Annotated corpora enable supervised machine learning and data analysis. To reduce the cost of manual annotation, tasks are often assigned to internet workers whose judgments are reconciled by crowdsourcing models. We approach the problem of crowdsourcing using a framework for learning from rich prior knowledge, and we identify a family of crowdsourcing models with the novel ability to combine annotations with differing structures: e.g., document labels and word labels. Annotator judgments are given in the form of the predicted expected value of measurement functions computed over annotations and the data, unifying annotation models. Our model, a specific instance of this framework, compares favorably with previous work. Furthermore, it enables active sample selection, jointly selecting annotator, data item, and annotation structure to reduce annotation effort.

In modern text annotation projects, crowdsourced annotations are often aggregated using item response models or by majority vote. Recently, item response models enhanced with generative data models have been shown to yield substantial benefits over those with conditional or no data models. However, suitable generative data models do not exist for many tasks, such as semantic labeling tasks. When no generative data model exists, we demonstrate that similar benefits may be derived by conditionally modeling documents that have been previously embedded in a semantic space using recent work in vector space models. We use this approach to show state-of-the-art results on a variety of semantic annotation aggregation tasks.

Probabilistic models are a useful means for analyzing large text corpora. Integrating such models with human interaction enables many new use cases. However, adding human interaction to probabilistic models requires inference algorithms which are both fast and accurate. We explore the use of Iterated Conditional Modes as a fast alternative to Gibbs sampling or variational EM. We demonstrate superior performance both in run time and model quality on three different models of text including a DP Mixture of Multinomials for web search result clustering, the Interactive Topic Model, and M OM R ESP , a multinomial crowdsourcing model.

Data annotation in modern practice often involves multiple, imperfect human annotators. Multiple annotations can be used to infer estimates of the ground-truth labels and to estimate individual annotator error characteristics (or reliability). We introduce MomResp, a model that incorporates information from both natural data clusters as well as annotations from multiple annotators to infer ground-truth labels and annotator reliability for the document classification task. We implement this model and show dramatic improvements over majority vote in situations where both annotations are scarce and annotation quality is low as well as in situations where annotators disagree consistently. Because MomResp predictions are subject to label switching, we introduce a solution that finds nearly optimal predicted class reassignments in a variety of settings using only information available to the model at inference time. Although MomResp does not perform well in annotation-rich situations, we show evidence suggesting how this shortcoming may be overcome in future work.

The task of corpus-dictionary linkage (CDL) is to annotate each word in a corpus with a link to an appropriate dictionary entry that documents the sense and usage of the word. Corpus-dictionary linked resources include concordances, dictionaries with word usage examples, and corpora annotated with lemmas or word-senses. Such CDL resources are essential in learning a language and in linguistic research, translation, and philology. Lemmatization is a common approximation to automating corpus-dictionary linkage, where lemmas are treated as dictionary entry headwords. We intend to use data-driven lemmatization models to provide machine assistance to human annotators in the form of pre-annotations, and thereby reduce the costs of CDL annotation. In this work we adapt the discriminative string transducer DirecTL+ to perform lemmatization for classical Syriac, a low-resource language. We compare the accuracy of DirecTL+ with the Morfette discriminative lemmatizer. DirecTL+ achieves 96.92% overall accuracy but only by a margin of 0.86% over Morfette at the cost of a longer time to train the model. Error analysis on the models provides guidance on how to apply these models in a machine assistance setting for corpus-dictionary linkage.

We describe an under-studied problem in language resource management: that of providing automatic assistance to annotators working in exploratory settings. When no satisfactory tagset already exists, such as in under-resourced or undocumented languages, it must be developed iteratively while annotating data. This process naturally gives rise to a sequence of datasets, each annotated differently. We argue that this problem is best regarded as a transfer learning problem with multiple source tasks. Using part-of-speech tagging data with simulated exploratory tagsets, we demonstrate that even simple transfer learning techniques can significantly improve the quality of pre-annotations in an exploratory annotation.

Manual annotation of large textual corpora can be cost-prohibitive, especially for rare and under-resourced languages. One potential solution is pre-annotation: asking human annotators to correct sentences that have already been annotated, usually by a machine. Another potential solution is correction propagation: using annotator corrections to bad pre-annotations to dynamically improve to the remaining pre-annotations within the current sentence. The research presented in this paper employs a controlled user study to discover under what conditions these two machine-assisted annotation techniques are effective in increasing annotator speed and accuracy and thereby reducing the cost for the task of morphologically annotating texts written in classical Syriac. A preliminary analysis of the data indicates that pre-annotations improve annotator accuracy when they are at least 60% accurate, and annotator speed when they are at least 80% accurate. This research constitutes the first systematic evaluation of pre-annotation and correction propagation together in a controlled user study.

We introduce CCASH (Cost-Conscious Annotation Supervised by Humans), an extensible web application framework for cost-efficient annotation. CCASH provides a framework in which cost-efficient annotation methods such as Active Learning can be explored via user studies and afterwards applied to large annotation projects. CCASH’s architecture is described as well as the technologies that it is built on. CCASH allows custom annotation tasks to be built from a growing set of useful annotation widgets. It also allows annotation methods (such as AL) to be implemented in any language. Being a web application framework, CCASH offers secure centralized data and annotation storage and facilitates collaboration among multiple annotations. By default it records timing information about each annotation and provides facilities for recording custom statistics. The CCASH framework has been used to evaluate a novel annotation strategy presented in a concurrently published paper, and will be used in the future to annotate a large Syriac corpus.

Expert human input can contribute in various ways to facilitate automatic annotation of natural language text. For example, a part-of-speech tagger can be trained on labeled input provided offline by experts. In addition, expert input can be solicited by way of active learning to make the most of annotator expertise. However, hiring individuals to perform manual annotation is costly both in terms of money and time. This paper reports on a user study that was performed to determine the degree of effect that a part-of-speech dictionary has on a group of subjects performing the annotation task. The user study was conducted using a modular, web-based interface created specifically for text annotation tasks. The user study found that for both native and non-native English speakers a dictionary with greater than 60% coverage was effective at reducing annotation time and increasing annotator accuracy. On the basis of this study, we predict that using a part-of-speech tag dictionary with coverage greater than 60% can reduce the cost of annotation in terms of both time and money.

Fixed, limited budgets often constrain the amount of expert annotation that can go into the construction of annotated corpora. Estimating the cost of annotation is the first step toward using annotation resources wisely. We present here a study of the cost of annotation. This study includes the participation of annotators at various skill levels and with varying backgrounds. Conducted over the web, the study consists of tests that simulate machine-assisted pre-annotation, requiring correction by the annotator rather than annotation from scratch. The study also includes tests representative of an annotation scenario involving Active Learning as it progresses from a naïve model to a knowledgeable model; in particular, annotators encounter pre-annotation of varying degrees of accuracy. The annotation interface lists tags considered likely by the annotation model in preference to other tags. We present the experimental parameters of the study and report both descriptive and inferential statistics on the results of the study. We conclude with a model for estimating the hourly cost of annotation for annotators of various skill levels. We also present models for two granularities of annotation: sentence at a time and word at a time.

This paper presents the French implementation of Amalgam, a machine-learned sentence realization system. It presents in some detail two of the machine-learned models employed in Amalgam and shows how linguistic intuition and knowledge can be combined with statistical techniques to improve the performance of the models.