Modelling human/machine generated text detection

• Anna (she/her)
• MA student
  • BA on modeling probabilistic reduction
• Involved in multiple projects across departments  
  • Speech to text for Kinyarwanda
  • Modelling probabilistic speech reduction
  • Morphological representations in embeddings
  • Teaching assistant for CL courses
• FOSS enthusiast
• Outside of university: bouldering, gaming, crocheting

$whoami

SemEval

• The 18th International Workshop on Semantic Evaluation
• NLP research challenges on issues in semantic analysis
• SemEval 2024 categories:
  • Semantic Relations​

  • Discourse and Argumentation

  • LLM Capabilities

  • Knowledge Representation and Reasoning

• ​​Focus on multilingual and multimodal approaches

SemEval taSk 8

• Classification tasks between human-written and machine-generated texts
• Tracks:
  • Subtask A: human vs. machine classification
  • Subtask B: human vs. specific machine classification
  • Subtask C: boundary detection in mixed texts  
• Monolingual (English) or multilingual

Subtask A: Binary classification for monolingual human- and machine-generated texts¹

¹Co-authors: Vittorio Ciccarelli, Cornelia Genz, Nele Mastracchio, Hanxin Xia and Wiebke Petersen

M4 dataset

• Multi-Generator, Multi-Domain, and Multi-Lingual Black-Box Machine-Generated Text Detection (Wang et al. 2023)
  • Five machines
  • Six domains
  • Seven languages: Chinese, Russian, Urdu, Indonesian, Arabic, English

M4 dataset

Data splits:

M4 dataset

Domains:

M4 dataset

LLMs:

• Davinci-text-003/GPT 3.5 (OpenAI 2023)
• chatGPT (OpenAI 2023)
• GPT4 (OpenAI 2023)
• Cohere 
• Dolly-v2 (Conover et al. 2023)
• BLOOMz 176B(Muenninghoff et al. 2022)

M4 dataset

LLM prompts:

• 2-8 different prompts for each LLM
• PeerRead example: 
  • "Please write a peer review for the paper + title" (Wang et al. 2023:A.3)
  • "Write a peer review by first describing what problem or question this paper addresses, then strengths and weaknesses, for the paper + title, its main content is as follows: + abstract" (Wang et al. 2023:A.3)

M4 dataset

LLM splits in data:

M4 dataset

Machine-generated text sample: 

{"text":"Building a Railroad Tie Retaining Wall can be a daunting task, but with the proper tools and techniques, it can be completed with ease. If you want to create a strong and durable retaining wall that is both functional and attractive, follow the steps below.\n\nBulldoze or Dig a Section of the Dirt from the Hill Out to Where You Want to Build a Railroad Tie Retaining Wall\n\nThe first step in building a Railroad Tie Retaining Wall is to determine where you want to build it. Once you have located the perfect spot, you will need to bulldoze or dig a section of the dirt from the hill out to this area. [...] But, by following these steps, you can create a strong, durable, and attractive Retaining Wall that will serve you for years to come.","label":1,"model":"chatGPT","source":"wikihow","id":7}

M4 dataset

{"text":" It is possible to become a VFX artist without a college degree, but the path is often easier with one. VFX artists usually major in fine arts, computer graphics, or animation. Choose a college with a reputation for strength in these areas and a reputation for good job placement for graduates. The availability of internships is another factor to consider.Out of the jobs advertised for VFX artists, a majority at any given time specify a bachelor\u2019s degree as a minimum requirement for applicants. [...] To build your specialization, start choosing jobs with that emphasis and attend additional training seminars.For example, some VFX specialists focus on human character\u2019s faces, animal figures, or city backgrounds.\n\n","label":0,"model":"human","source":"wikihow","id":56408}

Human-written text sample: 

Submission idea

• Establish baseline using traditional ML classifers like logistic regression and random forrest
• RoBERTa + MLP classifiers as an ensemble model
• Correctional MLP classifiers for RoBERTa output 

• Focus on label features, not domains
• Compute features of texts based on previous literature
• Feature selection by correlation with the labels 

Architecture

• RoBERTa AI detector (Soleiman et al. 2019)
  • 1.5B-parameter GPT-2 model
  • Fine-tuned with 10% of training data, no features
  • 0.89 accuracy on dev

• Multi-layer perceptron classifier
  • Hidden layer size = 100, ReLu activation
  • Trained on train and dev dataset, with features

features

Count-based features:

• Mean sentence length
• Mean word length 
• Ratio of punctuation to words
• Ratio of word types to tokens
• Ratio of vowels to words
• Number of hapax legomena
• Number of negation words 
• Type-to-token ratio (TTR)
• Number of unique words

features

features

features

Frequency features:

• Mean log. frequency of content words
• Ratio of frequent words to content words
• Ratio of hapax legomena to content words
• Ratio of content words to top 10% of high frequency words/fantasy words in Wikipedia lists

FEATURES

Syntactic features:

• Ratio of words to 
  • Nouns, verbs, adjectives, adverbs, adpositions, conjunctions, numerals, pronouns, determiners
• Ratio of adjectives to nouns
• Ratio of verbs to nouns
• Maximum dependency distance in a syntactic tree (text)
• Maximum dependency distance in a syntactic tree (sentence)
• Number of passive constructions

FEATURES

Word difficulty features:

• Ratio of content words to
  • A1-level words
  • A2-level words
  • B1-level words
  • B2-level words
  • C1-level words
  • C2-level words
• Both on the basis of stemmed and lemmatized words

FEATURES

Stylistic features:

• Ratio of words with negative opinion (Liu et al. 2005)
• Readability by Flesch-ease reading score¹

Sentiment features:

• Scores for "neutral", "anger", "disgust", "fear", "joy", "sadness", "surprise" by DistilRoBERTa-base (Hartmann et al. 2022)
• Score for "positive" or "negative" by Huggingface sentiment analysis pipeline 
• Score for "formal" and "informal" by a formality ranker (Babkov et al. 2023)
• Score for "toxic" and "non-toxic" by finetuned-RoBERTa model (s-nlp 2022) 

FEATURES

RoBERTa-based features:

• Logits of RoBERTa for each text
• Last hidden states of RoBERTa for each text, reduced by 
  • Principal component analysis (PCA)
  • Uniform manifold approximation and projection (UMAP), distance measures: cosine and jaccard

computational cost 

• Feature extraction: 
  • Ran locally with 24GB RAM, no GPU
  • ~6hours for 'expensive' spacy-based features for all datasets
  • ~4mins for all other features for all datasets

• RoBERTa fine-tuning:
  • Ran on Google Colab with T4 GPU
  • ~45 minutes

results

• Placement: 32nd of 141 with an accuracy of 0.85
• RoBERTa-base baseline accuracy: 0.74 

Error Analysis

Classification errors by model:

Error Analysis

Correction errors:

Conclusion

• Feature-informed MLP classifier helped adjust machine label bias in RoBERTa model
• MLP classifier was better at capturing human-like texts
  • Human > machine correction 64% wrong
  • Machine > human correction 28% wrong 

Ensemble model worked semi-well

Conclusion

• Combat bias
  • Fine-tune the RoBERTa AI detector with more data​
  • Analyze correction classifier for performance
    • Particularly wrong human > machine corrections
• Improve performance
  • Analyze distribution of features for all LLMs
  • Add more sophisticated features such as argument structure, contextual predictability
  • Different correctional classifier?

Future directions for this architecture:

References

assets

results

RoBERTa-base OpenAI detector on test (acc. 0.64):

Fine-tuned RoBERTa classifier used in our submission on dev: