How specialized are specialized corpora Behavioral evaluation of

How specialized are specialized corpora? Behavioral evaluation of corpus representativeness for Maltese Jerid Francom (Wake Forest University) Adam Ussishkin (University of Arizona) Amy La. Cross (University of Arizona) 19 May 2010: O 7 (Evaluation of Methodologies), 14. 45 -15. 05 LREC 2010, Mediterranean Conference Center Valletta, Malta

Acknowledgements Generous contribution of data to this project by Dr. Albert Gatt (Univ. of Malta) Statistical expertise from Jeff Berry (Univ. of Arizona) Funding from the United States National Science Foundation (BCS-0715500) to Adam Ussishkin 2

Goals Issue For many languages, the quality of available textual data is less than ideal for corpus creation in the light of standard sampling practices. Propose Behavioral data can provide a valuable metric to evaluate corpus resources otherwise considered ‘specialized’. Case Psy. Co. L Maltese Lexical Corpus Contribute Novel, cross-discipline metric for evaluating the quality of language resources 3

Sparse coverage Most of the world’s 57000 languages have no corpus resources Efforts to fill the gap, often exploit the availability of language data on the web An Crúbadán project, 446 languages (Scannell, 2007) Mc. Enery et al. , (2006) survey of recent work 4

Sparse coverage Low-density languages (Borin, 2009) Languages in which resources exist; but in limited quantity/quality Limited access to print and/or electronic data Available primary data may be less-thanrepresentative Weakens assurance that results from lowdensity language resources are credible 5

Corpus representativeness What is a ‘representative corpus’? An externally valid sample of language use A sample that approximates what the language is. Full range of structural types (language units) What are the characteristics of such a sample? Genre/register Modality 6

An issue for low-density languages Standard practice to achieve representativeness Apply rigorous sampling methods Collect large amounts of data Problematic for low-density languages: a representativeness bottleneck Lack large amounts of data Available data is often limited in register, modality, etc. Corpus resources are typically specialized 7

Assessing representativeness How do we know whether we have a ‘representative’ sample? We don’t, in an absolute sense. Faith in survey sampling practices Casting the net far and wide Can we be assured we don’t have a representative sample? Not exactly. • It is logically possible that smaller, less diverse samples are externally valid for linguistic units that appear in the collection. 8

Proposal Need for an external metric. Current proposal suggests findings from behavioral experimentation can provide a valuable metric to evaluate corpus resources. Exploit the correlation between derived frequency counts and elicited behavioral reactions Behavioral data and adjusted frequency (Gries 2008; 2009) Of particular importance for specialized corpora 9

Behavioral findings Well-known robust effects for relative frequency in language processing Word naming RTs (e. g. , Forster & Chambers, 1973) Lexical decision RTs (e. g. , Carroll & White, 1973) Sentence reading RTs (e. g. , Mac. Donald, 1994) Word familiarity ratings (e. g. , Gernsbacher 1984) Log frequency is a good predictor of behavior. 10

Approach Evaluating corpus representativeness through behavioral assessment 1. Derive frequency counts from a specialized corpus 2. Elicit behavioral response of participants from target population 3. Assess correlation strength: how well do behavioral responses correlate with corpus measures? 11

Case study and predictions Case study Calculate: log frequency of subset of items in a Maltese lexical corpus Measure: subjective word familiarity ratings of native speakers of Maltese Assess: relative distribution of the measures Prediction Congruence between relative distributions indicates a representative sample of the language Mismatches underscore potential sampling issues 12

The specialized corpus Psy. Co. L Maltese Lexical Corpus (PMLC) (Francom, Ussishkin, and Woudstra, 2009) http: //psycol. sbs. arizona. edu/resources/ Online Maltese newspapers, 1998 -1999; 2005 2007 Psy. Co. L lab (59. 8%) and Dr. Albert Gatt (40. 2%) 3, 325 total tokens (53, 000 unique) Token/type ratio of 1. 6% Typical for low-density languages Large corpus, still relatively small (cf. British National Corpus 100+million; Corpus of Contemporary American English 400+ million) Limited in register, modality 13

Linguistic variable to quantify Because there is little previous quantitative research on Maltese, the empirical focus of this investigation was narrowed to: Semitic-origin verbs/binyanim (also known as form) Semitic-origin verbs in Maltese conform to the classical Semitic binyan system (categories based on morphosyntactic and phonological properties) Question: How does frequency as measured in our corpus correlate with behavior? Can the binyan categories be exploited to provide correlations? 14

Maltese binyanim Binyan Function Prosodic shape Example 1 basic active (transitive or intransitive) CVCVC kiser ‘he broke’ 2 intensive of 1, transitive of 1 CVCCVC kisser ‘he smashed’ 3 transitive of 1 CV: CVC bi: rek ‘he blessed 5 passive of 2, reflexive of 2 t. CVCCVC tkisser ‘it got smashed’ 6 passive of 2, reflexive of 3 t. CV: CVC tki: teb ‘he corresponded’ 7 passive of 1, reflexive of 1 n. CVCVC nkiser ‘it got broken’ 8 passive of 1, reflexive of 1 Ct. VCVC ftakar ‘he remembered’ 9 inchoative, acquisition of a quality CCV: C hma: r ‘he blushed’ 10 originally inchoative st. VCCVC stenbah ‘to wake’ 15

A behavioral task: word familiarity • • • We devised three tests to measure corpus representativeness Each test measured a different aspect of our corpus counts and our behavioral task. The behavioral task involved native Maltesespeakers, who gave subjective word familiarity ratings for all Semitic-origin Maltese verbs taken from Aquilina (2000); n=1536. Scale from very unfamiliar to very familiar Shown to be a reliable predictor of lexical processing (Connine et al. 1990) 16

Word familiarity experiment Participants 107 native speakers of Maltese Task Subjective word familiarity task, online 17

Measuring frequency in the corpus • • • We then used the PMLC to calculate word frequency measures for the same set of verbs. Using regular expression-enabled searching, we counted token frequency for all verbs occurring in the PMLC (n=447). Frequency was then encoded as a log-based measure. 18

Three tests • Next, we conducted three distinct statistical analyses to assess correlation between these corpus measures and the results of our word familiarity experiment • • • 1. Statistical regression between corpus log frequency and behavioral data. 2. Binned groups by frequency to determine whether any correlation is found. 3. Binned items by binyan to determine whether any correlation is found. 19

1. Statistical regression • We found a weak correlation (r=. 14); these results show at best a trend toward correlation, but suggests that familiarity ratings likely do not predict word frequency given these results. 20

2. Binning by frequency Binning into two bands shows a correlation: • • Binning into three bands also shows a correlation: 21

2. Binning by frequency • An LMER analysis of each binning (2 groups and 3 groups) shows significance: • • All contrasts for two-bin intervals (High/Low=4. 2, t=2. 0) and three-bin intervals (High/Mid=7. 1, t=3. 9; Mid/Low=7. 0, t=2. 2) were significant. These results support the hypothesis that behavior and corpus measures are correlated. 22

3. Binning by binyan • • Earlier and ongoing work (Frost et al. 1997, 1998, 2000; Ussishkin et al. in progress) shows binyan effects in Hebrew in both visual and auditory modalities, so Maltese could be expected to show similar effects. Our goal here is to measure whether verbs, when grouped by binyan, show a correlation between word frequency measures and word familiarity ratings. 23

3. Binning by binyan • Only binyanim 1, 2, 5, 7 were analyzed; binyanim 3, 6, 8, 9, and 10 were not included in the analyses because they are so sparsely populated: 24

3. Binning by binyan • • Word frequency results: significant contrasts found between Binyanim 7 and 2 (β=. 54, t=6. 0); and between Binyanim 7 and 5 (β=1. 15, t=-2. 2). Word familiarity results: no significant contrasts found. Binyan by word frequency Binyan by word familiarity 25

General assessment • • • The results show that verb frequency distributions in the PMLC pattern to some degree with the psychological representations of native speakers (the representative population) On the surface suggests the PMLC is on the right track, but underscores the specialized nature of corpus However, a response bias in the word familiarity task may play a part in the mismatches • Ceiling effect may have contributed to lower correlation scores 26

General assessment • Reasons to be optimistic about the verb distributions in the PMLC: • • • Distribution of verb count/ frequency (Zipf, 1949) Distribution of word length/ frequency (Li, 1992) Both measures trend as expected for representative samples 27

Conclusion • • Novel methodology: direct comparison between corpus resource and behavior. Highlighting a robust effect from psycholinguistics (frequency of linguistic units predicts behavior). We predicted the opposite could occur; this provides a way to validate LDL resources. This approach encourages cross-discipline endeavors for resource development and theoretical investigation. 28

• Thank you very much! • Grazzi ħafna! 29