Cognitive Investigations of Comprehension Processes 曾玉村 Yuhtsuen Tzeng Learning & Cognition Lab, National Chung Cheng University Presented at GSME, NCYU, May, 04, 2011

To understand is human nature Comprehension/understanding rests on memory Comprehension promote reasoning, problem solving, creation… To understand is to activate old knowledge To understand is to a first step of learning Education should aim at understanding

What is comprehension/understanding? Language Non-language

What is language comprehension? Complex cognitive processes The construction of a coherent mental representation of the text (e.g., Kintsch, 1988; Graesser & Clark, 1985; Trabasso & van den Broek, 1985) The reader identifies meaningful relations between parts of the text between text and background knowledge

What is comprehension? Text input provides raw elements for comprehension

What is comprehension? Readers connect text elements to build a network representation

What do the LS models say? How do readers achieve comprehension? As readers proceed through the text, they attempt to maintain coherence for each new text segment (sentence) (Re)activate information from prior text or background knowledge E.g., referential and causal coherence Naming/lexical decision/speeded recognition as well as reading time studies (e.g., O’Brien & Myers, 1989; McKoon & Ratcliff, 1990) Inferences are crucial to go beyond text inputs

Formalization of the LS model Attentional/Working memory limitations Comprehension involves fluctuation of activation from Text input Carry over Reinstatement (from representation of prior text) Activation of background knowledge Cohort activation (see below)  landscape of activations

The LS model: Constructing a memory representation Connections are identified between text items that are activated simultaneously The episodic memory representation is updated as a result of each new activation vector, via an asymptotic (delta) rule

LS model: Dynamic interaction between activation and representation Co-activated elements form a cohort in the episodic representation Cohort activation provides additional activation during reading process Consistent with current models (e.g., Construction-Integration model, Kintsch 1988; Resonance model,O’Brien & Myers)

Mathematical Formalizations of LS model Three phases of LS model Determine input Compute cohort Update episodic memory

Determine Input

Compute Cohort Activation- Equation 1

Transform Connection Strengths-Equation 2

Sigmoid Transformation

Compute Expectancy- Equation 3

Update Connection Strengths- Equation 4

Partial Memory Representation at Reading Cycle 3

Partial Memory Representation at Reading Cycle 13

The LS program: Input

The LS program: Input

The LS program: Output of activation patterns

The LS program: Output of connection patterns

Offline memory representation of text comprehension (Tzeng & Chen, 2006) Textbase Situation model

Offline memory representation of text comprehension (曾玉村＆陳秋芬，修改中)

How does causal structure of text affect online reading processes? A minimalist view: readers make minimum amount inferences (McKoon & Ratcliff, 1992) A bottom up approach of inference makings Many “inferences” are passive memory activation processes Readers will make inferences only if No readily available information in WM Coherence breaks in texts

How does causal structure of text affect online reading processes? A constructionist view: readers would search for meaning (Graesser, Singer, & Trabasso, 1994) Readers would make many (but not all) types of inferences to satisfy coherence assumption and explanation assumption Coherence assumption: need local & global coherence to build situation models Explanation assumption: to explain why certain event occur

A contrasting prediction A minimalist view predicts no inferences will be made if texts have local coherence A constructionist view predicts that inferences will be made even if texts have local coherence but do not have global coherence

Label Hierarchical Version Sequential Version 1. Setting 2. Initiating 1 3. Goal 1 4. Action 1 5. Outcome 1 6. Reaction 1 7. Initiating 2 8. Goal 2 9. Action 2 10. Action 3 11. Outcome 2 12. Action 4 13. Action 5 14. Outcome 3 從前有個男孩叫吉米 有一天他看到他的朋友騎著一輛新腳踏車 吉米也想要買一輛新腳踏車 他去告訴媽媽 他媽媽拒絕幫他買腳踏車. 吉米很傷心 他媽媽告訴他應該要自己存一些錢 吉米想要賺一些錢. 他就在雜貨店找到工作 他替雜貨店外送貨品 他賺了不少錢. 他去一家百貨公司 他走到二樓 他買了一部新腳踏車. 他媽媽答應幫他買腳踏車. 吉米很高興 他買了一個新的籃球.

Do readers make online global inferences? Procedures 22 college students 32 stories were constructed out of 16 scenarios, half sequential & half hierarchical Participants engaged in self-paced reading while recording their eye movements using Eyelink II

Eye tracking system

Eye tracking system

Results of eye movement data Focusing on two types of sentences Major goals & consequences (i.e., 3th、5th、8th、11th & 14th sentences) Spill over effects of them (i.e., 4th、6th、9th & 12th sentences)

Results of eye movement data First Gaze: the sum of time a reader’s first look of a sentence to the first time s/he exits that sentence no differences on both versions except for outcome 3 longer duration for sequential than hierarchical stories (889 vs. 777 ms respectively, t = 2.56, p = .011) all other ps > .05

Results of eye movement data Second Gaze: the sum of time readers’ fixations on their second pass of a sentence longer for sequential version than hierarchical version on outcome 1 (1159 vs. 818 ms, t = 2.61, p = .009)

Results of eye movement data Total duration: the sum of first-gaze and second-gaze longer for sequential version than for hierarchical version on outcome 1 (2258 vs. 1789 ms, t = 2.65, p = .009) longer for Action 1 for sequential version (a spillover region of Goal 1) (2350 vs. 1968 ms, t = 2.06, p = .04)

Results of eye movement data Pattern of first Gaze Cognitive loading for both version is the same since the wordings and syntax are almost identical

Results of eye movement data Pattern of second gaze Longer fixation time for sequential versions on Outcome 1 Readers attempt to go back to integrate sentences to build a coherent representation since it start a new episode after this sentence (i.e., lack of global coherence)

Results of eye movement data Pattern of total gaze Outcome 1 marks the first difference between two versions of text therefore readers pay different amount of time this effect spill back to Action 1 There is a coherence break between the first and the second episode Lack of global coherence in sequential stories causes readers spend more time on the first episode

Conclusions of eye movement data Causal structure affect online reading times Fixation time increase if texts lack of cross-episode (global) coherence High causal coherent texts result in fewer looking back behaviors Support a constructionist view

Key Issues of comprehension research Successful decoding Form text-based representations extract propositions from texts (Kintsch & van Dijk, 1978) Form situation model: what the text is about Go beyond text Adding background kknowledge (van Dijk & Kintsch, 1983; Kintsch, 1988)

Inferences are crucial for constructing situation model Three turtles floating on a log. Several fish swimming under the log.

What are inferences Information not explicit in text & readers retrieve from memory（Kintsch, 1998; Graesser, Singer, & Trabasso, 1994; Till, Mross & Kintsch, 1988) 「他把鎖好的門打開」 「鑰匙」 「氣憤的老婆把杯子往地板上摔」 「碎掉」

閱讀理解的推論能力 Weaver 與 Kintsch (1991)估計讀者每讀一則文章之敘述，平均必須進行上打的推論才能真正 完全理解文意 理解需要多少推論以及何種推論有待釐清

閱讀理解的推論能力 最小量論（minimalist） ( McKoon & Ratcliff, 1992） 若文本具有局部連貫性，讀者則沒有進行整體推論連貫之必要，除非相關訊息可免費取得 建構論（constructivist） ( Graesser, Singer, & Trabasso, 1993） 讀者會進行必要之推論以獲取文章整體意義，滿足讀者尋求意義的需求

閱讀理解的推論能力發展 傳統閱讀發展停留於整體能力指標描述 傳統閱讀能力發展研究往往缺乏理論 推論能力發展研究較具特定性 推論能力發展研究有潛力提早診斷閱讀理解困難

兒童讀者推論能力的發展 Song 和Fisher(2005, 2007)發現，幼兒階段已開始發展代名詞使用與理解的能力。 針對兒童讀者所做的研究，發現代名詞推論(pronoun resolution)的能力可以有效預測兒童的閱讀理解表現。(Hickmann et al., 1995; Kail & Hickmann, 1992; Oakhill & Yuill, 1986)。

中文兒童推論能力發展 曾玉村、陳明蕾與陳沛嵐（Tzeng, Chen, & Chen, 2007）尋找中文讀者特有的理解特性 小六學生閱讀完整指稱詞的文章與極少指稱詞（正常文章盡量刪除指稱詞）的文章的理解並無差異 小四學生對於極少指稱詞文章的理解低於完整指稱詞的文章 黃秋華、曾玉村與陸偉明（Huang, Tzeng, & Lu, 2009）則進一步尋求年幼讀者指稱詞推論能力的發展的線上指標 指稱詞推論能力介於四到六年級似乎有明顯變化 六年級對於零代詞的處理能力似乎已經發展成熟

研究目的 了解兒童因果推論與指稱詞推論在理解歷程中之發展與可能的交互影響 同時收集線上與離線資料

研究假設與預測 邏輯 預測 短文前後句有無代名詞或不同強度之因果關係時，會影響讀者閱讀理解的心理表徵，進而影響回答問題之正確率與閱讀時間 六年級優於四年級 年級與短文連貫性有交互作用 高因果優於低因果短文 有無代名詞短文僅對四年級有影響

實驗方法 受試者： 台南市國小學生兩班 排除特殊兒童以及上學期國語科成績低於兩個標準差者 一班四年級學生23人 一班六年級學生30人

實驗材料 材料的形式為前後兩句的極短文 由大學生事先評定後再挑選與修改 短文共54篇：實驗短文36篇，filler18篇 操弄因果連貫性 高因果連貫：米奇和高菲穿著外套跳繩，他流很多汗 低因果連貫：米奇和高菲剪頭髮動不停，他喝到髒水 操弄指稱詞關連性 有指稱詞連貫：米奇和高菲穿著外套跳繩，他流很多汗 無指稱詞連貫：米奇和高菲穿著外套跳繩，流很多汗 短文主角均為兒童熟悉的八個卡通人物進行變換

實驗材料實例 高因果連貫＋有指稱詞連貫 高因果連貫＋無指稱詞連貫 低因果連貫＋有指稱詞連貫 低因果連貫＋無指稱詞連貫 米奇和高菲穿著外套跳繩，他流很多汗。 高因果連貫＋無指稱詞連貫 米奇和高菲穿著外套跳繩，流很多汗。 低因果連貫＋有指稱詞連貫 米奇和高菲剪頭髮動不停，他喝到髒水。 低因果連貫＋無指稱詞連貫 米奇和高菲剪頭髮動不停，喝到髒水

實驗程序 指導語 練習階段，熟悉程序與按鍵的操作 正式實驗 所需時間約為40分鐘 指導語 練習題 正式題（分成兩階段）

實驗程序 實驗材料用Experiment Builder呈現 每一短文分成三個畫面： 米奇和高菲穿著外套跳繩， 他流很多汗。 米奇穿外套作運動嗎？

實驗程序 受試者依自己的速度按空白鍵閱讀短文 受試者在第三個畫面必須回答問題 紀錄受試者反應正確率及每句閱讀時間 正確按「○」的按鍵 錯誤按「ㄨ」的按鍵 紀錄受試者反應正確率及每句閱讀時間

實驗設計 2(年級) x 2(因果連貫) x 2(代名詞連貫) 年級為受試者間設計 因果連貫與代名詞連貫為受試者內設計

結果：答對率 高因果 有代名詞 無代名詞 低因果 四年級 0.88(.12) 0.90(.08) 0.82(.11) 0.76(.19) 六年級 0.94(.08) 0.93(.09) 0.86(.12) 0.94(.07)

結果：答對率 三因子交互作用不顯著 六年級答對率高於四年級(F=22.601, p=.000 < .05) 六年級的部分

結果：答對率 四年級的部分 因果關係與代名詞交互作用沒有顯著 因果關係的主要效果達顯著(F=13.579, p=.000<.05) 高因果優於低因果 代名詞的主要效果沒有顯著 (F=.277, p=.600>.05)

結果：答對率 年級較大正確率較高 四年級完全受因果連貫強度影響 六年級也受因果連貫性影響 閱讀高因果連貫短文，不受代名詞有無之影響 閱讀低因果連貫短文，沒有代名詞之短文較佳

結果：第二句的閱讀時間 （單位：ms) 高因果 有代名詞 無代名詞 低因果 四年級 六年級 1250.54 (382.57) 1508.43 (568.58) 1739.73 (740.69) 2002.46 (923.61) 六年級 1016.93 (282.56) 1205.58 (377.34) 1464.08 (552.78) 1705.45 (430.44)

結果：第二句的閱讀時間 三因子交互作用不顯著 六年級閱讀時間短於四年級 （F=10.905, p= .001< .05） 六年級的部分 因果關係與代名詞沒有交互作用 因果關係的主要效果達顯著 (F=37.731, p=.000<.05) 高因果連貫閱讀時間較短 代名詞的主要效果達顯著(F=7.780, p=.006<.05) 有代名詞閱讀時間較短

結果：第二句的閱讀時間 四年級的部分 因果關係與代名詞交互作用沒有顯著 因果關係的主要效果達顯著 (F=11.882, p=.001<.05) 高因果連貫閱讀時間較短 代名詞的主要效果沒有顯著 (F=3.331, p=.071>.05)

結果：第二句的閱讀時間 年級大閱讀時間較短 具有較高因果連貫性文章，對四六年級均較易閱讀 具有明確代名詞連貫性文章，有助於六年級閱讀（四年級尚未定）

進行中 未來三年更多發展性資料涵蓋 二到六年級學生陸續收集中 進一步探討推論能力的大腦機制

Vocabulary instruction & comprehension Word generation program (Snow & Lawrence, 2009) Direct teach cross-subject vocabularies to 7t graders across different contexts

Science understanding Science learning is affected by language Good science learning require causal understanding Deep Knowing why A collaborated project on early science learning (Stein, 2010) Based on theories of coherence

Knowing what v.s. knowing why state changes of water Knowing what is only a first step Knowing why is true understanding How to make young children know why? Molecular concepts are crucial Can we teach molecular concepts to 4th graders? Knowing science also involve quantifications Units of measurement become important

Comparison of U.S. and Taiwan 4th Grade Measurement Scores at Pretest – Fall All Groups This is the pretest data from the 2010 Fall CESR study. 76

Comparison of U.S. and Taiwan 4th Grade Measurement Scores – Spring

Comparison of Fall and Spring 4th Grade Taiwan Measurement Scores with Typical Instruction

Comparison of Fall and Spring 4th Grade U. S Comparison of Fall and Spring 4th Grade U.S. Measurement Scores with Typical Instruction

U. S. and Taiwan Comparison: U. S U. S. and Taiwan Comparison: U.S. Instruction in Measurement versus Taiwan Natural Teaching

Comparison of U.S. Experimental and Control Conditions from Pre to Post Test

Comparison of Experimental and Control Conditions – Fall 4th Grade Taiwan Measurement Scores

Early Learning in the Physical Sciences What is to be taught? How can we show that core concepts (those that are chosen for instruction) are linked to extremely important issues and problems. Global warming and climate change? Nuclear reactors: the control of radioactive substances? Earth sciences: nature of earthquakes, volcanic eruption? The mining of coal, uranium, oil? Sustainable and clean energy sources?

感謝 國科會99-2420-H-194-003- 宗卓章基金會 中正大學前瞻計畫

中正大學認知與學習實驗室成員 連啟舜博士 陳明蕾博士 吳裕聖博士 蔡瓊賢 黃秋華 蔡幸錦 郭淑雅 國立嘉大附小師生 嘉義市博愛國小 嘉義市興嘉國小 嘉義縣三興國小 台南市公園國小 台南市裕文國小