ACL 2015 Papers

Online social networks nowadays have the worldwide prosperity, as they have revolutionized the way for people to discover, to share, and to diffuse information. Social networks are powerful, yet they still have Achilles Heel: extreme data sparsity. Individual posting documents, (e.g., a microblog less than 140 characters), seem to be too sparse to make a difference under various scenarios, while in fact they are quite different. We propose to tackle this specific weakness of social networks by smoothing the posting document language model based on social regularization. We formulate an optimization framework with a social regularizer. Experimental results on the Twitter dataset validate the effectiveness and efficiency of our proposed model.

Expert finding on social media benefits both individuals and commercial services. In this paper, we exploit a 5-level tree representation to model the posts on social media and cast the expert finding problem to the matching problem between the learned user tree and domain tree. We enhance the traditional approximate tree matching algorithm and incorporate word embeddings to improve the matching result. The experiments conducted on Sina Microblog demonstrate the effectiveness of our work.

Recurrent neural network (RNN) is recognized as a powerful language model (LM). We investigate deeper into its performance portfolio, which performs well on frequent grammatical patterns but much less so on less frequent terms. Such portfolio is expected and desirable in applications like autocomplete, but is less useful in social content analysis where many creative, unexpected usages occur (e.g., URL insertion). We adapt a generic RNN model and show that, with variational training corpora and epoch unfolding, the model improves its performance for the task of URL insertion suggestions.

Alcohol abuse may lead to unsociable behavior such as crime, drunk driving, or privacy leaks. We introduce automatic drunk-texting prediction as the task of identifying whether a text was written when under the influence of alcohol. We experiment with tweets labeled using hashtags as distant supervision. Our classifiers use a set of N-gram and stylistic features to detect drunk tweets. Our observations present the first quantitative evidence that text contains signals that can be exploited to detect drunk-texting.

We present and evaluate a method for automatically detecting sentence fragments in English texts written by non-native speakers. Our method combines syntactic parse tree patterns and parts-of-speech information produced by a tagger to detect this phenomenon. When evaluated on a corpus of authentic learner texts, our best model achieved a precision of 0.84 and a recall of 0.62, a statistically significant improvement over baselines using non-parse features, as well as a popular grammar checker.

Languages using Chinese characters are mostly processed at word level. Inspired by recent success of deep learning, we delve deeper to character and radical levels for Chinese language processing. We propose a new deep learning technique, called “radical embedding”, with justifications based on Chinese linguistics, and validate its feasibility and utility through a set of three experiments: two in-house standard experiments on short-text categorization (STC) and Chinese word segmentation (CWS), and one in-field experiment on search ranking. We show that radical embedding achieves comparable, and sometimes even better, results than competing methods.

How do we know which grammatical error correction (GEC) system is best? A number of metrics have been proposed over the years, each motivated by weaknesses of previous metrics; however, the metrics themselves have not been compared to an empirical gold standard grounded in human judgments. We conducted the first human evaluation of GEC system outputs, and show that the rankings produced by metrics such as MaxMatch and I-measure do not correlate well with this ground truth. As a step towards better metrics, we also propose GLEU, a simple variant of BLEU, modified to account for both the source and the reference, and show that it hews much more closely to human judgments.

This paper presents a novel task, namely the automatic identification of ageappropriate ratings of a musical track, or album, based on its lyrics. Details are provided regarding the construction of a dataset of lyrics from 12,242 tracks across 1,798 albums along with age-appropriate ratings obtained from various web resources, along with results from various text classification experiments. The best accuracy of 71.02% for classifying albums by age groups is achieved by combining vector space model and psycholinguistic features.

In this article, we discuss the challenges of document summarization for the blind and visually impaired people and then propose a new system called BrailleSUM to produce better summaries for the blind and visually impaired people. Our system considers the factor of braille length of each sentence in news articles into the ILPbased summarization method. Evaluation results on a DUC dataset show that BrailleSUM can produce shorter braille summaries than existing methods, meanwhile, it does not sacrifice the content quality of the summaries.

Pivot translation allows for translation of language pairs with little or no parallel data by introducing a third language for which data exists. In particular, the triangulation method, which translates by combining source-pivot and pivot-target translation models into a source-target model, is known for its high translation accuracy. However, in the conventional triangulation method, information of pivot phrases is forgotten and not used in the translation process. In this paper, we propose a novel approach torememberthe pivot phrases in the triangulation stage, and use a pivot language model as an additional information source at translation time. Experimental results on the Europarl corpus showed gains of 0.4-1.2 BLEU points in all tested combinations of languages 1 .

A joint-space model for cross-lingual distributed representations generalizes language-invariant semantic features. In this paper, we present a matrix cofactorization framework for learning cross-lingual word embeddings. We explicitly define monolingual training objectives in the form of matrix decomposition, and induce cross-lingual constraints for simultaneously factorizing monolingual matrices. The cross-lingual constraints can be derived from parallel corpora, with or without word alignments. Empirical results on a task of cross-lingual document classification show that our method is effective to encode cross-lingual knowledge as constraints for cross-lingual word embeddings.

Domain adaptation is an active field of research in statistical machine translation (SMT), but so far most work has ignored the distinction between the topic and genre of documents. In this paper we quantify and disentangle the impact of genre and topic differences on translation quality by introducing a new data set that has controlled topic and genre distributions. In addition, we perform a detailed analysis showing that differences across topics only explain to a limited degree translation performance differences across genres, and that genre-specific errors are more attributable to model coverage than to suboptimal scoring of translation candidates.

In Statistical Machine Translation, some complex features are still difficult to integrate during decoding and usually used through the reranking of the k-best hypotheses produced by the decoder. We propose a translation table partitioning method that exploits the result of this reranking to iteratively guide the decoder in order to produce a new k-best list more relevant to some complex features. We report experiments on two translation domains and two translations directions which yield improvements of up to 1.4 BLEU over the reranking baseline using the same set of complex features. On a practical viewpoint, our approach allows SMT system developers to easily integrate complex features into decoding rather than being limited to their use in one-time k-best list reranking.

In this paper we present the UNRAVEL toolkit: It implements many of the recently published works on decipherment, including decipherment for deterministic ciphers like e.g. the ZODIAC-408 cipher and Part two of the BEALE ciphers, as well as decipherment of probabilistic ciphers and unsupervised training for machine translation. It also includes data and example configuration files so that the previously published experiments are easy to reproduce.

Statistical models for reordering source words have been used to enhance hierarchical phrase-based statistical machine translation. There are existing word-reordering models that learn reorderings for any two source words in a sentence or only for two contiguous words. This paper proposes a series of separate sub-models to learn reorderings for word pairs with different distances. Our experiments demonstrate that reordering sub-models for word pairs with distances less than a specific threshold are useful to improve translation quality. Compared with previous work, our method more effectively and efficiently exploits helpful word-reordering information; it improves a basic hierarchical phrase-based system by 2.4-3.1 BLEU points and keeps the average time of translating one sentence under 10 s.

We propose a novel method for translation selection in statistical machine translation, in which a convolutional neural network is employed to judge the similarity between a phrase pair in two languages. The specifically designed convolutional architecture encodes not only the semantic similarity of the translation pair, but also the context containing the phrase in the source language. Therefore, our approach is able to capture context-dependent semantic similarities of translation pairs. We adopt a curriculum learning strategy to train the model: we classify the training examples into easy, medium, and difficult categories, and gradually build the ability of representing phrase and sentence level context by using training examples from easy to difficult. Experimental results show that our approach significantly outperforms the baseline system by up to 1.4 BLEU points.

The usefulness of translation quality estimation (QE) to increase productivity in a computer-assisted translation (CAT) framework is a widely held assumption (Specia, 2011; Huang et al., 2014). So far, however, the validity of this assumption has not been yet demonstrated through sound evaluations in realistic settings. To this aim, we report on an evaluation involving professional translators operating with a CAT tool in controlled but natural conditions. Contrastive experiments are carried out by measuring post-editing time differences when: i) translation suggestions are presented together with binary quality estimates, and ii) the same suggestions are presented without quality indicators. Translators’ productivity in the two conditions is analysed in a principled way, accounting for the main factors (e.g. differences in translators’ behaviour, quality of the suggestions) that directly impact on time measurements. While the general assumption about the usefulness of QE is verified, significance testing results reveal that real productivity gains can be observed only under specific conditions.

Recently, a variety of representation learning approaches have been developed in the literature to induce latent generalizable features across two domains. In this paper, we extend the standard hidden Markov models (HMMs) to learn distributed state representations to improve cross-domain prediction performance. We reformulate the HMMs by mapping each discrete hidden state to a distributed representation vector and employ an expectationmaximization algorithm to jointly learn distributed state representations and model parameters. We empirically investigate the proposed model on cross-domain part-ofspeech tagging and noun-phrase chunking tasks. The experimental results demonstrate the effectiveness of the distributed HMMs on facilitating domain adaptation.

Rumours on social media exhibit complex temporal patterns. This paper develops a model of rumour prevalence using a point process, namely a log-Gaussian Cox process, to infer an underlying continuous temporal probabilistic model of post frequencies. To generalize over different rumours, we present a multi-task learning method parametrized by the text in posts which allows data statistics to be shared between groups of similar rumours. Our experiments demonstrate that our model outperforms several strong baseline methods for rumour frequency prediction evaluated on tweets from the 2014 Ferguson riots.