Probing Classifiers: Promises, Shortcomings, and Advances

Probing classifiers have emerged as one of the prominent methodologies for interpreting and analyzing deep neural network models of natural language processing. The basic idea is simple —a classifier is trained to predict some linguistic property from a model's representations—and has been used to examine a wide variety of models and properties. However, recent studies have demonstrated various methodological limitations of this approach. This article critically reviews the probing classifiers framework, highlighting their promises, shortcomings, and advances.

Foreign Accent in Pre- and Primary School Heritage Bilinguals

Previous research has shown that the two languages of early bilingual children can influence each other, depending on the linguistic property, while adult bilinguals predominantly show influence from the majority language to the minority (heritage) language. While this observed shift in influence patterns is probably related to a shift in dominance between early childhood and adulthood, there is little data documenting it. Our study investigates the perceived global accent in the two languages of German-Russian bilingual children in Germany, comparing 4–6-year-old (preschool) children and 7–9-year-old (primary school) children. The results indicate that in German the older children sound less accented than the younger children, while the opposite is true for Russian. This suggests that the primary school years are a critical period for heritage language maintenance.

Crosslinguistic Interference in Simultaneous Acquisition of Turkish and Italian

This study compared the acceptability judgment task (AJT) and translation task (TT) in 10 simultaneous Turkish–Italian bilinguals with age-matched Turkish and Italian monolinguals as a control group. The position of subjects occurring with unaccusative verbs was investigated as the linguistic property. The results of the AJT in Turkish revealed no significant difference between the groups. However, in the Italian AJT, high variability was found in the younger bilinguals. Only bilinguals could participate in the TT. The participants, while translating into Turkish, always produced pragmatically appropriate sentences. In the case of translation to Italian, a significant effect correlated with the position of the subject in the Turkish stimulus was found.

Bilingual children and adult heritage speakers: The range of comparison

This paper compares the language of child bilinguals and adult unbalanced bilinguals (heritage speakers) against that of bilingual native speakers of their home language (baseline). We identify four major vectors of correspondence across the language spoken by these three groups. First, all varieties may represent a given linguistic property in a similar way (child bilinguals = adult heritage speakers = bilingual native speakers of their home language). This occurs when either (i) the property in question is highly robust and is acquired by learners without difficulty or (ii) the property is already in decline in the baseline. We illustrate scenario (i) with data from Russian count forms, which are morphologically quite complex. The preservation of these forms in child bilinguals and adult heritage speakers suggests that simplicity of encoding is not the only factor determining robustness of retention. Second, child and heritage speakers may share a linguistic structure that differs from the one found in the baseline (bilingual native speakers of their home language ≠ child bilinguals = adult heritage speakers). This scenario occurs when incipient structural changes in the baseline become amplified in the language of next-generation bilinguals, or when a given structure is rare, confined to a specific register, and/or reinforced through literacy. Third, a structure may be acquired by bilingual children faithfully, but undergo reanalysis/attrition in the adult heritage language (bilingual native speakers of their home language = child bilinguals ≠ adult heritage speakers). Russian relativization illustrates this scenario; child bilinguals show native-like performance on relative clauses but adult heritage speakers show an exaggerated subject preference in the interpretation of gaps. Finally, a structure that is not fully learned by child speakers may be reanalyzed by adult heritage speakers following general principles, thus bringing the adult heritage representation closer to that of the baseline (bilingual native speakers of their home language = adult heritage speakers ≠ child bilinguals). Heritage speakers’ production and comprehension of psychological predicates in Spanish illustrates this possibility.

Downward entailment in child Mandarin

ABSTRACTThere are three hallmarks of core linguistic properties. First, they are expected to be manifested in typologically different languages. Second, they should unify superficially unrelated linguistic phenomena. Third, they are expected to emerge early in the course of language development, all things being equal (Crain, 1991). The present study investigates a candidate for a core linguistic property, namely the semantic property of downward entailment. We report the findings of two experimental studies of children's knowledge of downward entailment. These experiments explore two different aspects of downward entailment, in a study with Mandarin-speaking children. Taken together with previous research findings, the results of the present study support the conclusion that downward entailment is a core property of human languages.

Phonology Matters: The Phonological Frequency Effect in Written Chinese

Does phonology play a role in silent reading? This issue was addressed in Chinese. Phonology effects are less expected in Chinese than in alphabetical languages like English because the basic units of written Chinese (the characters) map directly into units of meaning (morphemes). This linguistic property gave rise to the view that phonology could be bypassed altogether in Chinese. The present study, however, shows that this is not the case. We report two experiments that demonstrate pure phonological frequency effects in processing written Chinese. Characters with a high phonological frequency were processed faster than characters with a low phonological frequency, despite the fact that the characters were matched on orthographic (printed) frequency. The present research points to a universal phonological principle according to which phonological information is routinely activated as a part of word identification. The research further suggests that part of the classic word-frequency effect may be phonological.

Structural Regularities vs. Definition Patterns

In a paper dealing with the computational analysis of dictionaries, Boguraev ("Building a Lexicon: The Contribution of Computers", in International Journal of Lexicography, 4:3, Autumn 1991) claims that exploiting definition patterns in the Longman Dictionary of Contemporary English (PROCTER, 1978) to identify verbs participating in the causative/inchoative alternation misses a number of verb candidates. His contention is that this linguistic property of verbs is reflected in the structure of dictionary definitions and not in the language of definitions. However, the algorithm he suggests for extracting causative/inchoative verbs from LDOCE on the basis of purely structural regularities is not sufficiently selective insofar as it also identifies verbs which participate in other types of transitivity alternations, such as, for instance, the reciprocal alternation.