In the field of digital marketing, understanding the voice of the customer is paramount. Mining textual content written by visitors on websites or social media can offer new dimensions to marketers and CX executives. Traditional tasks in NLP like sentiment analysis, topic modeling etc. can solve only certain specific problems but don’t provide a generic solution to identifying/understanding the intention behind a text. In this paper we consider higher dimensional extensions to the sentiment concept by incorporating labels like product enquiry, buying intent, seeking help, feedback and pricing query which give us a deeper understanding of the text. We show how our model performs in a real-world enterprise use case. Word2Vec embeddings are used for word representations and later we compare three algorithms for classification. SVM’s provide us with a strong baseline. Two deep learning models viz. vanilla CNN and RNN’s with LSTM are compared. With no use of hard-coded or hand engineered features, our generic model can be used in a variety of use cases where text mining is involved with ease.

Consumers often rely on product reviews to make purchase decisions, but how consumers use review content in their decision making has remained a black box. In the past, extracting information from product reviews has been a labor-intensive process that has restricted studies on this topic to single product categories or those limited to summary statistics such as volume, valence, and ratings. This paper uses deep learning natural language processing techniques to overcome the limitations of manual information extraction and shed light into the black box of how consumers use review content. With the help of a comprehensive dataset that tracks individual-level review reading, search, as well as purchase behaviors on an e-commerce portal, we extract six quality and price content dimensions from over 500,000 reviews, covering nearly 600 product categories. The scale, scope, and precision of such a study would have been impractical using human coders or classical machine learning models. We achieve two objectives. First, we describe consumers’ review content reading behaviors. We find that although consumers do not read review content all the time, they do rely on review content for products that are expensive or of uncertain quality. Second, we quantify the causal impact of content information of read reviews on sales. We use a regression discontinuity in time design and leverage the variation in the review content seen by consumers due to newly added reviews. To extract content information, we develop two deep learning models: a full deep learning model that predicts conversion directly and a partial deep learning model that identifies content dimensions. Across both models, we find that aesthetics and price content in the reviews significantly affect conversion across almost all product categories. Review content information has a higher impact on sales when the average rating is higher and the variance of ratings is lower. Consumers depend more on review content when the market is more competitive or immature. A counterfactual simulation suggests that re-ordering reviews based on content can have the same effect as a 1.6% price cut for boosting conversion.

Interactive systems have taken over the web and mobile space with increasing participation from users. Applications across every marketing domain can now be accessed through mobile or web where users can directly perform certain actions and reach a desired outcome. Actions of user on a system, though, can be representative of a certain intent. Ability to learn this intent through user's actions can help draw certain insight into the behavior of users on a system. In this paper, we present models to optimize interactive systems by learning and analyzing user intent through their actions on the system. We present a four phased model that uses time-series of interaction actions sequentially using a Long Short-Term Memory (LSTM) based sequence learning system that helps build a model for intent recognition. Our system then provides an objective specific maximization followed by analysis and contrasting methods in order to identify spaces of improvement in the interaction system. We discuss deployment scenarios for such a system and present results from evaluation on an online marketplace using user clickstream data.

While there have been significant advances in detecting emotions from speech and image recognition, emotion detection on text is still under-explored and remained as an active research field. This paper introduces a corpus for text-based emotion detection on multiparty dialogue as well as deep neural models that outperform the existing approaches for document classification. We first present a new corpus that provides annotation of seven emotions on consecutive utterances in dialogues extracted from the show, Friends. We then suggest four types of sequence-based convolutional neural network models with attention that leverage the sequence information encapsulated in dialogue. Our best model shows the accuracies of 37.9% and 54% for fine- and coarse-grained emotions, respectively. Given the difficulty of this task, this is promising.

Machine learning techniques have shown great promises across domains, however they fail to impress when there is scarcity of training data. Work in the area of affective content analysis can not take complete advantage of machine learning techniques when there is a lack of sufficient training data. It is well known that recurrent neural networks (RNNs), particularly with long-short term memory (LSTM) units, perform better than feed-forward neural networks (FFNNs) on sequential data as they are architecturally designed to learn temporal relationships existing in the training data while FFNNs are not. But RNNs require sufficient training data to learn these temporal relationships. In this paper, we show that one can take advantage of a-priori knowledge about the temporal correlations in the training data even in a FFNN architecture. We call this the knowledge-driven FFNN or k-FFNN architecture. We show using the MediaEval dataset that the k-FFNN model not only outperforms FFNN, but also performs better than RNN models (i.e., Simple RNN, RNN with LSTM units and bi-directional RNN with LSTM units (BLSTM)), especially when the amount of training data is sparse.

Deep Neural Network (DNN) acoustic models often use discriminative sequence training that optimises an objective function that better approximates the word error rate (WER) than frame-based training. Sequence training is normally implemented using Stochastic Gradient Descent (SGD) or Hessian Free (HF) training. This paper proposes an alternative batch style optimisation framework that employs a Natural Gradient (NG) approach to traverse through the parameter space. By correcting the gradient according to the local curvature of the KL-divergence, the NG optimisation process converges more quickly than HF. Furthermore, the proposed NG approach can be applied to any sequence discriminative training criterion. The efficacy of the NG method is shown using experiments on a Multi-Genre Broadcast (MGB) transcription task that demonstrates both the computational efficiency and the accuracy of the resulting DNN models.

We present a new method of blackbox optimization via gradient approximation with the use of structured random orthogonal matrices, providing more accurate estimators than baselines and with provable theoretical guarantees. We show that this algorithm can be successfully applied to learn better quality compact policies than those using standard gradient estimation techniques. The compact policies we learn have several advantages over unstructured ones, including faster training algorithms and faster inference. These benefits are important when the policy is deployed on real hardware with limited resources. Further, compact policies provide more scalable architectures for derivative-free optimization (DFO) in high-dimensional spaces. We show that most robotics tasks from the OpenAI Gym can be solved using neural networks with less than 300 parameters, with almost linear time complexity of the inference phase, with up to 13x fewer parameters relative to the Evolution Strategies (ES) algorithm introduced by Salimans et al. (2017). We do not need heuristics such as fitness shaping to learn good quality policies, resulting in a simple and theoretically motivated training mechanism.

Traditionally, deep learning algorithms update the network weights whereas the network architecture is chosen manually, using a process of trial and error. In this work, we propose two novel approaches that automatically update the network structure while also learning its weights. The novelty of our approach lies in our parameterization where the depth, or additional complexity, is encapsulated continuously in the parameter space through control parameters that add additional complexity. We propose two methods: In tunnel networks, this selection is done at the level of a hidden unit, and in budding perceptrons, this is done at the level of a network layer; updating this control parameter introduces either another hidden unit or another hidden layer. We show the effectiveness of our methods on the synthetic two-spirals data and on two real data sets of MNIST and MIRFLICKR, where we see that our proposed methods, with the same set of hyperparameters, can correctly adjust the network complexity to the task complexity.

Training an artificial neural network involves an optimization process over the landscape defined by the cost (loss) as a function of the network parameters. We explore these landscapes using optimisation tools developed for potential energy landscapes in molecular science. The number of local minima and transition states (saddle points of index one), as well as the ratio of transition states to minima, grow rapidly with the number of nodes in the network. There is also a strong dependence on the regularisation parameter, with the landscape becoming more convex (fewer minima) as the regularisation term increases. We demonstrate that in our formulation, stationary points for networks with $N_h$ hidden nodes, including the minimal network required to fit the XOR data, are also stationary points for networks with $N_{h} +1$ hidden nodes when all the weights involving the additional nodes are zero. Hence, smaller networks optimized to train the XOR data are embedded in the landscapes of larger networks. Our results clarify certain aspects of the classification and sensitivity (to perturbations in the input data) of minima and saddle points for this system, and may provide insight into dropout and network compression.

Deep networks have achieved impressive results across a variety of important tasks. However a known weakness is a failure to perform well when evaluated on data which differ from the training distribution, even if these differences are very small, as is the case with adversarial examples. We propose Fortified Networks, a simple transformation of existing networks, which fortifies the hidden layers in a deep network by identifying when the hidden states are off of the data manifold, and maps these hidden states back to parts of the data manifold where the network performs well. Our principal contribution is to show that fortifying these hidden states improves the robustness of deep networks and our experiments (i) demonstrate improved robustness to standard adversarial attacks in both black-box and white-box threat models; (ii) suggest that our improvements are not primarily due to the gradient masking problem and (iii) show the advantage of doing this fortification in the hidden layers instead of the input space.