient-based rankings of attention weights better predict their effects than their magnitudes. We conclude that while attention noisily predicts input components' overall importance to a model, it is by no means a fail-safe indicator.
Deep neural networks have achieved great success in classification tasks during the last years. However, one major problem to the path towards artificial intelligence is the inability of neural networks to accurately detect samples from novel class distributions and therefore, most of the existent classification algorithms assume that all classes are known prior to the training stage. In this work, we propose a methodology for training a neural network that allows it to efficiently detect out-of-distribution (OOD) examples without compromising much of its classification accuracy on the test examples from known classes. We propose a novel loss function that gives rise to a novel method, Outlier Exposure with Confidence Control (OECC), which achieves superior results in OOD detection with OE both on image and text classification tasks without requiring access to OOD samples. Additionally, we experimentally show that the combination of OECC with state-of-the-art post-training OOD detection methods, like the Mahalanobis Detector (MD) and the Gramian Matrices (GM) methods, further improves their performance in the OOD detection task, demonstrating the potential of combining training and post-training methods for OOD detection.
Script knowledge consists of detailed information on everyday activities. Such information is often taken for granted in text and needs to be inferred by readers. Therefore, script knowledge is a central component to language comprehension. Previous work on representing scripts is mostly based on extensive manual work or limited to scenarios that can be found with sufficient redundancy in large corpora. We introduce the task of scenario detection, in which we identify references to scripts. In this task, we address a wide range of different scripts (200 scenarios) and we attempt to identify all references to them in a collection of narrative texts. We present a first benchmark data set and a baseline model that tackles scenario detection using techniques from topic segmentation and text classification.
In this work, we intrinsically and extrinsically evaluate and compare existing word embedding models for the Armenian language. Alongside, new embeddings are presented, trained using GloVe, fastText, CBOW, SkipGram algorithms. We adapt and use the word analogy task in intrinsic evaluation of embeddings. For extrinsic evaluation, two tasks are employed: morphological tagging and text classification. Tagging is performed on a deep neural network, using ArmTDP v2.3 dataset. For text classification, we propose a corpus of news articles categorized into 7 classes. The datasets are made public to serve as benchmarks for future models.
Obstacles hindering the development of capsule networks for challenging NLP applications include poor scalability to large output spaces and less reliable routing processes. In this paper, we introduce: 1) an agreement score to evaluate the performance of routing processes at instance level; 2) an adaptive optimizer to enhance the reliability of routing; 3) capsule compression and partial routing to improve the scalability of capsule networks. We validate our approach on two NLP tasks, namely: multi-label text classification and question answering. Experimental results show that our approach considerably improves over strong competitors on both tasks. In addition, we gain the best results in low-resource settings with few training instances.
We introduce VAMPIRE, a lightweight pretraining framework for effective text classification when data and computing resources are limited. We pretrain a unigram document model as a variational autoencoder on in-domain, unlabeled data and use its internal states as features in a downstream classifier. Empirically, we show the relative strength of VAMPIRE against computationally expensive contextual embeddings and other popular semi-supervised baselines under low resource settings. We also find that fine-tuning to in-domain data is crucial to achieving decent performance from contextual embeddings when working with limited supervision. We accompany this paper with code to pretrain and use VAMPIRE embeddings in downstream tasks.
We consider Large-Scale Multi-Label Text Classification (LMTC) in the legal domain. We release a new dataset of 57k legislative documents from EURLEX, annotated with ~4.3k EUROVOC labels, which is suitable for LMTC, few- and zero-shot learning. Experimenting with several neural classifiers, we show that BIGRUs with label-wise attention perform better than other current state of the art methods. Domain-specific WORD2VEC and context-sensitive ELMO embeddings further improve performance. We also find that considering only particular zones of the documents is sufficient. This allows us to bypass BERT's maximum text length limit and fine-tune BERT, obtaining the best results in all but zero-shot learning cases.
Classification of personal text messages has many useful applications in surveillance, e-commerce, and mental health care, to name a few. Giving applications access to personal texts can easily lead to (un)intentional privacy violations. We propose the first privacy-preserving solution for text classification that is provably secure. Our method, which is based on Secure Multiparty Computation (SMC), encompasses both feature extraction from texts, and subsequent classification with logistic regression and tree ensembles. We prove that when using our secure text classification method, the application does not learn anything about the text, and the author of the text does not learn anything about the text classification model used by the application beyond what is given by the classification result itself. We perform end-to-end experiments with an application for detecting hate speech against women and immigrants, demonstrating excellent runtime results without loss of accuracy.
Process model extraction (PME) is a recently emerged interdiscipline between natural language processing (NLP) and business process management (BPM), which aims to extract process models from textual descriptions. Previous process extractors heavily depend on manual features and ignore the potential relations between clues of different text granularities. In this paper, we formalize the PME task into the multi-grained text classification problem, and propose a hierarchical neural network to effectively model and extract multi-grained information without manually-defined procedural features. Under this structure, we accordingly propose the coarse-to-fine (grained) learning mechanism, training multi-grained tasks in coarse-to-fine grained order to share the high-level knowledge for the low-level tasks. To evaluate our approach, we construct two multi-grained datasets from two different domains and conduct extensive experiments from different dimensions. The experimental results demonstrate that our approach outperforms the state-of-the-art methods with statistical significance and further investigations demonstrate its effectiveness.
Topological Data Analysis (TDA) is a novel new and fast growing field of data science providing a set of new topological and geometric tools to derive relevant features out of complex high-dimensional data. In this paper we apply two of best methods in topological data analysis, "Persistent Homology" and "Mapper", in order to classify persian poems which has been composed by two of the best Iranian poets namely "Ferdowsi" and "Hafez". This article has two main parts, in the first part we explain the mathematics behind these two methods which is easy to understand for general audience and in the second part we describe our models and the results of applying TDA tools to NLP.
We describe and validate a metric for estimating multi-class classifier performance based on cross-validation and adapted for improvement of small, unbalanced natural-language datasets used in chatbot design. Our experiences draw upon building recruitment chatbots that mediate communication between job-seekers and recruiters by exposing the ML/NLP dataset to the recruiting team. Evaluation approaches must be understandable to various stakeholders, and useful for improving chatbot performance. The metric, nex-cv, uses negative examples in the evaluation of text classification, and fulfils three requirements. First, it is actionable: it can be used by non-developer staff. Second, it is not overly optimistic compared to human ratings, making it a fast method for comparing classifiers. Third, it allows model-agnostic comparison, making it useful for comparing systems despite implementation differences. We validate the metric based on seven recruitment-domain datasets in English and German over the course of one year.
Deep learning models such as convolutional neural networks and recurrent networks are widely applied in text classification. In spite of their great success, most deep learning models neglect the importance of modeling context information, which is crucial to understanding texts. In this work, we propose the Adaptive Region Embedding to learn context representation to improve text classification. Specifically, a metanetwork is learned to generate a context matrix for each region, and each word interacts with its corresponding context matrix to produce the regional representation for further classification. Compared to previous models that are designed to capture context information, our model contains less parameters and is more flexible. We extensively evaluate our method on 8 benchmark datasets for text classification. The experimental results prove that our method achieves state-of-the-art performances and effectively avoids word ambiguity.
Power-efficient CNN Domain Specific Accelerator (CNN-DSA) chips are currently available for wide use in mobile devices. These chips are mainly used in computer vision applications. However, the recent work of Super Characters method for text classification and sentiment analysis tasks using two-dimensional CNN models has also achieved state-of-the-art results through the method of transfer learning from vision to text. In this paper, we implemented the text classification and sentiment analysis applications on mobile devices using CNN-DSA chips. Compact network representations using one-bit and three-bits precision for coefficients and five-bits for activations are used in the CNN-DSA chip with power consumption less than 300mW. For edge devices under memory and compute constraints, the network is further compressed by approximating the external Fully Connected (FC) layers within the CNN-DSA chip. At the workshop, we have two system demonstrations for NLP tasks. The first demo classifies the input English Wikipedia sentence into one of the 14 ontologies. The second demo classifies the Chinese online-shopping review into positive or negative.
We introduce a lifelong language learning setup where a model needs to learn from a stream of text examples without any dataset identifier. We propose an episodic memory model that performs sparse experience replay and local adaptation to mitigate catastrophic forgetting in this setup. Experiments on text classification and question answering demonstrate the complementary benefits of sparse experience replay and local adaptation to allow the model to continuously learn from new datasets. We also show that the space complexity of the episodic memory module can be reduced significantly (~50-90%) by randomly choosing which examples to store in memory with a minimal decrease in performance. We consider an episodic memory component as a crucial building block of general linguistic intelligence and see our model as a first step in that direction.
Discourse structure is integral to understanding a text and is helpful in many NLP tasks. Learning latent representations of discourse is an attractive alternative to acquiring expensive labeled discourse data. Liu and Lapata (2018) propose a structured attention mechanism for text classification that derives a tree over a text, akin to an RST discourse tree. We examine this model in detail, and evaluate on additional discourse-relevant tasks and datasets, in order to assess whether the structured attention improves performance on the end task and whether it captures a text's discourse structure. We find the learned latent trees have little to no structure and instead focus on lexical cues; even after obtaining more structured trees with proposed model modifications, the trees are still far from capturing discourse structure when compared to discourse dependency trees from an existing discourse parser. Finally, ablation studies show the structured attention provides little benefit, sometimes even hurting performance.
Learning binary classifiers only from positive and unlabeled (PU) data is an important and challenging task in many real-world applications, including web text classification, disease gene identification and fraud detection, where negative samples are difficult to verify experimentally. Most recent PU learning methods are developed based on the conventional misclassification risk of the supervised learning type, and they require to solve the intractable risk estimation problem by approximating the negative data distribution or the class prior. In this paper, we introduce a variational principle for PU learning that allows us to quantitatively evaluate the modeling error of the Bayesian classifier directly from given data. This leads to a loss function which can be efficiently calculated without any intermediate step or model, and a variational learning method can then be employed to optimize the classifier under general conditions. In addition, the discriminative performance and numerical stability of the variational PU learning method can be further improved by incorporating a margin maximizing loss function. We illustrate the effectiveness of the proposed variational method on a number of benchmark examples.