NAACL 2015 Papers

We present an unsupervised model for inducing signed social networks from the content exchanged across network edges. Inference in this model solves three problems simultaneously: (1) identifying the sign of each edge; (2) characterizing the distribution over content for each edge type; (3) estimating weights for triadic features that map to theoretical models such as structural balance. We apply this model to the problem of inducing the social function of address terms, such as 'Madame', 'comrade', and 'dude'. On a dataset of movie scripts, our system obtains a coherent clustering of address terms, while at the same time making intuitively plausible judgments of the formality of social relations in each film. As an additional contribution, we provide a bootstrapping technique for identifying and tagging address terms in dialogue.

Vector space word representations are learned from distributional information of words in large corpora. Although such statistics are semantically informative, they disregard the valuable information that is contained in semantic lexicons such as WordNet, FrameNet, and the Paraphrase Database. This paper proposes a method for refining vector space representations using relational information from semantic lexicons by encouraging linked words to have similar vector representations, and it makes no assumptions about how the input vectors were constructed. Evaluated on a battery of standard lexical semantic evaluation tasks in several languages, we obtain substantial improvements starting with a variety of word vector models. Our refinement method outperforms prior techniques for incorporating semantic lexicons into word vector training algorithms.

Previous work has debated whether humans make use of hierarchic syntax when processing language (Frank and Bod, 2011; Fossum and Levy, 2012). This paper uses an eye-tracking corpus to demonstrate that hierarchic syntax significantly improves reading time prediction over a strong n-gram baseline. This study shows that an interpolated 5-gram baseline can be made stronger by combining n-gram statistics over entire eye-tracking regions rather than simply using the last n-gram in each region, but basic hierarchic syntactic measures are still able to achieve significant improvements over this improved baseline.

If speakers use language rationally, they should structure their messages to achieve approximately uniform information density (UID), in order to optimize transmission via a noisy channel. Previous work identified a consistent increase in linguistic information across sentences in text as a signature of the UID hypothesis. This increase was derived from a predicted increase in context, but the context itself was not quantified. We use microblog texts from Twitter, tied to a single shared event (the baseball World Series), to quantify both linguistic and non-linguistic context. By tracking changes in contextual information, we predict and identify gradual and rapid changes in information content in response to in-game events. These findings lend further support to the UID hypothesis and highlights the importance of nonlinguistic common ground for language production and processing.

Categories such as ANIMAL or FURNITURE are acquired at an early age and play an important role in processing, organizing, and conveying world knowledge. Theories of categorization largely agree that categories are characterized by features such as function or appearance and that feature and category acquisition go hand-in-hand, however previous work has considered these problems in isolation. We present the first model that jointly learns categories and their features. The set of features is shared across categories, and strength of association is inferred in a Bayesian framework. We approximate the learning environment with natural language text which allows us to evaluate performance on a large scale. Compared to highly engineered pattern-based approaches, our model is cognitively motivated, knowledge-lean, and learns categories and features which are perceived by humans as more meaningful.

User-generated passwords tend to be memorable, but not secure. A random, computergenerated 60-bit string is much more secure. However, users cannot memorize random 60bit strings. In this paper, we investigate methods for converting arbitrary bit strings into English word sequences (both prose and poetry), and we study their memorability and other properties.

Most of the current approaches to sentiment analysis of product reviews are dependent on lexical sentiment information and proceed in a bottom-up way, adding new layers of features to lexical data. In this paper, we maintain that a typical product review is not a bag of sentiments, but a narrative with an underlying structure and reoccurring patterns, which allows us to predict its sentiments knowing only its general polarity and discourse cues that occur in it. We hypothesize that knowing only the review’s score and its discourse patterns would allow us to accurately predict the sentiments of its individual sentences. The experiments we conducted prove this hypothesis and show a substantial improvement over the lexical baseline.

This paper proposes a novel approach to sentiment analysis that leverages work in sociology on symbolic interactionism. The proposed approach uses Affect Control Theory (ACT) to analyze readers’ sentiment towards factual (objective) content and towards its entities (subject and object). ACT is a theory of affective reasoning that uses empirically derived equations to predict the sentiments and emotions that arise from events. This theory relies on several large lexicons of words with affective ratings in a three-dimensional space of evaluation, potency, and activity (EPA). The equations and lexicons of ACT were evaluated on a newly collected news-headlines corpus. ACT lexicon was expanded using a label propagation algorithm, resulting in 86,604 new words. The predicted emotions for each news headline was then computed using the augmented lexicon and ACT equations. The results had a precision of 82%, 79%, and 68% towards the event, the subject, and object, respectively. These results are significantly higher than those of standard sentiment analysis techniques.

This paper introduces a task of identifying and semantically classifying lexical expressions in running text. We investigate the online reviews genre, adding semantic supersense annotations to a 55,000 word English corpus that was previously annotated for multiword expressions. The noun and verb supersenses apply to full lexical expressions, whether single- or multiword. We then present a sequence tagging model that jointly infers lexical expressions and their supersenses. Results show that even with our relatively small training corpus in a noisy domain, the joint task can be performed to attain 70% class labeling F1.

Translation models often fail to generate good translations for infrequent words or phrases. Previous work attacked this problem by inducing new translation rules from monolingual data with a semi-supervised algorithm. However, this approach does not scale very well since it is very computationally expensive to generate new translation rules for only a few thousand sentences. We propose a much faster and simpler method that directly hallucinates translation rules for infrequent phrases based on phrases with similar continuous representations for which a translation is known. To speed up the retrieval of similar phrases, we investigate approximated nearest neighbor search with redundant bit vectors which we find to be three times faster and significantly more accurate than locality sensitive hashing. Our approach of learning new translation rules improves a phrase-based baseline by up to 1.6 BLEU on Arabic-English translation, it is three-orders of magnitudes faster than existing semi-supervised methods and 0.5 BLEU more accurate.

This work presents a flexible and efficient discriminative training approach for statistical machine translation. We propose to use the RPROP algorithm for optimizing a maximum expected BLEU objective and experimentally compare it to several other updating schemes. It proves to be more efficient and effective than the previously proposed growth transformation technique and also yields better results than stochastic gradient descent and AdaGrad. We also report strong empirical results on two large scale tasks, namely BOLT Chinese!English and WMT German!English, where our final systems outperform results reported by Setiawan and Zhou (2013) and on matrix.statmt.org. On the WMT task, discriminative training is performed on the full training data of 4M sentence pairs, which is unsurpassed in the literature.

Given a (static) scene, a human can effortlessly describe what is going on (who is doing what to whom, how, and why). The process requires knowledge about the world, how it is perceived, and described. In this paper we study the problem of interpreting and verbalizing visual information using abstract scenes created from collections of clip art images. We propose a model inspired by machine translation operating over a large parallel corpus of visual relations and linguistic descriptions. We demonstrate that this approach produces human-like scene descriptions which are both fluent and relevant, outperforming a number of competitive alternatives based on templates, sentence-based retrieval, and a multimodal neural language model.

Solving the visual symbol grounding problem has long been a goal of artificial intelligence. The field appears to be advancing closer to this goal with recent breakthroughs in deep learning for natural language grounding in static images. In this paper, we propose to translate videos directly to sentences using a unified deep neural network with both convolutional and recurrent structure. Described video datasets are scarce, and most existing methods have been applied to toy domains with a small vocabulary of possible words. By transferring knowledge from 1.2M+ images with category labels and 100,000+ images with captions, our method is able to create sentence descriptions of open-domain videos with large vocabularies. We compare our approach with recent work using language generation metrics, subject, verb, and object prediction accuracy, and a human evaluation.

While the effect of various lexical, syntactic, semantic and stylistic features have been addressed in persuasive language from a computational point of view, the persuasive effect of phonetics has received little attention. By modeling a notion of euphony and analyzing four datasets comprising persuasive and non-persuasive sentences in different domains (political speeches, movie quotes, slogans and tweets), we explore the impact of sounds on different forms of persuasiveness. We conduct a series of analyses and prediction experiments within and across datasets. Our results highlight the positive role of phonetic devices on persuasion.

The subconscious influence of framing on perceptions of political issues is well-document in political science and communication research. A related line of work suggests that drawing attention to framing may help reduce such framing effects by enabling frame reflection, critical examination of the framing underlying an issue. However, definite guidance on how to identify framing does not exist. This paper presents a technique for identifying frame-invoking language. The paper first describes a human subjects pilot study that explores how individuals identify framing and informs the design of our technique. The paper then describes our data collection and annotation approach. Results show that the best performing classifiers achieve performance comparable to that of human annotators, and they indicate which aspects of language most pertain to framing. Both technical and theoretical implications are discussed.

Cross-lingual text classification is a major challenge in natural language processing, since often training data is available in only one language (target language), but not available for the language of the document we want to classify (source language). Here, we propose a method that only requires a bilingual dictionary to bridge the language gap. Our proposed probabilistic model allows us to estimate translation probabilities that are conditioned on the whole source document. The assumption of our probabilistic model is that each document can be characterized by a distribution over topics that help to solve the translation ambiguity of single words. Using the derived translation probabilities, we then calculate the expected word frequency of each word type in the target language. Finally, these expected word frequencies can be used to classify the source text with any classifier that was trained using only target language documents. Our experiments confirm the usefulness of our proposed method.

This paper presents CROWN, an automatically constructed extension of WordNet that augments its taxonomy with novel lemmas from Wiktionary. CROWN fills the important gap in WordNet’s lexicon for slang, technical, and rare lemmas, and more than doubles its current size. In two evaluations, we demonstrate that the construction procedure is accurate and has a significant impact on a WordNet-based algorithm encountering novel lemmas.

Our research aims to extract information about medication use from veterinary discussion forums. We introduce the task of categorizing information about medication use to determine whether a doctor has prescribed medication, changed protocols, observed effects, or stopped use of a medication. First, we create a medication detector for informal veterinary texts and show that features derived from the Web can be very powerful. Second, we create classifiers to categorize each medication mention with respect to six categories. We demonstrate that this task benefits from a rich linguistic feature set, domain-specific semantic features produced by a weakly supervised semantic tagger, and balanced self-training.

Lemmatization for the Sumerian language, compared to the modern languages, is much more challenging due to that it is a long dead language, highly skilled language experts are extremely scarce and more and more Sumerian texts are coming out. This paper describes how our unsupervised Sumerian named-entity recognition (NER) system helps to improve the lemmatization of the Cuneiform Digital Library Initiative (CDLI), a specialist database of cuneiform texts, from the Ur III period. Experiments show that a promising improvement in personal name annotation in such texts and a substantial reduction in expert annotation effort can be achieved by leveraging our system with minimal seed annotation.