ACL 2015 Papers

This paper explores a simple discriminative preordering model for statistical machine translation. Our model traverses binary constituent trees, and classifies whether children of each node should be reordered. The model itself is not extremely novel, but herein we introduce a new procedure to determine oracle labels so as to maximize Kendall’s τ . Experiments in Japanese-to-English translation revealed that our simple method is comparable with, or superior to, state-of-the-art methods in translation accuracy.

The inability to model long-distance dependency has been handicapping SMT for years. Specifically, the context independence assumption makes it hard to capture the dependency between translation rules. In this paper, we introduce a novel recurrent neural network based rule sequence model to incorporate arbitrary long contextual information during estimating probabilities of rule sequences. Moreover, our model frees the translation model from keeping huge and redundant grammars, resulting in more efficient training and decoding. Experimental results show that our method achieves a 0.9 point BLEU gain over the baseline, and a significant reduction in rule table size for both phrase-based and hierarchical phrase-based systems.

A language lexicon can be divided into four main strata, depending on origin of words: core vocabulary words, fullyand partiallyassimilated foreign words, and unassimilated foreign words (or transliterations). This paper focuses on translation of fullyand partially-assimilated foreign words, called “borrowed words”. Borrowed words (or loanwords) are content words found in nearly all languages, occupying up to 70% of the vocabulary. We use models of lexical borrowing in machine translation as a pivoting mechanism to obtain translations of out-of-vocabulary loanwords in a lowresource language. Our framework obtains substantial improvements (up to 1.6 BLEU) over standard baselines.

We exploit the visual properties of concepts for lexical entailment detection by examining a concept’s generality. We introduce three unsupervised methods for determining a concept’s generality, based on its related images, and obtain state-ofthe-art performance on two standard semantic evaluation datasets. We also introduce a novel task that combines hypernym detection and directionality, significantly outperforming a competitive frequencybased baseline.

We propose Imaginet, a model of learning visually grounded representations of language from coupled textual and visual input. The model consists of two Gated Recurrent Unit networks with shared word embeddings, and uses a multi-task objective by receiving a textual description of a scene and trying to concurrently predict its visual representation and the next word in the sentence. Like humans, it acquires meaning representations for individual words from descriptions of visual scenes. Moreover, it learns to effectively use sequential structure in semantic interpretation of multi-word phrases.

In this paper, we propose a novel query expansion approach for improving transferbased automatic image captioning. The core idea of our method is to translate the given visual query into a distributional semantics based form, which is generated by the average of the sentence vectors extracted from the captions of images visually similar to the input image. Using three image captioning benchmark datasets, we show that our approach provides more accurate results compared to the state-of-theart data-driven methods in terms of both automatic metrics and subjective evaluation.

Two recent approaches have achieved state-of-the-art results in image captioning. The first uses a pipelined process where a set of candidate words is generated by a convolutional neural network (CNN) trained on images, and then a maximum entropy (ME) language model is used to arrange these words into a coherent sentence. The second uses the penultimate activation layer of the CNN as input to a recurrent neural network (RNN) that then generates the caption sequence. In this paper, we compare the merits of these different language modeling approaches for the first time by using the same state-ofthe-art CNN as input. We examine issues in the different approaches, including linguistic irregularities, caption repetition, and data set overlap. By combining key aspects of the ME and RNN methods, we achieve a new record performance over previously published results on the benchmark COCO dataset. However, the gains we see in BLEU do not translate to human judgments.

A wide array of natural dialogue discourse can be found on the internet. Previous attempts to automatically determine disagreement between interlocutors in such dialogue have mostly relied on n-gram and grammatical dependency features taken from respondent text. Agreement-disagreement classifiers built upon these baseline features tend to do poorly, yet have proven difficult to improve upon. Using the Internet Argument Corpus, which comprises quote and response post pairs taken from an online debate forum with human-annotated agreement scoring, we introduce semantic environment features derived by comparing quote and response sentences which align well. We show that this method improves classifier accuracy relative to the baseline method namely in the retrieval of disagreeing pairs, which improves from 69% to 77%.

Discourse parsing is the process of discovering the latent relational structure of a long form piece of text and remains a significant open challenge. One of the most difficult tasks in discourse parsing is the classification of implicit discourse relations. Most state-of-the-art systems do not leverage the great volume of unlabeled text available on the web–they rely instead on human annotated training data. By incorporating a mixture of labeled and unlabeled data, we are able to improve relation classification accuracy, reduce the need for annotated data, while still retaining the capacity to use labeled data to ensure that specific desired relations are learned. We achieve this using a latent variable model that is trained in a reduced dimensionality subspace using spectral methods. Our approach achieves an F1 score of 0.485 on the implicit relation labeling task for the Penn Discourse Treebank.

This paper is the first to examine the effect of prosodic features on coreference resolution in spoken discourse. We test features from different prosodic levels and investigate which strategies can be applied. Our results on the basis of manual prosodic labelling show that the presence of an accent is a helpful feature in a machine-learning setting. Including prosodic boundaries and determining whether the accent is the nuclear accent further increases results.

We present an experiment to compare a standard, minimally distinguishing algorithm for the generation of relational referring expressions with two alternatives that produce overspecified descriptions. The experiment shows that discrimination which normally plays a major role in the disambiguation task is also a major influence in referential overspecification, even though disambiguation is in principle not relevant.

This paper describes an experiment to elicit referring expressions from human subjects for research in natural language generation and related fields, and preliminary results of a computational model for the generation of these expressions. Unlike existing resources of this kind, the resulting data set the Zoom corpus of natural language descriptions of map locations takes into account a domain that is significantly closer to real-world applications than what has been considered in previous work, and addresses more complex situations of reference, including contexts with different levels of detail, and instances of singular and plural reference produced by speakers of Spanish and Portuguese.

Simplification of lexically complex texts, by replacing complex words with their simpler synonyms, helps non-native speakers, children, and language-impaired people understand text better. Recent lexical simplification methods rely on manually simplified corpora, which are expensive and time-consuming to build. We present an unsupervised approach to lexical simplification that makes use of the most recent word vector representations and requires only regular corpora. Results of both automated and human evaluation show that our simple method is as effective as systems that rely on simplified corpora.

The validity of applying paraphrase rules depends on the domain of the text that they are being applied to. We develop a novel method for extracting domainspecific paraphrases. We adapt the bilingual pivoting paraphrase method to bias the training data to be more like our target domain of biology. Our best model results in higher precision while retaining complete recall, giving a 10% relative improvement in AUC.

We explore using relevant tweets of a given news article to help sentence compression for generating compressive news highlights. We extend an unsupervised dependency-tree based sentence compression approach by incorporating tweet information to weight the tree edge in terms of informativeness and syntactic importance. The experimental results on a public corpus that contains both news articles and relevant tweets show that our proposed tweets guided sentence compression method can improve the summarization performance significantly compared to the baseline generic sentence compression method.

There have been many recent advances in the structure and measurement of distributed language models: those that map from words to a vector-space that is rich in information about word choice and composition. This vector-space is the distributed language representation. The goal of this note is to point out that any distributed representation can be turned into a classifier through inversion via Bayes rule. The approach is simple and modular, in that it will work with any language representation whose training can be formulated as optimizing a probability model. In our application to 2 million sentences from Yelp reviews, we also find that it performs as well as or better than complex purpose-built algorithms.

Predicting the helpfulness of product reviews is a key component of many ecommerce tasks such as review ranking and recommendation. However, previous work mixed review helpfulness prediction with those outer layer tasks. Using nontext features, it leads to less transferable models. This paper solves the problem from a new angle by hypothesizing that helpfulness is an internal property of text. Purely using review text, we isolate review helpfulness prediction from its outer layer tasks, employ two interpretable semantic features, and use human scoring of helpfulness as ground truth. Experimental results show that the two semantic features can accurately predict helpfulness scores and greatly improve the performance compared with using features previously used. Cross-category test further shows the models trained with semantic features are easier to be generalized to reviews of different product categories. The models we built are also highly interpretable and align well with human annotations.

We present a general framework for incorporating sequential data and arbitrary features into language modeling. The general framework consists of two parts: a hidden Markov component and a recursive neural network component. We demonstrate the effectiveness of our model by applying it to a specific application: predicting topics and sentiments in dialogues. Experiments on real data demonstrate that our method is substantially more accurate than previ-

In this paper, we address semi-supervised sentiment learning via semi-stacking, which integrates two or more semi-supervised learning algorithms from an ensemble learning perspective. Specifically, we apply metalearning to predict the unlabeled data given the outputs from the member algorithms and propose N-fold cross validation to guarantee a suitable size of the data for training the meta-classifier. Evaluation on four domains shows that such a semi-stacking strategy performs consistently better than its member algorithms.