evo1: Vocalization Displays and the Recognition of Typed Object-Features

The evolution of language is difficult to explain. It is a complex faculty with various aspects or components, and it is difficult to see why rudimentary versions of those aspects would have selectional advantages for differential reproduction in hominids. It would be helpful to identify preadaptations, particular characters that developed complexity in some other function, that came to be used in a new function of language-based communication.

My working hypothesis is that the capacity for visual recognition of objects and scenes, specifically recognition of physical objects and behaving animals in the shared immediate social environment of hominids, gave rise to a complex of preadaptations for language. One preadaptation is 1) a shared conceptual scheme of individuation, where certain scenes were classified as involving significant happenings or objects. This scheme would eventually give rise to the mental lexicon, and the related capacity for semantic memory. Another is 2) the capacity to represent a scene and individual objects in terms of typed structures of object-features. By the hyphenated term object-feature, I refer specifically to higher level visual features like shape outline and component shape, that are distinguished from low level features like edge or color or texture. Both these preadaptations were refined and repurposed in the context of 3) recognizing social displays of individuals of the same species. Some displays, like pointing, have 4) a property of intentionality. Intentional displays are produced to trigger an observer's attunement to a relation between the structure of the display and a scene where some salient aspects are referenced (physical objects in the immediate environment are referred to by a shift of gaze or a hand movement, happenings are referenced in relation to a doer or an undergoer). Clearly vocalization displays are core to the evolution of language, but gestural displays that support the repeated emergence of sign languages (among deaf and hearing communities) are also of research interest.

How did the use of articulatory gestures in vocalization evolve into phonology? One possibility is song, perhaps evolved from gibbon-like vocalizations to mark territory. Tone is differentiated and evolves to be worth remembering, so song is born. The continuous articulation of vowels is another dimension of song, and this could eventually evolve into a phonemic set of distinguished vowel phonemes. Why would consonants and syllabification be added to the mix? Why would song recognition and articulation capacities get repurposed for referring gestures, replacing eye gaze and pointing with nouns or noun phrases?

From studying the semantics-syntax interface, we can theorize about two large layers in the structure of utterances. These layers define the linguitic types that an utterance (or sign language expression) is recognized to instantiate. One layer is the referring expression, where an concrete object or abstraction is referred to and a corresponding mental concept takes a place in a shared information structure where it could be referred to again. In other words, referring expressions introduce referential indices into discourse, and create the possibility of anaphoric resolution. We have three levels here:

1. the etic level of a complex nominal utterance in the context where the speaker picks out a particular situation for the listener,
2. the emic level of the types from the mental lexicon and language faculty that recognize the utterance as an instance of a structure phrasal type involving lexical items and a distinguished head item, and
3. the level of the shared information state including referential indices.
   

The other large layer is that of speech acts made with clausal utterances (or the sentences used in them), that introduce a semantics and pragmatics of what Searle calls conditions of satisfaction. In a request, the shared information state is about a state which does not obtain but where the listener can act to fulfill those conditions. A promise mentions a possible future state where the the speaker (or signer) will act to fulfill the conditions. And an assertive speech act calls attention to a state where the situation referred to (a visible scene, or some abstract situation) has the participants referred to with referential indexes, and they stand in a happening-relation mentioned by the head verb or predicator. Any modifiers also introduce properties of the participants or situation. Returning to the three levels, at

1. the etic level, a clausal utterance is used to perform a speech act between speaker and listener
2. the emic level, a sentence or clause is recognized as consisting of a head verb, arguments that have dependency relations to that verb, and additional modifiers
3. the shared information state level, of a discourse representation structure where referential indexes are resolved to salient items in discourse, creating shared attunement to a described situation, which may be a part of the concrete situation surrounding speaker and listener, or else a more abstract situation that does not (yet) exist in the physical surroundings.

This is the rudimentary model of language that I am using. I would need to demonstrate that a language faculty that supports this model is systematically related to existing visual recognition faculties (of objects, scenes and social displays) that provided some preadaptations for the relatively rapid emergence of complex language among hominids

First I will discuss typed feature structures in language and high-level vision, then I will discuss shared schemes of information in gestural displays and language. Future posts will cover:

• Feature structures for phrasal features and visual object-features and scenes
  
• Typing and semantic memory
  
• Schemes of individuation
  
• Shared schemes and the first-person character of perceptual qualia
  
• Types in the mental lexicon
  

A Minimalist Theory with Untyped Features

I read a review by Ash Asudeh and Ida Toivonen of two introductory textbooks on Minimalism. What interests me is the book by Adger:

David Adger, Core syntax: a Minimalist approach. Oxford: Oxford

University Press, 2003. Pp. xiii+424.

This is described as developing a specific coherent theory within the Minimalist Program, and it ends up being close to a lexicalist unification grammar. Like LFG, and unlike HPSG, the feature theory is untyped.

I don't know if there will be a lot of interest, but it may be useful to have a P&P (GB and/or MP) syntax module that can be plugged in to a Linguistic Exploration Environment. Adger's version may be a candidate, because it is more formalizable, and may have some compatibility at the level of feature theory. This allows a clearer connection to the lexicon.

I also have been reading the critique of GB and MP in Culicover and Jackendoff's Simpler Syntax. I am also interested in GB, because Paul Llido uses it in studying Case in Cebuano, mixing it with LFG.

forthcoming books by Stephen Wechsler

1. Wechsler,  Stephen (in progress). The Syntax-Lexicon Interface (working title). Oxford Surveys in Syntax and Morphology, General editor: Robert D. Van Valin, Jr.  Oxford University Press, Oxford. 

2. Asudeh, Ash, Joan Bresnan, Ida Toivonen, and Stephen Wechsler (in progress).   Lexical Functional Syntax.  Blackwell Publishers.  

It seems Wechsler is now doing LFG, although he has worked extensively in HPSG. I think the crossfertilization between the two is healthy, I Wayan Arka is also working on LFG for Indonesian, and he has collaborated with Wechsler on HPSG

Constructional Semantics

There has been extensive research on Lexical Semantics (e.g. Levin, Pustejovsky), often presuming a bottom-up compositionality. This is sometimes contrasted with a Phrasal Semantics that is often in the tradition of Montague grammar, focusing on issues like anaphora and scoping, often assuming that each lexical token contributes one atomic symbol to a semantic representation.

Rather than talk about Phrasal Semantics directly, I will assume that the semantical representation of a phrasal expression can be decomposed into both a lexical contribution and a constructional contribution. Compositionality is still possible, but it involves both bottom-up and top-down contributions.

An initial focus will be Argument Structure Constructions (ASCx) of verbs, such as those analyzed by Goldberg for English. However, I am also interested in modeling Austronesian morphosyntactic alignment (comparing with Llido's analysis of Cebuano, for example). I am assuming that related but distinct senses of a lexical entry are actually expression-level construal from combining fewer word senses with meaning-contributing constructions, especially ASCx's. It is possible that a set of related lexical entry senses are actually the same underlying word sense distinguished in usage by the constructions forming different expressions. It is further possible that the constructions are actually related constructions, perhaps linked by inheritance (Goldberg proposes several distinct types of Inheritance, some of which involve Lakoff-style metaphor).

Whether this approach captures a significant amount of generality about lexical senses can be explored empirically. If this proves productive, it could later be incorporated into a linguistic exploration environment.

A gold standard for distinguishing senses in a lexical entry using a corpus-based approach is the COBUILD dictionary of English. Each entry has a stylized definition, which generally picks out the typical arguments of the verb, including a common noun as a general type constraint for each argument. To some extent, it identifies other participants in the mentioned situation which may not be explicitly realized in arguments. These additional participants may be a mechanism for characterizing the connotations of a specific sense.

I propose to explore some of the most common verbs of English (those in the top tier of of 680+ most frequent used lexical entries). For the sample sentences of each sense (and perhaps additional sentences from the underlying corpus, the Bank of English), I would try to identify any ASCx, based on typical arguments for that sense. Closely related senses may have related constructions imposed on the same underlying verb sense.

Multilingual Explorations

It may be useful to compare several languages at this level of semantics, for example Chinese or German or Filipino or Cebuano. I am interested if there are potential applications in basic education, adult L2 education, human translation, machine assisted translation, etc. This may also be relevant to the documentation and description of less studied languages by people more familiar with another more widespread language (the SIL use cases for Fieldworks).

teasing apart "word sense"

Analysis of "word sense"

From the point of view of constructions, a c:lexical_sense contributes only part of the meaning of a c:word_usage. The construction itself contributes something additional, which also interacts with the larger discourse environment and the social context of successful communication.

We can divide the meaning of a c:word_usage into a purely semantic representation, which is the typical meaning of the sentence being used. Or even a simplified form of the basic sentence, with only head words and abstracting from embellishments that don't contribute to a specified lexical sense. The semantic representation includes the literal meaning, but may also introduce additional roles that follow from "typically" understood arguments and properties of the explicitly mentioned words. It covers denotation and the related intensional functions, as well as connotations relative to a "typical" mental lexicon (perhaps of an ideal listener/reader).

In addition to the purely semantic representation, competent language users can infer a more c:pragmatic_construal of the c:word_usage.

In the context of the COBUILD lexical word senses and the Bank of English used to generate them, a distinguished lexical sense classifies a class of sentences which share a similar word_usage. It should be possible to isolate the intuitions of this "gold standard" classification into a semantic representation. This abstracts from less reproducible aspects of pragmatic construal, and focuses on what is typically inferred from the sentence itself, as if it were used in isolation rather that in a larger discourse or social situation.

We would like to tease apart the characterization of each lexical sense as a purely semantic representation into two parts, the contribution of the word (e.g. a free morpheme verb, or the stem of an inflected verb) from the contribution of the construction or constructions (including morphological constructions, and argument structure constructions at least).

We would like to discover if the set of lexical senses of a word (initially, some of the more frequently used verbs) can be characterized with a set of constructions related by the inheritance relations distinguished by Goldberg. We also want to see to what extent these constructions are reused across different verbs, and if we can identify verb classes that share the same inheritance-related constructions.

Intelligent Agents for User Interface

I watched a TechTalk at Google (July 25, 2008) that may be relevant to designing an annotation interface for linguistic exploration. In "Intelligent Assistance for Desktop User Tasks" Anthony Tomasic talked about work with John Zimmerman integrating machine learning with UI Design: VIO (being commercialized for SalesForce.com) selects and pre-fills forms from email, and Mixer generates database queries from email. VIO intentionally avoids NLP or domain-specific explicit engineering, looking to evaluate a lightweight solution.

The experience highlights how little I know about machine learning. VIO uses K-way classification (form suggestion), Conditional Random Fields (field s.) and Reference Resolution over instances (instance s.), which is all greek to me. The work is an example of empirical methods, they train their analyzers and evaluate them with experiements on real users.

Regional Dictionary of Biography

I was browsing the Australian Dictionary of Biography, which is an ANU project that documents Australians with either 2000-6000 words, or 500-2000 words. It tries to be "representative" and has a number of biographies of Aborigines, assigning them occupations like "indigenous culture informant".

It would be great to create a similar resource about prominent Filipinos, and I suggest a regional focus to ensure that every ethnic group is covered. We could start immediately with a Cebuano region initiative, then work with the rest of UPV to cover the Visayas. 

I am sure many towns would be interested in their local history, so we could start with trying to document the most prominent individuals who were born or have lived in a given municipality. I am thinking of Todd Lucero Sales' recent publication, Argao: 400 Years in Legend and History (Argao, 2008). Perhaps I can interest Madrileña de la Cerna of UP Cebu in this project.

Case and voice systems in Cebuano

I have been planning to study the syntactic side of argument structure in Cebuano.

The trigger is Paul Llido's paper from ICAL 10 in Palawan, "Inflectional Case Assignment in Cebuano"

This references Yehuda Falk's paper "Case Typology and Case Theory"

I should look at the attitude towards VSO, etc. typology in WALS (World Atlas of Linguistic Structures).

There is also the following exchange from LFG98.

Introduction: The problem of voice and grammatical functions in Western Austronesian languages	Simon Musgrave	Uni. of Melbourne
Symmetrical Voice Systems and Precategoriality in Philippine Languages	Bill Foley	Uni. of Sydney
Nouns and Verbs in Tagalog: A Response to Foley	Paul Kroeger	SIL, Darwin