Association for Computers and the Humanities

Poster Sessions from ALLC/ACH '96

~ OVERVIEW ~

• Introduction
• Characteristics and Goals: Corpus Linguistics and MT
• Language Use and Construction Lexicons
• Corpus Development: Finding the Right Resource
• Analytical Challenges

~ Introduction ~

"Efforts in the development of natural language processing (NLP) and information technology are converging on the recognition of the importance of some sort of corpus-based research as part of the infrastructure for the development of advanced language processing applications." - Atkins, Clear and Ostler 1992

Why?

• Availability of on-line corpora
• Advances in storage capacity and computing power
• General renewal of interest in data-centered approaches

~ Corpus Linguistics ~

• Renewed Prestige - From Thirty Years Ago
  • Data was sparse then
  • Linguistic generalities could apply only to observed data
  • Intuitions were overlooked
  
  
• Characteristics - Study of Language Use
  • Historical lx'c events can predict potential lx'c events
  • Level of Language regularly employed
  • Understood by members of a linguistic community
  • Absent from most grammars
  
  
• Goals- Now
  Broad Coverage through
  • Observation-enhanced linguistic intuitions
  • Testing, confirming reformulated hypotheses
  • More effective computational grammars

~ Machine Translation ~

• Renewed Prestige - From Thirty Years Ago
  • ALPAC (1966) discredited MT efforts
  • Emphasis on formal linguistic theories
  • Major advances in computational power
  
  
• Characteristics - Study of Language Use
  • Limitations of formal theories: different focus
  • Direct and transfer approaches: syntactic level
  • Quality is leveling off: call for paradigmatic shift
  • Knowledge-based interlingua among new strategies
  
  
• Goals- Now
  Broad Coverage through
  • Robust, text-meaning, domain representations
  • Processing of non-literal language
  • Construction grammars

~ Language Use: Congruity of Efforts ~

• Usefulness and Limitations of Theoretical Models
  • Formal Grammars enhance processing
  • Rules are broken for communicative reasons
  • Much of linguistic expression based on different rules!
  
  
• Goals and Challenges of Broad Coverage
  • Identification of what is not covered
  • Use of morphologies, taggers and stop lists
  • Elimination of non-characteristic phenomena
  • Language Use
  
  
• Characterizing the Level of Language Use
  • MT research: non-literal language, constructions
  • Corpus linguistics: quantitative studies, characterizations
  • May vary according to text type
  • Normalcy and frequency of phenomena as criteria

~ Language Use: Traditions in Processing ~

• Interlingual MT Research
  • Top-down processing: Identify types then find tokens
  • Non-literal language: Types of deviation from literalness
  • Constructions: Types of non-compositional meaning
  
  
• Corpus Linguistic Linguistics Research
  • Bottom-up processing: Patterns in the text
  • Frequency studies: Account for what happens most often
  • Revise and refine hypotheses based on data
  
  
• Corpus Data to Inform MT Lexical-Semantic Design
  • Patterns, frequently-occurring phenomena and idiosyncratic expressions as representations of semantic and structural types
  • Efficiency demands that MT modules be designed to process phenomena which actually occur in text and not invented examples of it

~ Language Use: Description ~

• Linguistic Dichomoties
  • Grammatical v. ungrammatical structures
  • Competence v. performance models
  • Lexical v. syntactic representations
  
  
• Grammar Studies using the LOB Corpus - an example problem
  • NP structure types with one token apiece
  • Uncommon NP types in common constituent types
  
  
• Drawing the line in a grammar of language use - currency
  • Normative, no meta-grammatical or highly stylized language
  • Based on language recognition, not production skills
  
  
• Solutions
  • Lexical meaning can vary with form, e.g., "see" v. "seeing"
  • Language Use and Constructions

~ Language Use: Translation ~

• Translation Process
  • Distinguishes MT from other NLP applications
  • Translation theory emphasis:
    • Meaning of source language (SL) text
  • Interlingual MT research focus:
    • Representation of SL text meaning
  
  
• Organization of Present Project
  • Exploration of linguistic issue as MT research goal
  • Italian predicates of cognition, sensation and emotion
  • Selection of SL forms based on aspectual features
  • Observation of sense differences at lexical level in TL
  • Capturing of generalities
  
  
• Consequences
  • Some sense differences reflected in form alone
  • Morphological info included in semantic processing

~ Corpus Development ~

• Developing a Special-Purpose Corpus
  • Texts to Meet Linguistic Research Goal
    • Text types distinguished by internal structure
    • Informational texts containing Experiencer verbs
  • Texts to Meet MT Application Goal
    • Genres distinguished by external function
    • Material typically submitted for MT:
      • newspapers, journals, technical material
  
  
• Issues in Selecting Texts for Inclusion in the Corpus
  • Textual-linguistic features: What is being investigated?
  • Size: Are there enough tokens of the phenomenon?
  • Availability: Where does it reside? Protected?
  • Practical Issues: Copyright, format, tagging, etc.
  • Balance: Text types (synchronic), linguistic eras (diachronic)

~ Results of Survey of Available Italian Corpora ~

~ Corpus Development ~

• Solution for Present Study: The Italian Reference Corpus (IRC)
  • Linguistic Phenomenon
    • Ample tokens of Experiencer verb forms
  • Availability
    • KWIC runs available via ftp
    • Database Testuale (DBT) as interface at ILC
  • Restrictions: For research only
  • Reference Code tagged to every context
    • Permits token extraction from specific sources
    • Present study: non-informational texts excluded
  • Synchronic linguistic study permitted
    • Relevant texts in concise timeframe
    • Present study: contemporary language only

~ Italian Reference Corpus: A Subcorpus ~

Newspaper Subcorpus: Source Information

(Adapted from Bindi, et al. 1991:6-7)

~ Design of Analysis ~

• Analysis
• Phenomenon to be investigated
• Form/Sense matrix


• Italian SL Aspectual Changes
• Morphologically encoded - other levels not analyzed
• Signalling differences at lexical level in English TL


• Verb Forms Reflect Grammatical Aspect or Viewpoint Type
• Imperfect: continuous action - imperfective aspect
• Preterit: completed action - perfective aspect


• Senses Reflect Inherent or Situational Aspect: e.g., pensare
• cause to have a mental representation: "think up"
• have a mental representation: "think"
• Ha pensato v. pensava
• Complications: complementation, phrasals, constructions

~ Italian Experiencer Verbs: Frequency Study Data ~

Most Frequent Experiencer Verbs in Three Studies: Sciarone `77, Bortolini, et al `71, and Traversa `73

Hypothesis 1:

A verb sense which entails inherent continuous aspect will, in forms expressing the grammatical aspect of completed action, indicate the inception of the verb sense.

Hypothesis2:

A verb sense which entails inherent completed aspect will not appear in forms expressing the grammatical aspect of continuous action.

~ Verb Senses ~

Pensare:

1. Examine with thought or imagination; reason; think about; judge
2. Hold a fixed thought; believe; suppose
3. Invent; think up
4. Take care of; attend to

Sentirsi:

1. Be aware of a feeling or an emotive situation
2. Experience a psychic sensation
3. Be conscious of a physical sensation

Vedere:

1. Perceive with the eyes
2. Encounter; meet; consult; visit
3. Perceive with the mind; grasp; realize; understand; imagine

~ Pensare: Examples ~

H1.

Sense 1 of pensava (durative):
Giacomo Bove pensava a una possibile emigrazione

Sense 1 of pensò (inceptive):
facevano il giro delle corti europee. Quindi, pensò [...] di dimostrare...

H2.

Sense 3 of pensò:
Un giorno, pensò, finiremo per conoscerci [...]

Similar structure with pensava (Sense 2):
<<Io andrò piì lontano>> pensava Giuanin, ambizioso [...]

~ Sentirsi: Observations ~

H1.

Grammatical aspect of completed action is interpreted as temporary duration, often indicated by accompanying adverbial, rather than inception:

[...] allora sì che mi sentii ricco e felice [...]

H2.

Hypothesis not testable as no senses entail inherent completed aspect.

General Comments

~ Vedere: Observations ~

H1.

Grammatical aspect of completed action signals the mental act of envisioning.

Sense 3: videro
[...] la fantasia dei governanti che videro nell'iniziativa

Sense 3: vedevano
[...] i conservatori lo vedevano come socialista.

H2.

Hypothesis not testable: All senses present in data entail inherent continuous aspect.

General Comments

~ Analytical Challenges ~

• Language Use/Construction Criteria
• Flexible enough for lexical variation?
• Possibility of good structural description?
• High frequency of occurrence? Not too deviant?


• Careful Selection of Material
• Genres relevant to the application
• Text types with similar features
• Material which contains the phenomenon!


• Observe patterns in Form/Sense Matrix
• Grammatical v. inherent aspect
• Anomalies as candidates for unique treatment

Go back to the ACH Home Page

Any questions or bug reports regarding the ACH Web Pages should go to ACHWeb@brown.edu