The logistic normal distribution has recently been adapted via the transformation of multivariate Gaus- sian variables to model the topical distribution of documents in the presence of correlations among topics. In this paper, we propose a probit normal alternative approach to modelling correlated topical structures. Our use of the probit model in the context of topic discovery is novel, as many authors have so far con- centrated solely of the logistic model partly due to the formidable inefficiency of the multinomial probit model even in the case of very small topical spaces. We herein circumvent the inefficiency of multinomial probit estimation by using an adaptation of the diagonal orthant multinomial probit in the topic models context, resulting in the ability of our topic modelling scheme to handle corpuses with a large number of latent topics. An additional and very important benefit of our method lies in the fact that unlike with the logistic normal model whose non-conjugacy leads to the need for sophisticated sampling schemes, our ap- proach exploits the natural conjugacy inherent in the auxiliary formulation of the probit model to achieve greater simplicity. The application of our proposed scheme to a well known Associated Press corpus not only helps discover a large number of meaningful topics but also reveals the capturing of compellingly intuitive correlations among certain topics. Besides, our proposed approach lends itself to even further scalability thanks to various existing high performance algorithms and architectures capable of handling millions of documents.

Sentiment classification provides information about the author's feeling toward a topic through the use of expressive words. However, words indicative of a particular sentiment class can be domain-specific. We train a text classifier for Twitter data related to games using labels inferred from emoticons. Our classifier is able to differentiate between positive and negative sentiment tweets labeled by emoticons with 75.1% accuracy. Additionally, we test the classifier on human-labeled examples with the additional case of neutral or ambiguous sentiment. Finally, we have made the data available to the community for further use and analysis.

We investigate ways in which to improve the interpretability of LDA topic models by better analyzing and visualizing their outputs. We focus on examining what we refer to as topic similarity networks: graphs in which nodes represent latent topics in text collections and links represent similarity among topics. We describe efficient and effective approaches to both building and labeling such networks. Visualizations of topic models based on these networks are shown to be a powerful means of exploring, characterizing, and summarizing large collections of unstructured text documents. They help to "tease out" non-obvious connections among different sets of documents and provide insights into how topics form larger themes. We demonstrate the efficacy and practicality of these approaches through two case studies: 1) NSF grants for basic research spanning a 14 year period and 2) the entire English portion of Wikipedia.

We propose a parsimonious topic model for text corpora. In related models such as Latent Dirichlet Allocation (LDA), all words are modeled topic-specifically, even though many words occur with similar frequencies across different topics. Our modeling determines salient words for each topic, which have topic-specific probabilities, with the rest explained by a universal shared model. Further, in LDA all topics are in principle present in every document. By contrast our model gives sparse topic representation, determining the (small) subset of relevant topics for each document. We derive a Bayesian Information Criterion (BIC), balancing model complexity and goodness of fit. Here, interestingly, we identify an effective sample size and corresponding penalty specific to each parameter type in our model. We minimize BIC to jointly determine our entire model -- the topic-specific words, document-specific topics, all model parameter values, {\it and} the total number of topics -- in a wholly unsupervised fashion. Results on three text corpora and an image dataset show that our model achieves higher test set likelihood and better agreement with ground-truth class labels, compared to LDA and to a model designed to incorporate sparsity.

We describe a state-of-the-art sentiment analysis system that detects (a) the sentiment of short informal textual messages such as tweets and SMS (message-level task) and (b) the sentiment of a word or a phrase within a message (term-level task). The system is based on a supervised statistical text classification approach leveraging a variety of surface-form, semantic, and sentiment features. The sentiment features are primarily derived from novel high-coverage tweet-specific sentiment lexicons. These lexicons are automatically generated from tweets with sentiment-word hashtags and from tweets with emoticons. To adequately capture the sentiment of words in negated contexts, a separate sentiment lexicon is generated for negated words. The system ranked first in the SemEval-2013 shared task `Sentiment Analysis in Twitter' (Task 2), obtaining an F-score of 69.02 in the message-level task and 88.93 in the term-level task. Post-competition improvements boost the performance to an F-score of 70.45 (message-level task) and 89.50 (term-level task). The system also obtains state-of-the-art performance on two additional datasets: the SemEval-2013 SMS test set and a corpus of movie review excerpts. The ablation experiments demonstrate that the use of the automatically generated lexicons results in performance gains of up to 6.5 absolute percentage points.

Reinforcement learning-based spoken dialogue systems aim to compute an optimal strategy for dialogue management from interactions with users. They compare their different management strategies on the basis of a numerical reward function. Reward inference consists of learning a reward function from dialogues scored by users. A major issue for reward inference algorithms is that important parameters influence user evaluations and cannot be computed online. This is the case of task completion. This paper introduces Task Completion Transfer Learning (TCTL): a method to exploit the exact knowledge of task completion on a corpus of dialogues scored by users in order to optimise online learning. Compared to previously proposed reward inference techniques, TCTL returns a reward function enhanced with the possibility to manage the online non-observability of task completion. A reward function is learnt with TCTL on dialogues with a restaurant seeking system. It is shown that the reward function returned by TCTL is a better estimator of dialogue performance than the one returned by reward inference.

The ability to identify what humans are doing in the environment is a crucial element of successful responsive behavior in human-robot interaction. We examine new ways to perform plan recognition (PR) using natural language processing (NLP) techniques. PR often focuses on the structural relationships between consecutive observations and ordered activities that comprise plans. However, NLP commonly treats text as a bag-of-words, omitting such structural relationships and using topic models to break down the distribution of concepts discussed in documents. In this paper, we examine an analogous treatment of plans as distributions of activities. We explore the application of Latent Dirichlet Allocation topic models to human skeletal data of plan execution traces obtained from a RGB-D sensor. This investigation focuses on representing the data as text and interpreting learned activities as a form of activity recognition (AR). Additionally, we explain how the system may perform PR. The initial empirical results suggest that such NLP methods can be useful in complex PR and AR tasks.

Many Artificial Intelligence tasks need large amounts of commonsense knowledge. Because obtaining this knowledge through machine learning would require a huge amount of data, a better alternative is to elicit it from people through human computation. We consider the sentiment classification task, where knowledge about the contexts that impact word polarities is crucial, but hard to acquire from data. We describe a novel task design that allows us to crowdsource this knowledge through Amazon Mechanical Turk with high quality. We show that the commonsense knowledge acquired in this way dramatically improves the performance of established sentiment classification methods.

Probabilistic topic models have been widely used for sentiment analysis. However, most of existing topic methods only model the sentiment text, but do not consider the user, who expresses the sentiment, and the item, which the sentiment is expressed on. Since different users may use different sentiment expressions for different items, we argue that it is better to incorporate the user and item information into the topic model for sentiment analysis. In this paper, we propose a new Supervised User-Item based Topic model, called SUIT model, for sentiment analysis. It can simultaneously utilize the textual topic and latent user-item factors. Our proposed method uses the tensor outer product of text topic proportion vector, user latent factor and item latent factor to model the sentiment label generalization. Extensive experiments are conducted on two datasets: review dataset and microblog dataset. The results demonstrate the advantages of our model. It shows significant improvement compared with supervised topic models and collaborative filtering methods.

Recursive neural models have achieved promising results in many natural language processing tasks. The main difference among these models lies in the composition function, i.e., how to obtain the vector representation for a phrase or sentence using the representations of words it contains. This paper introduces a novel Adaptive Multi-Compositionality (AdaMC) layer to recursive neural models. The basic idea is to use more than one composition functions and adaptively select them depending on the input vectors. We present a general framework to model each semantic composition as a distribution over these composition functions. The composition functions and parameters used for adaptive selection are learned jointly from data. We integrate AdaMC into existing recursive neural models and conduct extensive experiments on the Stanford Sentiment Treebank. The results illustrate that AdaMC significantly outperforms state-of-the-art sentiment classification methods. It helps push the best accuracy of sentence-level negative/positive classification from 85.4% up to 88.5%.