Temporal Relations with Signals: the case of Italian Temporal Prepositions

Temporal Relations with Signals: the Case of Italian

Temporal Prepositions

Tommaso Caselli, Felice dell’Orletta and Irina Prodanof

{[email protected]}

ILC-CNR, Pisa

16th International Symposium on Temporal Representation and Reasoning TIME 2009

Bressanone/Brixen, July 24 2009

Different approach Application oriented NLP techniques Focus on: intuitions, knowledge and strategies people use in order

to

place events in time order events (encoding and decoding)

Query texts (corpora) and NOT structured knowledge

Introduction

Outline: Motivations

Temporal Signals in Italian: Theoretical background

Methodology

Corpus Study

A Maximum Entropy Model Feature Identification

Evaluation and Results

Conclusion and Future Work

Motivations Recovering temporal relations in text/discourse is essential to

improve the performance of many NLP systems (O.D-Q.A., Text Mining, Summarization, Reasoning)

Most temporal information in text/discourse is only IMPLICITLY stated

Need to develop procedures to maximize the role of the various sources of information

Temporal prepositions are a partially explicit source of information.

Determinig their meaning is part of a strategy to improve the extraction of temporal information

Motivations (2)

SIGNAL = cover term for a homogeneous class of words which express relations between textual entities

EXPLICIT = self-evident and stable meaning; Rel (X, Y)

IMPLICIT = abstract meaning which gets specialized in the co-text; Rel (λ(X), λ(Y))

Temporal signals express temporal relations.

Temporal signals can occur in 3 types of constructions involving temporal entities:

temporal expression – temporal expression

eventuality – temporal expression

eventuality - eventuality

Theoretical Background

Corpus Study: Data

To identify a large set of temporal signals realized by prepositions we have conducted a corpus study:

5 million shallow parsed word corpus (from the PAROLE corpus)

all PP chunks with their left and right contexts have been automatically extracted and imported into a database structure

automatically generated DB augmented with ontological information from the SIMPLE/CLIPS Ontology, by associating the head noun of each PP chunk to its ontological type

extraction of the noun head corresponding to type TIME + postprocessing to exclude false positives (e.g. incubation, school…)

Temporal relations coded by implicit signals:

annotation of temporal relations by means of paraphrase tests

e.g. [sono stato sposato] per [4 anni] (I’ve been married for four years)

The state of “being married” EQUALS four years

499 occurrences of construction of the type “eventuality + signal + temporal expressions”

9 temporal relations (compliant with TimeML and ISO-TimeML): overlap, simultaneous, before, after, no tlink, begin, end, before_ending, equals

the most frequent temporal relation/implicit signal is assumed to be the prototypical meaning of the signal

Corpus Study (2)

The corpus study together with theoretical statements have led to the identification of 16 features:

PREP: the signal lemma

3 sets of co-textual feature:

information about temporal expression

information about the eventuality

local contextual information

Feature Identification

Temporal expression features:

Ontological status: INSTANT, INTERVAL

Type of temporal expressions (TIMEX):

DATE: August 3; 1968; 01/12/1980…

DURATION: 3 hours; the last quarter…

SET: once every year…

TIME: 3 o’ clock; (in) the morning…

Presence of a quantifier: QUANTIFIER

Feature Identification – Temporal Expressions

Eventuality features:

Lemma (POTGOV_head);

POS of the eventuality: VERB, NOUN

Presence of negations (NEGATION)

Verb diatesis (DIATESIS)

Tense: PRESENT, IMPERFECT, FUTURE, PAST, INFINITIVE

(Viewpoint) Aspect: IMPERFECTIVE, PERFECTIVE, PROGRESSIVE, NONE

Lexical Aspect (AKTIONSAART): TRANSITION, PROCESS, STATE

Feature Identification - Eventuality

Local context features: features which accounts for the presence of further signals in the local context which influence the identification of the Rel value of the signal in analysis

FOLLOWED_SIGNAL+TIMEX

PRECEED_SIGNAL+TIMEX

FOLLOWED_SIGNAL+EVENT

Feature Identification – Local context

Feature annotation: manually conducted by one annotator + one of the author.

1000 instances of constructions of the type “eventuality + signal + timex”

• two interlinked criteria: semantic transparency of the signal + relative frequency of the signal in the 5 million shallow parsed corpus

Assigning the right temporal relation is (in essence) a tagging task.

Maximum Entropy algorithm: it provides a suitable solution to identify the set of possible values for each signal on the basis of the conditional probability distribution. No a priori constraints must be met other than those related to a set of features fi(a, c) of a context C, whose distribution is derived from the training data.

Building a M.E. Model

Evaluation

The data set has been split in test (100) and training (900) data

8 different models have been created to discover the most salient features. 10- cross fold validation/model.

All models outperforms the baseline relevance of the features

Evaluation (2)

1. PREP

2. INTERVAL3. INSTANT4. POTGOV_head5. VERB6. NOUN7. DIATESIS8. NEGATION

9. AKTIONSAART10. FOLLOWED_SIGNAL+TIMEX11. PRECEED_SIGNAL+TIMEX12. FOLLOWED_SIGNAL+EVENT13. TENSE14. ASPECT15. TIMEX16. QUANTIFIER

10 Feature Model

Performance = 90%

surface-based features

good performance without the AKTIONSAART feature

Evaluation (3)

Model Performance Features

9 features 89.8%

PREP, INTERVAL,

INSTANT, AKTIONSAART,

FOLLOWED_SIGNAL+TIMEX,

PRECEED_SIGNAL+TIMEX,

FOLLOWED_SIGNAL+EVENT,

TIMEX, QUANTIFIER

8 features 89.8%

PREP, INTERVAL,

INSTANT,

FOLLOWED_SIGNAL+TIMEX,

PRECEED_SIGNAL+TIMEX,

FOLLOWED_SIGNAL+EVENT,

TIMEX, QUANTIFIER

Evaluation (3)

Model Performance Features

8 features (No QUANTIFIER)

85%

PREP, INTERVAL,

INSTANT, AKTIONSAART,

FOLLOWED_SIGNAL+TIMEX,

PRECEED_SIGNAL+TIMEX,

FOLLOWED_SIGNAL+EVENT,

TIMEX

7 features 86.8%

PREP, INTERVAL,

INSTANT,

FOLLOWED_SIGNAL+TIMEX,

PRECEED_SIGNAL+TIMEX,

FOLLOWED_SIGNAL+EVENT,

TIMEX, QUANTIFIER

5 features 87.6%

PREP, INTERVAL,

INSTANT, TIMEX, QUANTIFIER

Mismatch between linguistic theory and features salience

Observations on the features:

5 core features: PREP, INSTANT, INTERVAL, TIMEX, QUANTIFIER (5 feature model)

AKTIONSAART influence in this task is almost null. It could be reduced with a set of features more surface-based e.g. presence of D.O., definiteness, cardinality, type of subject…

the remaining features could be activated in particular linguistic context and with particular signals; e.g. TENSE, ASPECT and AKTIONSAART (ot its subsitutes) with the signal IN; the local context features with the signals DA, A and TRA.

Conclusion & Future Work

Integration of the M.E. Model into a complete automatic system for temporal processing of text/discourse

Thanks