In this paper we discuss the task of dialogue act recognition as a part of interpreting user utterances in context. To deal with the uncertainty that is inherent in natural language processing in general and dialogue act recognition in particular we use machine learning techniques to train classifiers from corpus data. These classifiers make use of both lexical features of the (Dutch) keyboard-typed utterances in the corpus used, and context features in the form of dialogue acts of previous utterances. In particular, we consider probabilistic models in the form of Bayesian networks to be proposed as a more general framework for dealing with uncertainty in the dialogue modelling process.

In spite of difficulty in defining the syllable unequivocally, and controversy over its role in theories of spoken and written language processing, the syllable is a potentially useful unit in several practical tasks which arise in computational linguistics and speech technology. For instance, syllable structure might embody valuable information for building word models in automatic speech recognition, and concatenative speech synthesis might use syllables or demisyllables as basic units. In this paper, we first present an algorithm for determining syllable boundaries in the orthographic form of unknown words that works by analogical reasoning from a database or corpus of known syllabifications. We call this syllabification by analogy (SbA). It is similarly motivated to our existing pronunciation by analogy (PbA) which predicts pronunciations for unknown words (specified by their spellings) by inference from a dictionary of known word spellings and corresponding pronunciations. We show that including perfect (according to the corpus) syllable boundary information in the orthographic input can dramatically improve the performance of pronunciation by analogy of English words, but such information would not be available to a practical system. So we next investigate combining automatically-inferred syllabification and pronunciation in two different ways: the series model in which syllabification is followed sequentially by pronunciation generation; and the parallel model in which syllabification and pronunciation are simultaneously inferred. Unfortunately, neither improves performance over PbA without syllabification. Possible reasons for this failure are explored via an analysis of syllabification and pronunciation errors.

We present the source authoring facilities of a natural language generation system that produces personalised descriptions of objects in multiple natural languages starting from language-independent symbolic information in ontologies and databases as well as pieces of canned text. The system has been tested in applications ranging from museum exhibitions to presentations of computer equipment for sale. We discuss the architecture of the overall system, the resources that the authors manipulate, the functionality of the authoring facilities, the system's personalisation mechanisms, and how they relate to source authoring. A usability evaluation of the authoring facilities is also presented, followed by more recent work on reusing information extracted from existing databases and documents, and supporting the OWL ontology specification language.

Parsing unrestricted text is useful for many language technology applications but requires parsing methods that are both robust and efficient. MaltParser is a language-independent system for data-driven dependency parsing that can be used to induce a parser for a new language from a treebank sample in a simple yet flexible manner. Experimental evaluation confirms that MaltParser can achieve robust, efficient and accurate parsing for a wide range of languages without language-specific enhancements and with rather limited amounts of training data.

We address the problem of extracting bilingual chunk pairs from parallel text to create training sets for statistical machine translation. We formulate the problem in terms of a stochastic generative process over text translation pairs, and derive two different alignment procedures based on the underlying alignment model. The first procedure is a now-standard dynamic programming alignment model which we use to generate an initial coarse alignment of the parallel text. The second procedure is a divisive clustering parallel text alignment procedure which we use to refine the first-pass alignments. This latter procedure is novel in that it permits the segmentation of the parallel text into sub-sentence units which are allowed to be reordered to improve the chunk alignment. The quality of chunk pairs are measured by the performance of machine translation systems trained from them. We show practical benefits of divisive clustering as well as how system performance can be improved by exploiting portions of the parallel text that otherwise would have to be discarded. We also show that chunk alignment as a first step in word alignment can significantly reduce word alignment error rate.

In this paper we discuss a persistent problem arising from polysemy: namely the difficulty of finding consistent criteria for making fine-grained sense distinctions, either manually or automatically. We investigate sources of human annotator disagreements stemming from the tagging for the English Verb Lexical Sample Task in the SENSEVAL-2 exercise in automatic Word Sense Disambiguation. We also examine errors made by a high-performing maximum entropy Word Sense Disambiguation system we developed. Both sets of errors are at least partially reconciled by a more coarse-grained view of the senses, and we present the groupings we use for quantitative coarse-grained evaluation as well as the process by which they were created. We compare the system's performance with our human annotator performance in light of both fine-grained and coarse-grained sense distinctions and show that well-defined sense groups can be of value in improving word sense disambiguation by both humans and machines.

Natural Language Generation (NLG) can be used to generate textual summaries of numeric data sets. In this paper we develop an architecture for generating short (a few sentences) summaries of large (100KB or more) time-series data sets. The architecture integrates pattern recognition, pattern abstraction, selection of the most significant patterns, microplanning (especially aggregation), and realisation. We also describe and evaluate SumTime-Turbine, a prototype system which uses this architecture to generate textualsummaries of sensor data from gas turbines.

We present a simple and intuitive unsound corpus-driven approximation method for turning unification-based grammars, such as HPSG, CLE, or PATR-II into context-free grammars (CFGs). Our research is motivated by the idea that we can exploit (large-scale), hand-written unification grammars not only for the purpose of describing natural language and obtaining a syntactic structure (and perhaps a semantic form), but also to address several other very practical topics. Firstly, to speed up deep parsing by having a cheap recognition pre-flter (the approximated CFG). Secondly, to obtain an indirect stochastic parsing model for the unification grammar through a trained PCFG, obtained from the approximated CFG. This gives us an efficient disambiguation model for the unification-based grammar. Thirdly, to generate domain-specific subgrammars for application areas such as information extraction or question answering. And finally, to compile context-free language models which assist the acoustic model of a speech recognizer. The approximation method is unsound in that it does not generate a CFG whose language is a true superset of the language accepted by the original unification-based grammar. It is a corpus-driven method in that it relies on a corpus of parsed sentences and generates broader CFGs when given more input samples. Our open approach can be fine-tuned in different directions, allowing us to monotonically come close to the original parse trees by shifting more information into the context-free symbols. The approach has been fully implemented in JAVA.

We address the problem of improving the efficiency of natural language text input under degraded conditions (for instance, on mobile computing devices or by disabled users), by taking advantage of the informational redundancy in natural language. Previous approaches to this problem have been based on the idea of prediction of the text, but these require the user to take overt action to verify or select the system's predictions. We propose taking advantage of the duality between prediction and compression. We allow the user to enter text in compressed form, in particular, using a simple stipulated abbreviation method that reduces characters by 26.4%, yet is simple enough that it can be learned easily and generated relatively fluently. We decode the abbreviated text using a statistical generative model of abbreviation, with a residual word error rate of 3.3%. The chief component of this model is an n-gram language model. Because the system's operation is completely independent from the user's, the overhead from cognitive task switching and attending to the system's actions online is eliminated, opening up the possibility that the compression-based method can achieve text input efficiency improvements where the prediction-based methods have not. We report the results of a user study evaluating this method.

The recent scaling down of mobile device form factors has increased the importance of predictive text entry. It is now also becoming an important communication tool for the disabled. Techniques related to predictive text entry software are discussed in a generalized, language-independent manner. The essence of predictive text entry is twofold, consisting of (1) the design of codes for text entry, and (2) the use of adaptive language models for decoding. Code design is examined in terms of the information-theoretical efficiency. Four adaptive language models are introduced and compared, and experimental results on text entry with these models are shown for English, Thai and Japanese.

The PARC 700 dependency bank is a potentially very useful resource for parser evaluation that has, so to speak, a high barrier to entry, because of tokenisation that is quite different from the source of the data, the Penn Treebank, and because there is no representation of word order, producing an uncertainty factor of some 15%. There is also a small, but perhaps not insignificant, number of errors. When using the dependency bank for evaluation, it seems likely that these things will cause inflated counts for mismatches, so to obtain more accurate measurements, it is desirable to eliminate them. The work reported here consists of an automatic conversion of the dependency bank into a Prolog representation where the word order is explicit, as well as graphical representations of the dependency trees for all 700 sentences, automatically generated from the Prolog data. As a side effect of the transformation, errors were detected and corrected. It is hoped that this work will lead to more widespread use of the PARC 700 dependency bank for parser evaluation.

Problems in the area of text and document processing can often be described as text rewriting tasks: given an input text, produce a new text by applying some fixed set of rewriting rules. In its simplest form, a rewriting rule is given by a pair of strings, representing a source string (the “original”) and its substitute. By a rewriting dictionary, we mean a finite list of such pairs; dictionary-based text rewriting means to replace in an input text occurrences of originals by their substitutes. We present an efficient method for constructing, given a rewriting dictionary D, a subsequential transducer that accepts any text t as input and outputs the intended rewriting result under the so-called “leftmost-longest match” replacement with skips, t'. The time needed to compute the transducer is linear in the size of the input dictionary. Given the transducer, any text t of length |t| is rewritten in a deterministic manner in time O(|t|+|t'|), where t' denotes the resulting output text. Hence the resulting rewriting mechanism is very efficient. As a second advantage, using standard tools, the transducer can be directly composed with other transducers to efficiently solve more complex rewriting tasks in a single processing step.

In this paper we focus on a specific search-related query expansion topic, namely search on Danish compounds and expansion to some of their synonymous phrases. Compounds constitute a specific issue in search, in particular in languages where they are written in one word, as is the case for Danish and the other Scandinavian languages. For such languages, expansion of the query compound into separate lemmas is a way of finding the often frequent alternative synonymous phrases in which the content of a compound can also be expressed. However, it is crucial to note that the number of irrelevant hits is generally very high when using this expansion strategy. The aim of this paper is therefore to examine how we can obtain better search results on split compounds, partly by looking at the internal structure of the original compound, partly by analyzing the context in which the split compound occurs. In this context, we pursue two hypotheses: (1) that some categories of compounds are more likely to have synonymous ‘split’ counterparts than others; and (2) that search results where both the search words (obtained by splitting the compound) occur in the same noun phrase, are more likely to contain a synonymous phrase to the original compound query. The search results from 410 enhanced compound queries are used as a test bed for our experiments. On these search results, we perform a shallow linguistic analysis and introduce a new, linguistically based threshold for retrieved hits. The results obtained by using this strategy demonstrate that compound splitting combined with a shallow linguistic analysis focusing on the argument structure of the compound head as well as on the recognition of NPs, can improve search by substantially bringing down the number of irrelevant hits.

“Powerset Hype to Boiling Point”, said a February headline on TechCrunch. In the last installment of this column, I asked whether 2007 would be the year of question-answering. My query was occasioned by a number of new attempts at natural language question-answering that were being promoted in the marketplace as the next advance upon search, and particularly by the buzz around the stealth-mode natural language search company Powerset. That buzz continued with a major news item in the first quarter of this year: in February, Xerox PARC and PowerSet struck a much-anticipated deal whereby PowerSet won exclusive rights to use PARC's natural language technology, as announced in a VentureBeat posting. Following the scoop, other news sources drew the battle lines with titles like “Can natural language search bring down Google?”, “Xerox vs. Google?”, and “Powerset and Xerox PARC team up to beat Google”. An April posting on Barron's Online noted that an analyst at Global Equities Research had cited Powerset in his downgrading of Google from Buy to Neutral. And, all this on the basis of a product which, at the time of writing, very few people have actually seen. Indications are that the search engine is expected to go live by the end of the year, so we have a few more months to wait to see whether this really is a Google-killer. Meanwhile, another question remaining unanswered is what happened to the Powerset engineer who seemed less sure about the technology's capabilities: see the segment at the end of D7TV's PartyCrasher video from the Powerset launch party. For a more confident appraisal of natural language search, check out the podcast of Barney Pell, CEO of Powerset, giving a lecture at the University of California–Berkeley.

I've just come back from the 45th Annual Meeting of the Association for Computational Linguistics (ACL) in Prague; this was the biggest ever ACL conference, with more than 1,000 people attending for the first time. Attendance at ACL conferences has been growing year on year, and that is a sign of a healthy field. Another sign of health is industry sponsorship. For this year's conference, the Gold Sponsor was Google, and Microsoft and Yahoo! were Silver Sponsors, along with a few companies we have not seen as ACL sponsors before: −textkernel, News Tin, and – a name that seems now to pop up regularly in this column – Powerset. There are all sorts of reasons why companies sponsor conferences like this, but clearly a major purpose is to make themselves visible to potential employees. And, if companies are hiring, that's good news across the board: it gives us a way of attracting more students into the field, and more generally, it speaks to the industrial and commercial relevance of what we do. There is nothing like external validation – especially commercial validation – to wash away those niggling self-doubts about the utility of your research endeavours. I remember attending MT Summit VII in Singapore in 1999, when Jo Lernout, then of Lernout & Hauspie, gave an invited talk in which (I'm sure I'm remembering this correctly) he said his ambition was to hire everyone in the hall. There were around 250 attendees – big for an NLP conference at the time – so that created quite a buzz. Jo wanted to hire everyone, not just cherry pick those with the near-to-product big ideas; in his vision of the future, every teensy-weensy tightly-focused research contribution had a role to play. For just a moment, everybody felt wanted.

Back at the beginning of 2000, I was a member of a working group tasked to come up with some guidelines for revamping my University's website.During one of our meetings, someone made the suggestion thatdecisions about howto structure and present information on the website should be driven by the kinds of questions that users come to the site with.Suddenly a light went on, and there appeared an idea for data gathering that might provide us withsome useful information. To find out what people really wanted to know whenthey visited the website, we would replace the University's searchengine by a page that invited the user to type in his or her query as a full natural language question. Appropriately chosen examples would be given todemonstrate that using real questions delivered better pages as a result. The data gathered would tell us what people were really looking for, more than could be gleaned from conventional search queries, and would therefore help usto better structure the information available on the website.