Time-constrained Memory

A Reader-based Approach To Text Comprehension

Jean-Pierre Corriveau

• Publisher: Psychology Press
• Year: 2014
• Print ISBN: 9780805817119
• eBook ISBN: 9781317780106
• Topic: Personal Development
• Subtopic: History & Theory in Psychology

This book tries to answer the question posed by Minsky at the beginning of The Society of Mind: "to explain the mind, we have to show how minds are built from mindless stuff, from parts that are much smaller and simpler than anything we'd considered smart." The author believes that cognition should not be rooted in innate rules and primitives, but rather grounded in human memory. More specifically, he suggests viewing linguistic comprehension as a time-constrained process -- a race for building an interpretation in short term memory.

After reviewing existing psychological and computational approaches to text understanding and concluding that they generally rely on self-validating primitives, the author abandons this objectivist and normative approach to meaning and develops a set of requirements for a grounded cognitive architecture. He then goes on to explain how this architecture must avoid all epistemological commitments, be tractable both with respect to space and time, and, most importantly, account for the diachronic and non-deterministic nature of comprehension. In other words, a text may or may not lead to an interpretation for a specific reader, and may be associated with several interpretations over time by one reader.

Throughout the remainder of the book, the author demonstrates that rules for all major facets of comprehension -- syntax, reference resolution, quantification, lexical and structural disambiguation, inference and subject matter -- can be expressed in terms of the simple mechanistic computing elements of a massively parallel network modeling memory. These elements, called knowledge units, work in a limited amount of time and have the ability not only to recognize but also to build the structures that make up an interpretation.

Designed as a main text for graduate courses, this volume is essential to the fields of cognitive science, artificial intelligence, memory modeling, text understanding, computational linguistics and natural language understanding. Other areas of application are schema-matching, hermeneutics, local connectionism, and text linguistics. With its extensive bibliography, the book is also valuable as supplemental reading for introductory undergraduate courses in cognitive science and computational linguistics.

Part I

Foundations

Chapter 1

Overview

1.1 THE PROBLEM DOMAIN

Since the early 1970s, significant advances have been realized in the field of natural language processing (NLP). However, research has mainly focused on user interfaces and the parsing of isolated sentences; the processing of larger linguistic units has typically remained a stumbling block (Habel, 1983; Winograd, & Flores, 1986, chapter 9). There exists a multitude of models for the different facets (e.g., syntax, lexical disambiguation, inference, reference, etc.) of linguistic comprehension (or equivalently, “understanding”). In contrast, however, there are currently few computational models that tackle the comprehension of long, unrestricted, written text (henceforth, text). I use the term “text” in the broad sense in which Muller (1977, p. 5) defined it as:

any utterance or any succession of utterances, any use of speech or fragment of speech, with no restriction on its extent, produced by a single speaker or writer and displaying a certain unity.

This “certain unity” of text is taken to be central to comprehension: Subject matter (or equivalently, “aboutness”) is what gives a text this certain unity. It is generally accepted that if we fail to perceive the subject matter of a text, we find it difficult, if not impossible, to understand that text (Bransford, & Johnson, 1973). In this book, I address the problem of text comprehension, and in particular, the issue of the perception of subject matter. I view the expression, as opposed to the perception, of subject matter as a distinct problem that has more to do with the tasks of language generation (Jacobs, 1987; Ward, 1988) and memory recall (Baddeley, 1976; Kintsch, & van Dijk, 1978; van der Meer, 1987), which are not addressed here. The acquisition of reading and interpretative skills (Bertelson, 1987; Balota, Flores d’Arcais, & Rayner, 1990) is also ignored.

Of the computational models that have been proposed over the last two decades for text understanding (e.g., Schank, 1972, 1982; Cullingford, 1978; Dyer, 1983; Wilensky, 1978, 1983b; Norvig, 1987, 1989), most are symbolic conceptual analyzers primarily concerned with the problem of inference (Kass, 1986), and more specifically, with recognizing causal (van der Meer, 1987) relationships between the “facts” identified from the input text. Reference (see Kleiber, 1981) constitutes the other fundamental problem of linguistic comprehension. In essence, this task consists in retrieving from what has been memorized of the input text, the referent (e.g., a “character”, a “fact”, an “episode”, etc.) of a particular linguistic entity (such as a proper name, a pronoun, a definite noun phrase, a sentence, etc.). Both of these problems will be studied at length later in this book. For now, this intuitive introduction will suffice to understand a first commonly accepted hypothesis with respect to text understanding, namely that comprehension requires the use of knowledge beyond what information may be “in” the input text. For example, knowledge about committing suicide and hanging is required in the following example in order to understand the causal link between “rope” and “kill”: John bought a rope because he wanted to commit suicide:

Example 1.1.1 John wanted to kill himself. He went to the store to buy a rope.

Similarly, the reader must possess linguistic knowledge about pronouns in English (e.g., “her” should refer to a single female human being) in order to decipher this:

Example 1.1.2 Mary started distributing the course outline to the students. Each one of them got it from her.

Neither of these “pieces of knowledge” are “in” the text. Instead, they are part of the conceptual material a reader may bring to the act of interpreting a text.

Another fundamental assumption states that text comprehension entails information loss. For example, we do not remember the exact wording of the sentences of a book (see Gernsbacher, 1985; Phillips, 1985, pp. 3–4), and we often forget some of the facts we identify and some of the inferences we make while reading, much to the delight of mystery novelists. It follows that at the end of a reading, we have somehow assembled together a mental entity, which we will call a trace of the text, corresponding to what we have remembered of this text. Clearly, this trace is diachronic: It will change over time, during and after reading. Indeed, over the years, the vast majority of a trace may become totally unretrievable, if not completely forgotten.

It is generally acknowledged that the construction (e.g., Meutsch, 1986) of a trace is central to the perception of subject matter as well as to a posteriori tasks such as recall, summarizing, and question answering (see Graesser, & Clark, 1985). This construction process is taken to subsume several facets of linguistic comprehension such as word recognition, syntactic analysis, word sense disambiguation, reference resolution, inference generation and convergence, and perception of subject matter. Each of these tasks in itself constitutes a complex problem that is discussed later. For example, word sense disambiguation depends on a multitude of factors such as context, reader’s goals and expectations, prior processing, memory parameters, etc. However, it is not sufficient to address individually the different facets of linguistic comprehension; we must also consider their interactions: An integrated approach to interpretation is required, and although most of these facets are relevant at the level of the sentence (called the sentential level) or “lower” levels (such as the word or, equivalently, lexical level), some (e.g., perception of subject matter) are, however, specific to text understanding. In particular, the convergence problem (Corriveau, 1994c) consists in explaining how the multitude of inferences potentially generated by a text can be reduced to the small set that ends up in the trace. In other words, because a reader never draws all the possible inferences a text could suggest (for lack of time or lack of usefulness, and because there may be an infinity of such inferences), a model of text understanding must not only explain how inferences are created, but also how they are limited to a manageable number for further usage.

1.2 MOTIVATIONS AND GOALS

Minsky’s wrote the following at the beginning of The Society of Mind (1986, p.18):

To explain the mind, we have to show how minds are built from mindless stuff, from parts that are much smaller and simpler than anything we’d consider smart. Unless we can explain the mind in terms of things that have no thoughts or feelings of their own, we’ll only have gone around in a circle. But what could those simpler particles be—the ‘agents’ that compose our minds? There are many questions to answer… These questions all seem difficult, indeed, when we sever each one’s connection to the other ones. But once we see the mind as a society of agents, each answer will illuminate the rest.

My interpretation of this paragraph is that cognitive scientists must aim at building grounded (as opposed to “semantic” or “intelligent”) cognitive architectures, that is, architectures that are purely mechanistic and that do not embed within themselves any sort of knowledge, any type of epistemological commitment. In other words, an architecture that embodies either explicitly (in the form of rules or schemas) or implicitly (in the form of algorithms, procedures, nodes, and/or connections) any kind of knowledge, presents the problem of having this knowledge not reduced to “mindless stuff”, and thus, not explained per se, but rather granted an a priori existence.

Minsky’s paragraph also emphasizes the importance of an integrated approach to cognition, as opposed to a method that focuses on individual problems in isolation. Ideally, because the interactions between the different facets of cognition also need to be grounded, integration should become uniformization. In other words, within a grounded system, all knowledge should be processed in the same uniform way, that is, regardless of what kind of knowledge is in question (i.e., whether it is syntactic, lexical, semantic, pragmatic, etc.). Any other approach (e.g., one that presupposes the existence or the format of representation of schemas, one that prespecifies modules with specific cognitive tasks, etc.) suffers from depending on a priori mental entities and processes that are not grounded (by definition).

The present book proceeds essentially from this interpretation of Minsky’s challenge. More specifically, within the domain of text comprehension, my primary goal is to motivate and develop such a grounded uniform architecture and illustrate its relevance to the interpretation of written text. This goal is justified by observing that existing models of text understanding are not grounded and often quite incomplete.

On the one hand, the symbolic approach relies on arbitrarily complex data structures, rules, and algorithms that lack any sort of reduction to mindless stuff (Feldman, 1984). Most of these models have concentrated on inference generation and schema recognition and have ignored important problems such as syntax, disambiguation, reference resolution and so forth. Similarly, in local connectionist networks (see Preface) for text comprehension (Bookman, & Alterman, 1991; Bookman, 1992, 1994), a priori rules of interpretation (e.g., for case–role and thematic analysis) are embedded in the nodes and connections of these networks. However, such rules are not grounded, by definition.

On the other hand, the few existing PDP models of text comprehension typically oversimplify the problem by addressing only the learning and recognition of schemas (e.g., St. John, 1990; Golden, & Rumelhart, 1993), often from pre-processed texts. Miikkulainen (1993b, p. 258) remarked that such networks are not text understanders per se because they are too restricted; they do not offer any insight on the interactions between the different tasks of interpretation. Indeed, it appears that only his system constitutes an integrated PDP approach to text comprehension. However, as mentioned in the Preface, this is possible only through the use of prespecified structured representations (for interactions between a priori modules), which are not grounded.

Thus, without going in any further details (which are provided in the next chapter), it seems that there is indeed the need for a grounded uniform cognitive architecture for text interpretation. Bechtel (1994) confirmed this standpoint put forth by Minsky (1986) when he remarked:

The notion of levels has been widely used in discussion of cognitive science, especially in discussion of the relation of connectionist to symbolic modeling. I argue that many of the notions of levels employed are problematic for this purpose, and [I] develop an alternative notion grounded in the framework of mechanistic explanation. By considering the source of the analogies underlying both symbolic modeling and connectionist modeling, I argue that neither is likely to provide an adequate analysis of processes at the level at which cognitive theories attempt to function: one is drawn from too low a level, the other from too high a level.

The architecture I propose immediately presents the advantage of not depending on any sort of prespecified “levels” nor on any “universal” (i.e., innate) rules of interpretation, a notion that I will reject later.

I now want to submit that the use of a grounded architecture can lead to a standardization, not only in the processing underlying a model of text understanding, but also in the expression of this model. This claim stems from the blatant absence of such standardization in symbolic systems, systems which, I repeat, typically use arbitrarily complex structures and algorithms. Conversely, both structured connectionism and PDP have the virtue of using a standardized processing model (see Feldman, 1985a, 1985b). However, for the latter, there is still significant diversity (see Miikkulainen, 1993b, chapter 2 for details) with respect to architectural characteristics (e.g., number of hidden layers, number of networks and format of interactions between them, recurrent or non-recurrent network, etc.) and learning techniques (e.g., supervised versus unsupervised). For example, Miikkulainen (1993b, p. 9) abandoned the “standard” learning algorithm, namely backpropagation (see McClelland, & Rumelhart, 1986) for his model of memory, because it is not well suited for text comprehension.

Because it avoids epistemological commitments, a grounded architecture standardizes

1. the processing of knowledge: The system has no predefined knowledge “types”, structures, modules, or interactions. It merely follows what von der Malsburg (1985) calls a trivial algorithm that fixes the general form of operations: All knowledge is treated as data.

2. the expression of knowledge: All data, and thus all rules and schemas hypothesized for text interpretation, are expressed in a uniform fashion. Otherwise, the system would have to embed distinctions based on the different types of data.

Such a standardization obviously cannot be rigorously proven for all existing types of knowledge, but merely demonstrated through examples. Such a demonstration constitutes the second goal of my work and occupies a large portion of this book. The demonstration consists of two parts:

1. The technique of expression of knowledge, what I shall call the representational scheme, is motivated and described in details.

2. The applicability of this representational scheme to the most important facets of text understanding is illustrated at length.

As suggested above, a standardization in the processing of knowledge depends on a trivial algorithm. But this algorithm must be rooted in some underlying metaphor of cognition. For example, von der Malsburg (1985) defined the trivial algorithm as the operational foundation of the brain. I have already stated in the Preface that my architecture is to be rooted in the metaphor of human memory. This choice is discussed further at the end of the next chapter.

Additional processing standardization can be obtained by having the trivial algorithm controlled by a set of external parameters. In the case of my basic metaphor, this amounts to having the proposed architecture take the form of a model of memory whose parameters’ values are user-specified. As will be explained in chapter 4, these parameters may control both the structure (e.g., thresholds for temporal partitions, short term memory capacity) and the operations (e.g., decay rate, learning rate) of memory. In other words, structural and operational characteristics are not necessarily entrenched in the grounded architecture. Instead, some are controlled through parameters. Such a strategy presents the advantage of minimizing the static (or fixed) aspects of the system by avoiding embedding in it specific architectural decisions. In turn, this allows the system more flexibility in its processing through the use of these “standard” parameters: Only the parameters are predefined; their values are specified by the user of the model. As part of my second goal, the relevance of these parameters to text comprehension is...

Table of contents

• Cover
• Half Title
• Title Page
• Copyright Page
• Dedication
• Table of Contents
• Foreword
• Preface
• PART I Foundations
• PART II Design
• PART III Linguistic Comprehension with IDIoT
• PART IV Conclusions
• References
• Author Index
• Subject Index