Annotating clause boundary labels to the written composition corpus of Japanese elementary and junior high school students

To evaluate the development of children’s writing ability, it is necessary not only to examine quantitative indices such as the dependency distance, but also to inquiry the types of structures they use. We conducted clause boundary labeling using Support Vector Machine (SVM) on a corpus of Japanese students' compositions to investigate the change in the tendency of clause use with the progression of school age. The analysis of clause label frequency per sentence exhibited an increase in attributive clauses, nominal clauses, quotation clauses, and continuous clauses, and a decrease in parallel clauses, conditional clauses, reason clauses, time clauses, indirect interrogative clauses, and main clauses. The analysis of dependency distance demonstrated that most of the clauses that increased had short dependency distances, while most of the clauses that decreased had long dependency distances, and that the frequency of clauses with small dependency distances increased relatively with increasing school age. In addition, there was a shift in clause selection among functionally similar clauses, such as from “-te” to continuous forms, from “-tara” to “-ba”, and from “-kedo” and “-keredo” to “-ga”. These results suggest a change in the children’s lexical and grammatical choices, from coordinate to subordinate structures, and from spoken to written vocabulary.

Adversarial Transfer for Named Entity Boundary Detection with Pointer Networks

In this paper, we focus on named entity boundary detection, which aims to detect the start and end boundaries of an entity mention in text, without predicting its type. A more accurate and robust detection approach is desired to alleviate error propagation in downstream applications, such as entity linking and fine-grained typing systems. Here, we first develop a novel entity boundary labeling approach with pointer networks, where the output dictionary size depends on the input, which is variable. Furthermore, we propose AT-Bdry, which incorporates adversarial transfer learning into an end-to-end sequence labeling model to encourage domain-invariant representations. More importantly, AT-Bdry can reduce domain difference in data distributions between the source and target domains, via an unsupervised transfer learning approach (i.e., no annotated target-domain data is necessary). We conduct Formal Text to Formal Text, Formal Text to Informal Text and ablation evaluations on five benchmark datasets. Experimental results show that AT-Bdry achieves state-of-the-art transferring performance against recent baselines.

On the readability of leaders in boundary labeling

External labeling deals with annotating features in images with labels that are placed outside of the image and are connected by curves (so-called leaders) to the corresponding features. While external labeling has been extensively investigated from a perspective of automatization, the research on its readability has been neglected. In this article, we present the first formal user study on the readability of leader types in boundary labeling, a special variant of external labeling that considers rectangular image contours. We consider the four most studied leader types (straight, L-shaped, diagonal, and S-shaped) with respect to their performance, that is, whether and how fast a viewer can assign a feature to its label and vice versa. We give a detailed analysis of the results regarding the readability of the four models and discuss their aesthetic qualities based on the users’ preference judgments and interviews. As a consequence of our experiment, we can generally recommend L-shaped leaders as the best compromise between measured task performance and subjective preference ratings, while straight and diagonal leaders received mixed ratings in the two measures. S-shaped leaders are generally not recommended from a practical point of view.