Linguistic analysis of the word

5 Roots of a Mentalistic Conception of Language

Thus, Bloomfield insisted on a strictly antimentalistic program of linguistic analysis, according to his opinion that ‘Non-linguists … constantly forget that a speaker is making noise, and credit him, instead, with the possession of impalpable “ideas”.’ It remains for the linguist to show, in detail, that the speaker has no ‘ideas’ and that the noise is sufficient (1936). But while he defended this dogma by practically excluding semantics from the actual linguistic research agenda, he was, surprisingly, not able to exclude mentalistic concepts from the domain of his very central concern—the analysis of linguistic form: underlying forms, derivational rules, precedence order of operations, and even phonemes as parts of underlying forms are concepts that go beyond the mechanistic elements extracted from noise. This conflict between rigid methodological principles on the one side and explanatory insights on the other mirrors one aspect of the above mentioned tension between Bloomfield and Sapir, in the following sense: although both scholars agreed on the general principles of linguistics as a scientific enterprise, including insights in relevant detail such as the description of Indian languages, or the nature of sound change, Sapir always insisted on the psychological nature of linguistic facts. This is clear from his Sound patterns in language (1925), and even more programmatically from The psychological reality of phonemes (1933). As a matter of fact, the orientation of Sapir’s inspiring essays amounted to a mentalistic research strategy.

Thus, although Bloomfield in 1940 was Sapir’s successor as Sterling Professor of Linguistics at Yale University, it was Sapir’s notion of language as a psychological reality that dominated the subsequent period of post-Bloomfieldian linguistics, as shown, e.g., by the title and orientation of The Sound Pattern of Russian by Halle (1959) and The Sound Pattern of English by Chomsky and Halle (1968).

But the tension persists. While Sapir and Jakobson provided ideas and orientation of this development, it is still Bloomfield’s radical clarification of the conceptual framework and descriptive technology that is the indispensable foundation of modern linguistics. And perhaps more importantly, his Menomini morphophonemics is the paradigm case not only of a complete descriptive analysis given in terms of an explicit, complex rule system, but also of an impressive victory of clear insight over methodological orthodoxy. Conceding the amount of mentalistic machinery implied in this analysis of sound structure, strong theories of syntax and even semantics have in the meanwhile been shown to be possible and productive—in spite of Bloomfield’s antimentalistic bias.

Logical concepts and techniques are widely used (a) as tools for linguistic analysis and (b) in the study of the structure of grammatical theory. As to (a), the classical field of applying logic to linguistics is the analysis of meaning. The key concepts of linguistic semantics, including the very notion of meaning, are frequently defined in terms of, or in analogy with, concepts originating in formal logic. Moreover, algorithms translating natural language into logical symbolism are used to give precise accounts of, and allow comparison with, the semantic structure of natural language. Concerning (b), a central idea of generative grammar has been to specify logical characteristics of linguistic theory and study their repercussions on the complexity of the object under investigation, i.e., language.

1.1 Forensic linguistics

The number of computational approaches to forensic linguistics has increased significantly over the last decades because of increasing computer processing power as well as growing interest of computer scientists in forensic applications of Natural Language Processing (NLP). At the same time, forensic linguists faced the need to use computer resources in both their research and their casework, especially when dealing with large volumes of data (Coulthard, 2010).

These directions have been further advanced and extended to several other applications worldwide.

The written language of the law came to include applications other than studying the complexity of legal language; interaction in legal contexts has significantly evolved and now focuses on any kind of interaction in legal contexts. This also includes attempts to identify the use of language with misinterpretation (Gales, 2015), prove appropriate interpreting (Kredens, 2016), and manage language as evidence. Other research foci are on disputed meanings (Butters, 2012), the application of methods of authorship detection for cybercriminal investigations, and an attempt to develop authorship synthesis (Grant and MacLeod, 2018).

One of the main goals of forensic linguistics is to provide a careful and systemic analysis of syntactic and semantic features of legal language. The results of this analysis can be used by many different professionals. For example, police officers can use this evidence not only to interview witnesses and suspects more effectively but also to solve crimes more reliably. Lawyers, judges, and jury members can use these analyses to help evaluate questions of guilt and innocence more fairly. And translators and interpreters can use this research to communicate with greater accuracy. Forensic linguistics serves justice and helps people to find the truth when a crime has been committed.

In this chapter, we build a foundation for what one would call discourse forensic linguistics, that is, how legal writers organize their thoughts describing criminal intents and evidence, involving data items of various natures. Legal texts have peculiar discourse structures (Aboul-Enein, 1999); moreover, these texts are accompanied by structured numerical data like maps, logs, and various types of transaction records. Legal texts have DTs different from other written genres and a special form of logical implications. The role of argumentation-based discourse features in legal texts is important as well (Galitsky et al., 2009).

More than 40 years ago, Jan Svartvik showed dramatically just how helpful forensic linguistics could be. His analysis involved the transcript of a police interview with Timothy Evans, a man who had been found guilty of murdering his wife and his baby daughter in 1949. Svartvik demonstrated that parts of the transcript differed considerably in their grammatical style when he compared them to the rest of the recorded interview. Based on this research and other facts, the courts ruled that Evans had been wrongly accused. Unfortunately, Evans had already been executed in 1950. However, thanks to Svartvik’s work, 16 years later, he was formally acquitted. The study of Svartvik is considered today to be one of the first major cases in which forensic linguistics was used to achieve justice in a court of law. Today, forensic linguistics is a well-established, internationally recognized independent field of study.

7.2.1 Text Analysis

Text analysis transforms the input text into speakable forms using text normalization, linguistic and phonetic analysis. Text normalization converts numbers and symbols into words, replaces abbreviations by their corresponding whole phrases, among others. The goal of Linguistic analysis is to understand the content of the text. Statistical methods are used to find the most probable meaning of the text. A grapheme is a letter or a sequence of letters that represent a sound (i.e., phoneme) in a word. Phonetic analysis converts graphemes into phonemes. Generation of sequence of phonetic units for a given standard word is referred to as letter to phoneme or text to phoneme rule. The complexity of these rules and their derivation depends on the nature of the language.

Analysis

Fulford, like all other writers in this section except Boorse, aims to analyze an ‘everyday usage’ [1989, 27] of medical concepts largely common to physicians and the lay public. Unlike other writers, he identifies his method as the ‘linguistic analysis’ (22) practiced by such figures as Wittgenstein, Austin, Urmson, and Hare (xv, 22–3, 54, 121–22). One major goal is to illuminate conceptual problems in, specifically, psychiatry. Hence, Fulford begins by relating his project to the controversy over Szasz’s claim that mental illness, as a conceptual impossibility, is a myth. In his first chapter, he argues that we need to recast that controversy by rejecting two assumptions and a form of argument shared by both sides in this dispute, antipsychiatrists and psychiatrists. The assumptions are that mental illness is conceptually problematic, but that physical illness is not. The form of argument is to test alleged mental illnesses against properties thought essential to physical ones (5).

Fulford next seeks to undermine the ‘conventional’ view of medical concepts, as represented by Boorse. A brief identifies features of common medical usage that the BST allegedly fails to capture. Then chapter 3 offers a debate between a metaethical descriptivist and a nondescriptivist over ‘Boorse’s version of the medical model’ (37). One of the main conclusions of this debate is that ‘dysfunction’ is a ‘value term,’ i.e., has evaluative as well as descriptive content. The main argument for this thesis, in chapter 6, proceeds by comparing functions in artifacts and organisms. In artifacts, functioning is doing something in a special sense, ‘functional doing’ (92 ff). Fulford argues by a rich array of examples that the function of an artifact depends on its designer’s purposes in two ways, both as to end and as to means: ‘for a functional object to be functioning, not only must it be serving its particular ‘designed-for’ purpose, it must be serving that purpose by its particular ‘designed-for’ means’ (98). Moreover, ‘purpose’ is ‘an evaluative concept,’ i.e., ‘evaluation is … part of the very meaning of the term’ (106). From this Fulford concludes that claims of biological function also rest on value judgments.

Fulford’s own positive account can be summarized as follows. Not only are ‘illness’ and the various ‘malady’ terms such as ‘disease’ negative value terms, but they express a specifically medical kind of negative value involving action failure. As the ‘reverse-view’ strategy requires, ‘illness’ is the primary concept, from which all the others are definable. To motivate his account of illness, Fulford uses the point that bodies are dysfunctional, yet persons ill. To what, he asks, does illness of a person stand in the same relation as bodily, or artifact, dysfunction stands to functional doing? Since artifacts and their parts function by moving or not moving, Fulford considers illnesses consisting of movement or lack of movement. Using action theory’s standard example of arm-raising, he suggests that the concept of illness has ‘its origins in the experience of a particular kind of action failure’: failure of ordinary action ‘in the apparent absence of obstruction and/or opposition’ (109). By ordinary action, he means the ‘everyday’ kind of action which, as Austin said, we ‘just get on and do’ (116). He next suggests that this analysis also suits illnesses consisting in sensations or lack of sensations. Pain and other unpleasant sensations, when symptoms of illness, also involve action failure. That is because ‘pain-as-illness,’ unlike normal pain, is ‘pain from which one is unable to withdraw in the (perceived) absence of obstruction and/or opposition’ (138). Finally, the analysis covers mental illness as naturally as physical illness. As physical illness involves failure of ordinary physical actions, like arm-raising, so mental illness involves failure of ordinary mental actions, like thinking and remembering.

From this basic illness concept, Fulford suggests, all the more technical medical concepts are defined, and in particular a family of concepts of disease. He begins by separating, within the set I of conditions that ‘may’ be viewed as illness, a subset Id of conditions ‘widely’ viewed as such (59). ‘Disease’, he thinks, has various possible meanings in relation to Id. In its narrowest meaning (HDv), ‘disease’ may merely express the same value judgment as ‘illness’, but in relation only to the subset Id, not to all of I (61). A variant of this idea is a descriptive meaning of ‘disease’ (HDf1) as ‘condition widely viewed as illness,’ which, of course, is not an evaluation, as is HDv, but a description of one (63). Two other broader descriptive senses of ‘disease’ are then derivable from HDf1: condition causally [HDf2] or statistically [HDf3] associated with an HDf1 disease (65, 69). Fulford provides no comparable analyses for other ‘malady’ terms such as ‘wound’ and ‘disability’, nor for ‘dysfunction’. Nevertheless, he seems to hold that their meaning depends on the concept of illness in some fashion.

Fulford judges his account of illness a success by two outcome criteria. The first is that it find, if possible, a neutral general concept of illness, then explain the similarities and differences of ‘physical illness’ and ‘mental illness’ in ordinary usage (24). The main difference is that people disagree much more in evaluating mental features, e.g., anxiety compared with pain. Using the example of alcoholism, he tries to show how clinical difficulty in diagnosing mental illness results from the nature of the phenomena, which the concept, far from being defective, faithfully reflects (85,153).

Fulford’s account succeeds equally, he thinks, by his second criterion, clinical usefulness. In psychiatric nosology, it suggests several improvements: (1) to make explicit the evaluative, as well as the factual, elements defining any disease for which the former are clinically important; (2) to make psychiatric use of nondisease categories of physical medicine, such as wound and disability; and to be open to the possibilities that (3) mental-disease theory will look quite different from physical-disease theory and (4) in psychiatry, a second taxonomy of kinds of illness will also be required (182–3). Fulford claims three specific advantages for his view over conventional psychiatry. By revealing the essence of delusions, it vindicates the category of psychosis and explains the ethics of involuntary medical treatment. First, delusions can be seen as defective ‘reasons for action,’ whether cognitive or evaluative (215). Second, since this defect is structural, psychosis is the most radical and paradigmatic type of illness (239) — ‘not a difficulty in doing something, but a failure in the very definition of what is done’ (238). Third, that is what explains, as rival views cannot, both why psychotics escape criminal responsibility and why only mental, not physical, illness justifies involuntary medical treatment (240–3). Finally, Fulford thinks his theory promotes improvement in primary health care, as well as closer relations both between somatic and psychological medicine and between medicine and philosophy (244–54). He pursues the themes of “values-based” medicine in his [2004] and in Woodbridge and Fulford [2004].

4.3.3 The “process” block: how the collected data are analyzed and interpreted?

In addition, the underlying pedagogical usefulness acts as a strong criterion that drives the whole analysis process in order to produce valid and useful results. However, most cases draw the attention to boundaries and limitations related mostly to ethics and security issues on the data manipulation.

2.1.3 Pattern 3: such as

This pattern is proposed by Hearst (1992) as “such NP as NP”, but in the text of the ISA88 standard no occurrence of this pattern can be observed. Alternatively, a lot of “NP such as NP” are found. Thus, the pattern was adapted to this second case.

No linguistic analysis could be performed for this case because the analyser didn’t properly detect the “such as” construct to subsequently detect the two noun phrases surrounding it. A tagging process (Carreras et al., 2004) was performed in order to know the grammar category of each word and then nouns were detected before and after the “such as” construct. For example, in the phrase “Example 8: Process Management events such as allocation of equipment to a batch, creation of a control recipe, etc.”, allocation of equipment to a batch and creation of a control recipe are given as examples of Process Management events, thus the “Process Management event” concept is the parent of former concepts. Pattern 1.3 resulted in 305 candidate relations.

5 User intent recognizer

One of the essential capabilities of a chatbot is to discriminate between a request to commit a transaction (such as a patient treatment request) and a question to obtain some information. Usually, these forms of user activity follow each other.

Before a user wants a chatbot to perform an action (such as open a new bank account), the user would want to know the rules and conditions for this account. Once the user knowledge request is satisfied, the user makes a decision and orders a transaction. Once this transaction is completed by the chatbot, the user might want to know a list of options available and asks a question (such as how to fund this new account). Hence user questions and transactional requests are intermittent and need to be recognized reliably.

Errors in recognizing questions vs transactional requests are severe. If a question is misinterpreted and an answer to a different question is returned, the user can reformulate it and ask again. If a transactional request is recognized as a different (wrong) transaction, the user will understand it when the chatbot issues a request to specify inappropriate parameters. Then the user would cancel their request, attempt to reformulate it, and issue it again. Hence chatbot errors associated with wrongly understood questions and transactional requests can be naturally rectified. At the same time, chatbot errors recognizing questions vs transactional requests would break the whole conversation and the user would be confused as how to continue a conversation. Therefore, the chatbot needs to avoid these kinds of errors by any means.

Recognizing questions vs transactional requests must be domain independent. In any domain a user might want to ask a question or to request a transaction, and this recognition should not depend on the subject. Whereas a chatbot might need training data from a chatbot developer to operate in a specific domain (such as flu), recognizing questions vs transactional requests must be a capability built in advance by a chatbot vendor, before this chatbot will be adjusted to a particular domain.

We also target recognition of questions vs transactional requests to be in a context-independent manner. Potentially there could be any order in which questions are asked and requests are made. A user may switch from information access to a request to do something and back to information access, although this should be discouraged. Even a human customer support agent prefers a user to first receive information, make a decision, and then request an action (make an order).

A request can be formulated explicitly or implicitly. Could you do this may mean both a question about the chatbot capability as well as an implicit request to do this. Even a simple question like “what is my account balance?” may be a transactional request to select an account and execute a database query. Another way to express a request is via mentioning of a desired state instead of explicit action to achieve it. For example, the utterance “I am too cold” indicates not a question but rather a desired state that can be achieved by turning on the heater. If no available action is associated with “cold” this utterance is classified as a question related to “coldness.” To handle this ambiguity in a domain-independent manner, we differentiate between questions and transactional requests linguistically, not pragmatically.

Data, vocabularies, and configuration components include leading verbs indicating that an utterance is a request. They also include expressions used by an utterance author to indicate the author wants something from a peer, such as “Please do … for me.” These expressions also refer to information requests such as “Give me MY …,” for example, account information. For a question, this vocabulary includes the ways people address questions, such as “please tell me….”

Rule engine applies a sequence of rules, both keyword-based, vocabulary-based, and linguistic. The rules are applied in certain order, oriented to find indication of a transaction. If main cases of transactions are not identified, only then does the rule engine apply question rules. Finally, if question rules did not fire, we classify the utterance as unknown, but nevertheless treat it as default, a question. Most rules are specific to the class of requests; if none of them fire, then the decision is also a question.

Linguistic processor targets two cases: imperative leading verb and a reference to “my” object. Once parsing is done, the first word should be a regular verb in present tense and active voice, and be neither modal, mental, or a form of be. These constraints assure this verb is in the imperative form, for example, “Drop the temperature in the room.” The second case addresses utterances related to an object the author owns or is associated to, such as “my account balance” and “my car.” These utterances are connected with an intent to perform an action with these objects or a request for information on them (versus a question that expresses a request to share general knowledge, not about my object).

Decision former takes an output of the rule engine and outputs one out of three decisions, along with an explanation for each of them. Each fired keyword-based rule provides an explanation, as well as each linguistic rule. Thus when a resultant decision is produced, there is always a detailed backup of it. If any of the components failed while applying a rule, the resultant decision is unknown.

If no decision is made, the chatbot comes back to the user asking for explicit clarification: “Please be clearer if you are asking a question or requesting a transaction.”

We present examples for each class together with the rules that fired and delivered the decision (shown in brackets).

Questions

If I do not have my Internet Banking User ID and Password, how can I login? [if and how can I—prefix].

I am anxious about spending my money [mental verb].

I am worried about my spending [mental verb].

I am concerned about how much I used [mental verb].

I am interested how much money I lost on stock [mental verb].

How can my saving account be funded [How + my].

Domestic wire transfer [no transactional rule fired therefore question].

order replacement/renewal card not received [no transactional rule fired therefore question].

Requests from the chatbot to do something

Tell me… [leading imperative verb].

Confirm that…

Help me to … [leading imperative verb].

The Semantic Web

The underlying theme connecting most of the current work within information management is that of the Semantic Web. The general idea behind the Semantic Web is to create a system in which machines are able to ‘understand’ information presented on the Internet in the same way that humans can understand or comprehend information – thus allowing machines to better support standard processes such as finding, storing, translating or combining information from disparate sources. In short, the Semantic Web is a vision for the Internet in which more complex human requests are able to be intelligently interpreted by search engines and computers based on the semantics of the request – the relationships between words and phrases, their context and their intended meaning.

3 Empirical Research on the Hypothesis

Although the Sapir–Whorf proposal has had wide impact on thinking in the humanities and social sciences, it has not been extensively investigated empirically. Indeed, some believe it is too difficult, if not impossible in principle, to investigate. Further, a good deal of the empirical work that was first developed was quite narrowly confined to attacking Whorf’s analyses, documenting particular cases of language diversity, or exploring the implications in domains such as color terms that represent somewhat marginal aspects of language structure. In large part, therefore, acceptance or rejection of the proposal for many years depended more on personal and professional predilections than on solid evidence. Nonetheless, a variety of modern initiatives have stimulated renewed interest in mounting empirical assessments of the hypothesis.

Contemporary empirical efforts can be classed into three broad types, depending on which of the three key terms in the hypothesis they take as their point of departure: language, reality, or thought (Lucy 1997).

A structure-centered approach begins with an observed difference between languages, elaborates the interpretations of reality implicit in them, and then seeks evidence for their influence on thought. The approach remains open to unexpected interpretations of reality but often has difficulty establishing a neutral basis for comparison. The classic example of a language-centered approach is Whorf’s pioneering comparison of Hopi and English described above. The most extensive contemporary effort to extend and improve the comparative fundamentals in a structure-centered approach has sought to establish a relation between variations in grammatical number marking and attentiveness to number and shape (Lucy 1992b). This research remedies some of the traditional difficulties of structure-centered approaches by framing the linguistic analysis typologically so as to enhance comparison, and by supplementing ethnographic observation with a rigorous assessment of individual thought. This then makes possible the realization of the benefits of the structure-centered approach: placing the languages at issue on an equal footing, exploring semantically significant lexical and grammatical patterns, and developing connections to interrelated semantic patterns in each language.

A domain-centered approach begins with a domain of experienced reality, typically characterized independently of language(s), and asks how various languages select from, encode, and organize it. Typically, speakers of different languages are asked to refer to ‘the same’ materials or situations so the different linguistic construals become clear. The approach facilitates controlled comparison, but often at the expense of regimenting the linguistic data rather narrowly. The classic example of this approach, developed by Roger Brown and Eric Lenneberg in the 1950s, showed that some colors are more lexically encodable than others, and that more codable colors are remembered better. This line of research was later extended by Brent Berlin, Paul Kay, and their colleagues, but to argue instead that there are cross-linguistic universals in the encoding of the color domain such that a small number of ‘basic’ color terms emerge in languages as a function of biological constraints. Although this research has been widely accepted as evidence against the validity of linguistic relativity hypothesis, it actually deals largely with constraints on linguistic diversity rather than with relativity as such. Subsequent research has challenged Berlin and Kay’s universal semantic claim, and shown that different color-term systems do in fact influence color categorization and memory. (For discussions and references, see Lucy 1992a, Hardin and Maffi 1997, Roberson et al. 2000.) The most successful effort to improve the quality of the linguistic comparison in a domain-centered approach has sought to show cognitive differences in the spatial domain between languages favoring the use of body coordinates to describe arrangements of objects (e.g., ‘the man is left of the tree’) and those favoring systems anchored in cardinal direction terms or topographic features (e.g., ‘the man is east/uphill of the tree’) (Pederson et al. 1998, Levinson in press). This research on space remedies some of the traditional difficulties of domain-centered approaches by developing a more rigorous and substantive linguistic analysis to complement the ready comparisons facilitated by this approach.

A behavior-centered approach begins with a marked difference in behavior which the researcher comes to believe has its roots in and provides evidence for a pattern of thought arising from language practices. The behavior at issue typically has clear practical consequences (either for theory or for native speakers), but since the research does not begin with an intent to address the linguistic relativity question, the theoretical and empirical analyses of language and reality are often weakly developed. The most famous example of a behavior-centered approach is the effort to account for differences in Chinese and English speakers’ facility with counterfactual or hypothetical reasoning by reference to the marking of counterfactuals in the two languages (Bloom 1981). The interpretation of these results remains controversial (Lucy 1992a).