🧑‍🔬 Satanjeev Banerjee

Satanjeev Banerjee works on automatic measures in natural language processing. Their papers here, from 2003 to 2005, ask two related questions: how closely two concepts are related in meaning, and how well a machine-translated sentence matches human reference translations. Both lines of work use the same basic method. They match words between texts and turn the pattern of those matches into a score, drawing on stems, synonyms and dictionary definitions rather than exact spelling alone.

The most cited of these papers is METEOR: An Automatic Metric for MT Evaluation. It introduced a metric that aligns a system's output word by word with each reference translation. The alignment runs in stages: identical forms first, then shared stems, then synonyms. From that alignment the metric computes unigram precision and recall, gives recall more weight, and lowers the score when the matched words fall into many scattered chunks. The authors argue that this works better at the level of single sentences than BLEU does. The earlier paper, Extended Gloss Overlaps as a Measure of Semantic Relatedness, builds on the Lesk algorithm. Because WordNet definitions are too brief on their own, it also compares the glosses of concepts linked through relations such as hypernymy and meronymy. Longer shared word sequences earn disproportionately more credit.

The later work moves from lexical semantics to the evaluation of translation systems, though word-level matching remains central to both. This places Banerjee's papers here within machine translation and evaluation & benchmarks. Banerjee is affiliated with Carnegie Mellon University, the institution on papers from 2003 to 2018. Co-authors include Alon Lavie and Ted Pedersen.

Satanjeev Banerjee has 2 papers in our library, in 2003–2005, most of them on Machine translation, Evaluation & benchmarks; often with Alon Lavie, Ted Pedersen.