We show that jointly performing semantic role labeling (SRL) on bitext can improve SRL results on both sides. In our approach, we use monolingual SRL systems to produce argument candidates for predicates in bitext at first. Then, we simultaneously generate SRL results for two sides of bitext using our joint inference model. Our model prefers the bilingual SRL result that is not only reasonable on each side of bitext, but also has more consistent argument structures between two sides. To evaluate the consistency between two argument structures, we also formulate a log-linear model to compute the probability of aligning two arguments. We have experimented with our model on Chinese-English parallel Prop-Bank data. Using our joint inference model, F1 scores of SRL results on Chinese and English text achieve 79.53 % and 77.87 % respectively, which are 1.52 and 1.74 points higher than the results of baseline monolingual SRL combination systems respectively. 1

Most existing systems for Chinese Semantic Role Labeling (SRL) make use of full syntactic parses. In this paper, we evaluate SRL methods that take partial parses as inputs. We first extend the study on Chinese shallow parsing presented in (Chen et al., 2006) by raising a set of additional features. On the basis of our shallow parser, we implement SRL systems which cast SRL as the classification of syntactic chunks with IOB2 representation for semantic roles (i.e. semantic chunks). Two labeling strategies are presented: 1) directly tagging semantic chunks in onestage, and 2) identifying argument boundaries as a chunking task and labeling their semantic types as a classification task. For both methods, we present encouraging results, achieving significant improvements over the best reported SRL performance in the literature. Additionally, we put forward a rule-based algorithm to automatically acquire Chinese verb formation, which is empirically shown to enhance SRL. 1

Natural language systems trained on labeled data from one domain do not perform well on other domains. Most adaptation algorithms proposed in the literature train a new model for the new domain using unlabeled data. However, it is time consuming to retrain big models or pipeline systems. Moreover, the domain of a new target sentence may not be known, and one may not have significant amount of unlabeled data for every new domain. To pursue the goal of an Open Domain NLP (train once, test anywhere), we propose ADUT (ADaptation Using label-preserving Transformation), an approach that avoids the need for retraining and does not require knowledge of the new domain, or any data from it. Our approach applies simple label-preserving transformations to the target text so that the transformed text is more similar to the training domain; it then applies the existing model on the transformed sentences and combines the predictions to produce the desired prediction on the target text. We instantiate ADUT for the case of Semantic Role Labeling (SRL) and show that it compares favorably with approaches that retrain their model on the target domain. Specifically, this “on the fly ” adaptation approach yields 13 % error reduction for a single parse system when adapting from the news wire text to fiction. 1

system. We learn separate selectional preferences for noun phrases and prepositional phrases and we integrate them in a state-of-the-art SR classification system both in the form of features and individual class predictors. We show that the inclusion of the refined SPs yields statistically significant improvements on both in domain and out of domain data (14.07 % and 11.67 % error reduction, respectively). The key factor for success is the combination of several SP methods with the original classification model using metaclassification. 1

This paper presents a model that extends semantic role labeling. Existing approaches independently analyze relations expressed by verb predicates or those expressed as nominalizations. However, sentences express relations via other linguistic phenomena as well. Furthermore, these phenomena interact with each other, thus restricting the structures they articulate. In this paper, we use this intuition to define a joint inference model that captures the inter-dependencies between verb semantic role labeling and relations expressed using prepositions. The scarcity of jointly labeled data presents a crucial technical challenge for learning a joint model. The key strength of our model is that we use existing structure predictors as black boxes. By enforcing consistency constraints between their predictions, we show improvements in the performance of both tasks without retraining the individual models. 1

Study of feature selection to learn SVM

This paper presents a novel featurebased semantic role labeling (SRL) method which uses both constituent and dependency syntactic views. Comparing to the traditional SRL method relying on only one syntactic view, the method has a much richer set of syntactic features. First we select several important constituent-based and dependency-based features from existing studies as basic features. Then, we propose a statistical method to select discriminative combined features which are composed by the basic features. SRL is achieved by using the SVM classifier with both the basic features and the combined features. Experimental results on Chinese Proposition Bank (CPB) show that the method outperforms the traditional constituent-based or dependency-based SRL methods. 1

News tweets that report what is happening have become an important real-time information source. We raise the problem of Semantic Role Labeling (SRL) for news tweets, which is meaningful for fine grained information extraction and retrieval. We present a self-supervised learning approach to train a domain specific SRL system to resolve the problem. A large volume of training data is automatically labeled, by leveraging the existing SRL system on news domain and content similarity between news and news tweets. On a human annotated test set, our system achieves state-of-the-art performance, outperforming the SRL system trained on news. 1

We describe and analyze inference strategies for combining outputs from multiple question answering systems each of which was developed independently. Specifically, we address the DARPA-funded GALE information distillation Year 3 task of finding answers to the 5-Wh questions (who, what, when, where, and why) for each given sentence. The approach we take revolves around determining the best system using discriminative learning. In particular, we train support vector machines with a set of novel features that encode systems ’ capabilities of returning as many correct answers as possible. We analyze two combination strategies: one combines multiple systems at the granularity of sentences, and the other at the granularity of individual fields. Our experimental results indicate that the proposed features and combination strategies were able to improve the overall performance by 22 % to 36 % relative to a random selection, 16 % to 35 % relative to a majority voting scheme, and 15 % to 23 % relative to the best individual system. Index Terms: Question answering, Systems for spoken language understanding

This paper focuses on combining answers generated by a semantic parser that produces semantic role labels (SRLs) and those generated by syntactic parser that produces function tags for answering 5-W questions, i.e., who, what, when, where, and why. We take a probabilistic approach in which a system’s ability to correctly answer 5-W questions is measured with the likelihood that its answers are produced for the given word sequence. This is achieved by training statistical language models (LMs) that are used to predict whether the answers returned by semantic parse or those returned by the syntactic parser are more likely. We evaluated our approach using the OntoNotes dataset. Our experimental results indicate that the proposed LM-based combination strategy was able to improve the performance of the best individual system in terms of both F1 measure and accuracy. Furthermore, the error rates for each question type were also significantly reduced with the help of the proposed approach.