Skip to content

Instantly share code, notes, and snippets.

Show Gist options
  • Save shahbazsyed/aabd6579e428f85af6ee2a2b77fc0681 to your computer and use it in GitHub Desktop.
Save shahbazsyed/aabd6579e428f85af6ee2a2b77fc0681 to your computer and use it in GitHub Desktop.

Notes on Summarizing Information

This book by Brigitte Endres-Niggemeyer (1998) details the concept of summarizing information, its connection to cognitive pyschology, how professionals summarize information, and some computational approaches to automatic summarization.

Communication and Cognition

At its core, summarizing is the process of reducing textual information to its most essential parts. It is a situationally and communicatively bound cognitive task where three principal components of human communication are employed: the storage of knowledge in memory, understanding or learning knowledge from the environment, and the generation of utterances (imparting the learnt knowledge).

Communication is tied to the principal of relevance, i.e., one communication partner expects the statements of the other to influence their cognitive state in the current situation. This forms the communicative function of a discourse. Frequent functions are to inform, warn, instruct, convince, entertain, or amuse.

A text is a unit of meaning that is required in order to realize a communicative act.

Cognitive assumptions for discourse processing

  1. Constructivist assumption. Observing an event involves the construction of a mental representation of the event. Understanding a story about an event involves the construction of a separate mental representation of the story. The two representations are not identical.
  2. Interpretative assumption. The representation of a text is produced not by a simple conversion of the input text into an internal representation but through its interpretation. This includes the addition of the interpreter’s own knowledge.
  3. Online assumption. The meaning of input data is constructed more or less at the time of processing. Understanding takes place in parallel to the processing of input data (online), gradually and not post hoc.
  4. Presuppositional assumption. People who understand events are able to construct a mental representation, and especially a meaningful representation, only if they have prior knowledge about such events. In addition to this knowledge, the understander may have other cognitive information, such as beliefs, opinions, or attitudes regarding such events in general, or motivations, goals, or specific tasks in processing of these events.
  5. Strategic assumption. People are strategic in their information processing. In order to be as effective as possible in the construction of the mental representation, they flexibly make use of various kinds of information, they process information in several possible orders, and they cope with incomplete information.

Contextual assumptions for discourse processing

  1. Social functionality assumption. Discourse, and the process of understanding it, is functional in its social context. It follows that discourse producers and consumers (understanders) construct a representation not only of the discourse, but also of the social context, and that these two representations interact.
  2. Pragmatic assumption. Listeners evaluate a discourse on a number of points relative to the intended pragmatic functions. It may be found inadequate if it does not match contextual conditions or if it does not conform to its pragmatic aim. For e.g. a joke not being funny or told in a wrong context.
  3. Interactionist assumption. The interpretation of a discourse is embedded within an interpretation of the whole interaction process taking place between the discourse participants. This means among other things that communication partners construct a cognitive representation of the verbal and non-verbal interaction.
  4. Situational assumption. In a social situation, the interlocutors usually play a situational role. General norms, values, attitudes and conventions about the interaction in a particular situation are part of the presuppositions that must be accounted for during discourse processing.

Propositions

As we are able to remember not only single concepts (a bundle of properties), but also episodes from our personal experience, or theories that relate many knowledge items, our knowledge must be organized in structures that link individual concepts. Propositions are such a structure which combine concepts of different categories to characterize a more complex knowledge unit. The predicate typically specifies how the other concepts---the arguments---are related. For e.g. the proposition observe(Livia, plane) represents a situation “Livia observes a plane.” The predicate here represents the relation between the two arguments that can be true or false.

Larger meaning units

  1. Schemata. Their core function is to organize information into reusable packages that comprise a fixed core and some variable aspects. For e.g. a schema for buying something in a shop would have as a relatively fixed feature the exchange of money and goods, but as variable, the amount of money and the actual goods. Schemata help in establishing meaning, i.e., in understanding, learning and remembering, since they encapsulate what a subject knows about a particular aspect of the world. They exist at all levels of abstraction and concreteness.
  2. Frames. They resemble schemata in computational approaches. Frames bundle a set of features under a common name. Every feature is assigned a labeled slot.
  3. Scripts. They are the larger units of a schemata which store knowledge of commonly experienced social events. For example, a restaurant script states the common cultural assumptions that guide a customer at a restaurant. They state what we already know and are essentially conservative and bound to normal events.
  4. Dynamic model of memory. Here, knowledge is bundled into different sorts of structures including plans, scenes, memory organization packets (MOPs), and thematic organization packets (TOPs).

Integrated representation

The meaning of a document or a text is understood in terms of prior knowledge. Its representation is formed by selecting, modifying, and rearranging propositional elements from general knowledge. Meaning of a text is constructed in the knowledge net by relations. The immediate semantic neighbors and associates of a node (concepts or propositions) constitute the core meaning of a concept or proposition, more distant nodes contribute less. The more semantic relations have been installed, the deeper a proposition has been understood, and the more meaningful it is to the subject.

Understanding

In discourse comprehension, there is a continuous feedback between understanding less complex and more complex units. It is frequently the case that during a first reading of a text only some of the semantic relations between text units are recognized. Understanding remains local and incomplete, because the understander is too busy with small-scale comprehension to think about the global importance of meaning units. However, as soon as possible, meaning hypotheses are formed. Incrementally and in cooperation, the cognitive strategies work through the text base and improve its state of representation and integration.

Macrostructures and Superstructures

The holistic semantic structure or gist of a text is also called macrostructure and is made up of macropropositions which are derived from micropropositions of the text surface through macrostrategies. These macrostrategies summarize a group of propositions under one macroproposition, thus constructing a shorter and more abstract version of the text.

In good texts, many of these macropropositions appear on the text surface as topic sentences and thus facilitate understanding. Through repeated summarizing during understanding, we arrive at the topic sentence of the entire text, the highest macroproposition in the hierarchy (usually the title). As soon as all propositions of the text have been connected to the theme (or the thematic structure), the text has been understood. These schemata (the steps involved in understanding a discourse) can be text type-specific and are known as superstructures.

Text comprehension strategies

Strategies are pieces of goal-oriented behavior that try to reach their aims with appropriate means. They belong to the following functional classes:

  1. Propositional strategies. They interpret sentences and transfer them into propositional representation and are thus the counterpart to a parsing component.
  2. Strategies of local coherence. They work out how adjacent propositions are linked semantically with the most important condition being that propositions refer to coherent events in a possible world.
  3. Macrostrategies. They infer macropropositions from sequences of propositions and can be further subsumed to finally represent the global meaning of the text on various macrostructure levels.
  4. Schema strategies. They make use of the fact that many text types have a conventional structure (superstructure) that are familiar to the language users.
  5. Production strategies. They start with a mental representation of potential text knowledge, first of all constructing a text plan and then dealing with formulating the contents. The semantic macrostructure of the text to be produced is thus already known (as well as the pragmatics) at the start of formulation and serves as a global production plan for the text meaning. As soon as the propositions of the text are available and ordered in a linear sequence, their surface structure can be produced.
  6. Strategies of knowledge use. They work with the domain knowledge taken from memory.

Understanding as knowledge acquisition from text

This process is comprised of the following steps:

  1. Units of text are translated piece by piece into a propositional representation through a semantic-syntactic analysis. This representation yields the material for the mental model.
  2. Different kinds of prior knowledge are needed such as linguistic knowledge, factual knowledge and inferential knowledge. This prior knowledge is organized in the form of cognitive schemata. These schemata are activated by text information.
  3. During its incremental construction, the temporary mental model of the text is constantly being re-evaluated. It is checked for coherence, correspondence to prior knowledge, and completeness.
  4. An interaction develops between the information offered by the text and the information requirement of the reader defined by gaps in their mental model.
  5. Gaps in the text information are closed by elaborating the content of text where necessary. In case of acute differences, a temporary model may be rejected, the discourse regressed, and reconstruction of a better interpretation attempted.

Summarizing in Everyday Communication

The information to be summarized may either be available as a representation in the summarizer’s memory or come from external sources, such as a document. Then, the summarizer searches in their respective mental model for the most important pieces of information. Only these are included in the summary and passed on to its recipients. In the summary user’s memory (the recipient), a shortened representation of the object emerges, because only the knowledge gleaned from the summary is assimilated and can activate their own background knowledge.

The summarizer

Relevance assessment is the core issue of summarizing, which requires situation-specific knowledge in a given context. However, since the situation in which the summary is used is not known, the summarizer must consider several arguments for deciding about the importance of a specific piece of information, mixing speculation with rules of thumb and definite knowledge.

The summary users

They are often unknown to the summarizer. Communication problems between summarizer and the summary user often result from a lack of common knowledge. For e.g., they may not be familiar with the terminology used in the summary, or may be unfamiliar with the discourse type summary. In such cases, a summary reader expects a summary to explain a topic to the same extent as the original text.

The summarization process

It typically includes three main subtasks:

  1. Analyze the input information (understanding)
  2. Perform the reduction of information (condensation, abstraction)
  3. Present the details in appropriate form (presentation)

Condensation implies that the summary conveys more information than a normal document of the same size by avoiding redundancy, having a more concise formulation, and concentrating on key information.

Macrostrategies are the core operators of summarization that reduce and organize information to build higher-level semantic structures. They are of the following types for summarization:

  1. Strategies of knowledge use access to the individual’s knowledge store and make sure that meaning reduction takes the right direction.
  2. Syntactic strategies mark important elements of a discourse.
  3. Stylistic and rhetorical strategies are useful as well in choosing input for macrostrategies. For e.g. selecting propositions marked by rhetorical devices such as “I conclude that”, “most important thing to do is”, etc. These rhetorical strategies can note the author’s emphasis on important discourse passages.
  4. Schematic or superstructure strategies tell the summarizer what information contributes to the semantic core of the concept, event or text type.
  5. Situational strategies help summarizers identify the goals of discourse according to the situation.

Macrorules for information reduction

  1. Deletion. Given a sequence of propositions, delete each proposition that is not an interpretation condition for another proposition in the sequence.
  2. Generalization. Given a sequence of propositions, substitute the sequence by a proposition that is entailed by each of the propositions of the sequence.
  3. Construction. Given a sequence of propositions, replace it by a proposition that is entailed by the joint set of propositions of the sequence.

These can be empirically expanded as follows:

  • Delete trivial and redundant information.
  • Select a topic sentence already in the text.
  • Substitute a general term for a list of objects or a sequence of actions.
  • Invent a topic sentence if it does not appear in the text.

Likewise, the following aspects of an item can help decide its importance:

  • Fact: Relevant is what is important according to the domain knowledge.
  • Topic: Relevant is what relates to the text topic.
  • Purpose: Relevant is what serves the purpose.
  • Positive: Relevant is what is stated positively.
  • Contrast: Relevant is what differs from other things.
  • Stress: Relevant is what is characterized as relevant in the text.

For meaning reduction:

  • No reason: If you have the statement, do away with its reasons.
  • No void: Leave it out if it is not informative.
  • No comment: No comments and added explanations.
  • No example: Drop examples.

Assessing importance (relevance and interestingness)

Multiple dimensions of relevance may be considered:

  1. The structure of the topic. It determines what is to be included in a summary. Core features of the object of description are important.
  2. Adaptation to information needs. A summary should answer the summary user’s questions. Important is what fits their needs.
  3. Ease of knowledge assimilation. A summary is intended to present information that can be assimilated (absorbed, understood) by the audience. This is why it has to look at the prior knowledge of its recipients.
  4. Information value, innovation value, interestingness. Summaries have to fulfill a purpose by being effective as a means of communication. Therefore, from the original information, they must focus on what is new and interesting for the audience, i.e., what has information value.
  5. Amount of information. A summary which is constrained by length must consider its relevance decisions accordingly
  6. Focus and text design. A good summary should respect the author’s design of the original document wherever possible.
  7. The author’s intentions. A summary should convey the author’s reasons for writing.
  8. Limits imposed by cognitive capacity and own knowledge of the summarizer. What is beyond personal understanding is unimportant.
  9. Limits imposed by the media situation. A summary must consider the means of conveyance and be presented accordingly.

Good readers are better judges of importance than poor readers, when importance is defined as textual importance. Poor readers tend to select sentences as important if they are rich in visual and concrete detail. They choose information of high personal interest, not the kinds of information emphasized by the author as more important. Poor readers exhibit low consistency between what they rate as important and what they include in their summaries. Thus, the central role of importance judgments in summarizing is not self-evident to everybody.

Professional Summarizing

The information environment of professional summarizers differs to that of the casual summarizers in the following ways:

  • Professional summarizers are provided with explicit summarization methods. A corpus of domain-specific knowledge extraction methods is available for reference.
  • Professional summarization deals almost exclusively with professional and technical discourses. Both summarizers and summary users have access to the professional knowledge required to understand the terminology.
  • Efficiency and technical support. Software and intellectual guidelines are available.
  • Communication with summary users/summarizers is possible in a professional environment when a summary fails.

Instead of reading the entire document, professional summarizers search documents for useful information via dynamic or strategic reading. These document exploration techniques steer the basic reading strategies such that they acquire the desired information and neglect the remainder of the input.

*Sources

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment