Singular value decomposition (SVD) can be viewed as a method for unsupervised training of a network that associates two classes of events reciprocally by linear connections through a single hidden layer. SVD was used to learn and represent relations among very large numbers of words (20k-60k) and very large numbers of natural text passages (1k70k) in which they occurred. The result was 100-350 dimensional "semantic spaces" in which any trained or newly added word or passage could be represented as a vector, and similarities were measured by the cosine of the contained angle between vectors. Good accuracy in simulating human judgments and behaviors has been demonstrated by performance on multiple-choice vocabulary and domain knowledge tests, emulation of expert essay evaluations, and in several other ways. Examples are also given of how the kind of knowledge extracted by this method can be applied.

Metaphor comprehension involves an interaction between the meaning of the topic and vehicle terms of the metaphor. Meaning is represented by vectors in a high-dimensional semantic space. Predication modifies the topic vector by merging it with selected features of the vehicle vector. The resulting metaphor vector can be evaluated by comparing it with known landmarks in the semantic space. Thus, metaphorical predication is treated in the present model in exactly the same way as literal predication. Some experimental results concerning metaphor comprehension are simulated within this framework, such as the non-reversibility of metaphors, priming of metaphors with literal statements, and priming of literal statements with metaphors.

We present interpretation-based processing—a theory of sentence processing that builds a syntactic and a semantic representation for a sentence and assigns an interpretation to the sentence as soon as possible. That interpretation can further participate in comprehension and in lexical processing and is vital for relating the sentence to the prior discourse. Our theory offers a unified account of the processing of literal sentences, metaphoric sentences, and sentences containing semantic illusions. It also explains how text can prime lexical access. We show that word literality is a matter of degree and that the speed and quality of comprehension depend both on how similar words are to their antecedents in the preceding text and how salient the sentence is with respect to the preceding text. Interpretation-based processing also reconciles superficially contradictory findings about the difference in processing times for metaphors and literals. The theory has been implemented in ACT-R [Anderson and Lebiere, The

In Latent Semantic Analysis (LSA) the meaning of a word is represented as a vector in a high-dimensional semantic space. Different meanings of a word or different senses of a word are not distinguished. Instead, word senses are appropriately modified as the word is used in different contexts. In N-VP sentences, the precise meaning of the verb phrase depends on the noun it is combined with. An algorithm is described to adjust the meaning of a predicate as it is applied to different arguments. In forming a sentence meaning, not all features of a predicate are combined with the features of the argument, but only those that are appropriate to the argument. Hence, a different "sense" of a predicate emerges every time it is used in a different context. This predication algorithm is explored in the context of four different semantic problems: metaphor interpretation, causal inferences, similarity judgments, and homonym disambiguation.

Past research has explored the effectiveness of a Naive Bayesian classifier when filtering unsolicited bulk email (or "spare"). Results have shown that The degree of precision of this approach is generally superior to The degree of recall. This study evaluates The effectiveness of a classifier incorporating LaTent Semantic Indexing ("LSP') to filter spare email, using a corpus used in previous studies. Results show that using LSI as The basis for an email classifier to filter out spare enjoys a very high degree of recall as well as a high degree of precision, no matter if The corpus is treated using a stop list or a lemmadzer. While using LSI leads to precision roughly equal to That of using a Naive Bayesian approach, the LSI technique has a substantially highes' recall and is generally more effective under certain conditions.

We present CarmelTC, a novel hybrid text classification approach for analyzing essay answers to qualitative physics questions, which builds upon work presented in (Rose et al., 2002a). CarmelTC learns to classify units...

This paper summarizes three lines of research that are motivated by the practical problem of helping users find information from external data sources, most notably computers. The application areas include information retrieval, text categorization, and question answering. Acommon theme in these applications is that practical information access problems can be solved by analyzing the statistical properties of words in large volumes of real world texts. The same statistical properties constrain human performance, thus we believe that solutions to practical information access problems can shed light on human knowledge representation and reasoning.

The lexical semantic system is an important component of human language and cognitive processing. One approach to modeling semantic knowledge makes use of hand-constructed networks or trees of interconnected word senses (Miller, Beckwith, Fellbaum, Gross, & Miller, 1990; Jarmasz & Szpakowicz, 2003). An alternative approach seeks to model word meanings as high-dimensional vectors, which are derived from the cooccurrence of words in unlabeled text corpora (Landauer & Dumais, 1997; Burgess & Lund, 1997a). This paper introduces a new vector-space method for deriving word-meanings from large corpora that was inspired by the HAL and LSA models, but which achieves better and more consistent results in predicting human similarity judgments. We explain the new model, known as COALS, and how it relates to prior methods, and then evaluate the various models on a range of tasks, including a novel set of semantic similarity ratings involving both semantically and morphologically related terms.

Implicit knowledge is perhaps better understood as latent knowledge so that it is readily apparent that it contrasts with explicit knowledge in terms of the form of the knowledge representation, rather than by definition in terms of consciousness or awareness. We argue that as a practical matter any definition of the distinction between implicit and explicit knowledge further involves the notion of control. One advantage of the natural language meaning of the implicit-explicit distinction as applied to knowledge representations is that it provides a principled explanation for why the implicit is so quiet: it contrasts with the explicit by being in a form that cannot be expressed. Thus, rather than "unconsciousness" being a defining (and then yet to be explained) characteristic of implicit knowledge---as in "implicit knowledge is just like explicit knowledge, except it's quiet"---the "unconsciousness " associated with the implicit is a consequence of this indirect representation (s...

In this work, I report on a prototype system for technical support called the Frequently Asked Question Organizer (FAQO). This application enables technical support personnel to construct a knowledge base from email archives and other existing documents. Users can query the knowledge base using natural-language questions in order to find relevant documents. The prototype uses Latent Semantic Analysis (LSA) for query matching. A technical-support person at the Unidata Program Center tested the application for three weeks by querying the database with all technical questions that came in to him during that period, and rating the returned documents. About half the time emails were found that could help answer the question, and FAQO was found to be superior to the keyword-search tool previously used. Other experiments in matching questions with answers are reported here, along with preliminary precision/recall results that varied some of the key parameters of the LSA algorithm....