LINGUISTICALLY BASED REASONING

1
LINGUISTICALLY BASED REASONING WITH A DISCOURSE
MODEL Dario Bianchi, Rodolfo Delmonte
2

• Getaruns
• a system for text and reference understanding
  which is currently used for summarization and
  text generation
• has a highly sophisticated linguistically based
  semantic module which is used to build up the
  Discourse Model.
• semantic processing is strongly modularized and
  distributed amongst a number of different
  submodules which take care of
• Spatio-Temporal Reasoning,
• Discourse Level Anaphora Resolution
• Topic Hierarchy
• Relevance Scoring
• Reasoning based on terminological logics
  (KL-ONE), supporting inheritance

3
Discourse Level Semantic Parser
4
We shall be working through the creation of the
knowledge base on the basis of the following text
from Sanford and Garrod At the Restaurant John
went into a restaurant. There was a table in the
corner. The waiter took the order. The air was
nice and clean. He began to read his book.
5
Discourse Model

• We can divide up the Discourse Model processes
  into three parts or levels
• Level 1 takes care of Topic Hierarchy and
  Anaphora Resolution.
• Level 2 takes care of temporal reasoning and the
  creation of semantic individuals.
• Level 3 takes care of rhetorical structure
  information, relevance score and builds the
  Discourse Model.

6
Discourse Model
Informally, a DM may be described as the set of
entities "naturally evoked, or in Sidner's(1983)
terms, "specified" by a discourse, linked
together by the relations they participate in.
They are called discourse entities, but may also
be regarded as discourse referents or cognitive
elements. We want to keep referring to what
people do with language evoking and accessing
discourse entities are what texts/discourses do.
A discourse entity inhabits a speaker's discourse
model and represents something the speaker has
referred to. A speaker refers to something by
utterances that either evoke (if first reference)
or access (if subsequent reference) its
corresponding discourse entity.
7
Discourse Model
In order to build an adequate Discourse Model we
rely on Situation Semantics. Situations are
characterized in terms of infons . A basic infon
consists of a relation together with a polarity
which in turn consists of an assignment of
appropriate arguments to the argument roles of
that relation. Facts and sit have to do with
concrete ostensive entities which yield
information that is referential. Facts are
objective (viewed from the text author) Sit are
subjective (viewed from the Main Topic of the
discourse).
8
Discourse Model

• Main components of situation are
• ind unique individuals
• set for collection of individuals whith a given
  card
• class for generic sets
• Each entity has associated an unique id and infon
  numerical constants.
• Semantic roles are inherited from the lexical
  form associated to a given predicate in the
  lexicon and transferred into the f-structure of
  the utterance under analysis.
• Infons have a location which is made up of a
  couple of indices anchoring the
  event/state/processes to a given spatiotemporal
  location and a polarity.

9
Semantic Rules

• After collecting all modifier heads, if any, of
  the current predicate, the rule for the creation
  of semantic individuals separates previously
  resolved pronouns/nouns from non resolved ones.
• It uses some sort of reasoning in order to
  ascribe properties to already asserted semantic
  identifiers, by taking advantage of linguistic
  information encoded in Function/Role.
• New semantic individuals are added when needed.
• The creation of semantic individuals has as its
  fundamental task that of treating separately new
  individuals to be asserted in the DM from already
  asserted ones in which case, the semantic index
  should be inherited from properties belonging to
  previously asserted individuals.

10
DISCOURSE MODEL FOR TEXT UNDER ANALYSIS sentence(
r01.new,john, went, into, a, restaurant)
loc(infon3, id1, argmain_tloc,
argtr(f5_r01)) loc(infon4, id2, argmain_sloc,
argrestaurant) ind(infon5, id3) fact(infon6,
inst_of, indid3, classman, 1, univ,
univ) fact(infon7, name, john, id3, 1, univ,
univ) fact(infon9, isa, argid2,
argrestaurant, 1, id1, id2) fact(id4, go,
agenteid3, locatid2, 1, tes(f5_r01),
id2) fact(infon10, isa, argid4, argev, 1,
tes(f5_r01), id2) fact(infon11, isa, argid5,
argtloc, 1, tes(f5_r01), id2) fact(infon12,
past, argid5, 1, tes(f5_r01),
id2) overlap(tes(f5_r01), td(f5_r01))
11
DISCOURSE MODEL FOR TEXT UNDER ANALYSIS sentence(
r02.new, there, was, a, table, in, the, corner)
ind(infon21, id6) ind(infon22,
id7) fact(infon23, inst_of, indid7,
classthing, 1, univ, univ) fact(infon24, isa,
indid7, classcorner, 1, id1,
id2) fact(infon25, in, argid6, locativoid7,
1, id1, id2) fact(infon26, isa, indid6,
classtable, 1, id1, id2) fact(infon27, inst_of,
indid6, classthing, 1, univ, univ) fact(id8,
there_be, tema_nonaffid6, 1, tes(f4_free_r02),
id2) fact(infon31, isa, argid8, argst, 1,
tes(f4_free_r02), id2) fact(infon32, isa,
argid9, argtloc, 1, tes(f4_free_r02),
id2) fact(infon33, past, argid9, 1,
tes(f4_free_r02), id2) included(tr(f4_free_r02),
id1) contains(tes(f4_free_r02), tes(f5_r01))
12
DISCOURSE MODEL FOR TEXT UNDER ANALYSIS sentence(
r03.new, the, waiter, took, the, order,
.) ind(infon42, id10) fact(infon43, inst_of,
indid10, classsocial_role, 1, univ,
univ) fact(infon44, isa, indid10,
classwaiter, 1, id1, id2) fact(infon45, role,
waiter, id2, id10, 1, id1, id2) fact(id12,
take_order, actorid10, goalid3, 1,
tes(f2_free_aq), id2) fact(infon48, isa,
argid12, argpr, 1, tes(f2_free_aq),
id2) fact(infon49, isa, argid13, argtloc, 1,
tes(f2_free_aq), id2) fact(infon50, past,
argid13, 1, tes(f2_free_aq),
id2) included(tr(f2_free_aq), id1) after(tes(f2_fr
ee_aq), tes(f5_r01))
13
Conceptual Representations Conceptual
Representations(CR) have been introduced by
Jackendoff and others. CR may be considered as
the link from the semantics to the knowledge of
the world needed to represent meaning in a
general and uniform manner. The Discourse Model
only contains reference to semantic roles and
other semantic relations like Poss, which have a
correspondence in the CR.
14
MAPPING SEMANTIC REPRESENTATIONS INTO THE
KNOWLEDGE BASE Here below are CRs for some of
the verb predicates of the text under
analysis. exist gt BE(lttheme_unaffectgt(STAYposit(
AT))) entergtCAUSE(ltagentgt(GOposit(FROMx(INTO
ltlocat_into)))) take_ordergtCAUSE(ltactorgt(GO(FROM
ltgoalgt))) readgtLET(ltaddressgt(GO(REP(FROMltinfor
mtngt)))
15
Inferential Rules from the Conceptual
Representations a) if an agent X causes E than
E takes place, under the condition that reference
time be specific CAUSE (X,E) at t1 gt E
cond specific(t1) b) the condition is the
subinterval meeting relation which is cast into
J.Allens formalism for temporal reasoning STAY
(X,AT Y) from t1 to t2 gt BE (X,AT
Y) at t3 cond t1ltt3ltt2 c) a motion
predicate may be translated into a couple of
state predicates GO(X,FROM Y,TO Z) at
t1 gt BE (X,AT Y) at t2 BE
(X,AT Z) at t3 cond t2ltt1ltt3
16
Inferential Rules from the Conceptual
Representations d) GO implies NOT STAY GO
(X,(AWAY_)FROM Y,TO(WARD) Z from t1, to t2
gt NOT STAY (X,AT Y) from t1, to t2
NOT STAY (X,AT Z) from t1, to
t2 e)STAY in a time interval implies NOT GO in
every subinterval STAY (X,AT Y from t1, to
t2 gt NOT GO (X,(AWAY_)FROM Y,
TO(WARD)W) from t3, to t4 cond
t1ltt3ltt4ltt2
17
Temporal logic Temporal reasoning was developed
for applications such as updating databases,
understanding narratives and dialogues, planing,
diagnosis and explanation. We have used the
fist-order logic formalism of Allen (1983,1984)
based on time intervals. There are 13 mutual
exclusive relations that connect any two
intervals before, meets, overlaps, during,
starts, finishes, equal and the inverse of the
first six relations. These relations are
transitive. For example (t1 meets t2 AND t2
meets t3) gt t1 before t3. The relations amongst
a number of intervals constitute a network. The
addition of a new relation affects the entire
network for the effect of the transitivity.
18
Time representation in sentences Reichenbach
proposed that the interpretation of the tense
system requires three entities the speech time
(ST), the reference time (RT) and the event time
(ET). The relation between RT and ST depends
only on tense. A second relation connecting ET
and RT can be obtained taking in account both
tense and aspect. Between ST and ET there is no
direct relation. Temporal modifiers are also to
be considered.
19
Time representation in discourse Many authors
have pointed out that also tense is anaphoric. To
interpret tense it is necessary to refer to some
time or event present in the context. From these
grounds Webber (1988) introduced an analogue of
the discourse focus that she named the temporal
focus (TF). The dynamically changing TF is the
entity in the temporal structure of states or
events that is the centre of listeners attention
and has the higher probability to be in anaphoric
relation with the temporal intervals introduced
by the next clause. In particular, in the time
structure of a tensed clause, it is the RT that
has an anaphoric character. To assign a temporal
structure to a text composed by several sentences
it is necessary to account for the movement of
the temporal focus as the narrative progresses.
Backward and forward movements of the time focus
is possible.
20

• Knowledge representation by KL-ONE
• KL-ONE (Schmoltz and Brachman, 1982) is based on
  the idea of
• structured inheritance networks.
• The basic taxonomy is expressed by means of a
  hierarchy of generic concetpts defined by
• its subsuming concepts (superConcepts)
• its local internal structure expressed by Roles
  which describe potential relations between
  instance of the of the Concept and those of
  other associated concepts such as properties,
  parts etc.
• Individuals are described by IndividualConcepts
  with assiciated Iroles.
• KL-ONE provides a structural base for descriptive
  knowledge
• Prolog provides a clausal form for assertional
  knowledge

21
The basic Taxonomy
ANYTHING
PRIMITIVES
TIMES
LOCATIONS
PROPERTIES
EVENTS
THINGS
HUMANS
22
POSS John began to read his book
ANYTHING
HUMANS
Agent
Theme
John
poss1
id18
23
POSS John began to read his book gtgtgt poss1
describe have
and time id14 and agent id3
and theme id18 and
oneof(poss1)
24
QUERIES TO THE SYSTEM (Retrieving agent and
goal of the predicative form) Who took the
order? ?- who_took_order. take_order
id12 agent id10 the waiter took the
order Who ordered? ?- who_ordered. take_order
id12 goal id3 john ordered
25
QUERIES TO THE SYSTEM (Spatial
reasoning) Where was the table? ?-
where_was(id6). id6 was in id7 of id2 the table
was in the corner of the restaurant Where was
the corner? ?- where_was(id7). id7 part of
id2 the corner was part of the restaurant
26
QUERIES TO THE SYSTEM (Temporal
reasoning) "Where was john after being entered "
?- where_was_after(id4,id3). id2 in the
restaurant Where was John? ?-
where_was(id3). After tes(f5_r01) was in
id2 after entering john was in the
restaurant (From sitiation main location) Where
was the waiter? ?- where_is(waiter). location
id2 the waiter is in the restaurant
27
QUERIES TO THE SYSTEM (Who/What corresponds to
uman/thing taxonomy) Who was in the
restaurant? ?- who_was_there(restaurant). id1
0 the waiter was in the restaurant What was
in the restaurant? ?- what_was_there(restaurant
). id7,id6,id17 in the restaurant there
was a corner a table a book