Curt Burgess

PRODUCING HIGH-DIMENSIONAL SEMANTIC SPACES FROM LEXICAL CO-OCCURRENCE

A procedure that processes a corpus of text and produces numeric vectors containing information about its meanings for each word is presented. This procedure is applied to a large corpus of natural language text taken from Usenet, and the resulting vectors are examined to determine what information is contained within them. These vectors provide the coordinates in a high-dimensional space in which word relationships can be analyzed. Analyses of both vector similarity and multidimensional scaling demonstrate that there is significant semantic information carried in the vectors. A comparison of vector similarity with human reaction times in a single-word priming experiment is presented. These vectors provide the basis for a representational model of semantic memory, hyperspace analogue to language (HAL).

Explorations in context space: Words, sentences, discourse

Deriving representations of meaning, whether at the word, the sentence, or the discourse level, is a problem with a long history in cognitive psychology and psycholinguistics. In this article, we describe a computational model of high‐dimensional context space, the Hyperspace Analog to Language (HAL), and present simulation evidence that HALs vector representations can provide sufficient information to make semantic, grammatical, and abstract distinctions. Human participants were able to use the context neighborhoods that HAL generates to match words with similar items and to derive the word (or a similar word) from the neighborhood, thus demonstrating the cognitive compatibility of the representations with human processing. An experiment exploring the meaning of agent‐ and patient‐oriented verbs provided the context for discussing how the connotative aspects of word neighborhoods could provide cues in establishing a discourse model. Using the vector representations to build a sentence‐level representati...

Characterizing semantic space: Neighborhood effects in word recognition

A specification of the structural characteristics of the mental lexicon is a central goal in word recognition research. Of various word-level characteristics, semantics remains the most resistant to this endeavor. Although there are several theoretically distinct models of lexical semantics with fairly clear operational definitions (e.g., in terms of feature sharing, category membership, associations, or cooccurrences), attempts to empirically adjudicate between these different models have been scarce. In this paper, we present several experiments in which we examined the effects of semantic neighborhood size as defined by two models of lexical semantics—one that defines semantics in terms of associations, and another that defines it in terms of global co-occurrences. We present data that address the question of whether these measures can be fruitfully applied to examinations of lexical activation during visual word recognition. The findings demonstrate that semantic neighborhood can predict performance on both lexical decision and word naming.

The effect of corpus size in predicting reaction time in a basic word recognition task: Moving on from Kučera and Francis

Word frequency is one of the strongest determiners of reaction time (RT) in word recognition tasks; it is an important theoretical and methodological variable. The Kučera and Francis (1967) word frequency count (derived from the 1-million-word Brown corpus) is used by most investigators concerned with the issue of word frequency. Word frequency estimates from the Brown corpus were compared with those from a 131-million-word corpus (the HAL corpus; conversational text gathered from Usenet) in a standard word naming task with 32 subjects. RT was predicted equally well by both corpora for high-frequency words, but the larger corpus provided better predictors for low- and medium-frequency words. Furthermore, the larger corpus provides estimates for 97,261 lexical items; the smaller corpus, for 50,406 items.

FROM SIMPLE ASSOCIATIONS TO THE BUILDING BLOCKS OF LANGUAGE : MODELING MEANING IN MEMORY WITH THE HAL MODEL

This paper presents a theoretical approach of how simple, episodic associations are transduced into semantic and grammatical categorical knowledge. The approach is implemented in the hyperspace analogue to language (HAL) model of memory, which uses a simple global co-occurrence learning algorithm to encode the context in which words occur. This encoding is the basis for the formation of meaning representations in a high-dimensional context space. Results are presented, and the argument is made that this simple process can ultimately provide the language-comprehension system with semantic and grammatical information required in the comprehension process.

The effect of time course and context on the facilitation of semantic features in the cerebral hemispheres.

Two divided visual field priming experiments were designed to determine the nature of lexical retrieval in the cerebral hemispheres by studying the facilitation of semantic features of unambiguous nouns. Unambiguous nouns have a single meaning, yet semantic features associated with these nouns may vary in the degree to which they are compatible with this single meaning (e.g., LAMB-WOOL as compared with LAMB-CHOPS). Results suggest that the left hemisphere selects both strongly and weakly associated semantic features that are compatible with the dominant representation of the noun. Dominance compatibility, rather than association strength, seems to be the more important factor for deciding what features are maintained in the left hemisphere. In contrast, the right hemisphere maintains more varied information, including features that are less compatible with the dominant representation (Experiment 1) and context information (Experiment 2).

Cerebral hemispheric mechanisms linking ambiguous word meaning retrieval and creativity.

The deferral of ambiguity resolution has been thought to be an important component of creativity. The time course of priming of dominant and subordinate meanings of ambiguous words was investigated using a divided visual field priming paradigm with subjects that varied on a measure of creativity. The Wallach-Kogan similarities subtest was used to group 72 subjects into three levels of verbal creativity to compare their performance on the ambiguity resolution task (Burgess & Simpson, 1988a). Results suggest that both the left and right hemispheres contribute to the maintenance of multiple word meanings in highly creative subjects, while less creative subjects show sustained subordinate priming only in the right hemisphere or no sustained subordinate priming. These results support an interactive, collaborative theory of verbal creativity (Bogen & Bogen, 1969) and suggest that there are important individual differences that expand on the basic time course model of hemispheric processing (Burgess & Simpson, 1988a).

Mediated priming in the cerebral hemispheres.

The present experiment investigates hemispheric differences in mediated priming. Theories of lexical representation have argued for an asymmetrical coding between the right and left hemispheres ([Beeman, 1998]), claiming that the right hemisphere is more diffusely represented compared to the left hemisphere. Thus, the right hemisphere activates a larger semantic field compared to the left hemisphere. Mediated (two-step) priming is an ideal task to examine this representational claim, because of the distant nature of the prime-target pairs. Results showed no difference in the magnitude of priming (both mediated and direct) between the right and left hemispheres. These results suggest that the lexical representation of the two hemispheres is equivalent, not asymmetrical as Beeman suggests.

Large-scale databases of proper names

Few tools for research in proper names have been available—specifically, there is no large-scale corpus of proper names. Two corpora of proper names were constructed, one based on U.S. phone book listings, the other derived from a database of Usenet text. Name frequencies from both corpora were compared with human subjects’ reaction times (RTs) to the proper names in a naming task. Regression analysis showed that the Usenet frequencies contributed to predictions of human RT, whereas phone book frequencies did not. In addition, semantic neighborhood density measures derived from the HAL corpus were compared with the subjects’ RTs and found to be a better predictor of RT than was frequency in either corpus. These new corpora are freely available on line for download. Potentials for these corpora range from using the names as stimuli in experiments to using the corpus data in software applications.

Higher-order semantic structures in an African Grey parrot’s vocalizations: evidence from the hyperspace analog to language (HAL) model

Previous research has described the significant role that social interaction plays in both the acquisition and use of speech by parrots. The current study analyzed the speech of one home-raised African Grey parrot (Psittacus erithacus erithacus) across three different social contexts: owner interacting with parrot in the same room, owner and parrot interacting out of view in adjacent rooms, and parrot home alone. The purpose was to determine the extent to which the subject’s speech reflected an understanding of the contextual substitutability (e.g., the word street can be substituted in context for the word road) of the vocalizations that comprised the units in her repertoire (i.e., global co-occurrence of repertoire units; Burgess in Behav Res Methods Instrum Comput 30:188–198, 1998; Lund and Burgess in Behav Res Methods Instrum Comput 28:203–208, 1996). This was accomplished via the human language model hyperspace analog to language (HAL). HAL is contextually driven and bootstraps language “rules” from input without human intervention. Because HAL does not require human tutelage, it provided an objective measure to empirically examine the parrot’s vocalizations. Results indicated that the subject’s vocalization patterns did contain global co-occurrence. The presence of this quality in this nonhuman’s speech may be strongly indicative of higher-order cognitive skills.