ACL 2015 Papers

This paper studies the use of structural representations for learning relations between pairs of short texts (e.g., sentences or paragraphs) of the kind: the second text answers to, or conveys exactly the same information of, or is implied by, the first text. Engineering effective features that can capture syntactic and semantic relations between the constituents composing the target text pairs is rather complex. Thus, we define syntactic and semantic structures representing the text pairs and then apply graph and tree kernels to them for automatically engineering features in Support Vector Machines. We carry out an extensive comparative analysis of stateof-the-art models for this type of relational learning. Our findings allow for achieving the highest accuracy in two different and important related tasks, i.e., Paraphrasing Identification and Textual Entailment Recognition.

We focus on the task of interpreting complex natural language instructions to a robot, in which we must ground high-level commands such as microwave the cup to low-level actions such as grasping. Previous approaches that learn a lexicon during training have inadequate coverage at test time, and pure search strategies cannot handle the exponential search space. We propose a new hybrid approach that leverages the environment to induce new lexical entries at test time, even for new verbs. Our semantic parsing model jointly reasons about the text, logical forms, and environment over multi-stage instruction sequences. We introduce a new dataset and show that our approach is able to successfully ground new verbs such as distribute, mix, arrange to complex logical forms, each containing up to four predicates.

The Abstract Meaning Representation (AMR) is a representation for open-domain rich semantics, with potential use in fields like event extraction and machine translation. Node generation, typically done using a simple dictionary lookup, is currently an important limiting factor in AMR parsing. We propose a small set of actions that derive AMR subgraphs by transformations on spans of text, which allows for more robust learning of this stage. Our set of construction actions generalize better than the previous approach, and can be learned with a simple classifier. We improve on the previous state-of-the-art result for AMR parsing, boosting end-to-end performance by 3 F$_1$ on both the LDC2013E117 and LDC2014T12 datasets.

We introduce C-PHRASE, a distributional semantic model that learns word representations by optimizing context prediction for phrases at all levels in a syntactic tree, from single words to full sentences. C-PHRASE outperforms the state-of-theart C-BOW model on a variety of lexical tasks. Moreover, since C-PHRASE word vectors are induced through a compositional learning objective (modeling the contexts of words combined into phrases), when they are summed, they produce sentence representations that rival those generated by ad-hoc compositional models.

Existing methods for Japanese predicate argument structure (PAS) analysis identify case arguments of each predicate without considering interactions between the target PAS and others in a sentence. However, the argument structures of the predicates in a sentence are semantically related to each other. This paper proposes new methods for Japanese PAS analysis to jointly identify case arguments of all predicates in a sentence by (1) modeling multiple PAS interactions with a bipartite graph and (2) approximately searching optimal PAS combinations. Performing experiments on the NAIST Text Corpus, we demonstrate that our joint analysis methods substantially outperform a strong baseline and are comparable to previous work.

Selectional preferences (SPs) are widely used in NLP as a rich source of semantic information. While SPs have been traditionally induced from textual data, human lexical acquisition is known to rely on both linguistic and perceptual experience. We present the first SP learning method that simultaneously draws knowledge from text, images and videos, using image and video descriptions to obtain visual features. Our results show that it outperforms linguistic and visual models in isolation, as well as the existing SP induction approaches.

Many high level natural language processing problems can be framed as determining if two given sentences are a rewriting of each other. In this paper, we propose a class of kernel functions, referred to as type-enriched string rewriting kernels, which, used in kernel-based machine learning algorithms, allow to learn sentence rewritings. Unlike previous work, this method can be fed external lexical semantic relations to capture a wider class of rewriting rules. It also does not assume preliminary syntactic parsing but is still able to provide a unified framework to capture syntactic structure and alignments between the two sentences. We experiment on three different natural sentence rewriting tasks and obtain state-of-the-art results for all of them.

Most previous work of text normalization on informal text made a strong assumption that the system has already known which tokens are non-standard words (NSW) and thus need normalization. However, this is not realistic. In this paper, we propose a method for NSW detection. In addition to the information based on the dictionary, e.g., whether a word is out-ofvocabulary (OOV), we leverage novel information derived from the normalization results for OOV words to help make decisions. Second, this paper investigates two methods using NSW detection results for named entity recognition (NER) in social media data. One adopts a pipeline strategy, and the other uses a joint decoding fashion. We also create a new data set with newly added normalization annotation beyond the existing named entity labels. This is the first data set with such annotation and we release it for research purpose. Our experiment results demonstrate the effectiveness of our NSW detection method and the benefit of NSW detection for NER. Our proposed methods perform better than the state-of-the-art NER system.

In this paper, we propose a syllable-based method for tweet normalization to study the cognitive process of non-standard word creation in social media. Assuming that syllable plays a fundamental role in forming the non-standard tweet words, we choose syllable as the basic unit and extend the conventional noisy channel model by incorporating the syllables to represent the word-to-word transitions at both word and syllable levels. The syllables are used in our method not only to suggest more candidates, but also to measure similarity between words. Novelty of this work is three-fold: First, to the best of our knowledge, this is an early attempt to explore syllables in tweet normalization. Second, our proposed normalization method relies on unlabeled samples, making it much easier to adapt our method to handle non-standard words in any period of history. And third, we conduct a series of experiments and prove that the proposed method is advantageous over the state-of-art solutions for tweet normalization.

We investigate multiple many-to-many alignments as a primary step in integrating supplemental information strings in string transduction. Besides outlining DP based solutions to the multiple alignment problem, we detail an approximation of the problem in terms of multiple sequence segmentations satisfying a coupling constraint. We apply our approach to boosting baseline G2P systems using homogeneous as well as heterogeneous sources of supplemental information.

This paper presents the NL2KR platform to build systems that can translate text to different formal languages. It is freelyavailable1, customizable, and comes with an Interactive GUI support that is useful in the development of a translation system. Our key contribution is a userfriendly system based on an interactive multistage learning algorithm. This effective algorithm employs Inverse-λ, Generalization and user provided dictionary to learn new meanings of words from sentences and their representations. Using the learned meanings, and the Generalization approach, it is able to translate new sentences. NL2KR is evaluated on two standard corpora, Jobs and GeoQuery and it exhibits state-of-the-art performance on both of them.

We develop an approach for generating deep (i.e, high-level) comprehension questions from novel text that bypasses the myriad challenges of creating a full semantic representation. We do this by decomposing the task into an ontologycrowd-relevance workflow, consisting of first representing the original text in a low-dimensional ontology, then crowdsourcing candidate question templates aligned with that space, and finally ranking potentially relevant templates for a novel region of text. If ontological labels are not available, we infer them from the text. We demonstrate the effectiveness of this method on a corpus of articles from Wikipedia alongside human judgments, and find that we can generate relevant deep questions with a precision of over 85% while maintaining a recall of 70%.

Using natural language to write programs is a touchstone problem for computational linguistics. We present an approach that learns to map natural-language descriptions of simple “if-then” rules to executable code. By training and testing on a large corpus of naturally-occurring programs (called “recipes”) and their natural language descriptions, we demonstrate the ability to effectively map language to code. We compare a number of semantic parsing approaches on the highly noisy training data collected from ordinary users, and find that loosely synchronous systems perform best.

Stubs on Wikipedia often lack comprehensive information. The huge cost of editing Wikipedia and the presence of only a limited number of active contributors curb the consistent growth of Wikipedia. In this work, we present WikiKreator, a system that is capable of generating content automatically to improve existing stubs on Wikipedia. The system has two components. First, a text classifier built using topic distribution vectors is used to assign content from the web to various sections on a Wikipedia article. Second, we propose a novel abstractive summarization technique based on an optimization framework that generates section-specific summaries for Wikipedia stubs. Experiments show that WikiKreator is capable of generating well-formed informative content. Further, automatically generated content from our system have been appended to Wikipedia stubs and the content has been retained successfully proving the effectiveness of our approach.

Deception detection has been formulated as a supervised binary classification problem on single documents. However, in daily life, millions of fraud cases involve detailed conversations between deceivers and victims. Deceivers may dynamically adjust their deceptive statements according to the reactions of victims. In addition, people may form groups and collaborate to deceive others. In this paper, we seek to identify deceptive groups from their conversations. We propose a novel subgroup detection method that combines linguistic signals and signed network analysis for dynamic clustering. A social-elimination game called Killer Game is introduced as a case study1. Experimental results demonstrate that our approach significantly outperforms human voting and state-of-theart subgroup detection methods at dynamically differentiating the deceptive groups from truth-tellers.

The quality of statistical machine translation performed with phrase based approaches can be increased by permuting the words in the source sentences in an order which resembles that of the target language. We propose a class of recurrent neural models which exploit source-side dependency syntax features to re-order the words into a target-like order. We evaluate these models on the German-to-English language pair, showing significant improvements over a phrase-based Moses baseline, obtaining a quality similar or superior to that of hand-coded syntactical reordering rules.

We introduce into Bayesian decipherment a base distribution derived from similarities of word embeddings. We use Dirichlet multinomial regression (Mimno and McCallum, 2012) to learn a mapping between ciphertext and plaintext word embeddings from non-parallel data. Experimental results show that the base distribution is highly beneficial to decipherment, improving state-of-the-art decipherment accuracy from 45.8% to 67.4% for Spanish/English, and from 5.1% to 11.2% for Malagasy/English.

Modern statistical machine translation (SMT) systems usually use a linear combination of features to model the quality of each translation hypothesis. The linear combination assumes that all the features are in a linear relationship and constrains that each feature interacts with the rest features in an linear manner, which might limit the expressive power of the model and lead to a under-fit model on the current data. In this paper, we propose a non-linear modeling for the quality of translation hypotheses based on neural networks, which allows more complex interaction between features. A learning framework is presented for training the non-linear models. We also discuss possible heuristics in designing the network structure which may improve the non-linear learning performance. Experimental results show that with the basic features of a hierarchical phrase-based machine translation system, our method produce translations that are better than a linear model.

We achieve significant improvements in several syntax-based machine translation experiments using a string-to-tree variant of multi bottom-up tree transducers. Our new parameterized rule extraction algorithm extracts string-to-tree rules that can be discontiguous and non-minimal in contrast to existing algorithms for the tree-to-tree setting. The obtained models significantly outperform the string-to-tree component of the Moses framework in a large-scale empirical evaluation on several known translation tasks. Our linguistic analysis reveals the remarkable benefits of discontiguous and non-minimal rules.