These are the year's most interesting AI research articles. It merges breakthroughs in artificial intelligence (AI) and data science. It is arranged chronologically and has a link to a more extensive article.

Accurate Online Posterior Alignments for Principled Lexically-Constrained Decoding

Online alignment matches a target word to a source word in machine translation when only part of the target sequence has been decoded. Good online alignments make it easier to use essential applications like lexically constrained translation, in which user-defined dictionaries add lexical constraints to the translation model. The researchers propose a new posterior alignment method that can be done online and has a lower rate of alignment errors than other methods. Their proposed inference method takes alignment and token probabilities into account logically, and we can add it to existing constrained beam-search decoding algorithms without any problems. On five language pairs, including two that are very different, we see a consistent drop in the number of alignment errors. Furthermore, when the researchers use the system on seven lexically constrained translation tasks, they see significant improvements in BLEU around the constrained positions.

Alignment-Augmented Consistent Translation for Multilingual Open Information Extraction

Due to a lack of training data in other languages, most of the work on supervised Open Information Extraction (OpenIE) has been done in English. In this paper, researchers look at ways to automatically change English text into other languages so that we can train OpenIE systems in those other languages.

The researchers use the Alignment-Augmented Constrained Translation (AACTrans) model to translate English sentences and their extractions, so they match and don't modify vocabulary or semantic meaning. Using the data made by AACTrans, they train a new two-stage generative OpenIE model, which they call Gen2OIE. For each sentence, Gen2OIE gives out 

1) relations in the first stage and 

2) all extractions that contain the relation in the second stage. 

Furthermore, Gen2OIE increases relation coverage by using a training data transformation technique that can be applied to multiple languages. It is different from existing models, which use an English-specific training loss. Evaluations of five languages—Spanish, Portuguese, Chinese, Hindi, and Telugu—show that the Gen2OIE with AACTrans data is better than previous systems by 6-25% in F1.

Comprehensive Multi-Modal Interactions for Referring Image Segmentation

The Referring Image Segmentation (RIS) method, which generates a segmentation map that matches the natural language description, is the subject of the study. To effectively address RIS, it is necessary to consider interactions within and between the visual and language modalities. The limitations of current approaches stem from the fact that they either compute various forms of interactions sequentially or neglect intramodal interactions.

The researchers overcome this restriction by utilising a Synchronous Multi-Modal Fusion Module to carry out all three interactions simultaneously (SFM). Additionally, they suggest a unique Hierarchical Cross-Modal Aggregation Module (HCAM), where linguistic traits promote the flow of contextual information across the visual hierarchy to build refined segmentation masks. Finally, the researchers provide comprehensive ablation studies and evaluate the performance of our methodology on four benchmark datasets, demonstrating significant performance improvements over the current state-of-the-art (SOTA) approaches.