This paper describes a system that provides customer service by allowing users to retrieve identification numbers of parts for medical systems using spoken natural language dialogue. The paper also presents an evaluation of the system which shows that the system successfully retrieves the identification numbers of approximately 80% of the parts.
In this paper we tackle sentence boundary disambiguation through a part-of-speech (POS) tagging framework. We describe necessary changes in text tokenization and the implementation of a POS tagger and provide results of an evaluation of this System on two corpora. We also describe an extension of the traditional POS tagging by combining it with the document-centered approach to proper name identification and abbreviation handling. This made the resulting system robust to domain and topic shifts.
This paper presents a hybrid approach for named entity (NE) tagging which combines Maximum Entropy Model (MaxEnt), Hidden Markov Model (HMM) and handcrafted grammatical rules. Each has innate strengths and weaknesses; the combination results in a very high precision tagger. MaxEnt includes external gazetteers in the system. Sub-category generation is also discussed.
We compare the asymptotic time complexity of left-to-right and bidirectional parsing techniques for bilexical context-free grammars, a grammar formalism that is an abstraction of language models used in several state-of-the-art real-world parsers. We provide evidence that left-to-right parsing cannot be realised within acceptable time-bounds if the so called correct-prefix property is to be ensured. Our evidence is based on complexity results for the representation of regular languages.
Most studies on discourse markers implicitly assume that only one marker or discourse relation will occur in a sentence. In reality, more than one relation may hold between text spans and may be cued by multiple discourse markers. We describe here a method for hierarchically organising discourse markers. The hierarchies are intended for use by a generation system to enable the selection and placement of more than one marker in a single text span.
Although natural language is ambiguous, various linguistic and extra-linguistic factors often help determine a preferred reading. In this paper, we show that model generation can be used to model this process in the case of reciprocal statements. The proposed analysis builds on insights from Dalrymple et al. 98 and is shown to provide an integrated, computational account of the interplay between model theoretic interpretation, knowledge-based reasoning and preferences that characterises the interpretation of reciprocals.
A long-standing issue regarding algorithms that manipulate context-free grammars (CFGs) in a "top-down" left-to-right fashion is that left recursion can lead to nontermination. An algorithm is known that transforms any CFG into an equivalent nonleft-recursive CFG, but the resulting grammars are often too large for practical use. We present a new method for removing left recursion from CFGs that is both theoretically superior to the standard algorithm, and produces very compact non-left-recursive CFGs in practice.
This paper explores the usefulness of a technique from software engineering, namely code instrumentation, for the development of large-scale natural language grammars. Information about the usage of grammar rules in test sentences is used to detect untested rules, redundant test sentences, and likely causes of overgeneration. Results show that less than half of a large-coverage grammar for German is actually tested by two large testsuites, and that 10-30% of testing time is redundant. The methodology applied can be seen as a re-use of grammar writing knowledge for testsuite compilation.
We present an evaluation of domainindependent natural language tools for use in the identification of significant concepts in documents. Using qualitative evaluation, we compare three shallow processing methods for extracting index terms, i.e., terms that can be used to model the content of documents. We focus on two criteria: quality and coverage. In terms of quality alone, our results show that technical term (TT) extraction [Justeson and Katz 1995] receives the highest rating. However, in terms of a combined quality and coverage metric, the Head Sorting (HS) method, described in [Wacholder 1998], outperforms both other methods, keyword (KW) and TT.
A finite-state method, based on leftmost longest-match replacement, is presented for segmenting words into graphemes, and for converting graphemes into phonemes. A small set of hand-crafted conversion rules for Dutch achieves a phoneme accuracy of over 93%. The accuracy of the system is further improved by, using transformation-based learning. The phoneme accuracy of the best system (using a large rule and a 'lazy' variant of Brill's algoritm), trained on only 40K words, reaches 99%.
Compound noun segmentation is one of the crucial problems in Korean language processing because a series of nouns in Korean may appear without space in real text, which makes it difficult to identify its morphological constituents. This paper presents an effective method of Korean compound noun segmentation based on lexical data extracted from a corpus. The segmentation consists of two tasks: First, it uses a Hand-Build Segmentation Dictionary (HBSD) to segment compound nouns which frequently occur or need an exceptional process. Second, a segmentation algorithm using data from a corpus is proposed, where simple nouns and their frequencies are stored in a Simple Noun Dictionary (SND) for segmentation. The analysis is executed based on modified tabular parsing using min-max operation. Our experiments have shown a very effective accuracy rate of about 97.29%, which turns out to be very effective.
Trigrams'n'Tags (TnT) is an efficient statistical part-of-speech tagger. Contrary to claims found elsewhere in the literature, we argue that a tagger based on Markov models performs at least as well as other current approaches, including the Maximum Entropy framework. A recent comparison has even shown that TnT performs significantly better for the tested corpora. We describe the basic model of TnT, the techniques used for smoothing and for handling unknown words. Furthermore, we present evaluations on two corpora.
This paper presents a new approach to statistical sentence generation in which alternative phrases are represented as packed sets of trees, or forests, and then ranked statistically to choose the best one. This representation offers advantages in compactness and in the ability to represent syntactic information. It also facilitates more efficient statistical ranking than a previous approach to statistical generation. An efficient ranking algorithm is described, together with experimental results showing significant improvements over simple enumeration or a lattice-based approach.
Arabic inflectional morphology requires infixation, prefixation and suffixation, giving rise to a large space of morphological variation. In this paper we describe an approach to reducing the complexity of Arabic morphology generation using discrimination trees and transformational rules. By decoupling the problem of stem changes from that of prefixes and suffixes, we gain a significant reduction in the number of rules required, as much as a factor of three for certain verb types. We focus on hollow verbs but discuss the wider applicability of the approach.
The linear logic (Girard, 1987) provides a powerful framework to express categorial grammars (Ajdukiewicz, 1935) and Lambek calculus (Lambek, 1958), and a lot of work has presented proof nets uses for linguistic purposes, with a special look at proof nets for Lambek calculus (Roorda, 1991; Lamarche and Retore, 1996). But they have mainly explored the syntactic capabilities of proof nets, describing parsing processes. This paper wants to focus on the generation capabilities of proof nets thanks to their semantic readings as expressed in (de Groote and Retore, 1996). The main features of our proposal consist in the use of proof nets for Lambek calculus, of the Curry-Howard isomorphism (Howard, 1980; Girard et al., 1988), of semantic proof nets with semantic expressions a la Montague (Montague, 1974; Dowty et al., 1981), and in an algorithm for proof search with a target proof net. Unlike a previous proposal for generation in the Lambek calculus framework (Merenciano and Morrill, 1997), this point of view avoids the use of the -term unification to lead the generation process. And the algorithmic undecidability of this latter mechanism (from second order unification) does not occur any more. In this work, we do not consider the choice of lexical items from a given semantic expression the syntactic realization of which we want to generate, but rather the way we can associate given lexical entries to fit the given semantic expression and generate a syntactically correct expression. For this purpose, we express our problem as a proof search one in (multiplicative) linear logic which is decidable. Moreover, we characterize the semantic recipes of lexical items that provide a polynomial solution for the generation process.
This paper presents a parsing system for the detection of syntactic errors. It combines a robust partial parser which obtains the main sentence components and a finite-state parser used for the description of syntactic error patterns. The system has been tested on a corpus of real texts, containing both correct and incorrect sentences, with promising results.
This paper compares the efficiency of using a standard direct-manipulation graphical user interface (GUI) with that of using the QuickSet pen/voice multimodal interface for supporting a military task. In this task, a user places military units and control measures (e.g., various types of lines, obstacles, objectives) on a map. Four military personnel designed and entered their own simulation scenarios via both interfaces. Analyses revealed that the multimodal interface led to an average 3.5-fold speed improvement in the average entity creation time, including all error handling. The mean time to repair errors also was 4.3 times faster when interacting multimodally. Finally, all subjects reported a strong preference for multimodal interaction. These results indicate a substantial efficiency advantage for multimodal over GUI-based interaction during map-based tasks.
This paper describes an implementation of some key aspects of a theory of dialogue processing whose main concerns are to provide models of GROUNDING and of the role Of DISCOURSE OBLIGATIONS in an agent's deliberation processes. Our system uses the TrindiKit dialogue move engine toolkit, which assumes a model of dialogue in which a participant's knowledge is characterised in terms of INFORMATION STATES which are subject to various kinds of updating mechanisms.
We present a method for automatically detecting errors in a manually marked corpus using anomaly detection. Anomaly detection is a method for determining which elements of a large data set do not conform to the whole. This method fits a probability distribution over the data and applies a statistical test to detect anomalous elements. In the corpus error detection problem, anomalous elements are typically marking errors. We present the results of applying this method to the tagged portion of the Penn Treebank corpus.
This paper describes experiments to establish the performance of a named entity recognition system which builds categorized lists of names from manually annotated training data.Names in text are then identified using only these lists.This approach does not perform as well as state-of-the-art named entity recognition systems.However, we then show that by using simple filtering techniques for improving the automatically acquired lists, substantial performance benefits can be achieved, with resulting Fmeasure scores of 87% on a standard test set.These results provide a baseline against which the contribution of more sophisticated supervised learning techniques for NE recognition should be measured.
