Training a structured prediction model involves performing several loss-augmented inference steps. Over the lifetime of the training, many of these inference problems, although different, share the same solution. We propose AI-DCD, an Amortized Inference framework for Dual Coordinate Descent method, an approximate learning algorithm, that accelerates the training process by exploiting this redundancy of solutions, without compromising the performance of the model. We show the efficacy of our method by training a structured SVM using dual coordinate descent for an entityrelation extraction task. Our method learns the same model as an exact training algorithm would, but call the inference engine only in 10% – 24% of the inference problems encountered during training. We observe similar gains on a multi-label classification task and with a Structured Perceptron model for the entity-relation task.

The success of applying policy gradient reinforcement learning (RL) to difficult control tasks hinges crucially on the ability to determine a sensible initialization for the policy. Transfer learning methods tackle this problem by reusing knowledge gleaned from solving other related tasks. In the case of multiple task domains, these algorithms require an inter-task mapping to facilitate knowledge transfer across domains. However, there are currently no general methods to learn an inter-task mapping without requiring either background knowledge that is not typically present in RL settings, or an expensive analysis of an exponential number of inter-task mappings in the size of the state and action spaces. This paper introduces an autonomous framework that uses unsupervised manifold alignment to learn inter-task mappings and effectively transfer samples between different task domains. Empirical results on diverse dynamical systems, including an application to quadrotor control, demonstrate its effectiveness for cross-domain transfer in the context of policy gradient RL.

The aim of domain adaptation algorithms is to establish a learner, trained on labeled data from a source domain, that can classify samples from a target domain, in which few or no labeled data are available for training. Covariate shift, a primary assumption in several works on domain adaptation, assumes that the labeling functions of source and target domains are identical. We present a domain adaptation algorithm that assumes a relaxed version of covariate shift where the assumption that the labeling functions of the source and target domains are identical holds with a certain probability. Assuming a source deterministic large margin binary classifier, the farther a target instance is from the source decision boundary, the higher the probability that covariate shift holds. In this context, given a target unlabeled sample and no target labeled data, we develop a domain adaptation algorithm that bases its labeling decisions both on the source learner and on the similarities between the target unlabeled instances. The source labeling function decisions associated with probabilistic covariate shift, along with the target similarities are concurrently expressed on a similarity graph. We evaluate our proposed algorithm on a benchmark sentiment analysis (and domain adaptation) dataset, where state-of-the-art adaptation results are achieved. We also derive a lower bound on the performance of the algorithm.

Twitter, as a popular microblogging service, has become a new information channel for users to receive and exchange the mostup-to-date information on current events. However, since there is no control on how users can publish messages on Twitter, finding newsworthy events from Twitter becomes a difficult task like "finding a needle in a haystack". In this paper we propose a general unsupervised framework to explore events from tweets, which consists of a pipeline process of filtering, extraction and categorization. To filter out noisy tweets, the filtering step exploits a lexicon-based approach to separate tweets that are event-related from those that are not. Then, based on these event-related tweets, the structured representations of events are extracted and categorized automatically using an unsupervised Bayesian model without the use of any labelled data. Moreover, the categorized events are assigned with the event type labels without human intervention. The proposed framework has been evaluated on over 60 millions tweets which were collected for one month in December 2010. A precision of 70.49% is achieved in event extraction, outperforming a competitive baseline by nearly 6%. Events are also clustered into coherence groups with the automatically assigned event type label.

In this paper, we propose and address a novel task of recommending quotes for writing. Quote is short for quotation, which is the repetition of someone else’s statement or thoughts. It is a common case in our writing when we would like to cite someone’s statement, like a proverb or a statement by some famous people, to make our composition more elegant or convincing. However, sometimes we are so eager to make a citation of quote somewhere, but have no idea about the relevant quote to express our idea. Because knowing or remembering so many quotes is not easy, it is exciting to have a system to recommend relevant quotes for us while writing. In this paper we tackle this appealing AI task, and build up a learning framework for quote recommendation. We collect abundant quotes from the Internet, and mine real contexts containing these quotes from large amount of electronic books, to build up a dataset for experiments. We explore the particular features of this task, and propose a few useful features to model the characteristics of quotes and the relevance of quotes to contexts. We apply a supervised learning to rank model to integrate multiple features. Experiment results show that, our proposed approach is appropriate for this task and it outperforms other recommendation methods.

Most word embedding models typically represent each word using a single vector, which makes these models indiscriminative for ubiquitous homonymy and polysemy. In order to enhance discriminativeness, we employ latent topic models to assign topics for each word in the text corpus, and learn topical word embeddings (TWE) based on both words and their topics. In this way, contextual word embeddings can be flexibly obtained to measure contextual word similarity. We can also build document representations, which are more expressive than some widely-used document models such as latent topic models. In the experiments, we evaluate the TWE models on two tasks, contextual word similarity and text classification. The experimental results show that our models outperform typical word embedding models including the multi-prototype version on contextual word similarity, and also exceed latent topic models and other representative document models on text classification.

Automatic citation recommendation can be very useful for authoring a paper and is an AI-complete problem due to the challenge of bridging the semantic gap between citation context and the cited paper. It is not always easy for knowledgeable researchers to give an accurate citation context for a cited paper or to find the right paper to cite given context. To help with this problem, we propose a novel neural probabilistic model that jointly learns the semantic representations of citation contexts and cited papers. The probability of citing a paper given a citation context is estimated by training a multi-layer neural network. We implement and evaluate our model on the entire CiteSeer dataset, which at the time of this work consists of 10,760,318 citation contexts from 1,017,457 papers. We show that the proposed model significantly outperforms other state-of-the-art models in recall, MAP, MRR, and nDCG.

This research describes the development of a supervised classifier of English light verb constructions, for example, "take a walk" and "make a speech." This classifier relies on features from dependency parses, OntoNotes sense tags, WordNet hypernyms and WordNet lexical file information. Evaluation shows that this system achieves an 89% F1 score (four points above the state of the art) on the BNC test set used by Tu & Roth (2011), and an F1 score of 80.68 on the OntoNotes test set, which is significantly more challenging. We attribute the superior F1 score to the use of our rich linguistic features, including the use of WordNet synset and hypernym relations for the detection of previously unattested light verb constructions. We describe the classifier and its features, as well as the characteristics of the OntoNotes light verb construction test set, which relies on linguistically motivated PropBank annotation.

Recently, joint video-language modeling has been attracting more and more attention. However, most existing approaches focus on exploring the language model upon on a fixed visual model. In this paper, we propose a unified framework that jointly models video and the corresponding text sentences. The framework consists of three parts: a compositional semantics language model, a deep video model and a joint embedding model. In our language model, we propose a dependency-tree structure model that embeds sentence into a continuous vector space, which preserves visually grounded meanings and word order. In the visual model, we leverage deep neural networks to capture essential semantic information from videos. In the joint embedding model, we minimize the distance of the outputs of the deep video model and compositional language model in the joint space, and update these two models jointly. Based on these three parts, our system is able to accomplish three tasks: 1) natural language generation, and 2) video retrieval and 3) language retrieval. In the experiments, the results show our approach outperforms SVM, CRF and CCA baselines in predicting Subject-Verb- Object triplet and natural sentence generation, and is better than CCA in video retrieval and language retrieval tasks.

Easy-first, a search-based structured prediction approach, has been applied to many NLP tasks including dependency parsing and coreference resolution. This approach employs a learned greedy policy (action scoring function) to make easy decisions first, which constrains the remaining decisions and makes them easier. We formulate greedy policy learning in the Easy-first approach as a novel non-convex optimization problem and solve it via an efficient Majorization Minimizatoin (MM) algorithm. Results on within-document coreference and cross-document joint entity and event coreference tasks demonstrate that the proposed approach achieves statistically significant performance improvement over existing training regimes for Easy-first and is less susceptible to overfitting.