A Tale of Two Classifier Sets: Aphasia Treatment for a Cantonese Speaker

ABSTRACT

Background

Word retrieval deficits are common concerns for persons with aphasia (PWA). Treatment approaches can be divided into two processes, namely phonological and semantic, based on models suggesting that word activation consists of activating the semantic representation and phonological form of the target word. The role of syntax, however, has received less attention in models of retrieval and in the treatment of word retrieval deficits.

Aims

This study investigated whether a fixed noun phrase can support word finding in a PWA. Chinese is a classifier language where a classifier is obligatory in the quantitative noun phrase. The language uses two types of prenominal classifiers: sortal (serves to count individual nouns) and mensural (serves to count the units of nouns) classifiers.

Methods & Procedures

A Cantonese Chinese speaking PWA (LCH) with marked word retrieval deficits participated in this 15 bi-weekly, 90-minute treatment sessions, which focused on use of classifier cues to aid naming on confrontation naming. Twelve classifiers (6 sortal and 6 mensural) were selected to train target nouns. Target nouns were selected to pair with selected classifiers. Responses to cues were collected during each treatment session, documenting changes over time.

Outcomes & Results

Leveraging the relationship between target nouns and obligatory prenominal classifiers resulted in the participant’s improved picture naming for trained nouns as well as generalization to select untrained nouns. We report significant differences in the efficacy of sortal versus mensural classifiers to serve as primes for word retrieval. Sortal classifiers were significantly more effective than mensural classifiers, for both cueing trained and untrained target nouns.

Conclusions

For this speaker of a classifier language, use of classifiers served as viable cues for word retrieval and that sortal classifiers were more effective than mensural classifiers. Though word order does not distinguish between these two types of classifiers, their relative priming efficiency can be observed at the deep structure level. Findings suggest that a morphosyntactic approach to address word retrieval in a Chinese speaker may be beneficial but that great care must be taken in the creation of treatment materials. These factors must be further explored with additional participants and with a wider range of classifiers.

KEYWORDS:

• mensural classifiers
• sortal classifiers
• Cantonese
• aphasia treatment
• word retrieval

Introduction

Across aphasia types, word retrieval deficits are the most frequently reported challenges. Treatment approaches take advantage of phonological, semantic, and syntactic aspects of target words to facilitate retrieval. Research has shown that naming accuracy can be mediated by manipulation of these linguistic variables.

Phonological approaches provide spoken cues as a means of facilitating retrieval of target words. For example, presenting words that are phonologically related to the target word can facilitate retrieval by presenting the word ‘sheet’ to prime the target ‘sheep’ (Abdel Rahman & Melinger, 2008). These results imply that phonological cues can be used to activate the phonological representation of a target word. Fink et al. (2002) reported on the efficacy of a hierarchical phonological cueing protocol for persons with chronic aphasia presented via a computer-based treatment system. Similarly, Robson et al. (1998) presented a case report of an individual presenting with jargon aphasia and severe anomia. Treatment focused on strengthening access to phonological representations for target words, with notable success. Meta-analyses of aphasia treatment approaches revealed that those with moderate to severe deficits benefit from phonological treatments, as reported by Kristensson et al. (2021). These and other studies demonstrate the efficacy of phonological cues for some persons struggling with word retrieval.

Using semantic aspects of target words to gain access to target words is the premise of semantic approaches. These treatments, including answering yes/no questions, identifying category membership, word association tasks, and Cloze tasks, are used to elicit target words. Focusing on words that are semantically related to the target offers activation of the network that surrounds the target word. In this way, the target receives activation beyond its threshold and facilitates retrieval (Collins & Loftus, 1975). Boyle and Coelho (1995) reported improvement for trained and, to a limited extent, untrained words when they used a semantic feature analysis (SFA) approach to address word-finding deficits in a person presenting with mild non-fluent aphasia. Their participant, HW, was directed to describe semantic features of a target word, such as category membership, action, location, etc. The multiple exposures to target words and related semantic information are credited for improvement in naming trained and untrained targets.

In 2010, SFA was used to treat a person who sustained a closed head injury. In that case, not only was improvement observed for trained and untrained targets, but also in production of trained words in conversation. Findings suggested that SFA contributes to improved communication efficacy, at least for this participant (Coelho et al, 2010). Verb network strengthening treatment (VNeST) is another aphasia treatment program that directs attention to the semantic aspects of target words by focusing on verbs and their thematic roles (Edmonds, Nadeau, et al., 2009). For example, the verb ‘drive’ is associated with ‘car’, ‘truck’, ‘limo’, etc. Outcomes of this treatment have been promising with improved naming of trained and untrained objects as well as for relevant use of targets in discourse (Edmonds et al., 2014).

Phonological and semantic variables can be used to support those with word retrieval deficits. In addition, task demands seem to influence performance for these speakers. Persons with mild aphasia exhibited less difficulty naming target words when performing a narrative task, compared to performing a confrontation-naming task (Law et al., 2015; Pashek & Tompkins, 2002). Such differences suggest that lexical access in connected speech makes use of syntactic cues that are otherwise absent in confrontation naming. Unlike picture-naming tasks, producing a target word in a specific slot of a phrase or clause requires integrating semantic and syntactic information. These constraints provide a syntactic frame into which target words are placed (Gordon & Dell, 2003). That is, semantic and syntactic variables impose different but beneficial constraints on lexical access. In aphasia, a speaker may benefit from training in which syntactic cues are used to initiate an activation of the network for a target word. Gordon and Dell (2003) described this sequence as syntactic-sequential. Syntactic approaches are premised on the notion that lexical and syntactic aspects of sentences can be used to address production deficits. The lexical demands of verbs (number of arguments and the thematic roles of arguments) impose constraints on the resulting clause, which can be manipulated for treatment purposes (see Dickey and Thompson, 2007 for an extensive discussion).

Chinese and Chinese Classifiers

In Chinese, quantity is expressed by using a classifier (CL) in a fixed word order noun phrase (NP): “numeral+CL+noun” to indicate the number of objects being referenced. Discussions on the nature and dichotomy of Chinese classifiers are based on the work of Cheng and Sybesma (1998). Specifically, Chinese classifiers can be divided into two groups: those for counting individual objects (sortal classifiers) and those for counting measures of by using containers or units of objects or substances (mensural classifiers).

For example, the sortal classifier, ‘tiao’ 條, precedes nouns that are long and flowing such as rivers, scarves, and belts; the sortal classifier, ‘chang’ 場, precedes events such as dance performances, movies, operas, etc. Sortal classifiers group together based on semantic variables, including perceptual features or cognitive constructs. With the exception of the default classifier, ‘ge’ 個, which can be widely used (e.g., ‘ge’ 個can stand in for a pre-determined and well-established object-specific sortal classifier), most nouns are associated with specific classifiers. Sortal classifiers make up a closed class of obligatory morphemes to be used when expressing quantity. As stated, use of classifiers is obligatory in quantity NPs. In the following example, the NP “three silk scarves” must include the sortal classifier

Table

(1) a.				b.*
	三	條	絲巾		三	絲巾
	san	tiao	sijin		san	sijin
	three	sortal CL	silk scarves		three	silk scarves

As noun-selecting morphemes, sortal classifiers lexically constrain potential NPs that follow (see Zhang, 2007 for an extensive discussion). There are approximately 174 classifiers in Modern Mandarin (Huang et al., 1997) along with occasional idiomatic combinations that are historically rooted where the associations may not be transparent or logical to current speakers of Chinese. For example, though most animals (e.g., dogs, monkeys, birds, etc.) pair with the classifier, ‘zhi’ 隻, horse conventionally pairs with ‘pi’ 匹 while cow, hippo, and elephant usually pair with ‘tou’ 頭 (Guo & Zhong, 2005). Additionally, there are dialectal variations with respect to these pairings. ‘Zhi’ 隻 in Cantonese pairs with, in addition to many animals, ‘earrings’, ‘boat’, and one in a pair of objects (e.g., shoe, eye and glove).

Mensural classifiers refer to some containment or collection of an amount, and in some cases, a unit of some object or substance. For example, the mensural classifier ‘bei’ 杯, (cup, used as a unit of measure) is used to describe units of objects, such as three cups of coffee/sugar/tea and the mensural classifier, ‘di’ 滴 (‘drop of’) is used to describe nouns measured in drops, such as a drop of blood/medicine/water. Unlike sortal classifiers, which constitute a closed class of words, mensural classifiers are a subset of grammatical nouns that perform the function of measurement or containment; nouns are members of an open class of words. Hence, mensural classifiers are grammatical nouns functioning as classifiers in the quantitative NP construction. One example is ‘he’ 盒 (box), which functions as a grammatical noun in “I bought a blue box” but functions as a classifier in Chinese phrases such as “a box of candy” or “a box of socks” Like sortal classifiers, mensural classifiers are found in quantitative NPs. In the flowing example, the NP “three boxes of silk scarves” must include the mensural classifier

Table

(2) a.				b.*
	三	盒	絲巾		三	絲巾
	san	he	sijin		san	sijin
	three	mensural cl	scarves		three	scarves

When a mensural classifier is used, the sortal classifier associated with the target noun is omitted to allow for the mensural classifier as shown in (3)

Table

(3)
	三	盒	(條)	絲巾
	san	he	(tiao)	sijin
	three	mensural CL	(sortal CL)	silk scarves

While no word order distinction exists between sortal classifiers and mensural classifiers, the syntactic difference can be observed. Borrowing from Li (2013) and others, the sortal versus mensural distinction is found in the deep structure of the NP as shown in Figure 1:

Figure 1. Sortal versus Mensural Classifier Distinction

Despite the lack of a distinction between these NPs on the surface level, differences between sortal and mensural phrases are found in the underlying structure. Classifiers are central in the syntactic frame of the NP in that they provide distinct guidance for (target) word selection. Given pre-determined sortal classifier+noun pairings, sortal classifiers limit potential target words while mensural classifiers allow for a much wider range of potential target words. That is, mensural classifiers provide less constraint on NP selection compared to sortal classifiers, which impose greater constraint on NP selection. Given inherent differences, the relationship between classifier and noun warrants further consideration, especially its viability to serve as a prompt for word retrieval.

Huettig et al. (2010) found that listeners are sensitive to classifier information in predicting upcoming referents in noun phrases. This is demonstrated by eye gaze behaviors in favor of classifier-matched nouns versus classifier mismatched nouns. Similarly, Chou et al. (2014) reported that Mandarin speakers with aphasia are sensitive to the constraint imposed by classifier+noun pairing on a Chinese Cloze task, suggesting a preservation of morphosyntactic features—though to a lesser degree when compared to a neurotypical control group. These studies suggest that classifier information provides processing benefits for native Chinese speakers. However, it remains unclear whether these two sets of classifiers could facilitate lexical retrieval for speakers with compromised word finding abilities.

To date, there has not been a systematic exploration of using classifiers in the treatment of word retrieval deficits in a Chinese speaker. In this study, we examined the semantic contributions of an obligatory syntactic feature to facilitate word retrieval by investigating the feasibility of using classifiers in treating word retrieval deficits following aphasia in a Cantonese¹ speaker. Given that classifiers provide an argument constraint on the NP, we inquired about their efficacy as a prime for word finding. The purpose of this study was to document the use of a lexical treatment on aiding word retrieval in a Cantonese speaker with non-fluent aphasia. We tested the constraints of the fixed form, quantifier+CL, as a prime for target word and predicted that the closed-class sortal classifiers will have a different degree of impact on word finding compared to open-class mensural classifiers. Sortal classifiers offer global semantic constraints compared to the more local syntactic constraints offered by measure classifiers.

Therefore, we predicted that closed class classifiers (sortals) provide greater restrictions on its following potential target nouns and, therefore, reduce the number of possible collocations, leading to greater word finding success. We also argued that open class classifiers (mensural) provide less restrictions and, therefore, increase the number of potential NPs to follow. We surmised that this would lead to reduced word finding success. Using a single case design, we explored in detail the treatment effects of classifiers, which can be useful in aphasia by advancing current knowledge about treatment approaches for persons with aphasia by contrasting semantic and syntactic approaches.

Methods

Participant

LCH is a pre-morbidly, right-handed, 71-year-old Chinese-speaking male seen for an aphasia evaluation and treatment. His primary and preferred Chinese dialect is Cantonese. At the time of this intervention, LCH was 21 months post-stroke. There was no history of any developmental or learning disabilities, nor is there history of sensory impairment. LCH sustained a left-hemisphere cerebrovascular accident that resulted in non-fluent aphasia. A CT scan at the time revealed a large area of infarct involving the left frontal, parietal, and temporal lobes, as well as the left insula and deep gray and white matter regions of the left cerebral hemisphere. An old, smaller area of infarct within the right fronto-parietal region was noted. Mild to moderate scattered subcortical and periventricular white matter hypodensities, which are non-specific were also noted.

The family reported that LCH received two months of speech/language therapy at the time of the stroke, but they felt that he did not benefit from such services because the intervention was done in English, a language in which LCH has no functional skills. During these sessions, one of LCH’s adult sons interpreted for the therapist and his father. The son indicated that his father was asked to follow simple directions and to label pictures of objects. According to the son, LCH did not seem to understand what he was expected to do and seemed frustrated that he could not communicate. Family members state that though LCH has been in the United States for almost 20 years, he has not acquired any English, given the minimal need to interact in the second language and the fact that he had no formal educational experiences in the language. For example, he could conduct a simple conversation in English.

LCH completed a high school degree in Guangzhou, China and had worked many years in China as a salesperson before arriving in the United States in 2007. Upon his arrival in the U.S., LCH lived with his Cantonese-speaking spouse and worked in a predominantly Cantonese speaking community where there was no need for communication in English. His children were all born and educated in China. Therefore, the language of communication in the U.S. continued to be Cantonese among friends, work contacts, and family.

LCH is ambulatory but demonstrated a right-side hemiparesis observed as dragging the right lower extremity and no functional movement of the right upper extremity. Hearing was screened as part of an intake procedure and was normal at 30dB across speech frequencies.

Language Assessment

During the baseline evaluation sessions, several tests of language and cognition were adapted to Cantonese and administered, including the Western Aphasia Battery-Revised (WAB-R; Kertesz, 2007), the Boston Naming Test (BNT; Kaplan et al., 2001), and the Test of Non-Verbal Intelligence- 4 (TONI; Brown et al., 2010). All testing items were administered in Cantonese, which were translated by the first author, a clinician having extensive experience with Chinese speakers presenting with aphasia. It should be noted that the WAB-R has been adapted for Cantonese speakers by Yiu (1992) for the purposes of classifying “aphasic subjects into relatively similar patterns” (Yiu, 1992, p381). For this project, the authors chose to adapt the WAB to consider LCH’s language exposure and experiences here in the US. To evaluate LCH’s word retrieval skills, the Boston Naming Test was selected, given its demonstrated quantitative suitability for speakers of Chinese but at the same time, acknowledging its qualitative limitations; see Chen et al., 2014.

Baseline testing indicated that general non-verbal intelligence was unimpaired. However, language performance was clinically impaired. LCH’s performance on the WAB-R and BNT indicated marked difficulty with tasks involving word repetition, confrontation naming, and spontaneous language production. Impaired reading aloud and writing above the character/word level were also observed. His performance on the WAB suggests that auditory comprehension skills emerged as being stronger than language expression skills. For receptive language, LCH can execute simple 1- to 2-step commands, point to objects on command, respond to yes/no questions with head nods or head shakes, and able to detect and indicate false information by shaking his head. For expressive language, LCH is essentially non-verbal. He can repeat single words with multiple attempts, albeit with occasional perseveration.

On confrontation naming, LCH successfully identified 5/60 pictures presented on the BNT. Attempts to name common objects were limited to “um” and shaking his head. Successful automatic speech productions included counting to 20+. He can sing familiar Mandarin and Cantonese songs with minimal prompting from the examiner. Repetition skills are reduced to words and phrases. He is essentially non-fluent, exhibiting slow, effortful agrammatic productions. LCH is unable to sustain diadochokinetic rates. Reading skills were evaluated informally, using characters from a Chinese language third grade textbook. Here, LCH was able to match printed characters (words) to pictures of referents to with 100% accuracy (25/25). Additionally, he pointed to printed words during a sing-along task using familiar Chinese language songs. Due to hemiparesis, LCH attempted to write with his left hand, as he was pre-morbidly right-handed. He wrote his name in Chinese, the examiner’s Chinese name, along with words and numbers to dictation in Chinese. No English writing skills were seen.

Treatment

Individual speech/language therapy was initiated in Cantonese at approximately 21 months post-onset. A protocol using the two different types of classifiers was developed to address the marked word retrieval deficits, which is described below. Two 90-minute sessions were provided weekly for 15 weeks. After a one-month break from therapy, LCH was seen for post-treatment testing.

Stimuli

Twelve classifiers were selected for this training protocol: six sortal and six mensural classifiers. Selection was based on frequency of occurrence using the site https://lingua.mtsu.edu/ chinese-computing, which indirectly implied that each classifier would allow for a large number of potential NPs (see Appendix 1). Six target nouns were selected for each classifier for a total of 72 target nouns for training: 36 target (sortal) nouns for sortal classifiers and 36 target (mensural) nouns for mensural classifiers (see Appendices 2 and 3). In the absence of reliable, spoken Cantonese word frequency data, five neurotypical adult speakers of Cantonese from Guangzhou, China (three males, two females with ages ranging from 58 to 75 years of age (M = 65.0 years, SD = 7.3 years) and education level ranging from 9 to 12 years (M = 10.9 years, SD = 1.09 years), who have spent an average of 20 years in the United States (M=20.0 years, SD=3.16), were asked to name 20 nouns they could pair with each of the 12 classifiers (6 sortals and 6 mensurals) selected for this project. Given the constraints imposed by sortal classifiers, the noun lists generated for sortal classifiers exhibited less variability, whereas the noun lists generated for mensural classifiers showed greater variability. To reduce issues associated with word frequency, all nouns generated by these speakers were evaluated word frequency; the top 12 words for each classifier were selected this project. Results were ordered in frequency of occurrence, starting from the most frequently named to the least frequently named. Every other item was selected for training of each classifier (nouns numbered 1, 3, 5, 7, 9,11 and 13) starting with the most frequently named. Six nouns for each classifier were selected in this fashion. A total of 72 target nouns (36 sortal and 36 mensural nouns) were selected in this fashion. To assess treatment effects, a second set of nouns was selected by using the remaining items for each classifier (nouns numbered 2, 4, 6, 8,10 and 12) to determine whether there was any generalization from targeting word retrieval with classifiers. Therefore, six different nouns that are associated with each of the twelve classifiers selected for post-treatment evaluation. Six nouns for each classifier were selected in this fashion, giving a total of 72 untrained target nouns (36 sortal and 36 mensural nouns) were selected.

A total of 144 nouns were used for this treatment protocol; half were used for direct treatment while the second half were used to evaluate carryover of skills (see Table 1). A photograph of each object was used to create the confrontation-naming task. Photographs were presented one at a time, on an iPad tablet to elicit a naming response.

Table 1. Tests Administered

Test	Results
Western Aphasia Battery (Aphasia Quotient)	25/100
Spontaneous Speech (Subscore)	0
Auditory Verbal Comprehension (Subscore)	7
Repetition (Subscore)	2
Naming (Subscore)	3
Boston Naming Test	5/60
Test of Non-Verbal Intelligence-4	91st percentile

Two treatment blocks were created for this study: a sortal block (S) and a mensural block (M). For each block, each of the six classifiers in the block was paired with six different target nouns. A total of 36 nouns were selected for each treatment block. During each treatment session, each block of target nouns was presented four times: four times on S block runs (144 opportunities for naming) and four times on M block runs (144 opportunities for naming). LCH therefore had up to 288 opportunities to name target nouns during each treatment session. The order of block presentation alternated from one session to the next: SMSM for the first session followed by MSMS for the next. Within each block, the order of classifiers was randomly presented, and within each classifier the order of target nouns also changed from one session to the next.

LCH was given up to 10 seconds to name each target word. When he spontaneously and correctly named a target, the next picture stimulus was presented. Failed attempts included: no perceived response after 10 seconds, perseverative responses, vocalization, gestures of frustration, and attempts to relinquish his turn by shaking his head. In these cases, either a classifier cue or a phonological cue was offered, depending on the block (see below for details).

When LCH failed to name a target noun, two prompting conditions were used to facilitate word retrieval: a classifier condition and a phonological condition. In the classifier condition, the clinician presented a neutral clause with a target classifier: “Mr. Li + has + one + CL+ _____” (“李先生有一 + CL__”). Data was collected on the accuracy of his response to classifier cues; that is, whether classifier cues facilitated accurate retrieval of the target word. In the phonological condition, the clinician presented a phonological cue: “This + is + a + initial phoneme of the target word” (“這是一 initial phoneme of the target word”) when LCH failed to name an object spontaneously.

Data were collected on the accuracy of his response to phonological cues; that is, whether phonological cues facilitated accurate retrieval of the target word. Under both conditions, if LCH failed to name the target object following cues, the clinician modeled the response for him to repeat. No data were collected on whether the participant repeated the model, nor was data collected on the quality of his repetition attempts. Each session always started with classifier cues regardless of the type of target noun (sortal or mensural).

A pre-treatment/baseline was established before beginning this treatment, LCH was asked to name photographs of each of the 72 objects targeted for training. Like the training protocol, there were two presentations of these photos. During the first presentation, when LCH failed to name an object spontaneously, the examiner offered a classifier cue. During the second presentation, when LCH failed to name an object spontaneously, the examiner offered a phonological cue. His responses during this session served as the pre-training data.

Post-treatment testing was completed 1 month after completing 15 weeks of treatment. Here, he was presented with the set of 72 nouns (trained nouns) that were targeted for treatment along with the second set of 72 nouns (untrained nouns) and was asked to name photographs of the 144 objects. Consistent with the training protocol, there were two presentations of these photos. During the first presentation, when LCH failed to name an object spontaneously, the examiner offered a classifier cue. During the second presentation, when LCH failed to name an object spontaneously, the examiner offered a phonological cue. A pre-treatment session was conducted in the same manner; responses during that session served as pre-training data. The protocol for the SMSM sessions and the MSMS sessions followed the same process, but began with M nouns instead of S nouns (see Appendix 4 for detailed information). LCH’s responses were documented by the first author, who conducted all treatment sessions. The second author simultaneously collected data alongside of the first author for 15% of the treatment sessions to assure the accuracy and reliability of both protocol implementation and data collection and documentation; 100% agreement was observed for accuracy of data recording and execution of treatment trials.

Results

Response to treatment showed a differential effectiveness of the two types of classifiers and that responses to both trained and untrained items for sortal classifiers were improved compared to those for mensural classifiers. Pre-training baselines were collected to establish the participant’s word retrieval status. As a reminder, two separate sets of target nouns were selected: one set paired with sortal classifiers (sortal nouns) and the other paired with mensural classifiers (mensural nouns). Thirty-six nouns were selected for each set. Data were collected on LCH’s ability to name target objects spontaneously, which were comprised of LCH’s response to phonological cues for objects he failed to name spontaneously and his response to classifier cues for objects he failed to name spontaneously. These sets of nouns were presented only during the 15 weeks of speech/language therapy to address word retrieval deficits.

Following this training period, data were collected to evaluate whether there was a change in responsiveness from pre-training baseline data. In this case, data were collected using the same sets of trained common objects. Pre- and post-training data are presented in Table 2. In the following comparisons, a Bonferroni-corrected alpha of p < .0125 (.05/4) was used to control for family-wise type I error.

Table 2. Number of Trained Classifiers with Corresponding Target Nouns

	Sortal Classifier	Mensural Classifier	Total
Training Classifiers	6	6	12
Target Nouns (Trained)	36	36	72
Target Nouns (Untrained)	36	36	72

A paired t test showed a significant difference in overall naming accuracy rate of trained common objects between pre-training and post-training, t(7) = 5.69, p < .001. The overall naming accuracy rate at pre-training (M = 0.08, SD = 0.04) was significantly increased post-training (M = 0.34, SD = 0.14). This suggests that intensive treatment on a focused set of words improves naming accuracy.

In addition to documenting pre- and post-training differences, a second set of data was collected, using untrained target words. Here, the focus was on whether training with classifier and phonological cues produced any meaningful generalisation to untrained nouns. Table 3 includes the post training data for trained and untrained nouns.

Table 3. Pre- and Post- Training Word Naming Accuracy for Trained Nouns

Accuracy (Pre-training)	Ratio (Pre-training)	Condition	Accuracy (Post-training)	Ratio (Post-training)
0.08	3/36 (max=36 opportunities)	SponSortalN (spontaneous naming attempts prior to sortal classifier cues)	0.33	12/36 (max=36 opportunities)
0.06	2/33 (max=36-3 opportunities)	CLSortalN (response to sortal classifier cue after missed spontaneous naming attempt)	0.46	11/24 (max=36-12opportunities)
0.06	2/36 (max=36 opportunities)	SponSortalN (spontaneous naming attempts prior to phonological cues)	0.31	11/36 (max=36 opportunities)
0.15	5/34 (max=36-4 opportunities)	PhSortalN (response to phonological cue after missed spontaneous naming attempt)	0.64	16/25 (max=36-11opportunities)
0.03	1/36 (max=36-3 opportunities)	SponMensuralN (spontaneous naming attempts prior to mensural classifier cues)	0.28	10/36 (max=36 opportunities)
0.09	3/35 (max=36-1 opportunities)	CLMensuralN (response to mensural classifier cue after missed spontaneous naming attempts)	0.27	7/26 (max=36-10 opportunities)
0.06	2/36 (max=36-3 opportunities)	SponMensuralN (spontaneous naming attempts prior to mensural classifier cues)	0.19	7/36 (max=36 opportunities)
0.12	4/34 (max=36-2 opportunities)	PhMensuralN (response to phonological cue after missed spontaneous naming attempts)	0.24	7/29 (max=36-7 opportunities)

A significant difference was observed in overall naming accuracy rate between trained targets and untrained targets, t(7) = 3.71, p = .0076. The overall naming accuracy rate for trained targets (M = 0.34, SD = 0.14) was significantly higher than for untrained ones (M = 0.22, SD = 0.08). This suggests that intensive training improves naming accuracy of familiar words as compared to naming accuracy of unfamiliar ones.

We were also interested in whether training had any effect on LCH’s ability to name untrained targets. Here, we compared his performance on a set of untrained targets before starting treatment to his performance on a different set of untrained targets at the conclusion of treatment. Table 4 includes the participant’s responses to untrained targets prior to and after treatment.

Table 4. Post Training Performance on Trained and Untrained Target Nouns

Accuracy (Post-training Trained targets)	Ratio (Post-training Trained targets)	Condition	Accuracy (Post-training Untrained targets)	Ratio (Post-training Untrained targets)
0.33	12/36 (max=36 opportunities)	SponSortingN (spontaneous naming attempts prior to sortal classifier cues)	0.28	10/36 (max=36 opportunities)
0.46	11/24 (max=36-12opportunities)	CLSortingN (response to sortal classifier cue after missed spontaneous naming attempts)	0.27	7/26 (max=36-10 opportunities)
0.31	11/36 (max=36 opportunities)	SponSortingN (spontaneous naming attempts prior to phonological cues)	0.19	7/36 (max=36 opportunities)
0.64	16/25 (max=36-11opportunities)	PhSortingN (response to phonological cue after missed spontaneous naming attempts)	0.34	10/29 (max=36-7 opportunities)
0.28	10/36 (max=36 opportunities)	SponMensuralN (spontaneous naming attempts prior to measure classifier cues)	0.14	5/36 (max=36 opportunities)
0.27	7/26 (max=36-10 opportunities)	CLMensuralN (response to measure classifier cue after missed spontaneous naming attempts)	0.16	5/31 (max=36-5 opportunities)
0.19	7/36 (max=36 opportunities)	SponMensuralN (spontaneous naming attempts prior to measure classifier cues)	0.11	4/36 (max=36 opportunities)
0.24	7/29 (max=36-7 opportunities)	PhMensuralN (response to phonological cue after missed spontaneous naming attempts)	0.25	8/32 (max=36-8 opportunities)

A significant difference was observed in overall naming accuracy rate between pre-training naming ability and post-training naming ability, t(7) = 6.45, p < .001. The overall naming accuracy rate for the post-training period (M = 0.22, SD = 0.08) was significantly higher than that of pre-training (M = 0.08, SD = 0.04). This suggests that intensive training is generalised to untrained materials.

The participant’s response to treatment is documented in Table 5, which includes response accuracy to phonological cues and classifier cues by sortal and mensural noun sets over time. As in pre- and post-training data reporting, LCH was provided cues when he failed to name a target noun spontaneously. By reporting data from every third treatment session, we documented change over time, under four different prompting conditions; please see Table 6. Naming accuracy improved across both noun sets and with both types of cues. Sortal classifiers cues are more effective than mensural ones, and phonological cues provide about the same amount of support for both noun sets.

Table 5. Pre- and Post- Training Naming Accuracy for Untrained Words

Accuracy (Pre-training)	Ratio (Pre-training)	Condition	Accuracy (Post-training)	Ratio (Post-training)
0.08%	3/36 (max=36 opportunities)	SponSortingN (spontaneous naming attempts prior to sortal classifier cues)	0.28	10/36 (max=36 opportunities)
0.06%	2/33 (max=36-3 opportunities)	CLSortingN (response to sortal classifier cue after missed spontaneous naming attempts)	0.27	7/26 (max=36-10 opportunities)
0.06%	2/36 (max=36 opportunities)	SponSortingN (spontaneous naming attempts prior to phonological cues)	0.19	7/36 (max=36 opportunities)
0.15%	5/34 (max=36-4 opportunities)	PhSortingN (response to phonological cue after missed spontaneous naming attempts)	0.34	10/29 (max=36-7 opportunities)
0.03%	1/36 (max=36-3 opportunities)	SponMensuralN (spontaneous naming attempts prior to measure classifier cues)	0.14	5/36 (max=36 opportunities)
0.09%	3/35 (max=36-1 opportunities)	CLMensuralN (response to measure classifier cue after missed spontaneous naming attempts)	0.16	5/31 (max=36-5 opportunities)
0.06%	2/36 (max=36-3 opportunities)	SponMensuralN (spontaneous naming attempts prior to measure classifier cues)	0.11	4/36 (max=36 opportunities)
0.12%	4/34 (max=36-2 opportunities)	PhMensuralN (response to phonological cue after missed spontaneous naming attempts)	0.25	8/32 (max=36-8 opportunities)

Table 6. Effectiveness of Classifier Cues vs. Phonological Cues

	S-Classifier (accuracy of responses to phonological cues for the sortal noun set)	M-Classifier (accuracy of responses to phonological cues for the mensural noun set)	S-Phonological (accuracy of responses to phonological cues for the sortal noun set)	M-Phonological (accuracy of responses to phonological cues for the mensural noun set)
wk3	33%	21%	23%	28%
wk6	41%	21%	26%	36%
wk9	46%	23%	32%	29%
wk12	48%	23%	35%	36%
wk15	66%	19%	43%	25%

A paired t test showed a significant difference in overall naming accuracy rate between the use of sortal classifiers and mensural classifiers, t(4) = 4.35, p = .0121. The overall naming accuracy rate using sortal classifiers (M = 0.47, SD = 0.12) was significantly higher than for mensural classifiers (M = 0.21, SD = 0.02). No statistically significant difference was observed between using sortal (M = 0.32, SD = 0.08) and mensural (M = 0.31, SD = 0.05) phonological cues for naming, t(4) = 0.21, p =.844. This suggests that phonological cues provided a level of support regardless of the target noun.

Figure 2 documents the participant’s progress over the 15 weeks of intensive language therapy; response to treatment data is presented for every third week of treatment. Despite equal attention focused on the two types of cueing, differences in responsiveness to cue type are observed. LCH’s responses distinguish the efficacy of the two cue types offered to support word retrieval.

Figure 2. Percentage of Response Accuracy to Different Cues in Treatment

For the sortal noun set, sortal cues were more effective for LCH in eliciting target words compared to phonological cues. Consider the data points S-C (representing sortal nouns cued with its associated classifier cue) and S-P (representing sortal nouns cued with its associated phonological cue) that represent the percentage of accurate responses to cues over the course of treatment; they suggest that sortal classifier are more effective compared to phonological cues for sortal nouns.

For the mensural noun set, phonological cues were more effective for LCH in eliciting target words than classifier cues. Consider the data points M-C (representing mensural nouns cued with its associated classifier cue) and M-P (representing mensural nouns cued with its associated phonological cue) that represent the percentage of accurate responses to cues over the course of treatment; they suggest that phonological cues are more effective compared to mensural classifiers for mensural nouns.

Using a phonological cue does not seem to distinguish between the two sets of nouns. That is, phonological cues appear to have a similar impact on word retrieval regardless of noun set; consider the data points M-P and S-P. Data represents the percentage of accurate responses to cues over the course of treatment.

Using a classifier cue seems to distinguish between the two sets of nouns. This result suggests sortal cues appear to have a stronger impact on word retrieval than do mensural cues; consider data points S-C and M-C. Data present the percentage of accurate responses to cues over the course of treatment. Compared to mensural cues, sortal cues were more effective in facilitating word naming.

Figure 3 documents the progress of spontaneous naming over the course of treatment. Here, LCH’s ability to name target words in the two sets of nouns appears to improve over time. Towards the end of the treatment period, LCH was more successful in his attempts to name words treated with sortal cues compared to words treated with mensural cues.

Figure 3. Percentage of Accuracy in Naming Target Word Used in Treatment

Discussion

This study compared the effectiveness of two different levels of cueing (lexical versus phonological) in support of word retrieval for one Cantonese speaker (LCH) with severe non-fluent aphasia. Treatment-induced changes were observed in response to phonological as well as lexical cueing to facilitate word access. As predicted, LCH benefited from phonological prompts across all nouns presented. His performance was consistent with other reports supporting the use of phonological cueing to address word-finding deficits. Though LCH also benefited from lexical prompts from classifiers, differences in his response to sortal versus mensural classifier cues were observed. These differences suggested that despite the fixed word order of quantitative NPs in Cantonese, these two classifier morphemes influence word retrieval differently.

Language-specific features (in this case, morphosyntactic features) may be recruited to address lexical access deficits. LCH was significantly more responsive to semantic cues compared to syntactic cues. Sortal classifiers are pre-assigned and fixed to specific nouns compared to mensural classifiers, which collocate with nouns in a fluid, temporary, and unpredictable manner. As such, mensural classifiers freely associate with any number of different nouns. Additionally, because these are grammatical nouns functioning as classifiers in the quantitative NP frame, they serve to partition a measurement of some substance or objects but not as context to support word retrieval. In other words, the mensural classifier is the head of the NP, with the target noun serving as its relative clause. LCH’s responsiveness to categorical cues suggests a sparing of lexical association, which is handy for word retrieval, in contrast to the limited effectiveness of syntactic cues. His performance is proof of the concept that sortal and mensural classifiers behave differently, despite the fixed surface structure.

To explore the impact of classifiers as treatment options, we considered the post-training data and noted that LCH derived greater benefits from categorical (sortal) cues for both trained and untrained target nouns compared to benefits from syntactic (mensural) cues. Generalisation to untrained targets suggests that training typical items can positively impact word retrieval. Here we distinguished the benefits of sortal cues from mensural ones. Treatment afforded persistent activation of specific sortal classifiers shared between trained and untrained words. This may have unintentionally enhanced access to untrained words. By nature, generating potential nouns for mensural classifiers means that responses can be unpredictable, unlike generating potential nouns for sortal classifiers, which constrain the field of potential nouns, can result in retrieving predictable responses.

Recall that in selecting target nouns for classifiers used in this study, we asked neurotypical adult Cantonese speakers to generate nouns that associate with given classifiers. Greater conformity was observed on lists for sortal classifiers compared to lists in response to mensural classifiers. Divergent thinking supports the generation of a wide range of different responses, given weak task constraints. An example is listing potential functions of a pencil or potential nouns for mensural classifiers. Convergent thinking requires retrieving knowledge or facts. An example of this is listing potential animals in a zoo or potential nouns for sortal classifiers (Razumnikova, 2013). Syntactic cues did not provide the same level of efficacy as categorical ones for LCH. Mensural classifiers provide a syntactic frame into which a wide range of target nouns can be slotted. Satisfying this slot requires divergent processing as mensural classifiers do not impose enough constraints over this slot. Therefore, upcoming plausible target nouns are simply unpredictable. In contrast, categorical cues limit the range of plausible responses for the target noun slot. Satisfying this slot is dictated by the output of convergent processes – a strategy that proved less demanding for LCH.

Our results might be attributed to connectionist modeling whereby rehabilitation of targets is generalised to related but untreated ones (Plaut, 1996). The current study focused on Chinese classifiers, which allows for broad category membership based not solely on lexical-semantic variables only, but also cognitive ones (e.g., animacy, physical shape, function, and historical information). In EEG studies, violations of sortal classifier+noun agreement induced N400 responses in native Mandarin speakers (Zhang et al., 2012; Zhou et al., 2010). Such work supports the notion that sortal classifiers and their (noun) charges have a stronger link to each other compared to mensural classifiers, which provide only a temporary assignment to their charges.

Models of Word Retrieval

Longstanding models of word finding advance a two-step process in which semantic access leads to phonological access. Aphasia treatments using a semantic feature approach rely on network models in which activating a concept/node in the network spreads to other nodes and ultimately leads to lexical access (Dell et al., 1997). Quantifier phrases bound speakers of Chinese to the CL+NP frame so that nouns are selected with their associated sortal classifiers. Li (2016) argued for classifiers providing structural organization for nouns because speakers judge words that share classifiers as more similar than those that do not share them. In addition, speakers tend to group words by classifier when asked to recall a list of nouns.

As previously noted, robust event-related potential (ERP) data imply that accurate selection of the sortal classifier is paramount so that deviation from an expected form is quickly detected. We propose that sortal classifiers are critical conceptual nodes for Chinese nouns. Unlike other nodes in a network, classifiers likely enjoy a strong pairing relationship with target nouns given their fixed and obligatory characteristics. The consistency and frequency of occurrence of these pairings strengthens the sortal classifier+noun link (Collins & Loftus, 1975). For example, ‘san’傘 (‘umbrella’) may likely activate ‘rain’, ‘wet’ and ‘boots,’ and to some extent, ‘cold’, ‘open’, and ‘handle’ but it most certainly activates its paired classifier ‘ba’ 把 (see Figure 4).

Figure 4. Semantic Networks of Chinese Words: A Schematic

Clinical Implications

Can the principles of semantic feature analysis be adapted to take advantage of this structure? LCH’s response to treatment over time supports this possibility. If, as proposed here, the network of a Chinese noun phrase includes its assigned sortal classifier, then leveraging the strength of the classifier+noun link may explain LCH’s responsiveness to sortal cues relative to his responses to mensural cues. The relative sparing of sortal cues suggests that classifiers can be useful in supporting clinical treatment. We recognise the limitations inherent in single participant reports, given that information collected cannot be generalized since at the moment, the theoretical knowledge acquired still needs further investigation and applied in practical situations. However, individualizing a treatment approach for LCH has provided an opportunity for a focused investigation to describe, predict and manipulate the potential role of a grammatical class of words recruited to support word finding (Levelt, Roelofs, and Meyer, 1999;Yin, 2013). Along these veins, we strongly advocate for more exploratory studies to evaluate the feasibility of classifiers as cues for word retrieval for speakers of Chinese.

Despite the apparent limitations of a case study report, LCH’s gains underscore the importance of expanding the existing arsenal of treatment options, particularly for non-English speakers. We are keenly aware of the shortcomings of altering the content of formal assessment batteries, namely, the WAB-R, for purposes of characterizing and describing our participant’s language strengths and weaknesses. Future work in this area necessarily includes improving methods of evaluation of language loss in speakers of other languages (including Cantonese) and in different linguistic contexts (including the US) and in replication of the current findings with more Cantonese-speaking individuals with aphasia so as to both expand on and delve into the nature of classifiers and the organisation of semantic networks as well as ways of utilising and manipulating classifiers to repair damaged lexical access routes. Future work is also required to formalize this approach of pairing training targets and select classifiers are treatment purposes. With this, language tasks engaging divergent versus convergent processes for persons with aphasia can also be explored.

Conclusion

LCH’s performance suggests intact category (sortal) knowledge can have a positive impact on word retrieval. The obligatory function of sortal classifiers can be brought into treatment sessions in support of word retrieval attempts. This study confirmed the importance of systematically considering language-specific features that can be recruited as lexical cues to aid word retrieval. Although additional studies are needed, our preliminary results suggest that using classifiers in aphasia treatment might be a promising avenue that should be further explored in controlled clinical research.

Acknowledgments

We wish to thank the participant and his family for their support and commitment to this study. We also wish to thank Dr. Heather Harris Wright, who served as editor, and to the two anonymous reviewers who provided insightful comments and suggestions on a previous draft of this paper.

Disclosure Statement

The authors report no conflicts of interest.

Notes

¹ Though differences in how classifiers are managed in conversational Cantonese and Mandarin are noted, for the purposes of this study, a single expression of quantity was used in Cantonese. This expression is acceptable for both dialects of Chinese; see Cheng and Sybesma (2005) for a comparison of the two systems.

Appendix 1:

List of Classifiers Used for Training

Table

Sortal Classifiers	Mensural Classifiers
把	杯
粒	碗
條	盒
塊	排
隻	班
部	包

Appendix 2:

List of Sortal Classifiers and Target Nouns

Table

Sortal
把	掃帚	broom	塊	餅	cookie
	扇	fan		比薩	pizza
	劍	sword		麵包	bread
	刀	knife		牛肉	beef
	槍	gun		薯片	potato chip
	尺	tape		蛋糕	cake
	剪刀	scissors		玉	jade
	鑰匙	keys		木板	wood plank
	傘	umbrella		鏡	mirror
	牙刷	toothbrush		布	cloth
	梳	comb		唱片	music record
	火	fire		桌布	table cloth
粒	珍珠	pearl	隻/只^2	狗	dog
	丸	pill		貓	cat
	葡萄	grape		猴	monkey
	鑽石	diamond		龜	turtle
	沙	sand		杯	cup
	糖果	candy		手	hand
	飯	rice		腿	leg
	草莓	strawberry		碗	ear
	鈕	button		鞋	shoe
	蘑菇	mushroom		襪	sock
	蝦米	dry shrimp		眼	eye
	芝麻	sesame		手錶	wrist watch
條	皮帶	belt	部	電影	movie
	頭髮	hair		卡車	truck
	繩	rope		電腦	computer
	圍巾	scarf		電話	telephone
	裙	dress		飛機	airplane
	項鍊	necklace		冰箱	refrigerator
	蛇	snake		救護車	ambulance
	褲	pants		冷氣機	air conditioner
	絲帶	ribbon		汽車	automobile
	鞋帶	shoelace		故事書	story book
	毛巾	towel		電視	television
	魚	fish		洗衣機	washing machine

These classifiers are homophonous but orthographically, 隻 references animals whereas只references one in a pair of objects.

Appendix 3:

List of Mensural Classifiers and Target Nouns

Table

Mensural
杯	咖啡	coffee	排			娃娃	dolls
	茶	tea				衣服	clothes
	牛奶	milk				椅	chairs
	珍珠茶	boba tea				蘋果	apples
	冷水	tap water				卡車	trucks
	紅葡萄酒	red wine				行李	luggage
	白酒	white wine				屋	houses
	橙汁	orange juice				巴士	buses
	蘇打	soda				商店	stores
	蘋果汁	apple juice				南瓜	pumpkins
	熱水	hot water				瓶子	bottles
	奶茶	milk tea				座位	seats
碗	冬瓜湯	melon soup	班			小朋友	kids
	餛飩面	wonton soup				父母	parents
	粥	congee				朋友	friends
	白米	white rice				醫生	doctors
	炒飯	fried rice				空姐	stewardesses
	豆腐花	tofu snack				護士	nurses
	草藥湯	herbal soup				記者	reporters
	湯麵	soup noodle				耐心	patients
	草凍	grass jelly				藝術家	artists
	雞湯	chicken soup				廚師	chef
	餃子	dumplings				警察	police
	薑醋湯	ginger soup				老師	teachers
盒	玩具	toys	包			麵粉	flour
	紙巾	napkins				蘑菇	mushrooms
	藥	medicine				花生	peanuts
	衣服	clothes				乾蝦米	dried shrimp
	水果	fruit				橘子	oranges
	刀	knives				髒衣服	laundry
	書	books				香蕉	bananas
	鞋	shoes				杏仁	almonds
	麵包	bread				垃圾	trash
	眼鏡	glasses				錢	money
	毛巾	towels				尿布	diapers
	報紙	newspapers				紅蘿蔔	carrots

Appendix 4:

Treatment Protocol

1. Present Sortal (S) Noun set for confrontation naming

   1. If successful, continue to the next target

   2. If unsuccessful, present Sortal classifier prompt to assist the subject

   3. Regardless of whether the subject was successful or not, present the next target

   4. Complete all 36 Sortal nouns following step 1

2. Present Mensural (M) Noun set for confrontation naming

   1. If successful, continue to the next target

   2. If unsuccessful, present Mensural classifier prompt to assist the subject

   3. Regardless of whether the subject was successful or not, present the next target

   4. Complete all 36 Mensural nouns following step 2

3. Present Sortal (S) Noun set for confrontation naming

   1. If successful, continue to the next target

   2. If unsuccessful, present phonological prompt to assist the subject

   3. Regardless of whether the subject was successful or not, present the next target

   4. Complete all 36 Sortal nouns following step 3

4. Present Mensural (M) Noun set for confrontation naming

   1. If successful, continue to the next target

   2. If unsuccessful, present phonological prompt to assist the subject

   3. Regardless of whether the subject was successful or not, present the next target

   4. Complete all 36 Mensural nouns following step 4