Uninterpretable Features Research Articles

Interpretable Neuron Structuring with Graph Spectral Regularization.

While neural networks are powerful approximators used to classify or embed data into lower dimensional spaces, they are often regarded as black boxes with uninterpretable features. Here we propose Graph Spectral Regularization for making hidden layers more interpretable without significantly impacting performance on the primary task. Taking inspiration from spatial organization and localization of neuron activations in biological networks, we use a graph Laplacian penalty to structure the activations within a layer. This penalty encourages activations to be smooth either on a predetermined graph or on a feature-space graph learned from the data via co-activations of a hidden layer of the neural network. We show numerous uses for this additional structure including cluster indication and visualization in biological and image data sets.

Acquisition of Definiteness And Indefiniteness by Behdini Kurdish Learners of English

The acquisition of English grammatical articles by non-native speakers of English language has been the main concern of a wide range of research, especially in languages whose grammatical article system functions differently from the English grammatical article system. Behdini is a variety whose grammatical article system is different from that English grammatical parameter. Thus, the acquisition of L2 English articles among L1 Behdini speakers is investigated in this study along proposals based on the Interpretability Hypothesis (Hawkins & Hattori, 2006; Tsimpli & Dimitrakopoulou, 2007; Hawkins & Casillas, 2008), and the Full Transfer Full Access (FT/FA) hypothesis (Schwartz and Sprouse, 1994, 1996). The researcher developed a Judgement Elicitation Task (JET)   based on the following research question: Will Behdini learners’ English language proficiency level play a role in participants’ judgements? Will these learners be able to acquire English definite, indefinite, and zero articles systematically? In other words, is any development expected in Behdini learners’ acquisition of English articles? 40 Behdini L2 English learners took part in the Judgement Elicitation Task . English learners took an English language proficiency test based on which they were roughly grouped into three sub-groups: elementary (11 participants), intermediate (22 participants), and advanced (7 participants). All participants took two tests: an acceptability judgment test including 46  items (34 pure test items + 12 filler gaps) and a proficiency test with 40  items. The results of the study show that there is a transfer from L1 into the L2ers’ English interlanguage due to the finding that most participants, disregarding their proficiency level, failed to reject the ungrammatical sentences in almost all the categories. The study finds support for the FT/FA proposal, where first language transfer, second language input and access to universal grammar features are argued to have impacts on Learners of English article acquisition among L1 Behdini learners. The result  also show that learners of  English  fluctuate in their choice of articles, which is interpreted by the predictions made by the Fluctuation Theory. Uninterpretable features also, proved to be difficult to acquire by Behdini learners, a point that  can be interpreted the base form of the Interpretability Hypothesis.&nbsp

Uninterpretable features in learning and alternative grammars?

Uninterpretable features in learning and alternative grammars?

Are uninterpretable features vulnerable?

Are uninterpretable features vulnerable?

Complexity as L2-difficulty: Implications for syntactic change

AbstractRecent work has cast doubt on the idea that all languages are equally complex; however, the notion of syntactic complexity remains underexplored. Taking complexity to equate to difficulty of acquisition for late L2 acquirers, we propose an operationalization of syntactic complexity in terms of uninterpretable features. Trudgill’s sociolinguistic typology predicts that sociohistorical situations involving substantial late L2 acquisition should be conducive to simplification, i.e. loss of such features. We sketch a programme for investigating this prediction. In particular, we suggest that the loss of bipartite negation in the history of Low German and other languages indicates that it may be on the right track.

A document representation framework with interpretable features using pre-trained word embeddings

We propose an improved framework for document representation using word embeddings. The existing models represent the document as a position vector in the same word embedding space. As a result, they are unable to capture the multiple aspects as well as the broad context in the document. Also, due to their low representational power, existing approaches perform poorly at document classification. Furthermore, the document vectors obtained using such methods have uninterpretable features. In this paper, we propose an improved document representation framework which captures multiple aspects of the document with interpretable features. In this framework, a document is represented in a different feature space by representing each dimension with a potential feature word with relatively high discriminating power. A given document is modeled as the distances between the feature words and the document. To represent a document, we have proposed two criteria for the selection of potential feature words and a distance function to measure the distance between the feature word and the document. Experimental results on multiple datasets show that the proposed model consistently performs better at document classification over the baseline methods. The proposed approach is simple and represents the document with interpretable word features. Overall, the proposed model provides an alternative framework to represent the larger text units with word embeddings and provides the scope to develop new approaches to improve the performance of document representation and its applications.

Aquisição de pronomes clíticos de português europeu por falantes de chinês: dados sobre a colocação

This paper presents a pilot-study of a research which aims to investigate the acquisition of the clitic pronouns in European Portuguese (EP) by Chinese learners. Based on an acceptability judgement task, conducted with 20 Chinese students and a control group of 26 native speakers of EP, we would like, on the one hand, to make a better characterization of the development of placement properties of clitic pronouns in Chinese learners’ interlanguage grammars; on the other hand, to contribute to a better understanding of the theoretical proposals on the learnability problems regarding uninterpretable features not available in the first language (L1). Our preliminary results suggest that, as shown in previous works on L1 EP and L2 EP, the conditions which determine proclisis seem to be more difficult to acquire than those for enclisis and the different contexts of proclisis seem to develop gradually, following a similar path to that previously described for L1 EP.

Accounting for the asymmetrical interpretation of thematic and non-thematic verbs in L2 English

AbstractThis study presents data from an experiment on the interpretation of thematic and non-thematic verbs in second language (L2) English by three groups of adult native Chinese speakers and a group of native English speakers. English allows non-thematic verbs to raise but requires thematic verbs to remainin-situ. In contrast, neither thematic nor non-thematic verbs are allowed to raise in Chinese. The results indicate that there is a discrepancy between native and non-native mental representations of the grammars concerned; whereas native grammars require English thematic verbs to remainin-situbut allow non-thematic verbs to raise, neither thematic nor non-thematic verbs are allowed to raise in learners L2 English grammars. Results of the study argue against theValueless Features Hypothesis(Eubank 1993/94. On the transfer of parametric values in L2 development.Language Acquisition3. 183–208, 1994. Optionality and the initial state in L2 development. In T. Hoekstra & B. Schwartz (eds),Language acquisition studies in generative grammar: Papers in Honour of Kenneth Wexler from the 1991 GLOW Workshop, 369–388. Amsterdam: John Benjamins, 1996. Negation in early German-English interlanguage: More valueless features in the L2 initial state.Second Language Research12. 73–106.), which posits that the L1 syntactic features of INFL are initially inert and are not transferred. Instead, the results support theInterpretability Hypothesis(Tsimpli and Dimitrakopoulou 2007. The interpretability hypothesis: Evidence from wh-interrogatives in second language acquisition.Second Language Research23. 215–242.), which argues for the inaccessibility of uninterpretable syntactic features beyond a critical period. In particular, it is argued that uninterpretable syntactic features not selected during early stages of primary language acquisition become inaccessible in subsequent language acquisition. The results suggest that there may be cases where apparent target-like performance conceals non-target-like underlying competence.

Negative concord in Washo as negative agreement

Washo, a Native American isolate, displays negative concord morphology in the context of negation. Negative concord in Washo comes in the form of the morpheme -Na, which may be suffixed onto optionally many sentential elements in a single clause. Given the apparent lack of semantic contribution by this morpheme, I argue - building on accounts along the lines of Zeijlstra (2004) - that negative concord in Washo is best treated as the result of multiple agreement between an interpretable Neg feature high in the clause and optionally many uninterpretable Neg features present on the items it c-commands. Importantly, the Washo data shed light on an unexplored system of negative concord and the possible range of agreement phenomena that are sensitive to negation.

Object Omission in Contact: Object Clitics and Definite Articles in the West Thracian Greek (Evros) Dialect

We examine spontaneous production data from the dialect of Modern West Thracian Greek (mwtg) (the local dialect of Evros) with regard to a hypothesis of syntactic borrowing of verbal transitivity. We argue that mwtg allows omission of the direct object with specific reference, in contrast to Standard Modern Greek (smg) and other Modern Greek (mg) dialects (spoken in Greece), but similar to Turkish. Object omission in mwtg is possible only in contexts where smg and other mg dialects show obligatory use of the 3rd-person clitic. We argue that syntactic borrowing in the case of language contact follows the transfer with second language learners: the relevant elements that host uninterpretable features are used optionally. Moreover, the definite article, in contrast to the indefinite article, is also affected by language contact. The 3rd-person clitic and the definite article are affected by contact as uninterpretable clusters of features. We claim that interpretability plays a significant role in transitivity in cases of language contact.

The Syntactic Status of Arabic Clauses Introduced by ʔinna A Minimalist Study

A Standard Arabic (SA) complementizer known as ʔinna poses a restriction on word orders in the clause it introduces and induces accusative Case-marking on the otherwise nominative preverbal NPs (Note1). Following Chomsky’s (2001) account of the morphosyntax of Case, this paper argues that ʔinna is a Case assigner and thus it carries an uninterpretable Case feature that determines the value which it assigns to an unvalued Case feature concerning accessible goal within A-bar projection. The paper shows that this argument captures the asymmetrical word order between clauses introduced by ʔinna and those headed by null CPs. 

ɸ-Features at the Syntax-Semantics Interface: Evidence from Nominal Inflection

I argue for a novel model of feature valuation in the CI interface and explore under what circumstances a syntactic feature is semantically interpretable. As the groundwork for the investigation, I propose an explicit Distributed Morphology model of Italian nouns of profession. The data provide evidence that the morphology accesses the narrow-syntax representation at two different temporal points within a phase: the earlier point (Spell-Out) returns a morphological realization faithful to feature values present in narrow syntax, while the later point (Transfer) allows for a narrow-syntax representation to be enriched by the CI component. Thus, there is no syntactic distinction between interpretable and uninterpretable features: a syntactic feature appears to be interpretable only if it has been licensed by the CI interface.

Pre-emption of L1 properties in the L2 acquisition of Englishwh-interrogatives: Effects of L2 proficiency and age of onset

AbstractThis study investigates whether second language (L2) speakers can pre-empt a first language (L1) property which involves uninterpretable features, such as resumption. The Interpretability Hypothesis predicts persistent L1 effects in L2 grammars because uninterpretable features resist resetting beyond some critical period (Tsimpli and Dimitrakopoulou 2007). Unlike English, Saudi Arabic allows grammatical resumption in complexwh-interrogatives, which is highly preferred with (D)iscourse-linkedwh-forms (e. g.ʔayy-NP ‘which-NP’) but disallowed with non-D-linked ones (e. g.ʔeeʃ‘what’). The study was conducted with fifteen native English speakers and 34 (very)-advanced Saudi Arabic L2 speakers of English with age of onset (AO 1–13 years). In a bimodal, timed acceptability judgment task, their accuracy judgments of 32 (un)grammaticalwh-interrogatives were tested. As predicted, results show that L2 speakers of very advanced levels inaccurately accepted resumption especially with D-linkedwh-interrogatives. The results also show non-significant differences between AO 1–6 and 7–13 years in their rejection accuracy of resumption.

A compounding analysis of plural reduplication

The paper offers an analysis of plural reduplication in Japanese. I argue that reduplicated nouns are antonym compounds that have two opposite values of number features. Each member of a reduplicated noun is associated with an interpretable valued number feature, and values of the number features must not be identical to each other. I propose that two opposite values on the top node of a reduplicated noun becomes an uninterpretable valued feature in Japanese. As a result, reduplicated nouns in Japanese are not specified as singular or plural because uninterpretable features do not have any semantic import. The proposal makes the prediction that reduplicated nouns are similar to bare common nouns in the sense that they do not bear the specification of singular-plural distinction. I show this prediction is borne out. The proposal also leaves room for an analysis of typological variation of plural reduplication.

A compounding analysis of plural reduplication

The paper offers an analysis of plural reduplication in Japanese. I argue that reduplicated nouns are antonym compounds that have two opposite values of number features. Each member of a reduplicated noun is associated with an interpretable valued number feature, and values of the number features must not be identical to each other. I propose that two opposite values on the top node of a reduplicated noun becomes an uninterpretable valued feature in Japanese. As a result, reduplicated nouns in Japanese are not specified as singular or plural because uninterpretable features do not have any semantic import. The proposal makes the prediction that reduplicated nouns are similar to bare common nouns in the sense that they do not bear the specification of singular-plural distinction. I show this prediction is borne out. The proposal also leaves room for an analysis of typological variation of plural reduplication.

Syntax-Semantics Interface and Licensing of Nominals.

This paper inquires how Syntax-Semantic interface interacts with nominal licensing within the framework of the Minimalist Program (Chomsky 1995 et seq), a derivational model of grammar under which Syntax strictly precedes Semantics. Within this view, syntactic operations should not be motivated by semantic features even though they may be present in Syntax. The inquiry is pursued by investigating three seemingly independent case phenomena in Korean: Accusative adverbial nominals, Accusative Case alternation/stacking and Part-whole constructions. I demonstrate that Syntax-Semantics compatibility is crucial to licensing of nominals in Korean and propose that nominal licensing in general may even be reduced to LF legibility. The theoretical implication is that Legibility Conditions inspect not only the erasure of uninterpretable features but also non-feature-related, general semantic well-formedness.

C-T Inheritance and the left periphery in Old Japanese

This paper proposes an analysis of movement to the left periphery in Old Japanese within the framework of C-T Inheritance. Particular attention is given to the fact that nominative subjects precede focused constituents in focus concord constructions known as kakari-musubi (KM). I propose that this order is ensured by the nature of the features motivating these respective movements. Following Richards (2007), I propose that uninterpretable features – particularly those seeking a value – must be spelled out as soon as they are valued. Consequently, these features must be inherited so that they are in the domain of the phase head when spell out takes place. Building on earlier analyses of KM constructions as involving agreement between a focus particle and the verbal inflection which covaries with that particle (Ikawa 1998; Watanabe 2005; Kuroda 2007, and others), I propose that focused constituents must move to [Spec, TP], since their movement is motivated by an unvalued focus feature. In contrast to this, subject movement to value nominative case is not driven by an uninterpretable probe on the phase head like [uϕ], because Japanese lacks subject/verb agreement. Following Saito (2016), I propose that subjects undergo movement agnostically in order to value their own case features. Since there is no probe on C driving this movement, inheritance does not take place, and the subject moves to [Spec, CP], with the result that it precedes the focused constituent in surface word order.

Selective vulnerability in very advanced L2 grammars: Evidence from VPE constraints

This study examines whether the second language acquisition (L2A) of syntactic properties at the interfaces is problematic for L2 learners. English verb phrase ellipsis (VPE) was tested as an interface property which involves feature interpretability. Two subtle contrasts of VPE at different grammar-internal interfaces were examined: (a) copula be and lexical verb ellipsis at the syntax–lexicon interface (*John is here, and Mary will too vs. John slept, and Mary will too), and (b) progressive be and perfective have stranding at the syntax–semantics interface (*John slept, and Mary was too vs. Peter saw your parents last week, but he hasn’t since). Unlike (a), the contrast in (b) requires identifying (un)interpretable features on -ing and -en to recover the semantics of the elided material. Since Saudi Arabic does not license VPE, (very-)advanced English L2 speakers whose first language (L1) is Saudi Arabic were tested. The results from a bimodal timed acceptability judgment task showed they were more target-like on the contrast at the syntax–lexicon interface than they were on the syntax–semantics interface. They particularly deviated from target-like judgments on have stranding which requires recognizing the perfective uninterpretable feature on - en. These results suggest selective vulnerability at grammar-internal interfaces whenever uninterpretable features are involved.

Nump and Possp in Dialectal Brazilian Portuguese

In standard Brazilian Portuguese (BP), as well as in other Romance languages, possessives have uninterpretable number features, which are valued via nominal agreement. However, dialects of BP, especially the one spoken in Minas Gerais, have shown that 2nd person possessives, in postnominal position, do not have number agreement with the noun. In order to account for these facts, I will argue that, in this grammar, number features on 2nd person possessives are reanalyzed as being: (i) associated with the person (rather than the noun) and (ii) valued. From the frst postulation, ‘seu' is expected to be the possessive for 2nd person singular, and ‘seus' for 2nd person plural. From the second postulation, no number concord is expected to be triggered on the possessive. In addition, based on Danon (2011) and Norris (2014), I will argue that cardinals divide BP DPs into two domains in that phrases located above NumP are marked with the plural morpheme, while phrases below it are unmarked. In this sense, because prenominal possessives precede cardinals (NumP), they must be marked with the plural morpheme for nominal agreement; whereas postnominal possessives, which follow NumP, must be unmarked. Free from the plural marking associated with nominal agreement, postnominal 2nd person possessives favor the reanalysis of the morpheme ‘-s' as indicating the number associated with person features.

The Case of special qui

ABSTRACTFeature Inheritance (Richards 2007) entails that uninterpretable features originate on the phase head C or v* and are then transferred to the associated Agree head, T or V. In the present article, it is argued that the French que–qui alternation is the locus of a Case contrast, implying that nominative Case originates on the complementiser and only becomes associated with T as a consequence of feature transfer. Que–qui thus provides new, Case-based empirical support for the theory of Feature Inheritance. The article also suggests that the que–qui alternation has an important implication for Chomsky's recent application of dynamic antisymmetry, reinterpreted in terms of labelling, to the issue of subject extraction failure. Specifically, the alternation appears to indicate that Case-matching is required, in addition to phi-feature agreement, in order for extraction to be blocked by labelling.