As the quality of AI-generated text increases with the development of new Large Language Models, people use them to write in a variety of contexts. Human-AI collaborative writing poses a potential challenge for existing AI analysis techniques, which have been primarily tested either on human-written text only, or on samples independently generated by humans and AI. In this work, we investigate the extent to which existing AI detection and authorship analysis models can perform classification on data generated in human-AI collaborative writing sessions. Results show that, for AI text detection in the cowriting setting, classifiers based on authorship embeddings (Rivera-Soto et al., 2021) outperform classifiers used in prior work distinguishing AI vs. human text generated independently. However, these embeddings are not optimal for finer-grained authorship identification tasks: for authorship verification, n-gram based models are more robust to human-AI co-written text, and authorship attribution performance degrades compared to baselines that use human-written text only. Taken together, this suggests that the rise of human-AI co-written text will require adapting AI detection tools and authorship analysis techniques in the near future. We release our code at https://github.com/AARichburg/Human-AI_Authorship_Analysis.

Humans use document formatting to discover document and section titles, and important phrases. But when machines process a paper–especially documents OCRed from images–these cues are often invisible to downstream processes: words in footnotes or body text are treated as just as important as words in titles. It would be better for indexing and summarization tools to be guided by implicit document structure. In an ODNI-sponsored project, ARLIS looked at discovering formatting in OCRed text as a way to infer document structure. Most OCR engines output results as hOCR (an XML format), giving bounding boxes around characters. In theory, this also provides style information such as bolding and italicization, but in practice, this capability is limited. For example, the Tesseract OCR tool provides bounding boxes, but does not attempt to detect bold text (relevant to author emphasis and specialized fields in e.g. print dictionaries), and its discrimination of italicization is poor. Our project inferred font size from hOCR bounding boxes, and using that and other cues (e.g. the fact that titles tend to be short) determined which text constituted section titles; from this, a document outline can be created. We also experimented with algorithms for detecting bold text. Our best algorithm has a much improved recall and precision, although the exact numbers are font-dependent. The next step is to incorporate inferred structure into the output of machine translation. One way is to embed XML tags for inferred structure into the text extracted from the imaged document, and to either pass the strings enclosed by XML tags to the MT engine individually, or pass the tags through the MT engine without modification. This structural information can guide downstream bulk processing tasks such as summarization and search, and also enables building tables of contents for human users examining individual documents.

While collecting or generating more parallel data is necessary to improve machine translation (MT) in low-resource settings, we lack an understanding of how the limited amounts of existing data are actually used to help guide the collection of further resources. In this paper, we apply data cartography techniques (Swayamdipta et al., 2020) to characterize the contribution of training samples in two low-resource MT tasks (Swahili-English and Turkish-English) throughout the training of standard neural MT models. Our empirical study shows that, unlike in prior work for classification tasks, most samples contribute to model training in low-resource MT, albeit not uniformly throughout the training process. Furthermore, uni-dimensional characterizations of samples – e.g., based on dual cross-entropy or word frequency – do not suffice to characterize to what degree they are hard or easy to learn. Taken together, our results suggest that data augmentation strategies for low-resource MT would benefit from model-in-the-loop strategies to maximize improvements.

Byte-Pair Encoding (BPE) (Sennrich et al., 2016) has become a standard pre-processing step when building neural machine translation systems. However, it is not clear whether this is an optimal strategy in all settings. We conduct a controlled comparison of subword segmentation strategies for translating two low-resource morphologically rich languages (Swahili and Turkish) into English. We show that segmentations based on a unigram language model (Kudo, 2018) yield comparable BLEU and better recall for translating rare source words than BPE.