In recent years, both quantitative approaches to the implicative structure of paradigms (Wurzel 1989) and non-canonical properties of paradigms such as overabundance (Thornton 2011) have received growing attention from morphological theory. However, the most prominent framework to quantitatively assess implicative structure, the Paradigm Cell Filling Problem (PCFP, Ackerman et al. 2009), does not consider overabundance. In our paper, we present a strategy to measure paradigm predictability for overabundant paradigms and contrast two fine-grained measures, which we call predictive diversity and prediction reliability. We show that they capture conceptually different properties of paradigms. However, in an empirical study of Estonian nominal declension, we conclude that, once the relative token frequency of overabundant forms is taken into account, the two properties are strongly correlated. Our study confirms that overabundance affects quantitative approaches to the PCFP in non-trivial ways, and that token frequency should not be ignored in this context.
This study investigates the grammatical inference of phonological processes from featural representations, focusing on the class of k-Input Strictly Local (k-ISL) functions (Chandlee et al. 2014; Chandlee and Heinz 2018). We present the Factor-by-Feature (FxF) framework which decomposes the learning task into smaller ones along featural dimensions. The FxF framework is compatible with any grammatical inference algorithm designed to infer string-to-string functions in terms of finite-state transducers. The Structured Onward Subsequential Inference Algorithm (SOSFIA) is one such grammatical inference algorithm which has proven learnability results in linear time and data (Jardine et al. 2014). We empirically evaluate FxF on five phonological processes across four languages: English, Yawelmani, Chukchi, and Polish, representing cases of feature spreading, feature changing, and feature insertion. Comparisons are provided in each case between learning these processes over segments with SOSFIA alone and with SOSFIA in the FxF framework. To better measure data efficiency, we also introduce AMβA, a sampling algorithm that identifies smaller characteristic samples which, informally, are sets of data that guarantee successful generalization. Our results demonstrate that FxF consistently yields more compact and learnable data representations, significantly reducing the size of characteristic samples. These findings underscore the importance of phonological representation and support linguistically informed learners for phonological grammars.
We consider the possible role of current large language models (LLMs) in the study of human linguistic cognition. We focus on the use of such models as proxies for theories of cognition that are relatively linguistically-neutral in their representations and learning but differ from current LLMs in key ways. We illustrate this potential use of LLMs as proxies for theories of cognition in the context of two kinds of questions: (a) whether the target theory accounts for the acquisition of a given pattern from a given corpus; and (b) whether the target theory makes a given typologically-attested pattern easier to acquire than another, typologically-unattested pattern. For each of the two questions we show, building on recent literature, how current LLMs can potentially be of help, but we note that at present this help is quite limited.
Despite substantial progress of large language models (LLMs) for automatic poetry generation, the generated poetry lacks diversity while the training process differs greatly from human learning. Under the rationale that the learning process of the poetry generation systems should be more human-like and their output more diverse and novel, we introduce a framework based on social learning where we emphasize non-cooperative interactions besides cooperative interactions to encourage diversity. Our experiments are the first attempt at LLM-based multi-agent systems in non-cooperative environments for poetry generation employing both TRAINING-BASED agents (GPT-2) and PROMPTING-BASED agents (GPT-3 and GPT-4). Our evaluation based on 96k generated poems shows that our framework benefits the poetry generation process for TRAINING-BASED agents resulting in 1) a 3.0-3.7 percentage point (pp) increase in diversity and a 5.6-11.3 pp increase in novelty according to distinct and novel n-grams. The generated poetry from TRAINING-BASED agents also exhibits group divergence in terms of lexicons, styles and semantics. PROMPTING-BASED agents in our framework also benefit from non-cooperative environments and a more diverse ensemble of models with non-homogeneous agents has the potential to further enhance diversity, with an increase of 7.0-17.5 pp according to our experiments. However, PROMPTING-BASED agents show a decrease in lexical diversity over time and do not exhibit the group-based divergence intended in the social network. Our paper argues for a paradigm shift in creative tasks such as automatic poetry generation to include social learning processes (via LLM-based agent modeling) similar to human interaction.
This paper presents an HPSG analysis of Norwegian particle constructions with Ground promotion, as exemplified in (1), where av ‘off’ functions as a particle and the object bordet ‘the table’ is a Ground argument. (1) Jeg rydder AV bordet. I clear off table.DEF ‘I clear the table.’ It will be shown that the Ground promotion construction has similarities with both regular particle constructions and constructions with selected prepositions, and an analysis will be presented where Ground promotion is assumed to be a combination of these two constructions, viz. a complex particle construction. The analysis is implemented in a bidirectional typed feature structure grammar, and it will demonstrate that unification and type inheritance facilitates the merger of two constructions into one, while simultaneously allowing them to be kept apart.
Phrases such as burning question, digital waste or invasion of technology are relatively ordinary expressions understood by any speaker of English. While diverse in structure and meaning, they demonstrate a semantic tension between the basic meanings of a metaphoric and a non-metaphoric constituent. Albeit frequent in discourse, they remain a challenge for automatic language processing systems, especially for smaller, less represented languages. In this work, we inspect a broad array of language models to embed noun phrases in Slovene and investigate the potential of word embeddings to identify metaphoric phrases via the semantic distance of its constituents as measured via cosine similarity. The study shows both static and contextual monolingual embeddings encode relevant semantic information while multilingual embeddings demonstrate no significant effect in this experimental setting. Moreover, the study unravels the most effective layers for basic meaning representation and highlights the influence of other, non-semantic factors on cosine similarity. By shedding light on these mechanisms, the study provides new insights for both metaphor processing and our understanding of the inner workings of language models.
This paper proposes a novel analysis of extraction pathway marking in Type-Logical Grammar, taking advantage of proof-theoretic properties of logical proofs whose empirical application has so far been underexplored. The key idea is to allow certain linguistic expressions to be sensitive to the intermediate status of a syntactic proof. The relevant conditions can be stated concisely as constraints at the level of the proof term language, formally a special type of λ-calculus. The proposed analysis does not have any direct analog to either of the two familiar techniques for analyzing extraction pathway marking, namely, successive cyclic movement in derivational syntax and the SLASH feature percolation in HPSG. Moreover, the ‘meaning-centered’ perspective that naturally emerges from this new analysis is conceptually revealing: on this approach, extraction pathway marking essentially boils down to a strategy that certain languages employ to overtly flag the existence of a semantic variable inside a partially derived linguistic expression whose interpretation is dependent on a higher-order operator that is located in a larger structure.
This paper presents a subregular analysis of syntactic agreement patterns modeled using command strings over Minimalist Grammar (MG) dependency trees (Graf and Shafiei 2019), incorporating a novel MG treatment of agreement. Phenomena of interest include relativized minimality and its exceptions, direction of feature transmission, and configurations involving chains of agreeing elements. Such patterns are shown to fall within the class of tier-based strictly 2-local (TSL-2) languages, which has previously been argued to subsume the majority of long-distance syntactic phenomena, as well as those in phonology and morphology (Graf 2022a). This characterization places a tight upper bound on the range of configurations that are predicted to occur while providing parameters for variation which closely match the observed typology.
This paper presents a new computational implementation bridging several modules of grammar from phonetics to phonology to syntax. The system takes as input a speech signal annotated with syllables, interprets the phonetic data in phonological/prosodic terms, matches the data against a lexicon and makes the results available to a linguistically deep computational grammar. The system is showcased by means of syntactically ambiguous structures in German which can be disambiguated based on prosodic constituency information. A system evaluation with the German data showed good results for this new combination of automatic speech signal analysis and computational grammars, which takes a significant step towards a linguistically fine-grained computational analysis and hence towards real automatic speech understanding.
This article presents the structure of the ATLAs Alignment Module, a typological database designed to exhaustively capture languageinternal variation in argument marking (indexing and flagging). The flexible design of our database can be extended to cover further aspects of morphosyntactic alignment. We demonstrate with a small diversity sample how the database can be queried and the data aggregated at different levels of structure (e.g. for a language as a whole or for individual referential types in the form of alignment statements) for the purposes of cross-linguistic comparison. The database is made available in the Cross-Linguistic Data Formats (CLDF), and we provide code that generates an array of aggregations.
This paper studies the inflectional complexity of nouns, verbs and adjectives in 137 datasets, across 71 languages. I follow Ackerman and Malouf (2013) in distinguishing between E(numerative) complexity and I(ntegrative) complexity. The first one encompasses aspects of inflection, like the number of principal parts, paradigm size, and number of exponents, while the second one captures the implicative relations between paradigm cells (how difficult it is to predict one cell of a paradigm knowing a different cell). I provide a formalism and computational implementation to estimate both I- and E-complexity expressed through Word and Paradigm morphology (Blevins 2006, 2016), which is flexible and powerful enough for typological research. The results show that, as suggested by Ackerman and Malouf (2013), I-complexity is relatively low across the languages in the sample, with only two clear exceptions (Navajo and Yaitepec-Chatino). The results also show that E-complexity can vary considerably crosslinguistically. Finally, I show there is a clear correlation between I- and E-complexity.
In morphology, a distinction is commonly drawn between inflection and derivation. However, a precise definition of this distinction which reflects the way it manifests across languages remains elusive within linguistic theory, typically being based on subjective tests. In this study, we present 4 quantitative measures which use the statistics of a raw text corpus in a language to estimate to what extent a given morphological construction changes the form and distribution of lexemes. In particular, we measure both the average and the variance of this change across lexemes. Crucially, distributional information captures syntactic and semantic properties and can be operationalised by word embeddings. Based on a sample of 26 languages, we find that we can reconstruct 89±1% of the classification of constructions into inflection and derivation in UniMorph using our 4 measures, providing large-scale cross-linguistic evidence that the concepts of inflection and derivation are associated with measurable signatures in terms of form and distribution that behave consistently across a variety of languages. We also use our measures to identify in a quantitative way whether categories of inflection which have been considered noncanonical in the linguistic literature, such as inherent inflection or transpositions, appear so in terms of properties of their form and distribution. We find that while combining multiple measures reduces the amount of overlap between inflectional and derivational constructions, there are still many constructions near the model’s decision boundary between the two categories. This indicates a gradient, rather than categorical, distinction.
This study examines zero marking, i.e. the absence of an overt exponent, in adjectival, nominal, and verbal inflectional morphology across languages. The first part of the study provides an overview of the distribution of zero markers in inflection paradigms using the UniMorph dataset. The results show that there is a general preference against zero marking. The distribution of zero markers varies to a great extent across languages and lemmas, the only robust trend being that they are avoided in cells that express a high number of grammatical values. The second part of this study examines the association between marker frequencies and phonological length, using the Universal Dependencies treebanks. While token frequency is a good predictor for the length of overt markers, it does not account for the occurrence of zero markers. This is taken as evidence to support a differential non-development scenario of zero marking rather than a phonetic reduction scenario.
QRGS stands for the Question Responses Generation System. It is an online game-like framework designed for gathering various types of question responses. A QRGS user is asked to read a simple story and impersonate its main character. As the story unfolds the user is confronted with four questions and (s)he is expected to answer these in the way the main character would. In this way, we obtain responses to questions of a desired type. The data gathered via QRGS is a useful supplement to the linguistic data already present in language corpora – especially for languages for which such resources are sparse. As such, it opens the possibility for better understanding of the use of questions in natural language dialogues and analysing the response space of such questions. In this paper, we present the main idea of QRGS and the results of five studies (in Polish and in English) that test the framework. Our discussion addresses issues concerning the efficiency and accuracy of the proposed approach. We also discuss the availability of the QRGS and its potential future improvements.
We develop a three-part approach to Verb Sense Disambiguation (VSD) in German. After considering a set of lexical resources and corpora, we arrive at a statistically motivated selection of a subset of verbs and their senses from GermaNet. This sub-inventory is then used to disambiguate the occurrences of the corresponding verbs in a corpus resulting from the union of TüBa-D/Z, Salsa, and E-VALBU. The corpus annotated in this way is called TGVCorp. It is used in the third part of the paper for training a classifier for VSD and for its comparative evaluation with a state-of-the-art approach in this research area, namely EWISER. Our simple classifier outperforms the transformer-based approach on the same data in both accuracy and speed in German but not in English and we discuss possible reasons.
