Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
Automated Topical Component Extraction Using Neural NetworkAttention Scores from Source-based Essay Scoring
Haoran ZhangDepartment of Computer Science
University of PittsburghPittsburgh, PA 15260
Diane LitmanDepartment of Computer Science & LRDC
University of PittsburghPittsburgh, PA 15260
Abstract
While automated essay scoring (AES) can re-liably grade essays at scale, automated writingevaluation (AWE) additionally provides forma-tive feedback to guide essay revision. How-ever, a neural AES typically does not provideuseful feature representations for supportingAWE. This paper presents a method for link-ing AWE and neural AES, by extracting Top-ical Components (TCs) representing evidencefrom a source text using the intermediate out-put of attention layers. We evaluate perfor-mance using a feature-based AES requiringTCs. Results show that performance is compa-rable whether using automatically or manuallyconstructed TCs for 1) representing essays asrubric-based features, 2) grading essays.
1 Introduction
Automated essay scoring (AES) systems reliablygrade essays at scale, while automated writing eval-uation (AWE) systems additionally provide forma-tive feedback to guide revision. Although neuralnetworks currently generate state-of-the-art AESresults (Alikaniotis et al., 2016; Taghipour and Ng,2016; Dong et al., 2017; Farag et al., 2018; Jin et al.,2018; Li et al., 2018; Tay et al., 2018; Zhang andLitman, 2018), non-neural AES create feature rep-resentations more easily useable by AWE (Roscoeet al., 2014; Foltz and Rosenstein, 2015; Crossleyand McNamara, 2016; Woods et al., 2017; Madnaniet al., 2018; Zhang et al., 2019). We believe thatneural AES can also provide useful information forcreating feature representations, e.g., by exploitinginformation in the intermediate layers.
Our work focuses on a particular source-basedessay writing task called the response-to-text as-sessment (RTA) (Correnti et al., 2013). Recently,an RTA AWE system (Zhang et al., 2019) was builtby extracting rubric-based features related to the
use of Topical Components (TCs) in an essay. How-ever, manual expert effort was first required to cre-ate the TCs. For each source, the TCs consist ofa comprehensive list of topics related to evidencewhich include: 1) important words indicating theset of evidence topics in the source, and 2) phrasesrepresenting specific examples for each topic thatstudents need to find and use in their essays.
To eliminate this expert effort, we propose amethod for using the interpretable output of theattention layers of a neural AES for source-basedessay writing, with the goal of extracting TCs. Weevaluate this method by using the extracted TCsto support feature-based AES for two RTA sourcetexts. Our results show that 1) the feature-basedAES with TCs manually created by humans ismatched by our neural method for generating TCs ,and 2) the values of the rubric-based essay featuresbased on automatic TCs are highly correlated withhuman Evidence scores.
2 Related Work
Three recent AWE systems have used non-neuralAES to provide rubric-specific feedback. Woodset al. (2017) developed an influence estimation pro-cess that used a logistic regresion AES to identifysentences needing feedback. Shibani et al. (2019)presented a web-based tool that provides formativefeedback on rhetorical moves in writing. Zhanget al. (2019) used features created for a randomforest AES to select feedback messages, althoughhuman effort was first needed to create TCs froma source text. We automatically extract TCs usingneural AES, thereby eliminating this expert effort.
Others have also proposed methods for pre-processing source information external to an es-say. Content importance models for AES predictthe parts of a source text that students should in-clude when writing a summary (Klebanov et al.,
arX
iv:2
008.
0180
9v1
[cs
.CL
] 4
Aug
202
0
Source Excerpt: Today, Yala Sub-District Hospital has medicine, free of charge, for all of the most common diseases. Wateris connected to the hospital, which also has a generator for electricity. Bed nets are used in every sleeping site in Sauri...Essay Prompt: The author provided one specific example of how the quality of life can be improved by the Millennium VillagesProject in Sauri, Kenya. Based on the article, did the author provide a convincing argument that winning the fight against poverty isachievable in our lifetime? Explain why or why not with 3-4 examples from the text to support your answer.Essay: In my opinion I think that they will achieve it in lifetime. During the years threw 2004 and 2008 they made progress.People didnt have the money to buy the stuff in 2004. The hospital was packed with patients and they didnt have alot of treatmentin 2004. In 2008 it changed the hospital had medicine, free of charge, and for all the common dieases. Water was connectedto the hospital and has a generator for electricity. Everybody has net in their site. The hunger crisis has been addressed withfertilizer and seeds, as well as the tools needed to maintain the food. The school has no fees and they serve lunch. To me thatssounds like it is going achieve it in the lifetime.
Table 1: A source excerpt for the RTAMV P prompt and an essay with score of 3.
Prompt RTAMV P RTASpace
Score 1 852 538(29%) (26%)
Score 2 1197 789(40%) (38%)
Score 3 616 512(21%) (25%)
Score 4 305 237(10%) (11%)
Total 2970 2076
Table 2: The Evidence score distribution of RTA.
2014). Methods for extracting important keywordsor keyphrases also exist, both supervised (unlikeour approach) (Meng et al., 2017; Mahata et al.,2018; Florescu and Jin, 2018) and unsupervised(Florescu and Caragea, 2017). Rahimi and Litman(2016) developed a TC extraction LDA model (Bleiet al., 2003). While the LDA model considers allwords equally, our model takes essay scores intoaccount by using attention to represent word impor-tance. Both the unsupervised keyword and LDAmodels will serve as baselines in our experiments.
In the computer vision area, attention croppedimages have been used for further image classifi-cation or object detection (Cao et al., 2015; Yuxinet al., 2018; Ebrahimpour et al., 2019). In the NLParea, Lei et al. (2016) proposed to use a genera-tor to find candidate rationale and these are passedthrough the encoder for prediction. Our work issimilar in spirit to this type of work.
3 RTA Corpus and Prior AES Systems
The essays in our corpus were written by studentsin grades 4 to 8 in response to two RTA sourcetexts (Correnti et al., 2013): RTAMV P (2970 es-says) and RTASpace (2076 essays). Table 1 showsan excerpt from RTAMV P , the associated essaywriting prompt, and a student essay. The boldingin the source indicates evidence examples that ex-
perts manually labeled as important for studentsto discuss (i.e., TC phrases). Evidence usage ineach essay was manually scored on a scale of 1to 4 (low to high). The distribution of Evidencescores is shown in Table 2. The essay in Table 1received a score of 3, with the bolding indicatingphrases semantically related to the TCs from thesource text.
To date, two approaches to AES have been pro-posed for the RTA: AESrubric and AESneural. Tosupport the needs of AWE, AESrubric (Zhang andLitman, 2017) used a traditional supervised learn-ing framework where rubric-motivated featureswere extracted from every essay before model train-ing - Number of Pieces of Evidence (NPE) 1, Con-centration (CON), Specificity (SPC) 2, Word Count(WOC). The two aspects of TCs introduced in Sec-tion 1 (topic words, specific example phrases) wereused during feature extraction.
Motivated by improving stand-alone AES perfor-mance (i.e., when an interpretable model was notneeded for subsequent AWE), Zhang and Litman(2018) developed AESneural, a hierarchical neuralmodel with the co-attention mechanism in the sen-tence level to capture the relationship between theessay and the source. Neither feature engineeringnor TC creation were needed before training.
4 Attention-Based TC Extraction: TCattn
In this section we propose a method for extract-ing TCs based on the AESneural attention leveloutputs. Since the self-attention and co-attentionmechanisms were designed to capture sentenceand phrase importance, we hypothesize that theattention scores can help determine if a sentence or
1An integer feature based on the list of topic words foreach topic.
2A vector of integer values indicating the number of spe-cific example phrases (semantically) mentioned in the essayper topic.
No. Sentences attnsent attnphrase
1 People didn’t have the money tobuy the stuff in 2004.
0.00420 0.23372
2 The hunger crisis has been addressedwith fertilizer and seeds, as well asthe tools needed to maintain the food.
0.08709 0.62848
3 The school has no fees and theyserve lunch.
0.10686 0.63369
Table 3: Example attention scores of essay sentences.
phrase has important source-related information.To provide intuition, Table 3 shows examples
sentences from the student essay in Table 1. Boldedare phrases with the highest self-attention scorewithin the sentence. Italics are specific examplephrases that refer to the manually constructed TCsfor the source. Attnsent is the text to essay atten-tion score that measures which essay sentenceshave the closest meaning to a source sentence.Attnphrase is the self-attention score of the boldedphrase that measures phrase importance. A sen-tence with a high attention score tends to include atleast one specific example phrase, and vice versa.The phrase with the highest attention score tendsto include at least one specific example phrase ifthe sentence has a high attention score.
Based on these observations, we first extract theoutput of two layers from the neural network: 1)the attnsent of each sentence, and 2) the output ofthe convolutional layer as the representation of thephrase with the highest attnphrase in each sentence(denoted by cnnphrase). We also extract the plaintext of the phrase with the highest attnphrase ineach sentence (denoted by textphrase). Then, ourTCattn method uses the extracted information in3 main steps: 1) filtering out textphrase from sen-tences with low attnsent, 2) clustering all remain-ing textphrase based on cnnphrase, and 3) generat-ing TCs from clusters.
The first filtering step keeps all textphrase wherethe original sentences have attnsent higher thana threshold. The intuition is that lower attnsent
indicates less source-related information.The second step clusters these textphrase based
on their corresponding representations cnnphrase.We use k-medoids to cluster textphrase into Mclusters, where M is the number of topics in thesource text. Then, for textphrase in each topiccluster, we use k-medoids to cluster them into Nclusters, where N is the number of the specificexample phrases we want to extract from each topic.The outputs of this step are M ∗N clusters.
The third step uses the topic and example clus-
Layer Parameter Name ValueEmbedding Embedding dimension 50Word-CNN Kernel size 5
Number of filters 100Sent-LSTM Hidden units 100Modeling Hidden units 100Dropout Dropout rate 0.5Others Epochs 100
Batch size 100Initial learning rate 0.001
Momentum 0.9
Table 4: Hyper-parameters for neural training.
Figure 1: An overview of four TC extraction systems.
tering to extract TCs. As noted earlier, TCs in-clude two parts: topic words, and specific examplephrases. Since our method is data-driven and stu-dents introduce their vocabulary into the corpus,essay text is noisy. To make the TC output cleaner,we filter out words that are not in the source text.To obtain topic words, we combine all textphrasefrom each topic cluster to calculate the word fre-quency per topic. To make topics unique, we assigneach word to the topic cluster in which it has thehighest normalized word frequency. We then in-clude the top Ktopic words based on their frequencyin each topic cluster. To obtain example phrases,we combine all textphrase from each example clus-ter to calculate the word frequency per example,then include the top Kexample words based on theirfrequency in each example cluster.
5 Experimental Setup and Results
Figure 1 shows an overview of four TC extractionmethods to be evaluated. TCmanual (upper bound)uses a human expert to extract TCs from a sourcetext. TCattn is our proposed method and automat-ically extracts TCs using both a source text andstudent essays. TClda (Rahimi and Litman, 2016)(baseline) builds on LDA to extract TCs from stu-dent essays only, while TCpr (baseline) builds onPositionRank (Florescu and Caragea, 2017) to in-stead extract TCs from only the source text.
Since PositionRank is not designed for TC ex-
Prompt Component Parameter TClda TCpr TCattn
RTAMV P
Topic WordsNumber of Topics 9 19 16Number of Words 30 20 25
Example PhrasesNumber of Topics 20 1 18Number of Phrases 15 20 15
RTASpace
Topic WordsNumber of Topics 15 20 10Number of Words 10 10 20
Example PhrasesNumber of Topics 10 1 9Number of Phrases 20 50 20
Table 5: Parameters for different models.
traction, we needed to further process its output tocreate TCpr. To extract topic words, we extractall keywords from the output. Next, we map eachword to a higher dimension with word embedding.Lastly, we cluster all keywords using k-medoidsinto PRtopic topics. To extract example phrases,we put them into only one topic and remove allredundant example phrases if they are subsets ofother example phrases.
We configure experiments to test two hypotheses:H1) the AESrubric model for scoring Evidence(Zhang and Litman, 2017) will perform compara-bly when extracting features using either TCattn
or TCmanual, and will perform worse when us-ing TClda or TCpr; H2) the correlation betweenthe human Evidence score and the feature values(NPE and sum of SPC features)3 will be compa-rable when extracted using TCattn and TCmanual,and will be stronger than when using TClda andTCpr. The experiment for H1 tests the impact ofusing our proposed TC extraction method on thedownstream AESrubric task, while the H2 experi-ment examines the impact on the essay representa-tion itself.
Following Zhang and Litman (2017), we stratifyessay corpora: 40% for training word embeddingsand extracting TCs, 20% for selecting the best em-bedding and parameters, and 40% for testing. Weuse the hyper-parameters from Zhang and Litman(2018) for neural training as shown in Table 4. Ta-ble 5 shows all other parameters selected using thedevelopment set.
Results for H1. H1 is supported by the results inTable 6, which compares the Quadratic WeightedKappa (QWK) between human and AESrubric Ev-idence scores (values 1-4) when AESrubric usesTCmanual versus each of the automatic methods.TCattn always yields better performance, and evensignificantly better than TCmanual.
Results for H2. The results in Table 7 supportH2. TCattn outperforms the two automated base-
3These features are extracted based on TCs.
Prompt TCmanual (1) TClda (2) TCpr (3) TCattn (4)RTAMV P 0.643 (2,3) 0.614 (3) 0.525 0.648 (1,2,3)RTASpace 0.609 (3) 0.615 (3) 0.559 0.622 (1,3)
Table 6: The performance (QWK) of AESrubric us-ing different TC extraction methods for feature cre-ation. The numbers in the parentheses show the modelnumbers over which the current model performs signif-icantly better (p < 0.05). The best results betweenautomated methods in each row are in bold.
Prompt Feature TCmanual TClda TCpr TCattn
RTAMV PNPE 0.542 0.482 0.587 0.639SPC (sum) 0.689 0.585 0.365 0.679
RTASpaceNPE 0.484 0.513 0.494 0.625SPC (sum) 0.601 0.574 0.533 0.598
Table 7: Pearson’s r comparing feature values com-puted using each TC extraction method with human(gold-standard) Evidence essay scores. All correlationvalues are significant (p ≤ 0.05). The best results be-tween automated methods in each row are in bold.
lines, and for NPE even yields stronger correlationsthan the manual TC method.
Qualitative Analysis. The manually-createdtopic words for RTAMV P represent 4 topics,which are “hospital”, “malaria”, “farming” and“school”4. Although Table 5 shows that the au-tomated list has more topics for topic words andmight have broken one topic into separate topics,a good automated list should have more topics re-lated to the 4 topics above. We manually assign atopic for each of the topic words from the differentautomated methods. TClda has 4 related topics outof 9 (44.44%), TCpr has 6 related topics out of 19(31.58%), and TCattn has 10 related topics out of16 (62.50%). Obviously, TCattn preserves morerelated topics than our baselines.
Moving to the second aspect of TCs (specificexample phrases), Table 8 shows the first 10 spe-cific example phrases for a manually-created cat-egory that introduces the changes made by theMVP project5. This category is a mixture of dif-ferent topics because it talks about the “hospital”,“malaria”, “school”, and “farming” at the same time.TCattn has overlap with TCmanual on differenttopics. However, TClda mainly talks about “hospi-tal”, because the nature of the LDA model doesn’tallow mixing specific example phrases about dif-ferent topics in one category. Unfortunately, TCpr
4All Topic Words generated by different models can befound in the Appendix A.1.
5All Specific Example Phrases generated by different mod-els can be found in the Appendix A.2.
TCmanual TClda TCpr TCattn
progress just four years running water electricity brighter future hannah electricity running water irrigation setmedicine most common diseases water connected hospital generator electricity millennium villages project poor showed treatment school supplies
water connected hospital patients afford unpaved dirt road farmers could crops afford bedhospital generator electricity rooms packed patients probably bar sauri primary school electricity hospital
bed nets used every sleeping site share beds future hannah better fertilizer medicine enough alsohunger crisis addressed fertilizer seeds recieve treatment sauri primary school rooms packed patients
tools needed maintain food supply doctor clinical officer running hospital villages project food fertilizer crops get supplyno school fees doctors clinical millennium development goals five net costs 5
school attendance rate way up water fertilizer knowledge village leaders nets net bed freekids go school now receive treatment dirt road running water supplies schools almost
... ... ... ...
Table 8: Specific example phrases for the RTAMV P progress topic.
does not include any overlapped specific phrase inthe first 10 items; they all refer to some generalexample phrases from the beginning of the sourcearticle. Although there are some related specificexample phrases in the full list, they are mainlyabout school. This is because the PositionRankalgorithm tends to assign higher scores to wordsthat appear early in the text.
6 Conclusion and Future Work
This paper proposes TCattn, a method for using theattention scores in a neural AES model to automat-ically extract the Topical Components of a sourcetext. Evaluations show the potential of TCattn
for eliminating expert effort without degradingAESrubric performance or the feature represen-tations themselves. TCattn outperforms baselinesand generates comparable or even better resultsthan a manual approach.
Although TCattn outperforms all baselines andrequires no human effort on TC extraction, annota-tion of essay evidence scores is still needed. Thisleads to an interesting future investigation direc-tion, which is training the AESneural using thegold standard that can be extracted automatically.
One of our next steps is to investigate the im-pact of TC extraction methods on a correspondingAWE system (Zhang et al., 2019), which uses thefeature values produced by AESrubric to generateformative feedback to guide essay revision.
Currently, the TClda are trained on student es-says, while the TCpr only works on the sourcearticle. However, TCattn uses both student essaysand the source article for TC generation. It mightbe hard to say that the superior performance ofTCattn is due to the neural architecture and atten-tion scores rather than the richer training resources.Therefore, a comparison between TCattn and amodel that uses both student essays and the sourcearticle is needed.
Acknowledgments
We would like to show our appreciation to everymember of the RTA group for sharing their pearlsof wisdom with us. We are also immensely gratefulto all members of the PETAL group and reviewersfor their comments on an earlier version of thepaper.
The research reported here was supported, inwhole or in part, by the Institute of EducationSciences, U.S. Department of Education, throughGrant R305A160245 to the University of Pitts-burgh. The opinions expressed are those of theauthors and do not represent the views of the Insti-tute or the U.S. Department of Education.
ReferencesDimitrios Alikaniotis, Helen Yannakoudakis, and
Marek Rei. 2016. Automatic text scoring using neu-ral networks. In Proceedings of the 54th AnnualMeeting of the Association for Computational Lin-guistics (Volume 1: Long Papers), volume 1, pages715–725.
David M Blei, Andrew Y Ng, and Michael I Jordan.2003. Latent dirichlet allocation. Journal of ma-chine Learning research, 3(Jan):993–1022.
Chunshui Cao, Xianming Liu, Yi Yang, Yinan Yu,Jiang Wang, Zilei Wang, Yongzhen Huang, LiangWang, Chang Huang, Wei Xu, et al. 2015. Lookand think twice: Capturing top-down visual atten-tion with feedback convolutional neural networks.In Proceedings of the IEEE International Confer-ence on Computer Vision, pages 2956–2964.
Richard Correnti, Lindsay Clare Matsumura, LauraHamilton, and Elaine Wang. 2013. Assessing stu-dents’ skills at writing analytically in response totexts. The Elementary School Journal, 114(2):142–177.
Scott A Crossley and Danielle S McNamara. 2016.Adaptive educational technologies for literacy in-struction. Routledge.
Fei Dong, Yue Zhang, and Jie Yang. 2017. Attention-based recurrent convolutional neural network for au-tomatic essay scoring. In Proceedings of the 21stConference on Computational Natural LanguageLearning (CoNLL 2017), pages 153–162.
Mohammad K Ebrahimpour, Jiayun Li, Yen-Yun Yu,Jackson Reesee, Azadeh Moghtaderi, Ming-HsuanYang, and David C Noelle. 2019. Ventral-dorsal neu-ral networks: Object detection via selective attention.In 2019 IEEE Winter Conference on Applications ofComputer Vision (WACV), pages 986–994. IEEE.
Youmna Farag, Helen Yannakoudakis, and Ted Briscoe.2018. Neural automated essay scoring and coher-ence modeling for adversarially crafted input. InProceedings of the 2018 Conference of the NorthAmerican Chapter of the Association for Computa-tional Linguistics: Human Language Technologies,Volume 1 (Long Papers), pages 263–271.
Corina Florescu and Cornelia Caragea. 2017. Position-rank: An unsupervised approach to keyphrase ex-traction from scholarly documents. In Proceedingsof the 55th Annual Meeting of the Association forComputational Linguistics (Volume 1: Long Papers),pages 1105–1115.
Corina Florescu and Wei Jin. 2018. Learning featurerepresentations for keyphrase extraction. In Thirty-Second AAAI Conference on Artificial Intelligence.
Peter W Foltz and Mark Rosenstein. 2015. Analysisof a large-scale formative writing assessment systemwith automated feedback. In Proceedings of the Sec-ond (2015) ACM Conference on Learning@ Scale,pages 339–342. ACM.
Cancan Jin, Ben He, Kai Hui, and Le Sun. 2018.Tdnn: a two-stage deep neural network for prompt-independent automated essay scoring. In Proceed-ings of the 56th Annual Meeting of the Associationfor Computational Linguistics (Volume 1: Long Pa-pers), pages 1088–1097.
Beata Beigman Klebanov, Nitin Madnani, Jill Burstein,and Swapna Somasundaran. 2014. Content impor-tance models for scoring writing from sources. InProceedings of the 52nd Annual Meeting of the As-sociation for Computational Linguistics (Volume 2:Short Papers), pages 247–252.
Tao Lei, Regina Barzilay, and Tommi Jaakkola. 2016.Rationalizing neural predictions. In Proceedings ofthe 2016 Conference on Empirical Methods in Natu-ral Language Processing, pages 107–117.
Xia Li, Minping Chen, Jianyun Nie, Zhenxing Liu,Ziheng Feng, and Yingdan Cai. 2018. Coherence-based automated essay scoring using self-attention.In Chinese Computational Linguistics and NaturalLanguage Processing Based on Naturally AnnotatedBig Data, pages 386–397. Springer.
Nitin Madnani, Jill Burstein, Norbert Elliot,Beata Beigman Klebanov, Diane Napolitano,Slava Andreyev, and Maxwell Schwartz. 2018.Writing mentor: Self-regulated writing feedback forstruggling writers. In Proceedings of the 27th Inter-national Conference on Computational Linguistics:System Demonstrations, pages 113–117.
Debanjan Mahata, John Kuriakose, Rajiv Ratn Shah,and Roger Zimmermann. 2018. Key2vec: Auto-matic ranked keyphrase extraction from scientific ar-ticles using phrase embeddings. In Proceedings ofthe 2018 Conference of the North American Chap-ter of the Association for Computational Linguistics:Human Language Technologies, Volume 2 (Short Pa-pers), pages 634–639.
Rui Meng, Sanqiang Zhao, Shuguang Han, DaqingHe, Peter Brusilovsky, and Yu Chi. 2017. Deepkeyphrase generation. In Proceedings of the 55thAnnual Meeting of the Association for Computa-tional Linguistics (Volume 1: Long Papers), pages582–592.
Zahra Rahimi and Diane Litman. 2016. Automati-cally extracting topical components for a response-to-text writing assessment. In Proceedings of the11th Workshop on Innovative Use of NLP for Build-ing Educational Applications, pages 277–282.
Rod D Roscoe, Laura K Allen, Jennifer L Weston,Scott A Crossley, and Danielle S McNamara. 2014.The writing pal intelligent tutoring system: Usabilitytesting and development. Computers and Composi-tion, 34:39–59.
Antonette Shibani, Simon Knight, and Simon Bucking-ham Shum. 2019. Contextualizable learning ana-lytics design: A generic model and writing analyt-ics evaluations. In Proceedings of the 9th Interna-tional Conference on Learning Analytics & Knowl-edge, pages 210–219. ACM.
Kaveh Taghipour and Hwee Tou Ng. 2016. A neuralapproach to automated essay scoring. In Proceed-ings of the 2016 Conference on Empirical Methodsin Natural Language Processing, pages 1882–1891.
Yi Tay, Minh C Phan, Luu Anh Tuan, and Siu Che-ung Hui. 2018. Skipflow: incorporating neural co-herence features for end-to-end automatic text scor-ing. In Thirty-Second AAAI Conference on ArtificialIntelligence.
Bronwyn Woods, David Adamson, Shayne Miel, andElijah Mayfield. 2017. Formative essay feedbackusing predictive scoring models. In Proceedings ofthe 23rd ACM SIGKDD International Conferenceon Knowledge Discovery and Data Mining, pages2071–2080. ACM.
Peng Yuxin, He Xiangteng, and Zhao Junjie. 2018.Object-part attention model for fine-grained imageclassification. IEEE transactions on image process-ing: a publication of the IEEE Signal Processing So-ciety, 27(3):1487–1500.
Haoran Zhang and Diane Litman. 2017. Word embed-ding for response-to-text assessment of evidence. InProceedings of ACL 2017, Student Research Work-shop, pages 75–81.
Haoran Zhang and Diane Litman. 2018. Co-attentionbased neural network for source-dependent essayscoring. In Proceedings of the Thirteenth Workshopon Innovative Use of NLP for Building EducationalApplications, pages 399–409.
Haoran Zhang, Ahmed Magooda, Diane Litman,Richard Correnti, Elaine Wang, LC Matsmura,Emily Howe, and Rafael Quintana. 2019. erevise:Using natural language processing to provide forma-tive feedback on text evidence usage in student writ-ing. In Proceedings of the AAAI Conference on Arti-ficial Intelligence, volume 33, pages 9619–9625.
A Appendices
A.1 Topic Words ResultsTable 9 shows all topic words for the RTAMV P
from TCmanual. Table 10 shows all topic wordsfor the RTAMV P from TClda. Table 11 showsall topic words for the RTAMV P from TCpr. Ta-ble 12 shows all topic words for the RTAMV P
from TCattn.
A.2 Specific Example Phrases ResultsTable 13 shows all specific example phrases forthe RTAMV P from TCmanual. Table 14 shows allspecific example phrases for the RTAMV P fromTClda. Table 15 shows all specific example phrasesfor the RTAMV P from TCpr. Table 16 shows allspecific example phrases for the RTAMV P fromTCattn.
Topic 1 Topic 2 Topic 3 Topic 4care bed farmer school
health net fertilizer supplieshospital malaria irrigation fee
treatment infect dying studentdoctor bednet crop midday
electricity mosquito seed mealdisease bug water lunchwater sleeping harvest supplysick die hungry book
medicine cheap feed papergenerator infect food pencil
no biting energydie freekid childrenbed kid
patient goclinical attendofficer
running
Table 9: Topic words of TCmanual.
Topi
c1
Topi
c2
Topi
c3
Topi
c4
Topi
c5
Topi
c6
Topi
c7
Topi
c8
Topi
c9
help
keny
apo
vert
yfo
odm
oney
scho
olpe
ople
hosp
ital
year
spo
orlik
eth
ink
fert
ilize
rne
edki
dssa
uri
med
icin
eaf
rica
wor
ldbe
tter
auth
orcr
ops
nets
supp
lies
mal
aria
hosp
itals
proj
ect
good
know
lifet
ime
wat
erth
ing
child
ren
sick
wat
ervi
llage
sth
ings
life
artic
lefa
rmer
saf
ford
scho
ols
2008
free
saur
itim
ehe
lppo
ssib
lene
eded
dona
telu
nch
dise
ase
elec
tric
ityvi
llage
wor
kth
ink
conv
ince
dgr
owri
ght
educ
atio
n20
04di
seas
eshe
lped
hard
saur
ifig
htdy
ing
dolla
raf
ford
nets
med
icin
esch
ange
goin
gliv
epr
over
typr
oble
mtr
eatm
ent
ener
gym
osqu
itoes
doct
ors
lives
alot
clot
hes
said
fam
ilysu
rviv
ele
arn
getti
ng20
08go
als
reas
onst
ates
achi
evab
lefa
mili
esne
eds
stud
ents
says
gave
impr
oved
happ
enpl
ace
time
stop
stuf
fw
ent
year
sdo
ctor
2015
help
ing
heal
thco
nvin
cela
ckpe
rson
adul
tspr
ogre
ssex
ampl
eshe
lpgo
alim
port
ant
belie
vehu
nger
caus
efe
esdi
ed20
04ch
ange
dbe
lieve
feel
hann
ahto
ols
patie
nts
pare
nts
text
shap
eye
arpr
oble
ms
happ
ysh
ows
seed
spr
ovid
e20
04aw
aycu
rech
ange
sco
untr
ies
tell
reas
ons
plan
tsco
stlu
nche
sm
osqu
itos
runn
ing
star
ted
diff
eren
ceca
reco
nvin
cing
fert
ilize
rsbe
dsbo
oks
prev
ent
trea
tgr
eat
plac
essh
oes
fight
ing
farm
ing
mea
nsho
me
trea
ted
supp
ort
mill
enni
umch
ange
stor
yw
rote
able
dont
wan
ted
diei
ngco
mm
onpr
ogre
sslit
tleam
eric
ast
ory
solv
eddo
llars
chor
essa
idbe
dsca
me
impr
ove
way
sag
ree
supp
lym
edic
alm
eal
com
epa
tient
sgi
rlco
untr
yw
ants
sayi
ngir
riga
tion
jobs
woo
dni
ght
said
2025
achi
eve
mak
esop
inio
nw
ont
ever
yday
mat
eria
lsbi
tege
nera
tor
plac
eho
pecl
othi
ngw
inni
ngaf
ford
gone
lear
ning
deat
hcl
ean
prog
ram
help
sco
mm
unity
sach
shu
ngry
doct
ors
able
slee
pel
ectr
icty
tells
ever
ybod
yec
onom
ypr
ogre
sspl
ant
lots
supl
ies
impo
veri
shed
givi
ngsm
all
star
thi
stor
yco
nclu
sion
look
sick
ness
mea
lsliv
ing
drin
km
illen
ium
easy
para
grap
hsa
ysfa
rms
live
pape
ram
azin
gcu
res
read
mak
ing
that
sfu
ture
feed
fact
atte
ndan
ceea
sily
evid
ence
happ
ened
Tabl
e10
:Top
icw
ords
ofTC
lda.
Topi
c1
Topi
c2
Topi
c3
Topi
c4
Topi
c5
Topi
c6
Topi
c7
Topi
c8
Topi
c9
Topi
c10
Topi
c11
Topi
c12
Topi
c13
Topi
c14
Topi
c15
Topi
c16
Topi
c17
Topi
c18
Topi
c19
irri
gatio
nro
addi
seas
esad
ults
fight
deve
lopm
ent
joyf
ulpe
ople
mid
day
villa
gem
illen
nium
back
spl
enty
doct
orth
ing
pape
ren
dw
ork
slee
ping
fert
ilize
rbr
ight
erm
edic
ine
lifet
ime
villa
ges
dirt
kids
scho
olw
omen
acce
ssho
spita
lsu
pplie
sw
orld
bed
farm
ers
futu
rem
alar
iapr
ojec
tju
mp
fees
grou
ndca
resh
ape
chor
esne
tcr
ops
hann
ahdi
seas
ego
als
bar
stud
ents
bana
nas
med
icin
espa
tient
sbo
oks
nets
plan
tca
rm
osqu
itoes
plan
mus
icm
eal
clot
hsc
hool
str
eatm
ent
penc
ilssi
tese
eds
saur
ich
arge
econ
omy
sing
ing
ener
gym
othe
rsto
day
offic
erou
tcom
em
arke
tqu
ality
ever
yone
lunc
hfe
etsu
pply
wat
erla
ckye
arsu
ppor
ters
danc
ing
clot
hing
area
sel
ectr
icity
tool
stim
ehe
lpda
yki
ndge
nera
tor
plac
ehe
alth
room
sye
ars
advi
cefa
mily
pove
rty
item
slif
eta
rget
sco
mm
uniti
esde
ath
lead
ers
nigh
tgl
imps
eco
sts
afri
cadi
ech
emic
als
know
ledg
eso
lutio
nsfo
odm
illio
nspa
rent
s
Tabl
e11
:Top
icw
ords
ofTC
pr.
Topi
c1
Topi
c2
Topi
c3
Topi
c4
Topi
c5
Topi
c6
Topi
c7
Topi
c8
Topi
c9
Topi
c10
Topi
c11
Topi
c12
Topi
c13
Topi
c14
Topi
c15
Topi
c16
pove
rty
way
year
slu
nch
goal
sel
ectr
icity
supp
lies
affo
rdm
any
free
scho
olho
spita
lbe
dpr
ojec
tsu
pply
fert
ilize
rfig
htw
ould
four
serv
espr
oble
ms
wat
erfo
odlif
etim
epe
ople
med
icin
esc
hool
s20
04ne
tsw
orld
mai
ntai
nse
eds
win
ning
rate
villa
ges
pare
nts
day
gene
rato
rne
tco
uld
keny
acr
ops
fees
dise
ase
used
mill
enni
umdi
seas
esad
dres
sed
atte
ndan
ce80
atte
ndcl
oth
also
room
sac
hiev
able
saur
ich
arge
stud
ents
yala
ever
yvi
llage
hung
erir
riga
tion
help
prog
ress
pass
edth
ree
runn
ing
pack
edto
geth
erpe
ncils
farm
ers
slee
ping
acro
ssliv
esne
cess
ary
kids
last
mad
een
ergy
patie
nts
mal
aria
afri
cam
edic
ines
site
wor
kad
ults
tool
sen
ough
occu
rred
book
sco
nnec
ted
need
edta
keye
tm
idda
yen
dlif
ela
ckbe
tter
year
2015
5fu
ture
sach
sm
eal
wor
rydy
ing
plen
tygo
chan
ges
know
ledg
eke
epw
orke
dth
ough
dram
atic
supp
orte
rsde
ath
plan
tge
tou
tcom
ele
arn
poor
care
feed
chan
getim
eaw
ayco
mm
onpl
ace
toda
yon
efiv
efa
mily
two
clin
ical
2025
trea
ted
beco
me
solu
tions
first
like
hard
heal
thof
ficer
hist
ory
real
lyal
ong
com
ego
odse
tta
ttere
dse
lling
targ
ets
little
doct
orcr
isis
clot
hing
see
trea
tmen
tei
ther
area
sch
emic
als
die
min
imal
who
leite
ms
mal
aria
lhu
ngry
alm
ost
save
prev
enta
ble
danc
ing
harv
est
mill
ions
trea
tabl
ew
alke
dsh
owed
easy
cost
sba
rech
eap
met
feet
ever
hann
ahar
ound
impo
veri
shed
mos
quito
esen
cour
agin
gea
sily
prob
ably
Tabl
e12
:Top
icw
ords
ofTC
attn
.
Cat
egor
y1
Cat
egor
y2
Cat
egor
y3
Cat
egor
y4
unpa
ved
road
sun
ited
natio
nsin
terv
entio
nya
lasu
bdi
stri
ctho
spita
lm
alar
iaco
mm
ondi
seas
epr
even
tabl
etr
eata
ble
tatte
red
clot
hing
safe
rhea
lthie
rbet
terl
ife
thre
eki
dsbe
dtw
oad
ults
room
spa
cked
patie
nts
mos
quito
esca
rry
mal
aria
infe
ctpe
ople
bitin
gba
refe
etou
tpov
erty
stab
ilize
econ
omy
qual
itylif
eco
mm
uniti
esno
tmed
icin
etr
eatm
entc
ould
affo
rdki
dsdi
em
alar
iaad
ults
sick
2000
0da
yle
ssth
an1
dolla
rday
afri
cake
nya
saur
ino
doct
oron
lycl
inic
alof
ficer
runn
ing
hosp
ital
bed
nets
mos
quito
esaw
aype
ople
save
mill
ions
lives
goal
sm
et20
1520
25no
runn
ing
wat
erel
ectr
icity
bed
nets
cost
5do
llar
80vi
llage
sac
ross
sub-
saha
raaf
rica
sad
peop
ledy
ing
near
deat
hpr
even
tabl
ech
eap
med
icin
estr
eatm
alar
iaC
ateg
ory
5C
ateg
ory
6C
ateg
ory
7C
ateg
ory
8cr
ops
dyin
gki
dsno
tatte
ndgo
scho
olpr
ogre
ssju
stfo
urye
ars
prog
ress
enco
urag
ing
supp
orte
rsno
taff
ord
fert
ilize
rirr
igat
ion
nota
ffor
dsc
hool
fees
yala
sub
dist
rict
hosp
italh
asm
edic
ine
solu
tions
prob
lem
ske
eppe
ople
impo
veri
shed
outc
ome
poor
crop
ski
dshe
lpch
ores
fetc
hing
wat
erw
ood
med
icin
efr
eech
arge
chan
gepo
vert
yst
rick
enar
eas
good
lack
fert
ilize
rwat
ersc
hool
sm
inim
alsu
pplie
sbo
oks
pape
rpen
cils
med
icin
em
ostc
omm
ondi
seas
espo
vert
yhi
stor
yno
teas
yta
skha
rden
ough
food
crop
sha
rves
tfee
dw
hole
fam
ilyhu
ngry
sick
conc
entr
ate
note
nerg
yw
ater
conn
ecte
dho
spita
lw
inni
ngag
ains
tpov
erty
poss
ible
achi
evab
lelif
etim
eno
mid
day
mea
llun
chho
spita
lgen
erat
orel
ectr
icity
bed
nets
used
ever
ysl
eepi
ngsi
tehu
nger
cris
isad
dres
sed
fert
ilize
rsee
dsto
ols
need
edm
aint
ain
food
supp
lyki
dsgo
scho
olno
wno
scho
olfe
esno
wse
rves
lunc
hst
uden
tssc
hool
atte
ndan
cera
tew
ayup
Tabl
e13
:Spe
cific
exam
ple
phra
ses
ofTC
manual.
Cat
egor
y1
Cat
egor
y2
Cat
egor
y3
Cat
egor
y4
Cat
egor
y5
wor
kha
rdlif
etim
eau
thor
conv
ince
win
ning
fight
pove
rty
achi
evab
lelif
etim
ech
ildre
nad
ults
easy
task
bette
rpla
ceun
ited
natio
nsau
thor
conv
ince
dw
inni
ngfig
htpo
vert
yac
hiev
able
lifet
ime
mos
quito
esca
rry
mal
aria
lived
dolla
rbe
tterh
ealth
unite
dst
ates
auth
orw
ants
dise
ase
calle
dm
alar
iath
ing
hist
ory
brig
hter
futu
relif
eco
mm
uniti
esau
thor
conv
ince
win
ning
fight
prov
erty
com
eni
ght
stuf
fnee
dth
ings
like
like
book
spa
perp
enci
lsw
inni
ngfig
htpr
over
tyac
hiev
able
lifet
ime
mal
aria
lmos
quito
esea
rnm
oney
thin
gsne
edle
arn
life
keny
aw
inni
ngfig
htpo
vert
yac
hiev
able
life
time
easi
lyad
ults
sick
fight
ing
pove
rty
impo
rtan
tkid
sar
ticle
brig
hter
futu
reso
lutio
nspr
oble
ms
peop
leim
pove
rish
edw
ork
chan
geth
inks
impo
rtan
tw
inin
gfig
htpo
vert
yac
hiev
able
mos
quito
esaw
ayha
rdw
ork
wan
tskn
owar
ticle
stat
esin
fect
peop
lebi
ting
agre
eau
thor
win
ning
fight
pove
rty
ache
ivab
leaw
aysl
eepi
ngw
orki
ngha
rdau
thor
prov
ided
bette
rlif
e20
08au
thor
thin
ksbe
tterl
ife2
008
base
dar
ticle
auth
orco
nvin
cere
adin
gar
ticle
conv
ince
dpo
vert
yth
ings
chan
ged
pove
rty
ache
ivab
lelif
etim
eC
ateg
ory
6C
ateg
ory
7C
ateg
ory
8C
ateg
ory
9C
ateg
ory
10at
tend
ance
rate
amaz
ing
prog
ress
year
sgo
odsh
ape
kids
adul
tsdo
nate
mon
eym
idda
ym
eal
text
says
good
educ
atio
n20
1520
25ta
ttere
dcl
othe
sse
rves
lunc
hst
uden
tste
xtsa
idw
ents
choo
lhu
ngry
sick
tatte
red
clot
hing
mid
day
mea
lsye
argi
rlar
eas
good
chea
pm
edic
ines
bare
feet
serv
edlu
nch
year
2004
tryi
nghe
lpgo
als
supp
osed
dona
ting
mon
eyst
uden
tsw
ante
dle
arn
para
grap
hsa
ysw
orke
dha
rdsa
vem
illio
nsliv
esbo
oks
penc
ilspr
ogre
sssh
ows
win
ning
fight
pove
rty
achi
evab
lese
cond
reas
onki
dsat
tend
scho
oltr
eate
dch
emic
als
seco
ndex
ampl
esc
hool
sm
inim
alpa
ragr
aph
stat
esgi
rlw
ent
scho
ols
hosp
itals
prog
ress
enco
urag
ing
supp
orte
rsm
illen
nium
villa
ges
hann
ahsa
chs
conv
ince
dw
inni
ngsc
hool
scho
olfe
esw
entk
enya
prac
tical
item
ski
dssa
uria
ttend
scho
olpa
rent
saf
ford
scho
olfe
esat
tend
ence
rate
pare
nts
mon
eyC
ateg
ory
11C
ateg
ory
12C
ateg
ory
13C
ateg
ory
14C
ateg
ory
15cl
ean
wat
ergr
owcr
ops
mill
enni
umvi
llage
proj
ect
stop
pove
rty
runn
ing
wat
erel
ectr
icity
wat
erw
ood
feed
fam
ilym
illen
ium
villa
gepr
ojec
tlo
ngtim
ew
ater
conn
ecte
dho
spita
lgen
erat
orel
ectr
icity
fres
hw
ater
need
edhe
lpm
illen
nium
villa
ges
proj
ecth
elpe
dw
orld
wor
kch
ange
patie
nts
affo
rdne
eds
help
farm
ers
wor
rych
ange
dram
atic
ally
beat
pove
rty
room
spa
cked
patie
nts
prob
ably
med
icin
esfr
eech
arge
crop
sdy
ing
affo
rdne
cess
ary
fert
ilize
rirr
igat
ion
dram
atic
chan
ges
occu
red
villa
ges
subs
ahar
anaf
rica
endi
ngpo
vert
ysh
are
beds
chor
esfe
tchi
ngfe
rtili
zerk
now
ledg
epl
ace
live
wan
tlea
rnre
ciev
etr
eatm
ent
fetc
hing
wat
erhu
nger
cris
isad
dres
sed
fert
ilize
rsee
dsto
ols
need
edm
aint
ain
food
supp
lyha
ppen
edye
ars
plac
eslik
edo
ctor
clin
ical
offic
erru
nnin
gho
spita
lfe
edfa
mili
esdr
amat
icch
ange
soc
curr
edvi
llage
ssh
ows
win
ning
fight
pove
rty
achi
evab
lelif
etim
edo
ctor
scl
inic
alhu
nger
cris
isad
ress
edm
illen
nium
deve
lopm
entg
oals
wan
tkin
dpo
vert
yw
ater
fert
ilize
rkno
wle
dge
fam
ilypl
ants
eeds
outc
ome
poor
chan
gepo
vert
ystr
icke
nar
eas
good
pove
rty
assu
reac
cess
rece
ive
trea
tmen
tfa
rmer
sw
orri
edco
min
gye
ars
runn
ing
bare
enco
urag
ing
supp
orte
rsm
illen
nium
villa
ges
proj
ect
affo
rdtr
eatm
ent
occu
rred
villa
ges
subs
ahar
anC
ateg
ory
16C
ateg
ory
17C
ateg
ory
18C
ateg
ory
19C
ateg
ory
20ya
lasu
bdis
tric
thos
pita
lmed
icin
efr
eech
arge
com
mon
dise
ases
nets
slee
ping
site
saur
ipl
anpe
ople
pove
rty
achi
eve
goal
year
sla
ter
free
lunc
haf
ford
nets
stab
ilize
econ
omy
qual
itylif
eco
mm
uniti
esre
ach
goal
took
year
sya
ladi
stri
ctas
sure
acce
sshe
alth
care
help
goin
gsc
hool
star
ted
2004
prev
enta
ble
trea
tabl
epe
ople
peop
lest
ory
says
com
mon
afri
cane
arde
ath
achi
eve
goal
sdi
seas
eslik
epo
orcr
ops
lack
com
mon
dise
ase
afri
caho
mel
ess
peop
leho
spita
lgoo
dsh
ape
dist
rict
hosp
ital
Tabl
e14
:Spe
cific
exam
ple
phra
ses
ofTC
lda.
Category 1brighter future hannah
millennium villages projectunpaved dirt road
bar sauri primary schoolfuture hannah
sauri primary schoolvillages project
millennium development goalsvillage leaders
dirt roadcar jumplittle kids
preventable diseases peoplemany kids
diseases peoplekids die
school suppliesprimary school
school feesinfect people
Table 15: Specific example phrases of TCpr.
Cat
egor
y1
Cat
egor
y2
Cat
egor
y3
Cat
egor
y4
Cat
egor
y5
Cat
egor
y6
win
ning
fight
coul
dfe
edbe
dne
taff
ord
four
year
spr
ogre
sslif
etim
eye
arfe
esst
uden
tssc
hool
supp
lies
scho
ols
saur
ikno
wle
dge
supp
lies
med
icin
espo
vert
yw
inni
ngw
orld
villa
ges
peop
lesc
hool
wor
kha
rdbo
oks
villa
ges
occu
rred
80ac
ross
alon
gsc
hool
fees
supp
lies
affo
rdfe
rtili
zer
affo
rdsc
hool
fees
bette
rmed
icin
ew
ater
ener
gyw
inni
ngfig
htpo
vert
yal
soev
ery
dise
ases
kids
heal
thne
t5to
ols
crop
ssc
hool
fees
seed
sbe
dne
tshe
lpke
epho
spita
lele
ctri
city
conn
ecte
dw
inni
ngpo
vert
ypr
even
tabl
efa
mily
peop
leca
reye
ars
man
yvi
llage
ssa
urip
roje
ctfa
rmer
sro
oms
patie
nts
crop
spe
ople
food
atte
ndan
cero
oms
end
man
ybe
dne
ts5
also
fight
pove
rty
affo
rdsc
hool
fees
bed
nets
outc
ome
poor
crop
ssc
hool
lunc
hm
ealm
idda
ysu
pplie
spr
oble
ms
also
peop
leen
ergy
man
yw
ater
elec
tric
ityho
spita
lfer
tiliz
erpo
vert
yfig
htw
inni
ngal
sow
ould
ener
gyle
arn
help
prog
ress
year
ske
nya
afri
cato
day
lunc
hst
uden
tsse
rves
mid
day
food
supp
lym
aint
ain
elec
tric
itysu
pplie
sel
ectr
icity
wat
eren
ergy
fight
pove
rty
win
ning
peop
lefe
essc
hool
farm
ers
coul
dra
tepe
ople
med
icin
e20
045
year
ske
epsc
hool
fees
bed
show
edlu
nch
coul
dw
ork
elec
tric
itym
edic
ine
villa
ges
keny
a80
farm
ers
man
ysc
hool
lunc
hsc
hool
sal
sofe
es20
04al
soye
arra
tesc
hool
bed
nets
used
coul
daf
ford
fert
ilize
rfo
urye
ars
lifet
ime
pove
rty
year
year
ssc
hool
show
edho
spita
lwat
erfa
rmer
sne
eded
food
supp
lyvi
llage
sge
nera
tore
nerg
ysc
hool
supp
lies
little
affo
rden
ough
year
sfo
urla
stfiv
eda
ysc
hool
pare
nts
atte
ndbe
dne
tsfr
eefo
odal
sofa
rmer
stw
om
any
pove
rty
scho
olm
edic
ine
fert
ilize
rhos
pita
lbed
wat
erel
ectr
icity
also
fert
ilize
rsup
plie
sal
soto
ols
year
sch
ange
sfe
rtili
zera
ddre
ssed
scho
olsc
hool
sfe
esfr
eetw
oel
ectr
icity
wat
erru
nnin
gal
soge
nera
tor
supp
lym
aint
ain
food
also
tatte
red
year
svi
llage
ske
nya
proj
ecta
ttend
ance
scho
olfe
essc
hool
slu
nch
free
gene
rato
rele
ctri
city
ener
gypo
vert
yhu
nger
elec
tric
itylu
nch
scho
olcr
ops
food
farm
ers
fert
ilize
rbed
netw
ater
wat
erfe
rtili
zere
nerg
ysc
hool
med
icin
esfe
rtili
zera
ddre
ssed
scho
olsu
pplie
scr
isis
Cat
egor
y7
Cat
egor
y8
Cat
egor
y9
Cat
egor
y10
Cat
egor
y11
Cat
egor
y12
elec
tric
ityru
nnin
gw
ater
irri
gatio
nse
tbe
dsh
owed
dise
ases
help
stud
ents
supp
lies
peop
lesc
hool
sye
ars
four
free
scho
ols
med
icin
em
edic
ine
elec
tric
ityto
ols
fert
ilize
rmed
icin
essc
hool
sal
sosc
hool
stud
ents
atte
ndan
cepo
orsh
owed
trea
tmen
tsch
ools
uppl
ies
lunc
hm
eale
nerg
ype
ople
year
sfo
urth
ree
thou
ghsc
hool
scho
ols
free
supp
lies
fees
wat
erel
ectr
icity
conn
ecte
dsc
hool
sru
nnin
gfr
eech
arge
scho
olm
aint
ain
supp
lyfa
rmer
sco
uld
crop
saf
ford
bed
dram
atic
chan
gebe
dne
tsvi
llage
sye
ars
80po
vert
ym
any
crop
sfe
rtili
zerf
arm
ers
tool
spl
ant
stud
ents
lunc
hse
rves
scho
ol20
04cr
ops
farm
ers
2004
first
food
elec
tric
ityho
spita
lpo
vert
ybe
tterl
ives
mad
em
any
wor
ked
toge
ther
end
wat
erel
ectr
icity
supp
lies
scho
olen
ergy
med
icin
ecr
ops
free
hosp
itala
lso
lack
fert
ilize
rsch
oolb
edne
tsbe
tterf
ertil
izer
med
icin
een
ough
also
achi
evab
lelif
etim
esa
uri
penc
ilsst
uden
tssu
pplie
sye
tm
edic
ine
scho
olsu
pplie
sye
ars
hung
ersc
hool
supp
lies
farm
ers
atte
ndan
cecr
ops
bed
nets
year
sho
spita
lro
oms
pack
edpa
tient
sm
alar
iago
odbe
dne
tuse
dvi
llage
sm
any
keny
asa
uri8
0fe
rtili
zerc
rops
lack
farm
ers
wat
erw
ater
supp
lies
scho
ols
free
hosp
ital
hosp
itald
isea
sefo
urye
ars
2004
food
fert
ilize
rcro
psge
tsup
ply
bed
net
year
sfo
odsu
pply
hung
ercr
isis
fert
ilize
rirr
igat
ion
crop
sm
edic
ine
wat
ersc
hool
scr
ops
supp
lies
free
char
geev
ery
slee
ping
site
five
netc
osts
5co
mm
ondi
seas
essa
urin
etm
edic
ines
scho
olm
edic
ine
fert
ilize
rfre
esc
hool
scho
ols
lunc
hal
sofr
eesc
hool
bed
also
occu
rred
80ne
tsne
tbed
free
wor
kto
geth
erpo
vert
yne
t5fr
eech
arge
med
icin
esc
hool
med
icin
essc
hool
fees
scho
ols
year
sfo
ursc
hool
sla
stst
uden
tsru
nnin
gw
ater
supp
lies
scho
ols
alm
ost
hosp
italg
osc
hool
coul
daf
ford
scho
olsu
pplie
site
ms
seed
spl
antc
rops
fert
ilize
rsc
hool
fees
lunc
hsc
hool
supp
lies
scho
ols
also
2004
bed
supp
lies
know
ledg
em
edic
ines
affo
rdpr
ojec
tpro
gres
sm
ade
food
good
sach
sm
any
free
scho
ols
lunc
hsc
hool
char
gelu
nch
scho
ols
scho
olse
eds
food
crop
sfa
rmer
ssc
hool
spr
ojec
tals
osu
pplie
sfo
odsu
pply
farm
ers
wat
eral
soho
spita
ldoc
torc
linic
alsh
owed
bed
nets
wat
erfe
rtili
zerm
edic
ines
scho
olfe
essc
hool
sfr
eelu
nch
hosp
italy
ears
med
icin
esc
hool
wat
ersu
pplie
sm
idda
ysc
hool
food
hung
erye
ars
mad
em
alar
iata
kech
ange
sfr
eech
arge
med
icin
esc
hool
fert
ilize
rsc
hool
ssu
pplie
sel
ectr
icity
farm
ers
fert
ilize
rfr
eech
arge
scho
ols
year
sm
eal
man
yfo
odco
uld
bette
rfut
ure
peop
lelu
nch
crop
sfa
rmer
sfe
rtili
zere
lect
rici
tykn
owle
dge
stud
ents
lunc
hm
edic
ine
hosp
italm
ade
scho
olfe
essc
hool
sfr
eem
edic
ines
scho
ols
scho
olfa
rmer
scr
ops
bed
free
char
gesc
hool
year
shu
nger
Cat
egor
y13
Cat
egor
y14
Cat
egor
y15
Cat
egor
y16
Cat
egor
y17
Cat
egor
y18
bed
nets
villa
ges
afri
cam
illen
nium
80ac
ross
supp
lybo
oks
seed
sfe
rtili
zera
ddre
ssed
food
med
icin
een
ough
wou
ldw
ork
hard
bette
rw
ater
conn
ecte
dho
spita
lw
ater
runn
ing
med
icin
em
edic
ines
supp
lies
80vi
llage
sac
ross
elec
tric
ityw
ater
seed
ssu
pply
fert
ilize
rcro
pspl
enty
peop
lew
orld
saur
ikid
spo
vert
yne
tsbe
dus
edcr
ops
affo
rdbe
dne
tsm
edic
ine
crop
sel
ectr
icity
pove
rty
fight
peop
leke
nya
end
pove
rty
man
yliv
eshu
nger
ever
yfe
rtili
zers
eeds
crop
sm
any
peop
lepo
vert
yco
uld
take
mid
day
mea
lsa
urif
ree
bed
nets
wor
ld20
15di
seas
esla
ckw
ater
day
ever
yto
ols
fert
ilize
rke
nya
wou
ldbe
tterw
alke
dba
rem
idda
ym
eall
unch
crop
sfe
rtili
zerp
lant
food
irri
gatio
npo
orvi
llage
saur
iad
ults
one
bed
two
last
crop
sfa
rmer
sal
sow
ater
coul
dpo
vert
ypr
oble
ms
cris
isth
ough
man
ybe
dne
tsus
edbe
dne
tsev
ery
wat
erm
edic
ine
wel
lpro
ject
villa
ges
poor
end
peop
lefo
odw
ork
man
yen
ergy
crop
sse
eds
wat
erne
eded
peop
leke
nya
targ
ets
80vi
llage
sbe
dev
ery
slee
ping
site
net
fert
ilize
rcro
psw
ater
keep
tool
sac
hiev
able
keny
avi
llage
svi
llage
scho
olpe
ople
man
yad
dres
sed
fert
ilize
rsee
dsal
mos
tkid
sdi
epe
ople
bed
nets
ever
yus
edsc
hool
keny
abe
dne
tsm
any
villa
ges
peop
lepr
oble
ms
keny
asc
hool
food
scho
ols
hosp
italp
eopl
ese
eds
fert
ilize
rfoo
dal
sow
ater
rate
way
prog
ress
bette
rafr
ica
hosp
italw
ater
runn
ing
clin
ical
offic
erbe
dne
tsal
soad
ults
proj
ectv
illag
eske
nya
villa
gepe
ople
year
sch
ange
sfo
urfr
eeoc
curr
edse
eds
fert
ilize
rwat
erat
tend
ance
rate
way
wat
erho
spita
lbed
nets
saur
ibed
nets
goal
sfo
urye
ars
met
need
edw
ater
ever
yw
ork
scho
olfe
esfe
rtili
zerf
ood
see
wor
ldbe
dne
tsco
uld
keep
ever
ybe
dne
tspo
vert
yvi
llage
fight
afri
casa
uri
year
sho
spita
lvill
ages
char
geco
nnec
ted
fert
ilize
rirr
igat
ion
nece
ssar
yfa
rmer
sto
ols
gohu
ngry
getp
eopl
eco
uld
bed
nets
used
ever
ysl
eepi
ngdi
seas
esm
edic
ine
med
icin
esco
mm
onpr
even
tabl
eat
tend
ance
rate
way
selli
ngco
me
food
mai
ntai
nsu
pply
elec
tric
itysu
pplie
sfe
rtili
zers
eeds
irri
gatio
nfa
rmer
sla
ckge
tfoo
dw
ork
wou
ldpr
obab
lyho
spita
lcha
rge
bed
nets
prev
enta
ble
nets
bed
wat
ersa
uriy
ears
wor
kw
orld
help
last
toge
ther
2015
2025
dyin
ghu
nger
deat
hfe
rtili
zerl
ack
crop
sbe
com
esa
uri
wor
ldw
inni
ngfig
htw
aypl
ace
crop
sfe
rtili
zere
noug
hfa
rmer
spo
vert
ym
any
2015
mill
enni
umpr
ogre
ssdi
seas
esm
alar
iape
ople
easi
lysa
urih
isto
ryw
aysi
tesa
uri
help
peop
lepo
vert
ypl
ace
man
y
Tabl
e16
:Spe
cific
exam
ple
phra
ses
ofTC
attn
.