Conceptual modeling from natural language functional specifications

Abstract

In this paper we describe a structured method for developing a conceptual data model by starting from a functional model expressed in a natural language. We have used the Conceptual Dependency theory for mapping natural language descriptions to conceptual dependency diagrams. We have developed algorithms to convert these conceptual dependency diagrams into unit conceptual dependency tables, which are then merged to represent the whole context of the application. We also show how transactional requirements can be incorporated into the unit conceptual dependency table, and subsequently convert the unit conceptual dependency table into a corresponding conceptual model. We have developed an augmented transition network (ATN) parser to develop conceptual dependency diagrams from natural language descriptions. A prototype system has been implemented using Oracle8i and developer platforms.

Introduction

Many real-life applications including declarative Website designs [35], data warehousing [34], and traditional database applications require the ability to specify user requirements at one level of abstraction, frequently called the ‘conceptual level’, and develop detailed specifications at other levels, and somehow connect these levels using some metadata so as to provide communication between system designers and implementers and end-users, and also enable design changes in a seamless manner by decoupling these levels. In traditional database applications, this process has been divided into conceptual, logical, and physical design phases [14]. The conceptual design phase consists of two steps: view modeling, and view integration. In the view-modeling phase, each user group analyzes its data requirements and expresses them in terms of a local conceptual schema. In the view integration phase, these independently developed views are integrated into an overall enterprise-wide schema.

There are many methods for determining user requirements: interviewing users, studying existing environments, and analyzing functional specifications of the system [26]. In all of these cases, descriptions in natural languages are common and abundant. Interviews are held in natural languages and transcribed in natural language reports, existing environments can be described in natural language descriptions, and functional specifications have natural language definitions and glossary of terms. Developing the conceptual model involves identifying the various data elements from natural language descriptions, and determining their inter-relationships [9], [11], [12], [13]. Many of these descriptions are not geared towards database design, and thus may be incomplete in their specifications, and may contain redundancies. Hence it is the responsibility of the database designer to elicit the missing pieces of information that are implicit in these ‘informal’ descriptions. No systematic method has been suggested in the literature to identify missing pieces of information in the description of user requirements.

Redundancies in natural language descriptions may include repetition of the same concept in different names (synonyms) or description of different concepts using the same or similar names (homonyms). Typically, the homonym/synonym problem (described by the term interschema multiname anomaly in [4], [5]) is dealt with at the view integration stage [4], [9], [25], with the implicit assumption that within one single view this problem may not exist. However, if views are developed from natural language descriptions, then this problem may need to be addressed at the view modeling stage. Traditional view integration techniques take multiple (possibly independently developed) views v₁, v₂,…,v_n as input, and merge them into a single view, V. Several issues emerge in this context.

• Integrating two views v₁ and v₂ may require restructuring [3], [4], [21], [23], [25] during/prior to integration. Examples of such restructuring include transformation of an attribute into an entity and vice versa, and introducing subcategories of entities and relationships [33]. Most traditional view integration methods are binary [4], [5], [7], [8], [9], [15], [16], [19], [24] at any particular point in the process of integration, two views are merged into one. As a result, integrating N views may involve N−1 restructuring steps, in the worst case.

• A property of the binary integration method is that it is non-associative. In other words, the order in which the integration is carried out will determine the outcome of integration [6], [29].

• We have previously mentioned that the goal of view integration is to obtain an integrated schema, which is the least upper bound of the individual schemas. It can be shown [4], that view integration can introduce redundant cycles. In fact redundant cycles may also exist in the views prior to integration. Identification of redundancy is a non-trivial problem [10], and no systematic method has been suggested to address this issue.

• User transaction requirements have been addressed to a limited degree in the view integration literature. In a review of the database schema and view integration methodologies [6], only two works [30], [31], [32] that incorporate queries as processing requirements in the integration process, both of which use the functional data model. Other works that consider processing requirements (only to a limited extent) are Ref. [33], in which queries are dealt with in terms of mapping queries to the integrated schema, and not to determine the result of integration. Hence, this does not ensure that the integrated schema will be complete in terms of supporting user transaction requirements.

In this paper we develop a methodology for designing the conceptual model starting from a functional model expressed in natural language sentences. While we describe the methodology using a traditional database application, it can be extended to other real world design problems such as Web engineering and data warehousing as well. We adapt the following criteria for the design process from [17], [18], [20], [22]:

• Completeness. The data model must include all data elements that are referenced in the functional specifications of the system. Thus, a database user can request, modify, and store information about any data item that may be required to carry out the functions depicted in the functional model.

• Minimality. The database designer has to ascertain that no data item is represented more than once in the data model. Thus, avoiding data redundancy that leads to inconsistencies and waste of storage space.

• Sensitivity. In incremental development of the system, it would be necessary to determine if a new data item is already incorporated into the data model, in order to ensure the completeness and minimality criteria discussed above. Thus, there is a need to go from the data model to the functional model. On the other hand, if an existing system is being modified, the functional model itself may undergo changes, which would require the data model to be updated. Thus, it may also require us to traverse from the functional model to the data model. In either case, a linkage between the two models is necessary. In our proposed method this linkage is established while developing the data model, thereby eliminating the need for any further effort once the two models are created.

• We propose a simultaneous n-ary integration method, in which any number of views can be integrated simultaneously. This reduces the number of restructuring steps to one, as opposed to n−1, as in traditional view integration methods.

• We propose a method for identification of redundant relationships in a conceptual schema. Redundant relationships are those that can be derived from others present in the model, and are not required to support user transaction requirements.

• We propose a method for comparing the conceptual schema with user transaction requirements, and making suitable modifications. This could include adding additional data elements and eliminating redundant ones. This step ensures that the schema is complete and minimal with respect to the user transaction requirements.

We applied the theory of Conceptual Dependencies, developed by Schank [27], [28] to interpret natural language expressions. The theory of conceptual dependencies provides a method for developing conceptual representations of natural language sentences and has primarily been used for developing natural language understanding and generating systems. The methodology we describe identifies data elements from the functional model expressed as natural language description, locate missing pieces of information, combine the individual data elements into an overall conceptual schema, and establish object granularity in the data model.

The rest of the paper is organized as follows: In Section 2 we describe our methodologies and a prototype implementation, and in Section 3 we present our conclusions.