Detection of Neuropsychiatric States of Interest in Text

This paper provides an overview of a technique for measuring neuropsychiatric states such as anxiety, hostility, and depression and of a software implementation of the technique. The implementation uses both clause structure and a dictionary of semantically-tagged words and phrases to assign scores to individual clauses and to aggregate those scores over larger samples. The basic technique and software implementation have been used in a variety of settings, and form the basis of current research toward an psychodynamic psychotherapist. Obtaining Clinical Evaluations Before a dialogue system can respond to affect, it must detect it. Psychiatry and psychology have long experience in the detection and measurement of affect, both in characterizing subjects, behaviors, and diagnoses, and in selecting and measuring the efficacy of interventions. A clinician has several options in obtaining objective and valid clinical evaluations. For example, precision and accuracy may be avoided and impressionistic reactions and "gut feelings" can be relied on; some clinicians feel they are able to do competent clinical work with this approach. A clinician can spend considerable time and care in the diagnostic and therapeutic evaluation of children and adults with the goal of assessing accurately and precisely the magnitude of diverse psychopathological processes within patients at different times. Another approach is to use various observer psychiatric rating scales, such as the Brief Psychiatric Rating Scale, the Hamilton Anxiety or Depression Rating scales or various self-report measures, such as, various adjective checklists. Although such measures are widely used in many research projects, their use carries with them a false sense of security since quite often no inter-rater reliability tests are done with the rating scales, the assumption being that anybody can follow the instructions for rating and no measurement errors are likely to occur. With rating scales, however, raters vary widely on how much of the range of ratings they use with the same subjects. Some raters characteristically select the lower range of the ratings; whereas others habitually chose the higher range of the ratings. With self-report measures, the assumption is that self-raters are all, indeed, in good and equivalent contact with themselves and are not likely to be falsifying, consciously or unconsciously, their self-evaluations, though it is true that the self-rating comes directly from the individual being evaluated. These kinds of measurement errors in observer rating scales and self-report scales, usually disregarded by researchers and clinicians, are minimized in the measurement method of content analysis of verbal behavior. The subjects being rated are usually not aware what speech content or form is being analyzed, and they have difficulty covering up, even if they have some notions about such matters. Furthermore, the unstructured approach customarily used to elicit speech avoids the questionnaire or "prosecuting attorney" method, and allows the subject to elaborate and use free-will to the extent desired by the self on choice of topics to verbalize. Emotions, self-reflections, doubts, and defensive maneuvers are recorded, and these all contribute to the content analysis scores eventually calculated. The content analysis approach to the measurement of psychological dimensions includes the strengths of both the self-report approach and the observer rating scale approach, and minimizes the weaknesses of both in terms of measurement errors. Of particular interesting in the current setting is that content analysis is particularly wellsuited for use in dialogue systems. The Gottschalk-Gleser Scales While there are many content analysis scales and techniques, the Gottschalk-Gleser content analysis method (Gottschalk and Gleser, 1969) for measuring the magnitude of various psychobiological states and traits from the content analysis of verbal behavior has been successfully applied to many different neuropsychiatric dimensions. Extensive empirical research (Gottschalk et al, 1969) has established the validity and reliability of scales measuring a variety of emotional and psychobiological states including Anxiety, Hostility Outward, Hostility Inward, Ambivalent Hostility (hostility originating externally and directed towards the self), Social Alienation-Personal Disorganization, Cognitive Impairment, Hope, Depression, Human Relations, Achievement Strivings, Dependency Strivings, and Health/Sickness. Scores on the Gottschalk-Gleser content analysis scales are not simple word counts. The basis of analysis is the grammatical clause, requiring at least rudimentary syntactic analysis to determine clause boundaries. Assigning a tag to a clause may also require determining the agent and recipient of an action, and categorizing them as the self, other humans, subhuman, or inanimate. An example of a scale definition is given in Figure 1. Figure 1: Anxiety Scale 1. Death anxiety -references to death, dying, threat of death, or anxiety about death experienced by or