This paper describes our research into methods for inferring social and instrumental roles and relationships from document and discourse corpora. The goal is to identify the roles of initial authors and participants in internet discussions with respect to leadership, influence and expertise. Web documents, forums and blogs provide data from which the relationships between these concepts are empirically derived and compared. Using techniques from Natural Language Processing (NLP), characterizations of authority and expertise are hypothesized and then tested to see if these pick out the same or different participants as may be chosen by techniques based on social network analysis (Huffaker 2010) see if they pick out the same discourse participants for any given level of these qualities (i.e. leadership, expertise and influence). Our methods could be applied, in principle, to any domain topic, but this paper will describe an initial investigation into two subject areas where a range of differing opinions are available and which differ in the nature of their appeals to authority and truth: ‘genetic engineering' and a ‘Muslim Forum'. The available online corpora for these topics contain discussions from a variety of users with different levels of expertise, backgrounds and personalities.

Many applications of computational linguistics are greatly influenced by the quality of corpora available and as automatically generated corpora continue to play an increasingly common role, it is essential that we not overlook the importance of well-constructed and homogeneous corpora. This paper describes an automatic approach to improving the homogeneity of corpora using an unsupervised method of statistical outlier detection to find documents and segments that do not belong in a corpus. We consider collections of corpora that are homogeneous with respect to topic (i.e. about the same subject), or genre (written for the same audience or from the same source) and use a combination of stylistic and lexical features of the texts to automatically identify pieces of text in these collections that break the homogeneity. These pieces of text that are significantly different from the rest of the corpus are likely to be errors that are out of place and should be removed from the corpus before it is used for other tasks. We evaluate our techniques by running extensive experiments over large artificially constructed corpora that each contain single pieces of text from a different topic, author, or genre than the rest of the collection and measure the accuracy of identifying these pieces of text without the use of training data. We show that when these pieces of text are reasonably large (1,000 words) we can reliably identify them in a corpus.

This paper describes part of the corpus collection efforts underway in the EC funded Companions project. The Companions project is collecting substantial quantities of dialogue a large part of which focus on reminiscing about photographs. The texts are in English and Czech. We describe the context and objectives for which this dialogue corpus is being collected, the methodology being used and make observations on the resulting data. The corpora will be made available to the wider research community through the Companions Project web site.

This paper discusses how Information Extraction is used to understand and manage Dialogue in the EU-funded Companions project. This will be discussed with respect to the Senior Companion, one of two applications under development in the EU-funded Companions project. Over the last few years, research in human-computer dialogue systems has increased and much attention has focused on applying learning methods to improving a key part of any dialogue system, namely the dialogue manager. Since the dialogue manager in all dialogue systems relies heavily on the quality of the semantic interpretation of the user’s utterance, our research in the Companions project, focuses on how to improve the semantic interpretation and combine it with knowledge from the Knowledge Base to increase the performance of the Dialogue Manager. Traditionally the semantic interpretation of a user utterance is handled by a natural language understanding module which embodies a variety of natural language processing techniques, from sentence splitting, to full parsing. In this paper we discuss the use of a variety of NLU processes and in particular Information Extraction as a key part of the NLU module in order to improve performance of the dialogue manager and hence the overall dialogue system.

We present recent work in the area of Cross-Domain Dialogue Act (DA) tagging. We have previously reported on the use of a simple dialogue act classifier based on purely intra-utterance features - principally involving word n-gram cue phrases automatically generated from a training corpus. Such a classifier performs surprisingly well, rivalling scores obtained using far more sophisticated language modelling techniques. In this paper, we apply these automatically extracted cues to a new annotated corpus, to determine the portability and generality of the cues we learn.

Data sparsity is a large problem in natural language processing that refers to the fact that language is a system of rare events, so varied and complex, that even using an extremely large corpus, we can never accurately model all possible strings of words. This paper examines the use of skip-grams (a technique where by n-grams are still stored to model language, but they allow for tokens to be skipped) to overcome the data sparsity problem. We analyze this by computing all possible skip-grams in a training corpus and measure how many adjacent (standard) n-grams these cover in test documents. We examine skip-gram modelling using one to four skips with various amount of training data and test against similar documents as well as documents generated from a machine translation system. In this paper we also determine the amount of extra training data required to achieve skip-gram coverage using standard adjacent tri-grams.

In this paper we present a new approach to ontology learning. Its basis lies in a dynamic and iterative view of knowledge acquisition for ontologies. The Abraxas approach is founded on three resources, a set of texts, a set of learning patterns and a set of ontological triples, each of which must remain in equilibrium. As events occur which disturb this equilibrium various actions are triggered to re- establish a balance between the resources. Such events include acquisition of a further text from external resources such as the Web or the addition of ontological triples to the ontology. We develop the concept of a knowledge gap between the coverage of an ontology and the corpus of texts as a measure triggering actions. We present an overview of the algorithm and its functionalities.

As web searches increase, there is a need to represent the search results in the most comprehensible way possible. In particular, we focus on search results from queries about people and places. The standard method for presentation of search results is an ordered list determined by the Web search engine. Although this is satisfactory in some cases, when searching for people and places, presenting the information indexed by time may be more desirable. We are developing a system called Cronopath, which generates a timeline of web search engine results by determining the time frame of each document in the collection and linking elements in the timeline to the relevant articles. In this paper, we propose evaluation guidelines for judging the quality of automatically generated timelines based on a set of common features.

The paper examines briefly the impact of the “statistical turn” in machine translation (MT) R&D in the last decade, and particularly the way in which it has made large scale language resources (lexicons, text corpora etc.) more important than ever before and reinforced the role of evaluation in the development of the field. But resources mean, almost by definition, co-operation between groups and, in the case of MT, specifically co-operation between language groups and states. The paper then considers what alternatives there are now for MT R&D. One is to continue with interlingual methods of translation, even though those are not normally thought of as close to statistical methods. The reason is that statistical methods, taken alone, have almost certainly reached a ceiling in terms of the proportion of sentences and linguistic phenomena they can translate successfully. Interlingual methods remain popular within large electronics companies in Japan, and in a large US Government funded project (PANGLOSS). The question then discussed is what role there can be for interlinguas and interlingual methods in co-operation in MT across linguistic and national boundaries. The paper then turns to evaluation and asks whether, across national and continental boundaries, it can become a co-operative or a “hegemonic” enterprise. Finally the paper turns to resources themselves and asks why co-operation on resources is proving so hard, even though there are bright spots of real co-operation.

February 13-25, 1991

ULTRA (Universal Language TRAnslator) is a multilingual, interlingual machine translation system currently under development at the Computing Research Laboratory at New Mexico State University. It translates between five languages (Chinese, English, German, Japanese, Spanish) with vocabularies in each language based on approximately 10,000 word senses. The major design criteria are that the system be robust and general purpose with simple to use utilities for customization to suit the needs of particular users. This paper describes the central characteristics of the system: the intermediate representation, the language components, semantic and pragmatic processes, and supporting lexical entry tools.

PREMO is a knowledge-based Preference Semantics parser with access to a large, lexical semantic knowledge base and organized along the lines of an operating system. The state of every partial parse is captured in a structure called a language object, and the control structure of the preference machine is a priority queue of these language objects. The language object at the front of the queue has the highest score as computed by a preference metric that weighs grammatical predictions, semantic type matching, and pragmatic coherence. The highest priority language object is the intermediate reading that is currently most preferred (the others are still “alive,” but not actively pursued); in this way the preference machine avoids combinatorial explosion by following a “best-first” strategy for parsing. The system has clear extensions into parallel processing.