19/3/1997

Writer
Wendy G. Lehnert

Source
Morgan Kaufmann Publishers Inc. Readings In Natural Language Processing P651-658

Date
1986

Keywords
Natural Language Processing, Conceptual Question Answering, Conceptual Dependency Answering, QUALM

Abstract

This reference introduces a computer program (QUALM) which is based on a theory of conceptual question answering. The conceptual theory of question answering, a theory of natural language processing, relies on the ideas in conceptual information processing and theories of human memory organisation. The reference describes how QUALM answers the questions about what was read in the following four parts: conceptual categorisation, inferential analysis, content specification and searching heuristics.

Questions

1.0 What problems does a theory of conceptual question answering have to solve in order to enable a computer to answer questions in a manner which is natural for human interaction?

The following four questions based on a theory of human question answering should be addressed by a conceptual theory of question answering:

(1) What does understanding a question mean?

(2) How understanding of a question is affected by the context?

(3) What kind of responses to a question are appropriate?

(4) How to find an answer from memory?

Only when a computer has gained the knowledge of above mentioned four aspects, i.e. knows how people ask and understand questions and what kind of answers are expected to give, can it be made to answer questions like a human being.

2.0 What is QUALM?

QUALM is a computer program which implements a conceptual theory of question answering. The theory consists of two parts, memory processing and generation. It is based on Conceptual Dependency which is a robust system for Q/A.

QUALM uses two story understanding systems, SAM and PAM, to answer the question in a manner like humans. QUALM understands the question by running a conceptual analysis program, a parser, designed by Christopher Riesbeck, then it creates an answer by using a generator program designed by Neil Goldman. The former parses a question into a Conceptual Dependency representation, and the latter translates the Conceptual Dependency representation into a language. As this process is language independent, you will not be surprised that questions can be asked in English, and answers can be given in any language, say, Chinese.

There are four phases in QUALM process: conceptual categorisation, inferential analysis, content specification and searching heuristics. The first two phases represent the understanding stage and the last two the finding answer stage.

3.0 What is the function of Conceptual Categories of questions?

Conceptual categorisation is considered as a high level of understanding. It can decide what the questioner really means. The following Conceptual Categories are used by QUALM for understanding questions:

"(1) Causal Antecedent
(2) Goal Orientation
(3) Ennoblement
(4) Causal Consequent
(5) Verification
(6) Disjunctive
(7) Instrumental/Procedural
(8) Concept Completion
(9) Expectational
(10) Judgemental
(11) Quantification
(12) Feature Specification
(13) Request"

A question represented as a conceptualisation should be categorised into one of the above Conceptual Categories to get the exact meaning of the question. But wrong Conceptual Categorisation leads to misunderstanding of the question. For example:

Right: Q1: How could you write the paper?

(an Ennoblement question)

A1a: I reviewed a lot of references.

(an Ennoblement answer)

Wrong: Q1 How could you write the paper?

(an Ennoblement question)

A1b: I used a pen and paper.

(Instrumental/Procedural answer)

The questioner intends to know the enabling conditions for writing a paper. The right answer A1a addresses the necessary enabling condition (He has read a large number of references). Answer A1b uses a wrong category (Instrumental/Procedural). Paper and pen are instruments. People have a pen and paper are not necessarily able to write a paper. Answer A1b is not what the questioner expects and is a low level understanding of the question and therefore gets the wrong answer.

However, even the right category is assigned to the question, it doesn’t mean a complete understanding. An inferential analysis is needed to complete the understanding of the question.

4.0 Why do we need an inferential analysis for understanding? What is the Universal Set Inference?

Conceptual Categorisation can answer questions correctly, but more often than not it doesn’t mean questions are completely understood. You will need an inferential analysis as well. For example:

Q2: What haven’t I done?

(What haven’t I done that I should do?)

In this case, an appropriate constraints is produced by an inference. An answer should be given to this understanding.

The Universal Set Inference is a general inference mechanism that examines the context of a question and determines appropriate constraining factors. It can only apply to all questions where the following conditions are met:

(1) the question is categorised as a Concept Completion question, and

(2) the conceptual question has MODE = NEG

(3) there is an active script

If there is no active script, the implicit constraints can hardly be inferred. One way to design without being context-specific is based on general knowledge structures (Schank & Abelson ‘77).

5.0 What is the Content Specification? What is the Elaboration Options?

Content Specification determines how many or what kind of an answer should be given. It produces a system of descriptive instructions to guide memory retrieval processes.

Elaboration Options are one type of Content Specification, which specifies exactly how to retrieve information from memory. There are four parts for each Elaboration Option: an Intentionality Threshold, a Question Criterion, an Initial Answer Criterion, and Elaboration Instructions.

The Intentionality Threshold describes what system intentionality refers to, such as "talkative," "co-operative," "minimally responsive," etc. The Question Criterion is a Conceptual Category assigned to the question. The Answer Criterion determines the type of conceptual answer which the memory search must provide. The Elaboration Instructions is a main part that specifies how to extract an elaboration from memory and how to integrate it into the conceptual answer.

6.0 How to answer an Expectational question which cannot be answered on the basis of a story representation alone?

Answering an Expectational question needs some new information which is derived from the story. Integrative Memory Processing is a process that extracts information from different sources to produce new information. The new information can help to answer an Expectational question.

SAM and PAM can encode information about events that probably happened but not explicitly described in the story. To answer Expectational questions which ask about things that didn’t occur, the Ghost path generation, an integrative memory process, is developed. It reconstructs the failed expectations during the story understanding stage.

7.0 How to answer a Causal Antecedent question?

A Causal Antecedent question asks about the reason behind an event. The question is difficult to answer if there is no additional information to the story. Answer Selection first assumes what the questioner knows and doesn’t know, then determines which answer is the best to the question. If there is no information about the questioner, it is impossible to select the best from various answers.

8.0 Why do we need Conceptual Organisation of Knowledge?

People have their own form and organisation of conceptual information in the memory. Answers from people will make sense to show more or less about that. For example, consider the following:

The lecturer told the students not to delay the coursework.

Suppose we ask:

Q3: Where was the lecturer?

The common answer is:

A3a: In a classroom.

It is almost impossible to answer like this:

A3b: On the ground.

A3a is an acceptable answer while A3b is not. It shows something about a human memory organisation. To get a satisfactory answer, the memory representation will imitate a human memory organisation to produce the answer.

9.0 What can QUALM do?

QUALM is now available to answer questions about the stories it has read. The theoretical model QUALM is designed to answer questions in its most general form. The answer depends on the knowledge base from the text understanding. If some new scripts and plans are added, the new knowledge structure can automatically be used to the answer.

1.0 Introduction

A theory based on human question answering system is one of conceptual question answering. QUALM uses this theory to answer questions about the stories. QUALM is composed of two story understanding systems, SAM [Cullingford ‘77] and PAM [Wilensky ‘76].

Theories of natural language processing, Parsing [Riesbeck & Schank ‘77] and generation [Goldman ‘75] in Conceptual Dependency representation, were used in QUALM. QUALM differs from other question answering systems which are information retrieval oriented. Being independent of language, QUALM can cope with different languages without changing anything.

2.0 Understanding the Question

First, a question is understood naively by a conceptual parse, then the result, the Conceptual Dependency conceptualisation, is categorised to learn what the questioner really means. There are thirteen possible Conceptual Categories. QUALM will assign each question a category. If the wrong category is assigned, an inappropriate answer will be given.

However, Conceptual Categorisation is not enough to totally understand a question. Inferential analysis is needed. The categorisation system can indicate the application of inference mechanisms. When to use an interpretative inference mechanism is determined by several conditions. For example, the successfully used Universal Set Inference, a general inference mechanism, will be determined by the question category, question criteria and contextual criteria.

Based on general knowledge structures, a contextual sensitive processing mechanism can be used in different contexts to understand the questions.

3.0 Finding an Answer

QUALM will begin finding an answer after a question is fully understood. Content Specification will specify what kind of answer should be given. It can create a system of descriptive instructions to instruct how to find an answer. The difficulty is to specify how memory retrieval produces a minimally correct answer and comes up with everything that’s relevant. The Elaboration Options, one type of Content Specification, has four parts to determine what an answer will be. The four parts, an Intentionality Threshold, a Question Criterion, an Initial Answer Criterion, and Elaboration Instructions, guide the retrieval heuristics to find the information, maybe more information than the question concept.

The Conceptual Category of questions will determine which retrieval heuristics is to be applied. Integrative Memory Processing is designed for answering the Expectational questions. It can mix information from different sources to create the new information for answering.

A Causal Antecedent question, which asks about the reason behind a question, is difficult to answer although it doesn’t need the new information from the story. Answer Selection can deal with this problem. It assumes what the questioner knows, then select the best answer to the question.

Conceptual Organisation of Knowledge is important to give an acceptable answer. When a memory representation is close to the human memory organisation, the answer will be reasonable to people.

4.0 Summary

QUALM is a representative of question answering system which is based on natural language processing. The whole procedure to answer a question from a story includes two stages: understanding the question and finding an answer. And Conceptual Categorisation and Inferential Analysis phases are in the former stage, Content Specification and Searching Heuristics in the latter.

Conceptual Categorisation is the base of the following process. Inferential Analysis can grasp the exact means of the questioner. Content Specification specifies how much of an answer should be given. Searching Heuristics is a process to retrieve the answer from the memory.

The process of QUALM is language independent and its presentation is conceptual. QUALM uses a general knowledge structure to understand questions. Any new scripts and plans will be added to the knowledge base which is used to answer the new questions without any change to QUALM.

5.0 Future

QUALM is currently able to answer questions about the story, which is understood, like people. But a theoretical model [Lehnert ‘77] QUALM is designed to answer questions in the most general form which is regarded as a verbal communication device between people.