NAACL 2015 Papers

Monolingual alignment models have been shown to boost the performance of question answering systems by ”bridging the lexical chasm” between questions and answers. The main limitation of these approaches is that they require semistructured training data in the form of question-answer pairs, which is difficult to obtain in specialized domains or lowresource languages. We propose two inexpensive methods for training alignment models solely using free text, by generating artificial question-answer pairs from discourse structures. Our approach is driven by two representations of discourse: a shallow sequential representation, and a deep one based on Rhetorical Structure Theory. We evaluate the proposed model on two corpora from different genres and domains: one from Yahoo! Answers and one from the biology domain, and two types of non-factoid questions: manner and reason. We show that these alignment models trained directly from discourse structures imposed on free text improve performance considerably over an information retrieval baseline and a neural network language model trained on the same data.

Research on entity linking has considered a broad range of text, including newswire, blogs and web documents in multiple languages. However, the problem of entity linking for spoken language remains unexplored. Spoken language obtained from automatic speech recognition systems poses different types of challenges for entity linking; transcription errors can distort the context, and named entities tend to have high error rates. We propose features to mitigate these errors and evaluate the impact of ASR errors on entity linking using a new corpus of entity linked broadcast news transcripts.

We present an intuitive and effective method for inducing style scores on words and phrases. We exploit signal in a phrase’s rate of occurrence across stylistically contrasting corpora, making our method simple to implement and efficient to scale. We show strong results both intrinsically, by correlation with human judgements, and extrinsically, in applications to genre analysis and paraphrasing.

This work improves monolingual sentence alignment for text simplification, specifically for text in standard and simple Wikipedia. We introduce a method that improves over past efforts by using a greedy (vs. ordered) search over the document and a word-level semantic similarity score based on Wiktionary (vs. WordNet) that also accounts for structural similarity through syntactic dependencies. Experiments show improved performance on a hand-aligned set, with the largest gain coming from structural similarity. Resulting datasets of manually and automatically aligned sentence pairs are made available.

A portmanteau is a type of compound word that fuses the sounds and meanings of two component words; for example, “frenemy” (friend + enemy) or “smog” (smoke + fog). We develop a system, including a novel multitape FST, that takes an input of two words and outputs possible portmanteaux. Our system is trained on a list of known portmanteaux and their component words, and achieves 45% exact matches in cross-validated experiments.

We present a novel response generation system that can be trained end to end on large quantities of unstructured Twitter conversations. A neural network architecture is used to address sparsity issues that arise when integrating contextual information into classic statistical models, allowing the system to take into account previous dialog utterances. Our dynamic-context generative models show consistent gains over both context-sensitive and non-context-sensitive Machine Translation and Information Retrieval baselines.

Inferring latent attributes of online users has many applications in public health, politics, and marketing. Most existing approaches rely on supervised learning algorithms, which require manual data annotation and therefore are costly to develop and adapt over time. In this paper, we propose a lightly supervised approach based on label regularization to infer the age, ethnicity, and political orientation of Twitter users. Our approach learns from a heterogeneous collection of soft constraints derived from Census demographics, trends in baby names, and Twitter accounts that are emblematic of class labels. To counteract the imprecision of such constraints, we compare several constraint selection algorithms that optimize classification accuracy on a tuning set. We find that using no user-annotated data, our approach is within 2% of a fully supervised baseline for three of four tasks. Using a small set of labeled data for tuning further improves accuracy on all tasks.

Content analysis, a widely-applied social science research method, is increasingly being supplemented by topic modeling. However, while the discourse on content analysis centers heavily on reproducibility, computer scientists often focus more on scalability and less on coding reliability, leading to growing skepticism on the usefulness of topic models for automated content analysis. In response, we introduce TopicCheck, an interactive tool for assessing topic model stability. Our contributions are threefold. First, from established guidelines on reproducible content analysis, we distill a set of design requirements on how to computationally assess the stability of an automated coding process. Second, we devise an interactive alignment algorithm for matching latent topics from multiple models, and enable sensitivity evaluation across a large number of models. Finally, we demonstrate that our tool enables social scientists to gain novel insights into three active research questions.

We address the problem of automatically aligning natural language sentences with corresponding video segments without any direct supervision. Most existing algorithms for integrating language with videos rely on handaligned parallel data, where each natural language sentence is manually aligned with its corresponding image or video segment. Recently, fully unsupervised alignment of text with video has been shown to be feasible using hierarchical generative models. In contrast to the previous generative models, we propose three latent-variable discriminative models for the unsupervised alignment task. The proposed discriminative models are capable of incorporating domain knowledge, by adding diverse and overlapping features. The results show that discriminative models outperform the generative models in terms of alignment accuracy.

We extend the SKIP-GRAM model of Mikolov et al. (2013a) by taking visual information into account. Like SKIP-GRAM, our multimodal models (MMSKIP-GRAM) build vector-based word representations by learning to predict linguistic contexts in text corpora. However, for a restricted set of words, the models are also exposed to visual representations of the objects they denote (extracted from natural images), and must predict linguistic and visual features jointly. The MMSKIP-GRAM models achieve good performance on a variety of semantic benchmarks. Moreover, since they propagate visual information to all words, we use them to improve image labeling and retrieval in the zero-shot setup, where the test concepts are never seen during model training. Finally, the MMSKIP-GRAM models discover intriguing visual properties of abstract words, paving the way to realistic implementations of embodied theories of meaning.

We present a novel method for aligning a sequence of instructions to a video of someone carrying out a task. In particular, we focus on the cooking domain, where the instructions correspond to the recipe. Our technique relies on an HMM to align the recipe steps to the (automatically generated) speech transcript. We then refine this alignment using a state-of-the-art visual food detector, based on a deep convolutional neural network. We show that our technique outperforms simpler techniques based on keyword spotting. It also enables interesting applications, such as automatically illustrating recipes with keyframes, and searching within a video for events of interest.

Automatic headline generation is a sub-task of document summarization with many reported applications. In this study we present a sequence-prediction technique for learning how editors title their news stories. The introduced technique models the problem as a discrete optimization task in a feature-rich space. In this space the global optimum can be found in polynomial time by means of dynamic programming. We train and test our model on an extensive corpus of financial news, and compare it against a number of baselines by using standard metrics from the document summarization domain, as well as some new ones proposed in this work. We also assess the readability and informativeness of the generated titles through human evaluation. The obtained results are very appealing and substantiate the soundness of the approach.

We present a statistical framework to extract information-rich citation sentences that summarise the main contributions of a scientific paper. In a first stage, we automatically discover salient keywords from a paper’s citation summary, keywords that characterise its main contributions. In asecond stage, exploitingthe results of the first stage, we identify citation sentences that best capture the paper’s main contributions. Experimental results show that our approach using methods rooted in quantitative statistics and information theory outperforms the current state-of-the-art systems in scientific paper summarisation.

Syntactic linearization algorithms take a bag of input words and a set of optional constraints, and construct an output sentence and its syntactic derivation simultaneously. The search problem is NP-hard, and the current best results are achieved by bottom-up bestfirst search. One drawback of the method is low efficiency; and there is no theoretical guarantee that a full sentence can be found within bounded time. We propose an alternative algorithm that constructs output structures from left to right using beam-search. The algorithm is based on incremental parsing algorithms. We extend the transition system so that word ordering is performed in addition to syntactic parsing, resulting in a linearization system that runs in guaranteed quadratic time. In standard evaluations, our system runs an order of magnitude faster than a state-of-the-art baseline using best-first search, with improved accuracies.

Convolutional neural network (CNN) is a neural network that can make use of the internal structure of data such as the 2D structure of image data. This paper studies CNN on text categorization to exploit the 1D structure (namely, word order) of text data for accurate prediction. Instead of using low-dimensional word vectors as input as is often done, we directly apply CNN to high-dimensional text data, which leads to directly learning embedding of small text regions for use in classification. In addition to a straightforward adaptation of CNN from image to text, a simple but new variation which employs bag-of-word conversion in the convolution layer is proposed. An extension to combine multiple convolution layers is also explored for higher accuracy. The experiments demonstrate the effectiveness of our approach in comparison with state-of-the-art methods.

Character n-grams have been identified as the most successful feature in both singledomain and cross-domain Authorship Attribution (AA), but the reasons for their discriminative value were not fully understood. We identify subgroups of charactern-grams that correspond to linguistic aspects commonly claimed to be covered by these features: morphosyntax, thematic content and style. We evaluate the predictiveness of each of these groups in two AA settings: a single domain setting and a cross-domain setting where multiple topics are present. We demonstrate that characterngrams that capture information about affixes and punctuation account for almost all of the power of character n-grams as features. Our study contributes new insights into the use of n-grams for future AA work and other classification tasks.

In this paper, we apply a weakly-supervised learning approach for slot tagging using conditional random fields by exploiting web search click logs. We extend the constrained lattice training of T¨¨ om et al. (2013) to non-linear conditional random fields in which latent variables mediate between observations and labels. When combined with a novel initialization scheme that leverages unlabeled data, we show that our method gives significant improvement over strong supervised and weakly-supervised baselines.

Text documents of varying nature (e.g., summary documents written by analysts or published, scientific papers) often cite others as a means of providing evidence to support a claim, attributing credit, or referring the reader to related work. We address the problem of predicting a document’s cited sources by introducing a novel, discriminative approach which combines a content-based generative model (LDA) with author-based features. Further, our classifier is able to learn the importance and quality of each topic within our corpus – which can be useful beyond this task – and preliminary results suggest its metric is competitive with other standard metrics (Topic Coherence). Our flagship system, Logit-Expanded, provides state-of-the-art performance on the largest corpus ever used for this task.

Parsing full-fledged predicate-argument structures in a deep syntax framework requires graphs to be predicted. Using the DeepBank (Flickinger et al., 2012) and the Predicate-Argument Structure treebank (Miyao and Tsujii, 2005) as a test field, we show how transition-based parsers, extended to handle connected graphs, benefit from the use of topologically different syntactic features such as dependencies, tree fragments, spines or syntactic paths, bringing a much needed context to the parsing models, improving notably over long distance dependencies and elided coordinate structures. By confirming this positive impact on an accurate 2nd-order graph-based parser (Martins and Almeida, 2014), we establish a new state-of-the-art on these data sets.