We introduce the Situated Corpus Of Understanding Transactions (SCOUT), a multi-modal collection of human-robot dialogue in the task domain of collaborative exploration. The corpus was constructed from multiple Wizard-of-Oz experiments where human participants gave verbal instructions to a remotely-located robot to move and gather information about its surroundings. SCOUT contains 89,056 utterances and 310,095 words from 278 dialogues averaging 320 utterances per dialogue. The dialogues are aligned with the multi-modal data streams available during the experiments: 5,785 images and 30 maps. The corpus has been annotated with Abstract Meaning Representation and Dialogue-AMR to identify the speaker’s intent and meaning within an utterance, and with Transactional Units and Relations to track relationships between utterances to reveal patterns of the Dialogue Structure. We describe how the corpus and its annotations have been used to develop autonomous human-robot systems and enable research in open questions of how humans speak to robots. We release this corpus to accelerate progress in autonomous, situated, human-robot dialogue, especially in the context of navigation tasks where details about the environment need to be discovered.

We explore methods of combining the probability distributions generated by two LLM prompts in order to generate a continuation that is appropriate for both prompts at once. This is a new capability that extends the possibilities for branching and rejoining narratives in games.

NLP systems have shown impressive performance at answering questions by retrieving relevant context. However, with the increasingly large models, it is impossible and often undesirable to constrain models’ knowledge or reasoning to only the retrieved context. This leads to a mismatch between the information that the models access to derive the answer and the information that is available to the user to assess the model predicted answer. In this work, we study how users interact with QA systems in the absence of sufficient information to assess their predictions. Further, we ask whether adding the requisite background helps mitigate users’ over-reliance on predictions. Our study reveals that users rely on model predictions even in the absence of sufficient information needed to assess the model’s correctness. Providing the relevant background, however, helps users better catch model errors, reducing over-reliance on incorrect predictions. On the flip side, background information also increases users’ confidence in their accurate as well as inaccurate judgments. Our work highlights that supporting users’ verification of QA predictions is an important, yet challenging, problem.

Supervised event extraction models require a substantial amount of training data to perform well. However, event annotation requires a lot of human effort and costs much time, which limits the application of existing supervised approaches to new event types. In order to reduce manual labor and shorten the time to build an event extraction system for an arbitrary event ontology, we present a new framework to train such systems much more efficiently without large annotations. Our event trigger labeling model uses a weak supervision approach, which only requires a set of keywords, a small number of examples and an unlabeled corpus, on which our approach automatically collects weakly supervised annotations. Our argument role labeling component performs zero-shot learning, which only requires the names of the argument roles of new event types. The source codes of our event trigger detection1 and event argument extraction2 models are publicly available for research purposes. We also release a dockerized system connecting the two models into an unified event extraction pipeline.

We evaluate an annotation schema for labeling logical fallacy types, originally developed for a crowd-sourcing annotation paradigm, now using an annotation paradigm of two trained linguist annotators. We apply the schema to a variety of different genres of text relating to the COVID-19 pandemic. Our linguist (as opposed to crowd-sourced) annotation of logical fallacies allows us to evaluate whether the annotation schema category labels are sufficiently clear and non-overlapping for both manual and, later, system assignment. We report inter-annotator agreement results over two annotation phases as well as a preliminary assessment of the corpus for training and testing a machine learning algorithm (Pattern-Exploiting Training) for fallacy detection and recognition. The agreement results and system performance underscore the challenging nature of this annotation task and suggest that the annotation schema and paradigm must be iteratively evaluated and refined in order to arrive at a set of annotation labels that can be reproduced by human annotators and, in turn, provide reliable training data for automatic detection and recognition systems.

Modern large-scale Pre-trained Language Models (PLMs) have achieved tremendous success on a wide range of downstream tasks. However, most of the LM pre-training objectives only focus on text reconstruction, but have not sought to learn latent-level interpretable representations of sentences. In this paper, we manage to push the language models to obtain a deeper understanding of sentences by proposing a new pre-training objective, Sparse Latent Typing, which enables the model to sparsely extract sentence-level keywords with diverse latent types. Experimental results show that our model is able to learn interpretable latent type categories in a self-supervised manner without using any external knowledge. Besides, the language model pre-trained with such an objective also significantly improves Information Extraction related downstream tasks in both supervised and few-shot settings. Our code is publicly available at https://github.com/renll/SparseLT.

Generative language models trained on large, diverse corpora can answer questions about a passage by generating the most likely continuation of the passage followed by a question/answer pair. However, accuracy rates vary depending on the type of question asked. In this paper we keep the passage fixed, and test with a wide variety of question types, exploring the strengths and weaknesses of the GPT-3 language model. We provide the passage and test questions as a challenge set for other language models.

To combat COVID-19, both clinicians and scientists need to digest the vast amount of relevant biomedical knowledge in literature to understand the disease mechanism and the related biological functions. We have developed a novel and comprehensive knowledge discovery framework, COVID-KG to extract fine-grained multimedia knowledge elements (entities, relations and events) from scientific literature. We then exploit the constructed multimedia knowledge graphs (KGs) for question answering and report generation, using drug repurposing as a case study. Our framework also provides detailed contextual sentences, subfigures, and knowledge subgraphs as evidence. All of the data, KGs, reports.

Event schemas encode knowledge of stereotypical structures of events and their connections. As events unfold, schemas are crucial to act as a scaffolding. Previous work on event schema induction focuses either on atomic events or linear temporal event sequences, ignoring the interplay between events via arguments and argument relations. We introduce a new concept of Temporal Complex Event Schema: a graph-based schema representation that encompasses events, arguments, temporal connections and argument relations. In addition, we propose a Temporal Event Graph Model that predicts event instances following the temporal complex event schema. To build and evaluate such schemas, we release a new schema learning corpus containing 6,399 documents accompanied with event graphs, and we have manually constructed gold-standard schemas. Intrinsic evaluations by schema matching and instance graph perplexity, prove the superior quality of our probabilistic graph schema library compared to linear representations. Extrinsic evaluation on schema-guided future event prediction further demonstrates the predictive power of our event graph model, significantly outperforming human schemas and baselines by more than 17.8% on HITS@1.

We adopt, evaluate, and improve upon a two-step natural language understanding (NLU) pipeline that incrementally tames the variation of unconstrained natural language input and maps to executable robot behaviors. The pipeline first leverages Abstract Meaning Representation (AMR) parsing to capture the propositional content of the utterance, and second converts this into “Dialogue-AMR,” which augments standard AMR with information on tense, aspect, and speech acts. Several alternative approaches and training datasets are evaluated for both steps and corresponding components of the pipeline, some of which outperform the original. We extend the Dialogue-AMR annotation schema to cover a different collaborative instruction domain and evaluate on both domains. With very little training data, we achieve promising performance in the new domain, demonstrating the scalability of this approach.

This paper examines how Abstract Meaning Representation (AMR) can be utilized for finding answers to research questions in medical scientific documents, in particular, to advance the study of UV (ultraviolet) inactivation of the novel coronavirus that causes the disease COVID-19. We describe the development of a proof-of-concept prototype tool, InfoForager, which uses AMR to conduct a semantic search, targeting the meaning of the user question, and matching this to sentences in medical documents that may contain information to answer that question. This work was conducted as a sprint over a period of six weeks, and reveals both promising results and challenges in reducing the user search time relating to COVID-19 research, and in general, domain adaption of AMR for this task.

This paper describes a schema that enriches Abstract Meaning Representation (AMR) in order to provide a semantic representation for facilitating Natural Language Understanding (NLU) in dialogue systems. AMR offers a valuable level of abstraction of the propositional content of an utterance; however, it does not capture the illocutionary force or speaker’s intended contribution in the broader dialogue context (e.g., make a request or ask a question), nor does it capture tense or aspect. We explore dialogue in the domain of human-robot interaction, where a conversational robot is engaged in search and navigation tasks with a human partner. To address the limitations of standard AMR, we develop an inventory of speech acts suitable for our domain, and present “Dialogue-AMR”, an enhanced AMR that represents not only the content of an utterance, but the illocutionary force behind it, as well as tense and aspect. To showcase the coverage of the schema, we use both manual and automatic methods to construct the “DialAMR” corpus—a corpus of human-robot dialogue annotated with standard AMR and our enriched Dialogue-AMR schema. Our automated methods can be used to incorporate AMR into a larger NLU pipeline supporting human-robot dialogue.

Most of the current cross-lingual transfer learning methods for Information Extraction (IE) have been only applied to name tagging. To tackle more complex tasks such as event extraction we need to transfer graph structures (event trigger linked to multiple arguments with various roles) across languages. We develop a novel share-and-transfer framework to reach this goal with three steps: (1) Convert each sentence in any language to language-universal graph structures; in this paper we explore two approaches based on universal dependency parses and complete graphs, respectively. (2) Represent each node in the graph structure with a cross-lingual word embedding so that all sentences in multiple languages can be represented with one shared semantic space. (3) Using this common semantic space, train event extractors from English training data and apply them to languages that do not have any event annotations. Experimental results on three languages (Spanish, Russian and Ukrainian) without any annotations show this framework achieves comparable performance to a state-of-the-art supervised model trained from more than 1,500 manually annotated event mentions.

We present the first comprehensive, open source multimedia knowledge extraction system that takes a massive stream of unstructured, heterogeneous multimedia data from various sources and languages as input, and creates a coherent, structured knowledge base, indexing entities, relations, and events, following a rich, fine-grained ontology. Our system, GAIA, enables seamless search of complex graph queries, and retrieves multimedia evidence including text, images and videos. GAIA achieves top performance at the recent NIST TAC SM-KBP2019 evaluation. The system is publicly available at GitHub and DockerHub, with a narrated video that documents the system.

Event schemas can guide our understanding and ability to make predictions with respect to what might happen next. We propose a new Event Graph Schema, where two event types are connected through multiple paths involving entities that fill important roles in a coherent story. We then introduce Path Language Model, an auto-regressive language model trained on event-event paths, and select salient and coherent paths to probabilistically construct these graph schemas. We design two evaluation metrics, instance coverage and instance coherence, to evaluate the quality of graph schema induction, by checking when coherent event instances are covered by the schema graph. Intrinsic evaluations show that our approach is highly effective at inducing salient and coherent schemas. Extrinsic evaluations show the induced schema repository provides significant improvement to downstream end-to-end Information Extraction over a state-of-the-art joint neural extraction model, when used as additional global features to unfold instance graphs.

We detail refinements made to Abstract Meaning Representation (AMR) that make the representation more suitable for supporting a situated dialogue system, where a human remotely controls a robot for purposes of search and rescue and reconnaissance. We propose 36 augmented AMRs that capture speech acts, tense and aspect, and spatial information. This linguistic information is vital for representing important distinctions, for example whether the robot has moved, is moving, or will move. We evaluate two existing AMR parsers for their performance on dialogue data. We also outline a model for graph-to-graph conversion, in which output from AMR parsers is converted into our refined AMRs. The design scheme presented here, though task-specific, is extendable for broad coverage of speech acts using AMR in future task-independent work.

In this paper, we tackle the problem of “root extraction” from words in the Semitic language family. A challenge in applying natural language processing techniques to these languages is the data sparsity problem that arises from their rich internal morphology, where the substructure is inherently non-concatenative and morphemes are interdigitated in word formation. While previous automated methods have relied on human-curated rules or multiclass classification, they have not fully leveraged the various combinations of regular, sequential concatenative morphology within the words and the internal interleaving within templatic stems of roots and patterns. To address this, we propose a constrained sequence-to-sequence root extraction method. Experimental results show our constrained model outperforms a variety of methods at root extraction. Furthermore, by enriching word embeddings with resulting decompositions, we show improved results on word analogy, word similarity, and language modeling tasks.

The identification of complex semantic structures such as events and entity relations, already a challenging Information Extraction task, is doubly difficult from sources written in under-resourced and under-annotated languages. We investigate the suitability of cross-lingual structure transfer techniques for these tasks. We exploit relation- and event-relevant language-universal features, leveraging both symbolic (including part-of-speech and dependency path) and distributional (including type representation and contextualized representation) information. By representing all entity mentions, event triggers, and contexts into this complex and structured multilingual common space, using graph convolutional networks, we can train a relation or event extractor from source language annotations and apply it to the target language. Extensive experiments on cross-lingual relation and event transfer among English, Chinese, and Arabic demonstrate that our approach achieves performance comparable to state-of-the-art supervised models trained on up to 3,000 manually annotated mentions: up to 62.6% F-score for Relation Extraction, and 63.1% F-score for Event Argument Role Labeling. The event argument role labeling model transferred from English to Chinese achieves similar performance as the model trained from Chinese. We thus find that language-universal symbolic and distributional representations are complementary for cross-lingual structure transfer.

This paper demonstrates a state-of-the-art end-to-end multilingual (English, Russian, and Ukrainian) knowledge extraction system that can perform entity discovery and linking, relation extraction, event extraction, and coreference. It extracts and aggregates knowledge elements across multiple languages and documents as well as provides visualizations of the results along three dimensions: temporal (as displayed in an event timeline), spatial (as displayed in an event heatmap), and relational (as displayed in entity-relation networks). For our system to further support users’ analyses of causal sequences of events in complex situations, we also integrate a wide range of human moral value measures, independently derived from region-based survey, into the event heatmap. This system is publicly available as a docker container and a live demo.

This paper presents a research platform that supports spoken dialogue interaction with multiple robots. The demonstration showcases our crafted MultiBot testing scenario in which users can verbally issue search, navigate, and follow instructions to two robotic teammates: a simulated ground robot and an aerial robot. This flexible language and robotic platform takes advantage of existing tools for speech recognition and dialogue management that are compatible with new domains, and implements an inter-agent communication protocol (tactical behavior specification), where verbal instructions are encoded for tasks assigned to the appropriate robot.

Most previous efforts toward video captioning focus on generating generic descriptions, such as, “A man is talking.” We collect a news video dataset to generate enriched descriptions that include important background knowledge, such as named entities and related events, which allows the user to fully understand the video content. We develop an approach that uses video meta-data to retrieve topically related news documents for a video and extracts the events and named entities from these documents. Then, given the video as well as the extracted events and entities, we generate a description using a Knowledge-aware Video Description network. The model learns to incorporate entities found in the topically related documents into the description via an entity pointer network and the generation procedure is guided by the event and entity types from the topically related documents through a knowledge gate, which is a gating mechanism added to the model’s decoder that takes a one-hot vector of these types. We evaluate our approach on the new dataset of news videos we have collected, establishing the first benchmark for this dataset as well as proposing a new metric to evaluate these descriptions.

Most previous supervised event extraction methods have relied on features derived from manual annotations, and thus cannot be applied to new event types without extra annotation effort. We take a fresh look at event extraction and model it as a generic grounding problem: mapping each event mention to a specific type in a target event ontology. We design a transferable architecture of structural and compositional neural networks to jointly represent and map event mentions and types into a shared semantic space. Based on this new framework, we can select, for each event mention, the event type which is semantically closest in this space as its type. By leveraging manual annotations available for a small set of existing event types, our framework can be applied to new unseen event types without additional manual annotations. When tested on 23 unseen event types, our zero-shot framework, without manual annotations, achieved performance comparable to a supervised model trained from 3,000 sentences annotated with 500 event mentions.

ScoutBot is a dialogue interface to physical and simulated robots that supports collaborative exploration of environments. The demonstration will allow users to issue unconstrained spoken language commands to ScoutBot. ScoutBot will prompt for clarification if the user’s instruction needs additional input. It is trained on human-robot dialogue collected from Wizard-of-Oz experiments, where robot responses were initiated by a human wizard in previous interactions. The demonstration will show a simulated ground robot (Clearpath Jackal) in a simulated environment supported by ROS (Robot Operating System).

This position paper argues that, while prior work in spatial language understanding for tasks such as robot navigation focuses on mapping natural language into deep conceptual or non-linguistic representations, it is possible to systematically derive regular patterns of spatial language usage from existing lexical-semantic resources. Furthermore, even with access to such resources, effective solutions to many application areas such as robot navigation and narrative generation also require additional knowledge at the syntax-semantics interface to cover the wide range of spatial expressions observed and available to natural language speakers. We ground our insights in, and present our extensions to, an existing lexico-semantic resource, covering 500 semantic classes of verbs, of which 219 fall within a spatial subset. We demonstrate that these extensions enable systematic derivation of regular patterns of spatial language without requiring manual annotation.

Computational visual storytelling produces a textual description of events and interpretations depicted in a sequence of images. These texts are made possible by advances and cross-disciplinary approaches in natural language processing, generation, and computer vision. We define a computational creative visual storytelling as one with the ability to alter the telling of a story along three aspects: to speak about different environments, to produce variations based on narrative goals, and to adapt the narrative to the audience. These aspects of creative storytelling and their effect on the narrative have yet to be explored in visual storytelling. This paper presents a pipeline of task-modules, Object Identification, Single-Image Inferencing, and Multi-Image Narration, that serve as a preliminary design for building a creative visual storyteller. We have piloted this design for a sequence of images in an annotation task. We present and analyze the collected corpus and describe plans towards automation.

We describe a resource derived through extraction of a set of argument realizations from an existing lexical-conceptual structure (LCS) Verb Database of 500 verb classes (containing a total of 9525 verb entries) to include information about realization of arguments for a range of different verb classes. We demonstrate that our extended resource, called STYLUS (SysTematicallY Derived Language USe), enables systematic derivation of regular patterns of language usage without requiring manual annotation. We posit that both spatially oriented applications such as robot navigation and more general applications such as narrative generation require a layered representation scheme where a set of primitives (often grounded in space/motion such as GO) is coupled with a representation of constraints at the syntax-semantics interface. We demonstrate that the resulting resource covers three cases of lexico-semantic operations applicable to both language understanding and language generation.

We propose a method to improve human activity recognition in video by leveraging semantic information about the target activities from an expert-defined linguistic resource, VerbNet. Our hypothesis is that activities that share similar event semantics, as defined by the semantic predicates of VerbNet, will be more likely to share some visual components. We use a deep convolutional neural network approach as a baseline and incorporate linguistic information from VerbNet through multi-task learning. We present results of experiments showing the added information has negligible impact on recognition performance. We discuss how this may be because the lexical semantic information defined by VerbNet is generally not visually salient given the video processing approach used here, and how we may handle this in future approaches.

In this paper, we explore the challenges of building a computational lexicon for Moroccan Darija (MD), an Arabic dialect spoken by over 32 million people worldwide but which only recently has begun appearing frequently in written form in social media. We raise the question of what belongs in such a lexicon and start by describing our work building traditional word-level lexicon entries with their English translations. We then discuss challenges in translating idiomatic MD text that led to creating multi-word expression lexicon entries whose meanings could not be fully derived from the individual words. Finally, we provide a preliminary exploration of constructions to be considered for inclusion in an MD constructicon by translating examples of English constructions and examining their MD counterparts.

This paper identifies stylistic differences in instruction-giving observed in a corpus of human-robot dialogue. Differences in verbosity and structure (i.e., single-intent vs. multi-intent instructions) arose naturally without restrictions or prior guidance on how users should speak with the robot. Different styles were found to produce different rates of miscommunication, and correlations were found between style differences and individual user variation, trust, and interaction experience with the robot. Understanding potential consequences and factors that influence style can inform design of dialogue systems that are robust to natural variation from human users.

Robot-directed communication is variable, and may change based on human perception of robot capabilities. To collect training data for a dialogue system and to investigate possible communication changes over time, we developed a Wizard-of-Oz study that (a) simulates a robot’s limited understanding, and (b) collects dialogues where human participants build a progressively better mental model of the robot’s understanding. With ten participants, we collected ten hours of human-robot dialogue. We analyzed the structure of instructions that participants gave to a remote robot before it responded. Our findings show a general initial preference for including metric information (e.g., move forward 3 feet) over landmarks (e.g., move to the desk) in motion commands, but this decreased over time, suggesting changes in perception.

Recent computational work on Arabic dialect identification has focused primarily on building and annotating corpora written in Arabic script. Arabic dialects however also appear written in Roman script, especially in social media. This paper describes our recent work developing tweet corpora and a token-level classifier that identifies a Romanized Arabic dialect and distinguishes it from French and English in tweets. We focus on Moroccan Darija, one of several spoken vernaculars in the family of Maghrebi Arabic dialects. Even given noisy, code-mixed tweets,the classifier achieved token-level recall of 93.2% on Romanized Arabic dialect, 83.2% on English, and 90.1% on French. The classifier, now integrated into our tweet conversation annotation tool (Tratz et al. 2013), has semi-automated the construction of a Romanized Arabic-dialect lexicon. Two datasets, a full list of Moroccan Darija surface token forms and a table of lexical entries derived from this list with spelling variants, as extracted from our tweet corpus collection, will be made available in the LRE MAP.

Custom machine translation (MT) engines systematically outperform general-domain MT engines when translating within the relevant custom domain. This paper investigates the use of the Jensen-Shannon divergence measure for automatically routing new documents within a translation system with multiple MT engines to the appropriate custom MT engine in order to obtain the best translation. Three distinct domains are compared, and the impact of the language, size, and preprocessing of the documents on the Jensen-Shannon score is addressed. Six test datasets are then compared to the three known-domain corpora to predict which of the three custom MT engines they would be routed to at runtime given their Jensen-Shannon scores. The results are promising for incorporating this divergence measure into a translation workflow.

Progress in the Machine Translation (MT) research community, particularly for statistical approaches, is intensely data-driven. Acquiring source language documents for testing, creating training datasets for customized MT lexicons, and building parallel corpora for MT evaluation require translators and non-native speaking analysts to handle large document collections. These collections are further complicated by differences in format, encoding, source media, and access to metadata describing the documents. Automated tools that allow language professionals to quickly annotate, translate, and evaluate foreign language documents are essential to improving MT quality and efficacy. The purpose of this paper is present our research approach to improving MT through pre-processing source language documents. In particular, we will discuss the development and use of MTriage, an application environment that enables the translator to markup documents with metadata for MT parameterization and routing. The use of MTriage as a web-enabled front end to multiple MT engines has leveraged the capabilities of our human translators for creating lexicons from NFW (Not-Found-Word) lists, writing reference translations, and creating parallel corpora for MT development and evaluation.

Voss et al. (2006) analyzed newswire translations of three DARPA GALE Arabic-English MT systems at the segment level in terms of subjective judgmen+F925t scores, automated metric scores, and correlations among these different score types. At this level of granularity, the correlations are weak. In this paper, we begin to reconcile the subjective and automated scores that underlie these correlations by explicitly “grounding” MT output with its Reference Translation (RT) prior to subjective or automated evaluation. The first two phases of our approach annotate {MT, RT} pairs with the same types of textual comparisons that subjects intuitively apply, while the third phase (not presented here) entails scoring the pairs: (i) automated calculation of “MT-RT hits” using CMU aligner from METEOR, (ii) an extension phase where our Buckwalter-based Lookup Tool serves to generate six other textual comparison categories on items in the MT output that the CMU aligner does not identify, and (iii) given the fully categorized RT & MT pair, a final adequacy score is assigned to the MT output, either by an automated metric based on weighted category counts and segment length, or by a trained human judge.

When response metrics for evaluating the utility of machine translation (MT) output on a given task do not yield a single ranking of MT engines, how are MT users to decide which engine best supports their task? When the cost of different types of response errors vary, how are MT users to factor that information into their rankings? What impact do different costs have on response-based rankings? Starting with data from an extraction experiment detailed in Voss and Tate (2006), this paper describes three response-rate metrics developed to quantify different aspects of MT users’ performance identifying who/when/where-items in MT output, and then presents a loss function analysis over these rates to derive a single customizable metric, applying a range of values to correct responses and costs to different error types. For the given experimental dataset, loss function analyses provided a clearer characterization of the engines’ relative strength than did comparing the response rates to each other. For one MT engine, varying the costs had no impact: the engine consistently ranked best. By contrast, cost variations did impact the ranking of the other two engines: a rank reversal occurred on who-item extractions when incorrect responses were penalized more than non-responses. Future work with loss analysis, developing operational cost ratios of error rates to correct response rates, will require user studies and expert document-screening personnel to establish baseline values for effective MT engine support on wh-item extraction.

Task-based machine translation (MT) evaluation asks, how well do people perform text-handling tasks given MT output? This method of evaluation yields an extrinsic assessment of an MT engine, in terms of users’ task performance on MT output. While this method is time-consuming, its key advantage is that MT users and stakeholders understand how to interpret the assessment results. Prior experiments showed that subjects can extract individual who-, when-, and where-type elements of information from MT output passages that were not especially fluent. This paper presents the results of a pilot study to assess a slightly more complex task: when given such wh-items already identified in an MT output passage, how well can subjects properly select from and place these items into wh-typed slots to complete a sentence-template about the passage’s event? The results of the pilot with nearly sixty subjects, while only preliminary, indicate that this task was extremely challenging: given six test templates to complete, half of the subjects had no completely correct templates and 42% had exactly one completely correct template. The provisional interpretation of this pilot study is that event-based template completion defines a task ceiling, against which to evaluate future improvements on MT engines.

Even with recent, renewed attention to MT evaluation—due in part to n-gram-based metrics (Papineni et al., 2001; Doddington, 2002) and the extensive, online catalogue of MT metrics on the ISLE project (Hovy et al., 2001, 2003), few reports involving task-based metrics have surfaced. This paper presents our work on three parts of task-based MT evaluation: (i) software to track and record users' task performance via a browser, run from a desktop computer or remotely over the web, (ii) factorial experimental design with replicate observations to compare the MT engines, based on the accuracy of users' task responses, and (iii) the use of chi-squared and generalized linear models (GLMs) to permit finer-grained data analyses. We report on the experimental results of a six-way document categorization task, used for the evaluation of three Korean-English MT engines. The statistical models of the probabilities of correct responses yield an ordering of the MT engines, with one engine having a statistically significant lead over the other two. Future research will involve testing user performance on linguistically more complex tasks, as well as extending our initial GLMs with the documents' Bleu scores as variables, to test the scores as independent predictors of task results.