Está en la página 1de 13

International Journal of Approximate Reasoning 54 (2013) 22–34

Contents lists available at SciVerse ScienceDirect

International Journal of Approximate Reasoning


journal homepage: www.elsevier.com/locate/ijar

Towards linguistic descriptions of phenomena


Gracian Trivino ∗ , Michio Sugeno
European Centre for Soft Computing, Mieres, Asturias, Spain

ARTICLE INFO ABSTRACT

Article history: This paper deals with the development of computational systems that are able to provide
Received 4 January 2011 users with meaningful linguistic descriptions of phenomena, i.e., of occurrences, circum-
Revised 8 April 2012 stances, or facts that are perceptible by their senses. Current technology allows computers
Accepted 17 July 2012
to provide users with a great amount of data about a phenomenon in few seconds, never-
Available online 24 July 2012
theless the problem consists of interpreting this overwhelming amount of information in a
Keywords: practical and meaningful way. This contribution is organized around the concept of Gran-
Linguistic summarization ular Linguistic Model of a Phenomenon where we combine ideas from two fields, namely,
Computing with words and perceptions computational theory of perceptions and systemic functional linguistics.
Systemic functional linguistics We describe the main components of a computational system for generating meaningful
linguistic descriptions of phenomena. We illustrate the use of this new type of computational
application with a practical example. In our opinion, these ideas are the starting point of
an encouraging research line based on merging knowledge taken from the two mentioned
disciplines.
© 2012 Elsevier Inc. All rights reserved.

1. Introduction

Natural language (NL) allows human beings to make imprecise but practical representations of complex phenomena in
their environment. Using NL, we choose the most adequate degree of granularity in each circumstance to remark the relevant
and to hide the irrelevant aspects of a phenomenon.
Inspired in the way humans use NL, our aim is to develop computational systems to provide users with meaningful
linguistic descriptions of available data. Here, the computer takes the role of a virtual assistant that uses input data to
provide the specific user with relevant information about monitored phenomena in specific contexts.
This work is part of a research line devoted to the development of the Computational Theory of Perceptions (CTP). CTP
was introduced in the Zadeh’s seminal paper “From computing with numbers to computing with words - from manipulation
of measurements to manipulation of perceptions” [1] and further developed in subsequent papers, e.g., [2–7]. It is inspired
by the human capability to perform tasks without using crisp measurements and numerical computations, e.g., driving
in heavy traffic and playing golf. This capability plays a fundamental role in human recognition, decision and execution
processes. CTP provides a framework to develop computational systems with the capacity of computing with the meaning of
NL expressions, i.e., with the capacity of computing with imprecise descriptions of phenomena in a similar way that humans
do it. According to CTP, our perception of world is granular [8] and is articulated using NL. A granule underlies the concept of
a linguistic variable, i.e., a variable whose values are words or sentences in NL. In a first approach, a fuzzy linguistic label can
be viewed as a linguistic summary of numerical data, e.g., a set of temperature values are labeled as Medium. The definition
of linguistic label includes the concept of degree of validity of this label to describe each element in the set [9].

∗ Corresponding author.
E-mail addresses: gracian.trivino@softcomputing.es (G. Trivino), michio.sugeno@softcomputing.es (M. Sugeno).

0888-613X/$ - see front matter © 2012 Elsevier Inc. All rights reserved.
http://dx.doi.org/10.1016/j.ijar.2012.07.004
G. Trivino, M. Sugeno / International Journal of Approximate Reasoning 54 (2013) 22–34 23

Fig. 1. Phenomena evolve in time. The on-line process is performed by the computer to describe the current available data. The off-line process is performed by
the designer to configure the Experience database.

The basic concept of fuzzy linguistic summary has the general form [10,11]:

(w, “Q objects in database are S” )

where Q is called the quantifier, S is the summarizer and w is the degree of validity of the linguistic clause for representing
the meaning in the specific context. For example, (0.7, “most of the clients have a low efficiency in their consumption of
energy”).
In recent years, this basic concept has been extended in different ways [12] and used for different applications, e.g.,
data mining [13], database query [14–16], and for describing temporal series [17,18]. See in [19] a review on the state of the
research on this field. There are also interesting contributions to linguistic summarization based on type-2 fuzzy sets [20,21].
During the last years, researchers in the field of computing with words and perceptions have developed an important set of
resources to represent the meaning of perceptions for making decisions in specific applications [22–25].
The meaning of NL expressions have several dimensions that are not explicitly considered in these previous works.
The meaning of NL expressions depends on the context they are used [26]. This context of use includes, e.g., who is the
sender of the description, who is the receiver, their personal intentions and experience, the communication channel and
the form of expression applied (e.g., voice intonation, use of capital letters, and so on). Specifically, according to Systemic
Functional Linguistics (SFL), the meaning of a NL expression should be analyzed accounting with four principia, namely,
Strata, Meta-functions, Instantiation and Composition. For example, from this point of view, current approaches to linguistic
summarization of data are focused on the Ideational meta-function of NL and they do not consider the Interpersonal and
Textual meta-functions. In the next section, we introduce some details about these concepts.
Our research explores possibilities of generating linguistic summaries with more complex semantic and lexico-grammar
structures than the current ones. With this aim, we use the concept of Granular Linguistic Model of a Phenomenon (GLMP)
introduced in previous works. For example, in [27], we generate linguistic descriptions of the behavior of traffic in a round-
about and in [28], we generate linguistic reports of financial data. Our effort in this direction is closely related with the use
of ontologies to represent fuzzy knowledge [29–31].
In this paper, we extend the concept of GLMP by combining ideas from CTP and SFL. We claim that SFL is a useful tool
to fuel the development of the research line of CTP. We highlight relationships between these two disciplines that can be
useful for researchers in both fields. Merging ideas from both fields, we propose the basic architecture of a computational
system to generate linguistic summaries of data.
Generating meaningful linguistic descriptions requires to interpret data using the available experience. Fig. 1 shows how
the process of linguistic description of data is divided into off-line and on-line processes. In Section 3, we explain how,
during the off-line process, the designer analyzes different situation types in the application domain of language to produce
a knowledge database that we call Experience database. In Section 4, we explain that, during the on-line process, the Data
Acquisition module obtains the current data about the monitored phenomenon; then using the Experience Database, the Text
Generation module interprets these data to generate a linguistic description.
In Section 5, we include a demonstration in the context of a real practical project. This project consists of automatically
generating textual advices to the clients of an energy supplier company. Added in the bi-monthly invoice, these linguistic
advices aim to modify the clients’ habits of consumption [32].

2. Systemic functional linguistics

This introduction is the result of our understanding of the approach proposed by Halliday in two fundamental books
[33,34]. An interesting source of information about SFL, including a tool for drawing System Networks, can be found in the
SFL web site [35].
SFL is arguably the most powerful linguistics in dealing with meaning, as it provides a comprehensive model of language
in context. We use the model of language provided by SFL to perform a goal oriented approach to the challenge of linguistic
description of data to specific users in specific contexts.
24 G. Trivino, M. Sugeno / International Journal of Approximate Reasoning 54 (2013) 22–34

SFL was initiated by M.A.K. Halliday, a linguist in stark contrast to N. Chomsky [36]. According to Halliday, there are two
different approaches to analyze the meaning of NL: a rhetorical-ethnographic approach where language is seen as a resource
and a logical-philosophical approach, where language is seen as a system of rules. The former school approaches meaning
from context, in the belief that meaning resides inside the use of language and language always functions in context. The
latter school claims that meaning is located outside of language, either in the world or in the human mind, and approaches
meaning from grammar. Halliday takes the rhetorical-ethnographic approach while Chomsky takes the logical-philosophical
approach.
In conventional NL processing, concepts such as Shanks’s script [37], situational semantics, syntax and pragmatics are often
used. In SFL these are treated in a unified way in three strata, namely, Context, Semantics and Lexico-grammar. For example,
a concept like script is described as a generic structure in Context (this paper does not discuss it); situation semantics is
described as a register in Semantics: a semantic region constrained with a situation type in context which also specifies a
region in Lexico-grammar; syntax is of course included in Lexico-grammar; pragmatics belongs to Context and so on. See,
for details, in chapter 10 of [33] a brief survey of alternative approaches to meaning.
In recent years, several authors have considered SFL as an interesting tool to be applied in computational text generation
[38]. The following are some meaningful works belonging to this line of research. One of the earliest and best-known parsing
systems using SFL was Winograd’s SHRDLU [39]. Patten uses SFL to focus text generation as a typical “general problem solving
project” in AI [40]; Teich proposes to see text generation as a process of “navigation in a tree of choices” [41]. SFL has been
used successfully for text generation in several practical applications: Bateman (KPML) [42], Fawcett and Cardiff (Genesys)
and O’Donnell (WAG) [38]. Interestingly, Halliday, the father of SFL, has highlighted a possible use of Fuzzy Logic in SFL
by asserting that partitions in the NL-system are fuzzy rather than crisp [43], nevertheless this facet of SFL has not been
considered yet. Additionally, our approach is inspired in the research by Sugeno and collaborators on the application of SFL
in creating computational systems based on NL [44–46]. The possibility of merging ideas from SFL and Fuzzy Logic has been
already remarked in [47].
In the remainder of this section, we explain briefly the four principia of SFL, namely, Stratification, Meta-functions,
Instantiation and Composition. We focus on introducing several concepts that will be used later on.

2.1. Stratification

NL-system is stratified into four strata, namely, Context, Semantic, Lexico-grammar and Expression.

Context. The meaning of NL depends on the domain of experience (application domain) where is used. The Context stratum
defines the limits of the specific domain of experience under consideration. The basic unit of context is called a situation
type, and context is globally described by a network of situation types. For example, a situation type “cooking” can be
considered close to a situation type “shopping food”. However, a situation type “working on mathematics” is definitively
far from “cooking”.
Semantic. The Semantic stratum describes the meaning resources stored in the NL-system that are acquired through ex-
periences in terms of meaning features. Human experiences are chunked into figures; a figure is interpreted as a unit of
experience. A figure is classified into “doing”, “sensing”, “saying”, and “being”. For example in a domain of cooking, figures
of “doing” are, namely, writing recipe, making soup, adding spices; figures of “saying” are, namely, giving an instruction of
chef, a response of assistant and so on.
Lexico-grammar. It contains the grammatical resources in the NL-system to realize the meanings at the Semantic stratum as
a text. For example, figures are realized by a lexico-grammatical units called clauses which are, in short, simple sentences.
Corresponding to figures “doing”, “sensing”, “saying” and “being”, there are clauses “material”, “mental”, “verbal” and
“relational”.
Expression. It contains the resources to realize the final expression of texts, i.e., phonology or graphology. For example, in
computational applications, the so called TTS (text to speech generator) is a resource in this stratum that could include
possibilities to express meaning using voice intonation and volume.

Text generation goes from Context to Expression and text understanding follows the reverse path. In our approach, we use
these four strata to organize the analysis of the application domain and to design the Experience database (see Section 3).
The following is an example of a brief analysis of the simple clause “The oven has a high temperature” aimed to show the
multiple variables belonging to each stratum.

• Context
A situation type defined by, namely, the domain of experience of conversations about cooking in the kitchen, the chef
speaking to his assistant.
• Semantic
Type of figure: Being-having.
Type of process: Present, phasal, polar.
Type of participants: (i) simple thing-unconscious; (ii) simple quality-attribution, sense measure.
G. Trivino, M. Sugeno / International Journal of Approximate Reasoning 54 (2013) 22–34 25

• Lexico-grammar
Carrier: “The oven”, nominal group.
Process (intensive): “has”, verbal group.
Attribute: “a high temperature”, nominal group. “high” quality-sense measure + “temperature” thing.
• Expression
It is an assertive spoken statement. The voice intonation is assertive.

2.2. Meta-functions

NL-system provides three meta-functions to express meaning, namely, ideational, interpersonal and textual which function
at each stratum above.

Ideational. This meta-function, based on a view of language as reflection, is concerned with interpreting and using experience
as a resource for reflecting on the world. It allows the interpretation of what is taking place in a certain context by using
meaning resources as well as lexico-grammatical resources of the culturally recognized repertory of social practices and
concerns in the specific domain. For example, the phenomena belonging to the specific domain of experience of cooking.
Interpersonal. This meta-function, based on a view of language as action, is concerned with enacting interpersonal relations
through language. It accounts for who is participating in the communication process by using the culturally recognized
repertory of role relationships and interactive patterns corresponding with typical specific speaker - listener relationships
in the domain of experience. For example, Interpersonal meta-function deals with the different meanings depending on
either if the chef is speaking to a client, e.g. using very polite expressions, or if he is speaking to his assistant.
Textual. This meta-function, based on a view of language as information, is concerned with how to transmit the information
as a text, i.e., the channel of communication being used by considering both the medium (written, spoken, visual slides
in a presentation) and the rhetorical structure (persuasive, didactical, informative). Textual meta-function organizes the
discursive flow and it creates cohesion and continuity as it moves along by taking an account of ideational and interpersonal
meta-functions.

2.3. Instantiation

Instantiation is understood as the bi-directional relation between meaning potential and text instances where the mean-
ing potential consists of a set of meaning features contained in the NL-system. A process from meaning potential to text
instances is called text generation and the reverse process from instances to meaning potential is called text understanding.
In the process of text generation, the meaning specified by a situation type is realized as a text, e.g., a voice utterance, by
using the corresponding semantical and lexico-grammatical resources.
Each domain of experience has typical lexico-grammatical structures that are applied to express meaning in that domain.
For example, in the general NL-system, some lexico-grammar rules determine the typical structure of “assertive clauses”.
A template, e.g., “Set the oven at {low, medium, high} temperature” is an instantiation of this type of statements. Finally,
the clause “Set the oven at low temperature” is the output in the instantiation process. Additionally, the chef has a set of
possibilities to communicate meaning using his voice intonation. Finally, he chooses, e.g., a loud imperative intonation to
say the statement.
In our approach, we represent the potential of meaning of a computational application using a generic tree of choices.
In the current stage of development of our research, this tree of choices is designed by an expert in the application domain
that we call the designer.

2.4. Composition

NL-system is organized by the Composition principle as a network of interrelated categories using relations such as “a
part of” and “a kind of”. SFL uses System Networks to represent this taxonomy. System Networks are applied to describe the
use of the three meta-functions in each stratum including the representation of different levels of instantiation. Figs. 3, 4,
8–10 are examples of System Networks.
In our approach, we contribute to the implementation of this principium introducing the concept of GLMP.

3. Design of the experience database (off-line process)

3.1. Analysis of the application domain

We have used the four strata of NL to split up into four phases the analysis of the application domain. The designer follows
the process depicted in Fig. 2 to create the Experience database. During this analysis, the designer explores the application
domain identifying the features that correspond with each one of the three meta-functions of NL (Ideational, Interpersonal
and Textual). The content of the four phases in the analysis process is summarized as follows:
26 G. Trivino, M. Sugeno / International Journal of Approximate Reasoning 54 (2013) 22–34

Fig. 2. The designer follows four steps to create the Experience database. As result of each step, he/she creates the different elements in a tree of choices, namely,
situation types, figures, clauses and report template.

Fig. 3. In the Semantics stratum, this system network describes, namely, the available components for describing phenomena, the types of figure and the types of
element. The marked numbers indicate the choices made by the designer in the application example.

3.1.1. Context analysis


This phase consists of exploring the available documentation and interviewing the final user to identify a set of relevant
situation types. According with the three meta-functions of language, the analysis is performed as follows:

• The designer uses the available documentation about the application to obtain a set of relevant concepts and wordings
commonly used in the application domain. In this step, typically, the designer uses techniques of clustering to identify
different types of users and different situation types.
• In the communication process with the user, the computational system takes the role of a virtual assistant. The designer
must consider that users of specific applications have their specific experience and a personal use of NL.
• Depending on the computational resources, the designer may plan different forms of the NL utterances to be generated
as descriptions. Here, the designer analyzes the user’s requirements to define the format of the final report.

3.1.2. Semantic analysis


This phase consists of identifying a set of figures for each situation type. The designer organizes these figures forming
a GLMP for each situation type (see Section 3.2). Fig. 3 shows the available elements to describe a phenomenon, namely,
element, figure and sequence. Using the information belonging to the defined context, the designer must define the most
suitable type of figures to model the phenomenon meaning in the application context.
In addition, Fig. 3 shows the elements of figures, namely, Process, Participants and Circumstances. SFL identifies different
types for each one of these elements (see [33] for the details). The designer must define the most suitable types of these
elements for representing the desired meaning.

3.1.3. Lexico-grammar analysis


This phase consists of identifying a set of possible clauses related with each figure defined above. Fig. 4 shows how the
elements of figures are realized by elements belonging to the Lexico-grammar stratum. In the application example (in Section
5), we show different types of clauses. Fig. 9 shows how the clause-process type are the realization of process in the Semantics
stratum. The designer uses the words and linguistic expressions in the language of the application domain to define the
most suitable clauses to express the desired meaning.

3.1.4. Expression analysis


The designer must choose among a set possible forms of expression, e.g., between written and spoken utterances. In
the first case, he/she has several options to organize the text in a written report (format, choices of font, paper size and
so on). In voice generation, there are other interesting options including emotional intonation. See in [48] a first work to
G. Trivino, M. Sugeno / International Journal of Approximate Reasoning 54 (2013) 22–34 27

Fig. 4. Components of a Figure including (in rectangles) their possible realizations in the Lexico-grammar stratum.

include emotional intonation in spoken utterances. Other options include multi-modal expression where the voice can be
complemented with the face or body gestures of an avatar.
Here, this phase includes to design a report template for each situation type that allows to combine the clauses defined
above.

3.2. Granular linguistic model of a phenomenon as structure of meaning

We have seen that the designer creates a tree structure following several steps: (1) in the application domain selects a
set of situation types, (2) in each situation type selects a set of figures, (3) for each figure (depending on the interpersonal and
textual meta-functions) selects a set of clauses.
In our approach, the central element to merge SFL and CTP is a computational model of a situation type called GLMP.
According with the Composition principium, a GLMP consists of a structure of Computational Perceptions. These Computational
Perceptions are related among them using Perception Mappings, including figures, as defined below.

3.2.1. Computational perception (CP)


A CP is the computational model of a unit of meaning about the phenomenon to be modeled. In general, CPs correspond
with specific parts of the phenomenon at certain degree of granularity. A CP is a couple (A, W ) where:

A= {a1 , a2 , . . . , an } is a set of linguistic expressions (words or sentences in NL) that represents the whole linguistic domain
of CP. Each ai describes a possible value of CP with specific degree of granularity. During the off-line process, the designer
defines the linguistic domain A. These linguistic expressions can be either simple, e.g., a1 = “the consumption is low”,
a2 = “the consumption is medium”, etc., or more complex, e.g., a1 = “I am afraid that your efficiency of energy consumption
has decreased during the last semester”, a2 = “Congratulations, your efficiency of energy consumption has increased during the
last semester”, etc.
W = {w1 , w2 , . . . , wn } is a set of validity degrees wi ∈ [0, 1] associated to each ai . The values of wi are instantiated during
the on-line process, i.e., they change according with the state of the monitored phenomenon. The concept of validity
depends on the truthfulness and relevancy of each sentence in its context of use.

For example, in the context of cooking, during the off-line process, the designer defines the set of possible linguistic
expressions A, i.e., a1 = “Take care, the oven is very high”, a2 = “The oven is still low”, etc. Note that these sentences could
have different meaning depending on each specific cooking recipe. During the on-line process, if the thermometer in the
oven provides a measure of 100o C then the degrees of validity (e.g., in the case of baking bread) could be assigned to be
w1 = 0, w2 = 0.8, etc.
Typically,
 the domain of existence of a CP is modeled with a set of linguistic labels forming a strong fuzzy partition and
therefore wi = 1.

3.2.2. Perception mapping (PM)


We use PMs to create new CPs by aggregating CPs. There are many types of PM and this paper explores several of them.
A PM is a tuple (U , y, g , T ) where:

U is a set of input CP’s, U = (u1 , u2 , . . . , un ), where ui = (Ai , Wi ). In the special case of first order Perception Mapping
(1PM), U is a variable defined in the input data domain, e.g., the value z ∈ R provided by a thermometer.
y is the output CP, y = (Ay , Wy ) = {(a1 , w1 ), (a2 , w2 ), . . . , (any , wny )}.
g is the aggregation function Wy = g (Wu1 , Wu2 , . . . , Wun ) where Wy is a vector (w1 , w2 , . . . , wny ) of degrees of validity
assigned to each element in y and Wui are the degrees of validity of the input perceptions. In Fuzzy Logic many different
types of aggregation functions have been developed. For example g could be implemented using a set of fuzzy rules. In
case of 1PM, g is built using a set of membership functions Wy = (μa1 (z ), μa2 (z ), . . . , μany (z )) = (w1 , w2 , . . . , wny )
where Wy is the vector of degrees of validity assigned to each ay and z is the input data.
28 G. Trivino, M. Sugeno / International Journal of Approximate Reasoning 54 (2013) 22–34

Fig. 5. Example of a simple GLMP. We include several valid sentences for specific values of sensors.

T is a text generation algorithm that allows generating the linguistic expressions in Ay . T has associated a figure and
uses the input data to choose the most suitable clauses to describe the current state of the monitored phenomenon.
In simple cases, T can be implemented using a linguistic template, e.g., “The temperature in the room is {high | medium
| low}”.

3.2.3. GLMP
The GLMP consists of a network of PMs. Each PM receives a set of input CPs and transmits upwards a CP. We say that each
output CP is explained by the PM using a set of input CPs. In the network, each CP covers specific aspects of the phenomenon
with certain degree of granularity.
We call first order perception mapping (1PM) to those which are input to the GLMP. We call first order computational
perceptions (1CP) to the output of 1PM’s. PM’s which input are CP’s are called 2PM and their outputs are 2CP. This classification
is inspired on the definition of the three worlds by Popper, namely, the world-1 of physical objects (phenomena) , the world-2
of the perceived objects (1CP) and the world-3 of the mental objects built by using the objects in the world-2 (2CP) [49].
Fig. 5 shows an example of a simple GLMP that explains several 2CPs using data obtained from sensors. Also, we can see
several examples of clauses that describe linguistically the current state of the phenomenon at different degrees of granu-
larity. Using different aggregation functions and different linguistic expressions, the GLMP paradigm allows the designer to
model computationally his/her perceptions of complex phenomena.

4. Text generation (on-line process)

In the terminology of SFL, the Experience database contains the potential of meaning. During the on-line process, the
computer will instantiate this generic structure to generate the most suitable linguistic clauses to describe the current state
of the monitored phenomenon.
In the terminology of CTP, using the input data, the constraint propagation process allows selecting the CP values with
the highest degree of validity and expressing this meaning with the most relevant linguistic values.
The computer performs the on-line process using three modules, namely, Data acquisition, Validity calculation and Report
template instantiation.
The Data acquisition module either uses sensors or information from a database to provide the GLMP with input variables.
The Validity calculation module uses the aggregation functions in the GLMP to calculate the degree of validity of each CP
and the Text generation algorithm to select the most suitable clause to represent that meaning.
The Report template instantiation module uses a report template to merge valid clauses in instances of the final report. In
the next section, the application example uses a simple report template to provide causal explanations with different levels
of granularity.

5. Application example. Design of the Experience database

This application is a prototype built as part of a real ongoing project. Here, we limit the scope of the description to some
useful details to illustrate the ideas described above.
An electrical energy company wants to send their clients a text included into the bimonthly invoice. This text should
provide information about the clients’ consumption behavior. An objective of this text is to give advices about how to improve
the efficiency of the client’s consumption profile. The company wants to motivate the clients to move their consumption
G. Trivino, M. Sugeno / International Journal of Approximate Reasoning 54 (2013) 22–34 29

Fig. 6. This GLMP provides different linguistic expressions to explain at different levels of granularity the efficiency of energy consumption of a client during
weekends.

towards a period of the day when the electricity distribution network has the highest available capacity. See in [32] a previous
work on this project.
We obtained the input data from watt-meters and a set of forms fulfilled by these clients. Using this information is
possible to perform a work of clustering to identify different types of clients and situation types.

5.1. Context analysis

The details of this analysis considering the three meta-functions of NL can be summarized as follows:

Ideational. Using the available documentation about the project and after several technical interviews, the designer analyzed
the set of concepts and terminology typically used in the application domain. See in [28] an example of implementation
of this type of analysis. Here, we focus our description in one of the obtained situation types.

Interpersonal. In the application context there are two partners: the virtual assistant and the energy consumer, i.e., the client.
The virtual assistant acts as an advisor and the messages sent to clients need to present content in accordance to this role.
The target reader of the summary is a medium class house keeper. Therefore the concepts and associated wordings were
limited to the domain of experience typical for this type of people.

Textual. The virtual advisor has the role of an expert. The message is transmitted as a statement from an impersonal expert.
The rhetorical mode should be informative, semi-technical. Here, several possible rhetorical modes are applicable to
transmit the desired meaning, e.g., giving advices, giving warnings, and congratulating.

As example, we have selected a situation type that consists of generating a linguistic report about the energy consumption
of each client during weekends. The goal is to generate a sequence realized with a main clause and several subordinate clauses
explaining its meaning. We describe the details in the following subsections.

5.2. GLMP

The situation type was entitled “Reporting to clients their efficiency during weekends”, i.e., the GLMP was applied to
generate sentences of the type: “During weekends, your efficiency is high with respect to other clients in your category”.
Fig. 6 shows a GLMP designed to explain these CP’s. This GLMP includes several types of PM. Firstly, we describe these
PMs including a simple version of the Text generation algorithm. Later on, we provide a more elaborated example.
30 G. Trivino, M. Sugeno / International Journal of Approximate Reasoning 54 (2013) 22–34

Fig. 7. Fuzzy sets of hours belonging to the label Weekend and Working days.

5.2.1. Aggregation functions based on fuzzification


This type of aggregation function is typically used to define 1PMs. Here, the function g is constructed with respect
to a strong fuzzy partition on the universe of discourse of the input z. The degree of validity Wy is calculated as: Wy =
(μA1 (z), μA2 (z), . . . , μAny (z)), where z is the input numerical value, and μAi (z) are the membership functions associated
with the linguistic labels {A1 , A2 , . . . , Any } that cover the domain of z. The following is an example of 1PM.

Hourly consumption of client A (1PM1 ) provides the linguistic description of the hourly value of consumption of each
client. It is a four-tuple (U , y, g , T ) where:

U: They are the numerical measures (z) provided each hour (t) by a watt-meter installed in the house of client A.
y: It is a 1CP y11 = (A11 , W11 ) where A11 = {low, medium, high} and W11 = (w1low1
, w1medium
1
, w1high
1
). Note that, in A11 the
term low represents the complete clause “During the hour t, the consumption of client A was low”.
g: This aggregation function is calculated using a uniform strong fuzzy partition formed by a set of three trapezoidal
membership functions, namely, low, medium and high. The interval of definition of these functions is adapted by using
the minimum and maximum of the available data about the type of client. W11 = (μlow (z ), μmedium (z ), μhigh (z )).
T: It is a basic template: “During the hour t, the consumption of client A was {low | medium | high}”.

5.2.2. Aggregation of CPs of the same type


These PM’s allow the designer to aggregate CP’s of the same type generating output CP’s with lower level of granularity.
In the following, we show how to aggregate temporal series changing from a granularity of hours to days and weeks.

Consumption of client A during each weekend (2PM1 ) provides a reduction of granularity from hours to weeks, i.e.,
the description of the consumption of client A each weekend (wk). It is a four-tuple (U , y, g , T ) where:

U: It is a temporal series obtained from the hourly value of 1CP1 each week.
y: It is a 2CP y12 = (A21 , W12 ) where A21 = {low, medium, high} and W12 (wk) = (w1low
2
, w1medium
2
, w1high
2
). Note that, e.g.,
high is used to represent the whole sentence “The weekend wk the consumption of client A was high”.
g: We aggregate the elements of the temporal series W11 (t ) by using the membership function μweekend (t ) defined over
the domain of hours in the day (see Fig. 7). For example, we calculate the degree of validity of the clause “The weekend
wk the consumption of client A was low” as follows:
t =wk 1
t =1 w1low (t ) μweekend (t )
w1low (wk) =
2
t =wk
t =1 μweekend (t )
T: It is the template: “During weekend wk, the consumption of client A was {low | medium | high}”.

Consumption of client A in the low-cost period (2PM2 ) provides the perception of daily consumption of client A in
the low cost period. Note that here we create a temporal series with daily granularity from a temporal series of hourly
granularity. 2PM2 is a four-tuple (U , y, g , T ) where:

U: It is a temporal series obtained from the hourly value of 1CP1 .


y: It is a temporal series with a granularity of days. It is y22 = (A22 , W22 ) where A22 = {low, medium, high} and W22 (d) =
(w2low
2
, w2medium
2
, w2high
2
).
g: Here, (lowCost, normalCost) is a strong fuzzy partition defined over the hours of each day. For example, we calculate
the degree of validity of the clause “During the day d, in the low cost period, the consumption of client A is high” as follows:
t =d
t =0 w1high (t )μlowCost (t )
1
2
w2high (d) = t =d
t =0 μlowCost (t )
T: It is a basic template: “During the day d, in the low-cost period, the consumption of client A was {low | medium | high}”.
G. Trivino, M. Sugeno / International Journal of Approximate Reasoning 54 (2013) 22–34 31

Consumption of client A in the low-cost period during each weekend (2PM3 ) provides reduction of granularity from
days in CP22 to weeks. 2PM3 is a four-tuple (U , y, g , T ) where:

U: It is the temporal series of values provided by 2PM2 each week.


y: It is a 2CP y32 = (A23 , W32 ) where: A23 = {low, medium, high} and W32 (wk) = (w3low
2
, w3medium
2
, w3high
2
).
g: This aggregation function consists of adding the daily values of consumption weighted by the membership function
weekend2 (defined over the domain of the week in days) as follows:
d=7 2
w (d) μweekend2 (d)
2
w3low (wk) = d=1d2low =7 μ
d=1 weekend2 (d)

T: It is a template: “During weekend wk the consumption of client A was {low | medium | high} in the low-cost period”.

5.2.3. Aggregation based on rules


This type of aggregation allows the designer to combine CPs of different type. It can be implemented using a set of fuzzy
rules.

Efficiency of client A each weekend (2PM4 ) is a four-tuple (U , y, g , T ) where:

U: They are 2CP1 and 2CP3 .


y: It is y42 = (A24 , W42 ) expressing the designer interpretation of the efficiency of each client A each weekend, where:
A24 = {low, medium, high} and W42 (wk) = (w4low 2
, w4med
2
, w4high
2
).
g: It is formulated using a set of fuzzy rules, e.g.:
r1 : IF 2CP1 is low AND 2CP3 is low THEN 2CP4 is medium
r2 : IF 2CP1 is low AND 2CP3 is medium THEN 2CP4 is high
r3 : IF 2CP1 is low AND 2CP3 is high THEN 2CP4 is medium
...
That is to say: r1 : IF “Client A consumed low energy at weekend wk” AND “Client A consumed low energy in low cost
zone at weekend w” THEN “Efficiency of weekend w was medium”, . . .
Here, we have implemented this set of fuzzy rules using the minimum for AND, singletons for the consequents and
by applying:

low ωi
2
w4low (wk) = 
ωi

where ωi is the degree of firing of each rule ri and low ωi adds the contributions of rules having 2CP42 is low as
consequent.
T: The set of output sentences can be obtained from the template: “During weekend wk, the efficiency of client A is {low |
medium | high}”.

5.2.4. Aggregation based on quantification


The outputs of this type of PM are 2CPs based on the concept of fuzzy quantifier mentioned in Section 1. Typically, the
input is a temporal series of CPs of the same type, e.g., the consumption of energy of a set of clients. The designer can
summarize this information using sentences such as: “Most of clients consume much energy”. The validity of this type of
sentences can be calculated
 n much as follows: Firstly, we calculate the cardinality of the fuzzy set of clients consuming much energy:
Card(much) = 1n ii= =1 wci , where wcmuch
i
is the weight of the subordinate CP, i.e., “Client-i consumes much energy”. Sec-
ondly, we calculate the degree of membership of this cardinality to the quantifier most: w(most ,much) = μmost (Card(much)).

Efficiency of client A during weekends (2PM5 ) is a four-tuple (U , y, g , T ) where:

U: The temporal series of values of 2CP4 during the whole invoice period, i.e, weekends in the invoice period.
y: It is y52 = (A25 , W52 ) where: A25 = {(few, low), (few, medium), (few, high), . . . , (much, high)} and
W52 = (w5few2
,low , w5few,med , w5few,high , . . . , w5much,high ).
2 2 2

g: We calculate the degree of validity of, e.g., the clause “Few weekends the efficiency of client A is high” as follows:

=nwk
wk
1
Card(high) = 2
w4high (wk)
nwk wk=1
2
w5few ,high = μfew (Card(high))

being nwk the number of weekends in the invoice period.


32 G. Trivino, M. Sugeno / International Journal of Approximate Reasoning 54 (2013) 22–34

Fig. 8. Elements and choices of Participants and Circumstances. Note that this diagram extends the diagram shown in Fig. 3
.

Fig. 9. In the Lexico-grammar stratum, this System Network describes the elements and choices of Clause simplex.

Fig. 10. Choices of Mood. Note that this diagram extends the System Network shown in Fig. 9.

T: The set of output sentences can be obtained from the template: “{Few | Some | Most} weekends the efficiency of client
A is {low | medium | high}”.

5.3. Semantic analysis

First, the designer associated a figure with each CP in the GLMP. Hereinafter, as an example, we describe how the designer
defined the elements of a Text generation algorithm associated with 2PM5 .
Using the System Networks defined in SFL, the designer chose a figure of type being/ascribing (see marks 1-2-3 in Fig. 3).
Then the designer defined two Participants and one Circumstance. Fig. 8 shows the next marks in the System Networks.
The designer defined a type of participant 4-simple thing/non conscious that must be the carrier, that corresponds with “your
efficiency”.
The other participant was defined to be a 5-simple quality, i.e., the attribute “high / medium / low”. Note that the definitive
choice of this quality is performed during the on-line process using the sensor data.
The process was defined to be, polar, phasal, and present (these choices are not included in the diagrams).
Finally, the designer included a circumstance (mark 6) of type duration (mark 7), i.e., “During weekends/working days”.

5.4. Lexico-grammar analysis

Fig. 9 shows a sequence of definitions in this stratum, namely, 8-Process type, 9-relational, 10-true relational and 11-
attributive.
The box under the mark 11 is called realization statement. In this system network, this box shows that attributive processes
in the Semantic stratum are realized using the verb “to be” with a carrier and an attribute in the Lexico-grammar stratum.
Regarding with the implementation of the Interpersonal meta-function, the designer defined indicative/declarative mood
(see marks 12 and 13 in Fig. 10). Here, the realization statement indicates that the subject must precede the verb.
Therefore, an example of clause built in this way is “During weekends, your efficiency is low”.
G. Trivino, M. Sugeno / International Journal of Approximate Reasoning 54 (2013) 22–34 33

Table 1
Weights calculated for 2CP5 for the first five clients in the database.
Client 2CP5
Low Med High
1 0.7778 0.2222 0
2 0.2791 0.5271 0.1938
3 0.3199 0.5067 0.1734
4 0.2683 0.5847 0.1470
5 0.3670 0.4790 0.1540

Depending on the situation type, the designer can easily create more complex structures which deliver the final clauses
depending on sensor data. For example, the computational system could select clauses expressing different moods. For
example, If the profile of consumption is efficient, the sentence could be used to congratulate the client. If the efficiency is
medium and it could be improved, the sentence could take the form of a soft advise. If the efficiency is very low, the sentence
could take the form of a critical warning, e.g., “Regrettably, your efficiency has decreased during the last months”.

5.5. Expression analysis phase

In the application example, the designer used the current invoice format and content as starting point. The final realiza-
tion of the report consisted of two lines of written printed text, selecting a main topic and some subtopics. This text was
accompanied with a diagram with special symbols where bi-monthly continuity with date is specified.

6. Experimental results

The input data were obtained with watt-meters installed in 35 householder’s houses measuring the cumulative energy
consumption within 60 minutes intervals during several months. We used these sensor data to calculate the validity degree
associated with each possible clause in the GLMP.
Table 1 shows the validity degrees obtained for 2CP52 for the five first clients in the database. Using these results, the
instantiation process delivers a linguistic report, based on a sequence of figures, as follows:

“During weekends, your efficiency is low. For example, the first weekend in the invoice period, your consumption was high and
your use of the low-cost zone was low”.

7. Conclusions

We have described a model to represent complex phenomena with several interesting properties, namely, it is based
on NL and therefore easy to understand, it is granular, it provides representation for uncertainty and it delivers relevant
descriptions in function of the input data.
The model merges concepts belonging to two apparently separated disciplines, namely, CTP and SFL. In our opinion, the
relationship between these two disciplines can provide interesting insights for researchers in both areas.
This new approach to linguistic summarization provides opportunities for generating fuzzy linguistic summaries that
have richer semantic and lexico-grammatical structures than the current ones.
This is an introductory work in a long term research project. The scope of this paper is limited to introduce the basic ideas
and to provide a simple application example that shows partially their potentiality.
From the theoretical point of view, this work opens a number of opportunities for future research. In practical applications,
the current approach can already be used as a starting point for generating a new type of linguistic descriptions of data.

Acknowledgements

The authors thank HC Energia for the opportunity of applying this research in a practical project. This work has been
funded by the Spanish government under project TIN2011-29827-C02-01.

References

[1] L.A. Zadeh, From computing with numbers to computing with words – from manipulation of measurements to manipulation of perceptions, IEEE Transactions
on Circuits and Systems 45 (1) (1999) 105–119.
[2] L.A. Zadeh, A new direction in AI. towards a computational theory of perceptions of measurements to manipulation of perceptions, AI Magazine 22 (1) (2001)
73–84.
[3] L.A. Zadeh, Granular computing as a basis for a computational theory of perceptions, in: Proceedings of the IEEE International Conference on Fuzzy Systems,
Honolulu, USA, 2002, pp. 564–565.
[4] J.M. Mendel, Computing with words and its relationships with fuzzistics, Information Sciences 177 (2007) 988–1006.
[5] L.A. Zadeh, Toward human level machine intelligence - is it achievable? the need for a paradigm shift, IEEE Computational Intelligence Magazine (2008)
11–22.
34 G. Trivino, M. Sugeno / International Journal of Approximate Reasoning 54 (2013) 22–34

[6] J. Lawry, A methodology for computing with words, International Journal of Approximate Reasoning 28 (2-3) (2001) 51–89.
[7] R.R. Yager, Reasoning with doubly uncertain soft constraints, International Journal of Approximate Reasoning 52 (4) (2011) 554–561.
[8] A. Bargiela, W. Pedrycz, Granular Computing: An Introduction, Kluwer Academic Publishers., 2003.
[9] L.A. Zadeh, The concept of linguistic variable and its application to approximate reasoning, Information Sciences 8 (1975) 199–249.
[10] R.R. Yager, A new approach to the summarization of data, Information Sciences 28 (1982) 69–86.
[11] L.A. Zadeh, A computational approach to fuzzy quantifiers in natural languages, Computing and Mathematics with Applications 9 (1983) 149–184.
[12] L. Lietard, A new definition for linguistic summaries of data, in: FUZZ-IEEE International Conference on Fuzzy Systems, 2008.
[13] R.R. Yager, Fuzzy summaries in database mining, in: In Proceedings of the 11th Conference on Artificial Intelligence for Applications, 1995.
[14] J. Kacprzyk, R.R. Yager, Linguistic summaries of data using fuzzy logic, International Journal of General Systems 30 (2001) 133–154.
[15] J. Kacprzyk, R. Yager, S. Zadrozny, A fuzzy logic based approach to linguistic summaries of databases, International Journal of Applied Mathematics and
Computer Science (10) (2000) 813–834.
[16] D.E. Losada, F. Díaz-Hermida, A. Bugarín, S. Barro, Experiments on using fuzzy quantified sentences in adhoc retrieval, in: Proceedings of the 2004 ACM
symposium on Applied computing, New York, NY, USA, 2004, pp. 1059–1064.
[17] J. Kacprzyk, A. Wilbik, S. Zadrozny, Linguistic summarization of time series using a fuzzy quantifier driven aggregation, Fuzzy Sets and Systems 159 (12)
(2008) 1485–1499.
[18] R. Castillo-Ortega, N. Marín, D. Sánchez, Linguistic summary-based query answering on data cubes with time dimension, in: Proceedings of the 8th Inter-
national Conference on Flexible Query Answering Systems, FQAS ’09, 2009, pp. 560–571.
[19] J. Kacprzyk, S. Zadrozny, Computing with words is an implementable paradigm: Fuzzy queries, linguistic data summaries and natural language generation,
IEEE Transactions on Fuzzy Systems 18 (3) (2010) 461–472.
[20] D. Wu, J.M. Mendel, Linguistic summarization using if-then rules and interval type-2 fuzzy sets, IEEE Transactions on Fuzzy Systems 19-1 (2011) 136–151.
[21] A. Niewiadomski, A type-2 fuzzy approach to linguistic summarization of data, IEEE Transactions on Fuzzy Systems 16-1 (2008) 198–212.
[22] R. Degani, G. Bortolan, The problem of linguistic approximation in clinical decision making, International Journal of Approximate Reasoning 2 (2) (1988)
143–162.
[23] L. Martnez, Sensory evaluation based on linguistic decision analysis, International Journal of Approximate Reasoning 44 (2) (2007) 148–164.
[24] E. Herrera-Viedma, E. Peis, J.M.M. del Castillo, S. Alonso, K. Anaya, A fuzzy linguistic model to evaluate the quality of web sites that store xml documents,
International Journal of Approximate Reasoning 46 (1) (2007) 226–253.
[25] C. Mencar, C. Castiello, R. Cannone, A.M. Fanelli, Interpretability assessment of fuzzy knowledge bases: A cointension based approach, International Journal
of Approximate Reasoning 52 (2011) 501–518.
[26] L. Wittgenstein, Philosophical Investigations, Blackwell Publishing, 1953, 2001.
[27] G. Trivino, A. Sanchez, A.S. Montemayor, J.J. Pantrigo, R. Cabido, E.G. Pardo, Linguistic description of traffic in a roundabout, in: Proceedings of the IEEE Fuzzy
2010, Barcelona, Spain, 2010.
[28] S. Mendez-Nunez, G. Trivino, Combining semantic web technologies and computational theory of perceptions for text generation in financial analysis, in:
Proceedings of the IEEE Fuzzy 2010, Barcelona, Spain, 2010.
[29] F. Bobillo, U. Straccia, Fuzzy ontology representation using owl 2, International Journal of Approximate Reasoning (2011) 1073–1094.
[30] M. Reformat, C. Ly, Ontological approach to development of computing with words based systems, International Journal of Approximate Reasoning (2009)
72–91.
[31] S. Calegari, D. Ciucci, Granular computing applied to ontologies, International Journal of Approximate Reasoning (2010) 391–409.
[32] A. van der Heide, G. Trivino, Automatically generated linguistic summaries of energy consumption data, in: ISDA ’09: Proceedings of the 2009 Ninth
International Conference on Intelligent Systems Design and Applications, Pisa, Italy, 2009, pp. 553–559.
[33] M.A.K. Halliday, M.I.M. Matthiessen, Construing Experience through Meaning: A Language-based Approach to Cognition, Continuum; Study ed edition (June
3, 2006), 1999.
[34] M.A.K. Halliday, M.I.M. Matthiessen, An Introduction to Functional Grammar, Oxford University, New York, 2004.
[35] http://www.isfla.org/Systemics/Software/index.html Systemic functional linguistics information site. http://www.isfla.org/Systemics/Software/index.html.
[36] N. Chomsky, Aspects of the Theory of Syntax, MIT Press, 1965.
[37] K. Lamberts, D.R. Shanks, Knowledge Concepts and Categories, Mit Press, 1997.
[38] M. Odonnell, J. Bateman, SFL in computational context: A contemporary history, in: Continuing Discourse on Language: A Functional Perspective, Equinox,
2005, pp. 343–382.
[39] T. Winograd, Understanding Natural Language, Academic Press, 1972.
[40] T. Patten, Systemic text generation as problem solving, Cambridge University Press, 1988.
[41] E. Teich, Systemic functional grammar un natural language generation, Cassell, London, 1999.
[42] J. Bateman, Enabling technology for multilingual natural language generation: the kpml development environment, Cambridge University Press, 1997.
[43] M.A.K. Halliday, On Language and Linguistics (Collected Works of M.a.K. Halliday), Continuum, London, 2006.
[44] T. Sugimoto, N. Ito, H. Fujishiro, M. Sugeno, A study on meaning processing of dialogue with an example of development of travel consultation system,
Information Sciences 144 (1–4) (2002) 45–74.
[45] I. Kobayashi, M. Sugeno, An approach to a dynamic system simulation based on human information processing, International Journal of Uncertainty, Fuzziness
and Knowledge-Based Systems 10 (6) (2002) 611–633.
[46] S. Iwashita, N. Ita, I. Kobayashi, T. Sugimoto, M. Sugeno, Everyday language help system based on software manual, in: IEEE International Conference on
Systems, Man and Cybernetics, 2004.
[47] J. Kacprzyk, S. Zadrozny, Computing with words and Systemic Functional Linguistic: Linguistic data summaries and natural language generation, in: V.N.
Huynh (Ed.), et al., Integrated Uncertainty Management and Applications, AISC, Springer-Verlag, Berlin, 2010, pp. 23–36.
[48] A. van der Heide, G. Trivino, A prototype of a machine speaking with emotion, in: ESTYLF ’10: Proceedings of the XV Spanish Congress on Technologies and
Fuzzy Logic, 2010.
[49] K. Popper, J. Eccles, The Self and Its Brain, Springer Verlag, 1977.

También podría gustarte